test(wireguard): drive a real AmneziaWG tunnel through ClientBind

The unit tests pin the reserved-byte gate on each side in isolation, which would still pass if the two halves disagreed about when to apply it. This wires two real wireguard-go devices together over loopback UDP through ClientBind on both ends — the bind the detour path actually uses — configures ranged h1-h4 plus s4 and junk, and asserts an inner IP packet reaches the peer's TUN. It is red against the unconditional clear and green with the gate, so it covers the failure the field hit rather than the code we happened to write. Tagged with_awg, so it runs under the shipped router tag set. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
fix(logsink): collapse interleaved floods, not just consecutive lines
2026-07-25 23:25:05 +03:00 · 2026-07-25 23:23:24 +03:00 · 2026-07-25 23:23:24 +03:00 · 2026-07-25 23:23:24 +03:00 · 2026-07-25 21:35:06 +03:00 · 2026-07-25 21:35:06 +03:00
140 changed files with 15445 additions and 1692 deletions
@@ -1,51 +1,62 @@
-# Shater v0.2 — build the 4-package signed opkg feed and publish it as a rolling
-# Gitea release consumable as an `src/gz` feed.
+# Shater v0.2 — build the 4-package signed **apk** feed and publish it as
+# per-arch Gitea releases consumable as an apk repository.
 #
-# WHAT CHANGED FROM v0.1
-#   v0.1 shipped 3 packages: xrayctl (SDK-compiled Go) + shater-core +
-#   luci-app-shater (hand-packed data .ipk). v0.2 collapses the runtime into ONE
-#   forked binary and ships 4 packages, all built the canonical SDK way:
+# WHAT WE SHIP
+#   ONE forked binary plus its OpenWrt glue, 4 packages, all built the canonical
+#   SDK way:
 #     - shaterd          PREBUILT static-musl + SPA-embedded + UPX binary. Built
 #                        OUT OF TREE by scripts/build-shaterd.sh (Go + Node + UPX)
 #                        and staged into openwrt/shaterd/files/ BEFORE the SDK
 #                        build; the openwrt/shaterd package just $(INSTALL_BIN)s
-#                        the arch-matched artifact. (arch-specific .ipk)
+#                        the arch-matched artifact. (arch-specific .apk)
 #     - shater-core      data glue, PKGARCH=all
 #     - luci-app-shater  LuCI thin launcher, PKGARCH=all (uses feeds/luci/luci.mk)
 #     - byedpi           ciadpi, C cross-compiled from source by the SDK (arch-specific)
 #
 # TARGET HARDWARE / ARCH MATRIX
 #   x86_64             -> the QEMU testbed VM (generic x86-64).
-#   aarch64_cortex-a53 -> BOTH production routers (BPI-R3 + BPI-R4, mediatek/filogic).
+#   aarch64_cortex-a53 -> BOTH production routers (BPI-R3 mini + BPI-R4,
+#     mediatek/filogic), both on 25.12 with apk-tools 3.
 #     Only shaterd + byedpi are arch-specific; shater-core + luci-app-shater are
-#     PKGARCH=all, so one build of each covers every device. opkg filters by
-#     Architecture at install time, so a single combined feed URL serves all.
+#     PKGARCH=all, so one build of each covers every device — but the RELEASES
+#     are still per-arch (see the release-apk job for why).
 #
-# FEED SIGNING (opkg / usign — OpenWrt 24.10 is opkg, not apk; apk lands at 25.12)
-#   The feed index (Packages) is usign-signed with the SECRET key in the Gitea
-#   repo secret KEY_BUILD; routers verify it with the committed public key
-#   dist/shater-feed.pub (fingerprint 5ac4b177689cb8e0). Do NOT regenerate the
-#   key — that invalidates every deployed router's trust.
+# FORMAT: apk ONLY (25.12+)
+#   The fleet runs OpenWrt/ImmortalWrt 25.12, where opkg is replaced by Alpine
+#   apk (.apk files, binary packages.adb index, EC keys in /etc/apk/keys/). The
+#   old .ipk lane was removed in 2026-07 (docs-shater/DECISIONS.md D22): no
+#   device we serve has an opkg binary at all, so building and signing a second
+#   feed served nobody.
+#
+# FEED SIGNING (EC / apk)
+#   packages.adb is signed with the EC (prime256v1) SECRET key in the Gitea repo
+#   secret KEY_APK; routers verify it with the committed public key
+#   dist/shater-apk.pem (ci/gen-apk-key.sh). Do NOT regenerate the key — that
+#   invalidates every deployed router's trust.
 #
 # AUTO-RELEASE
-#   push a tag `vX.Y.Z` -> versioned release. workflow_dispatch / (optional) main
-#   -> rolling `latest` pre-release (always-fresh feed). Publish uses the Gitea
-#   API via curl (ci/gitea-release.sh) — no external action needed.
+#   push a tag `vX.Y.Z` -> versioned per-arch releases `apk-vX.Y.Z-<arch>`.
+#   workflow_dispatch -> rolling per-arch `apk-latest-<arch>` (always-fresh
+#   feed). Publish uses the Gitea API via curl (ci/gitea-release.sh) — no
+#   external action needed. NOTE: the apk release tags deliberately do NOT start
+#   with `v` so publishing them cannot re-trigger this workflow's `v*` filter.
 #
-# APK LANE (25.12+, ADDITIVE — T2)
-#   The fleet is migrating to BananaWRT 25.12-mtk-vendor (= ImmortalWrt 25.12
-#   base), where opkg is replaced by Alpine apk (.apk, binary packages.adb
-#   index, EC keys in /etc/apk/keys/). The `build-apk` + `release-apk` jobs
-#   below build the SAME 4 packages through the ImmortalWrt 25.12 apk-SDK and
-#   publish PER-ARCH apk repos as releases `apk-latest-<arch>` (rolling) /
-#   `apk-<tag>-<arch>` (versioned). Per-arch because apk filenames carry no
-#   architecture (shaterd-0.2.0-r1.apk would collide across arches in one flat
-#   release) and apk fetches packages relative to the packages.adb URL.
-#   Signed with the EC key in the Gitea secret KEY_APK; trust anchor
-#   dist/shater-apk.pem (ci/gen-apk-key.sh). The usign/opkg lane above is
-#   UNCHANGED and keeps serving the 24.10 fleet. NOTE: the apk release tags
-#   deliberately do NOT start with `v` so publishing them cannot re-trigger
-#   this workflow's `v*` tag filter.
+# PACKAGE VERSIONING (bug B4)
+#   PKG_VERSION/PKG_RELEASE are NOT hand-written in the Makefiles any more. They
+#   used to be, and nobody bumped them: v0.2.2…v0.2.6 all shipped as
+#   `shaterd 0.2.0-r3` with different binaries inside, so `apk update` never saw
+#   a new version and routers could not be updated at all. Now `ci/version.sh`
+#   derives them from the git tag ONCE per job (the "Compute version" step,
+#   exported via $GITHUB_ENV):
+#       tag `vX.Y.Z`         -> X.Y.Z-r1
+#       anything else        -> <nearest tag>-r<commits since it + 1>
+#   and hands them to the SDK build as SHATER_PKG_VERSION/SHATER_PKG_RELEASE;
+#   $SHATER_VERSION (the same numbers, plus the short sha off-tag) is stamped
+#   into the binary's constant.Version. ci/sdk-build-apk.sh then ASSERTS that the
+#   built .apk really carry that version, so the failure can never be silent
+#   again. This is also why the build job checks out with fetch-depth: 0
+#   — `git describe` needs tags and ancestry. `byedpi` is excluded: it keeps
+#   upstream ByeDPI's own PKG_VERSION (see openwrt/byedpi/Makefile).

 # CACHING (T3 — fast CI)
 #   All caches use actions/cache pinned to v3.3.2: the LAST release speaking the
@@ -63,34 +74,33 @@
 #       (PKG_VERSION/PKG_HASH live there). Stale-safe: the buildroot verifies
 #       PKG_HASH on every dl/ file and re-downloads on mismatch, so restore-keys
 #       prefix fallback is allowed.
-#     - Go module + build cache — key = hash of go.sum; shared by all 4 build
+#     - Go module + build cache — key = hash of go.sum; shared by both build
 #       jobs (each builds both GOARCHes).
 #     - panel/node_modules — key = hash of panel/package-lock.json, exact-only
 #       (a lockfile change MUST miss); on hit build-shaterd.sh gets --fast.
-#     - apt .deb archives for the apk lane's debian:bookworm host-deps
-#       (.cache/apt) — key = hash of ci/sdk-build-apk.sh (the apt list is in it).
-#     - usign binary (.cache/tools) — static helper, fixed key.
+#     - apt .deb archives for the debian:bookworm host-deps of the apk SDK
+#       container (.cache/apt) — key = hash of ci/sdk-build-apk.sh (the apt list
+#       is in it).
 #     - SDK feeds/ git checkouts (.cache/feeds) — the single biggest recurring
 #       cost: `scripts/feeds update -a` cloned base+packages+luci+routing+
 #       telephony EVERY run (~7 min/job; github.com is ~1 MB/s from this
 #       runner — run 51 evidence). The feeds dir is symlinked into the SDK
 #       container from the workspace cache; `feeds update` on an existing clone
 #       is a fast fetch+checkout of the pinned revs. Correctness-safe: update
-#       always checks out feeds.conf's pins, and ci/sdk-build*.sh wipes the
+#       always checks out feeds.conf's pins, and ci/sdk-build-apk.sh wipes the
 #       cache + re-clones fresh if update ever fails on a cached checkout.
-#       Key = lane + SDK release (shared across the two arch jobs of a lane —
-#       same release pins identical feed revs; the sequential runner means the
-#       second arch restores what the first saved). restore-keys lets an SDK
-#       version bump start from the old clones (git fetch delta, not re-clone).
+#       Key = lane + SDK release (shared across the two arch jobs — the same
+#       release pins identical feed revs; the sequential runner means the second
+#       arch restores what the first saved). restore-keys lets an SDK version
+#       bump start from the old clones (git fetch delta, not re-clone).
 #   Act_runner facts this design leans on (verified in run 51 logs):
 #     - the cache backend works: restores/saves confirmed, hashFiles() works;
-#     - docker images (openwrt/sdk, debian:bookworm, runner-images) live on the
-#       PERSISTENT host daemon — "Image is up to date" each run, no re-download;
+#     - docker images (debian:bookworm, runner-images) live on the PERSISTENT
+#       host daemon — "Image is up to date" each run, no re-download;
 #     - each actions/cache SAVE is followed by an exact 3-minute act_runner
 #       stall (node process lingers; hit→no-save→no stall). Steady state saves
 #       nothing, so adding cache entries is fine, but keys that change every
 #       run (e.g. github.sha) would cost +3 min/entry/run — do NOT do that.
-
 name: release

 on:
@@ -108,137 +118,11 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  build:
-    name: ${{ matrix.arch }}
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - { arch: x86_64,             sdk: x86_64-24.10.4 }            # testbed VM (generic x86-64)
-          - { arch: aarch64_cortex-a53, sdk: mediatek-filogic-24.10.4 }  # BPI-R3 + BPI-R4 (mediatek/filogic)
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      # scripts/build-shaterd.sh builds the engine via a go.mod
-      # `replace => ./submodules/wireguard-go` (AmneziaWG fork), so that submodule
-      # must be present or `go build` dies with "no such file or directory".
-      # actions/checkout does not fetch submodules by default; init ONLY this one
-      # (clients/apple+android are large and unused here).
-      - name: Init wireguard-go submodule (awg)
-        run: git submodule update --init --depth 1 submodules/wireguard-go
-
-      # Toolchain for scripts/build-shaterd.sh: Go (daemon), Node (Vite SPA), UPX.
-      - name: Set up Go
-        uses: actions/setup-go@v5
-        with:
-          go-version-file: go.mod   # pins Go 1.24.7 (go.mod `go` line)
-          cache: false              # explicit actions/cache@v3.3.2 below (setup-go's
-                                    # built-in cache uses the new API act_runner lacks)
-
-      - name: Set up Node
-        uses: actions/setup-node@v4
-        with:
-          node-version: '20'        # Vite 5 needs Node 18+; 20 LTS
-
-      # ---- caches (see the header comment for keys + version pin rationale) ----
-      - name: Cache Go modules + build cache
-        uses: actions/cache@v3.3.2
-        with:
-          path: |
-            ~/go/pkg/mod
-            ~/.cache/go-build
-          key: go-${{ hashFiles('go.sum') }}
-          restore-keys: |
-            go-
-
-      - name: Cache panel node_modules
-        id: npm-cache
-        uses: actions/cache@v3.3.2
-        with:
-          path: panel/node_modules
-          key: npm-${{ hashFiles('panel/package-lock.json') }}
-          # NO restore-keys: node_modules must exactly match the lockfile;
-          # on any lockfile change this misses and `npm ci` runs fresh.
-
-      - name: Cache SDK dl/ (package sources)
-        uses: actions/cache@v3.3.2
-        with:
-          path: .cache/dl
-          key: dl-${{ hashFiles('openwrt/*/Makefile') }}
-          restore-keys: |
-            dl-
-
-      # feeds git checkouts (see header): both 24.10.4 arch jobs share one entry
-      # (same release = same feeds.conf.default pins), so derive the release
-      # from the matrix sdk tag (x86_64-24.10.4 -> 24.10.4).
-      - name: Compute feeds cache key
-        id: feedskey
-        run: echo "ver=$(echo '${{ matrix.sdk }}' | sed 's/.*-//')" >> "$GITHUB_OUTPUT"
-
-      - name: Cache SDK feeds checkouts
-        uses: actions/cache@v3.3.2
-        with:
-          path: .cache/feeds
-          key: feeds-opkg-${{ steps.feedskey.outputs.ver }}
-          restore-keys: |
-            feeds-opkg-
-
-      - name: Cache CI tools (usign)
-        uses: actions/cache@v3.3.2
-        with:
-          path: .cache/tools
-          key: tools-usign-v1
-
-      - name: Install UPX
-        run: sudo apt-get update -qq && sudo apt-get install -y -qq upx-ucl
-
-      # Build the SPA-embedded, static-musl, UPX'd shaterd for BOTH arches and
-      # stage dist/shaterd-<a>.upx into openwrt/shaterd/files/. MUST run before
-      # the SDK package build (the openwrt/shaterd package installs the staged
-      # artifact). VERSION is stamped into constant.Version. On an exact
-      # node_modules cache hit, --fast skips the redundant `npm ci`.
-      - name: Build & stage shaterd artifact
-        env:
-          NPM_CACHE_HIT: ${{ steps.npm-cache.outputs.cache-hit }}
-        run: |
-          set -eu
-          if [ "${GITHUB_REF#refs/tags/}" != "$GITHUB_REF" ]; then
-            V="${GITHUB_REF#refs/tags/}"
-          else
-            V="v0.2.0-dev"
-          fi
-          FAST=""
-          if [ "${NPM_CACHE_HIT:-}" = "true" ]; then FAST="--fast"; fi
-          echo "shaterd version: $V (npm cache hit: ${NPM_CACHE_HIT:-false})"
-          bash scripts/build-shaterd.sh "$V" $FAST
-
-      # Compile the 4 packages through the arch-matched OpenWrt SDK and produce a
-      # signed per-arch opkg feed (Packages + Packages.gz + Packages.sig + .ipk).
-      - name: Build signed feed (SDK)
-        env:
-          KEY_BUILD: ${{ secrets.KEY_BUILD }}
-        run: bash ci/build-feed.sh "${{ matrix.arch }}" "${{ matrix.sdk }}" "out/${{ matrix.arch }}"
-
-      - name: Show feed
-        run: ls -l "out/${{ matrix.arch }}" && cat "out/${{ matrix.arch }}/Packages"
-
-      - name: Upload feed artifact
-        # v4 uses an artifact backend Gitea Actions does not implement
-        # (GHESNotSupportedError); v3 works on Gitea's act_runner.
-        uses: actions/upload-artifact@v3
-        with:
-          name: shater-${{ matrix.arch }}
-          path: out/${{ matrix.arch }}/*
-          if-no-files-found: error
-
  # ---------------------------------------------------------------------------
-  # APK lane (additive): the same 4 packages through the ImmortalWrt 25.12
-  # apk-SDK for the 25.12/apk fleet (BananaWRT 25.12-mtk-vendor routers + the
-  # future 25.12 VM). Produces a per-arch apk repo dir: *.apk + EC-signed
-  # packages.adb + shater-apk.pem. Artifact prefix `apkfeed-` (NOT `shater-`)
-  # so the opkg release job's `artifacts/shater-*` glob never picks these up.
+  # Build the 4 packages through the ImmortalWrt 25.12 apk-SDK for the 25.12/apk
+  # fleet (BPI-R3 mini on BananaWRT 25.12-mtk-vendor, BPI-R4 on OpenWrt 25.12,
+  # and the testbed VM). Produces a per-arch apk repo dir: *.apk + EC-signed
+  # packages.adb + shater-apk.pem, uploaded as the artifact `apkfeed-<arch>`.
  build-apk:
    name: apk ${{ matrix.arch }}
    runs-on: ubuntu-latest
@@ -253,8 +137,14 @@ jobs:
          - arch: aarch64_cortex-a53  # BPI-R3 mini (BananaWRT 25.12-mtk-vendor) + BPI-R4
            sdk_url: https://downloads.immortalwrt.org/releases/25.12.1/targets/mediatek/filogic/immortalwrt-sdk-25.12.1-mediatek-filogic_gcc-14.3.0_musl.Linux-x86_64.tar.zst
    steps:
+      # fetch-depth: 0 — the package version is DERIVED from the git tag
+      # (ci/version.sh: nearest `vX.Y.Z` + commits since it). The default
+      # shallow checkout has neither tags nor ancestry, so `git describe` would
+      # fail and every dispatch build would fall back to 0.0.0.
      - name: Checkout
        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0

      # scripts/build-shaterd.sh builds the AmneziaWG-patched wireguard-go via a
      # go.mod `replace => ./submodules/wireguard-go`, so that submodule must be
@@ -263,6 +153,13 @@ jobs:
      - name: Init wireguard-go submodule (awg)
        run: git submodule update --init --depth 1 submodules/wireguard-go

+      # THE version step (bug B4). One computation, used by both the binary
+      # (constant.Version) and the three tag-versioned packages, exported to
+      # every later step of this job:
+      #   tag vX.Y.Z -> X.Y.Z-r1 ; off-tag -> <last tag>-r<commits+1>
+      - name: Compute version from git tag
+        run: bash ci/version.sh --env >> "$GITHUB_ENV"
+
      - name: Set up Go
        uses: actions/setup-go@v5
        with:
@@ -336,25 +233,36 @@ jobs:
          restore-keys: |
            feeds-apk-

+      # D23 — the shipped tag set is a TRIMMED subset (scripts/router-tags.sh);
+      # everything else in CI builds with the full upstream set, so without this
+      # step the one combination we actually ship is never exercised. That is how
+      # `with_gvisor` was trimmed while `with_wireguard` stayed and every shipped
+      # binary answered a WireGuard node with "gVisor is not included in this
+      # build" (2026-07-25). The check runs the declared-feature/tag comparison
+      # and then constructs one node of every declared protocol through box.New
+      # UNDER THE SHIPPED TAGS. It runs before the artifact build so a tag trim
+      # that breaks a feature fails the release instead of shipping.
+      - name: Verify the shipped build-tag set (D23)
+        run: bash scripts/check-router-tags.sh
+
      - name: Install UPX
        run: sudo apt-get update -qq && sudo apt-get install -y -qq upx-ucl

-      # Same artifact-order contract as the opkg lane: the SPA-embedded shaterd
-      # binary is built OUT of the SDK and staged before the package build.
+      # Artifact-order contract: the SPA-embedded shaterd binary is built OUT of
+      # the SDK and staged into openwrt/shaterd/files/ BEFORE the package build
+      # (the openwrt/shaterd package only installs the staged artifact).
+      # $SHATER_VERSION (from the version step above) is stamped into
+      # constant.Version, so the binary and the package agree. On an exact
+      # node_modules cache hit, --fast skips the redundant `npm ci`.
      - name: Build & stage shaterd artifact
        env:
          NPM_CACHE_HIT: ${{ steps.npm-cache.outputs.cache-hit }}
        run: |
          set -eu
-          if [ "${GITHUB_REF#refs/tags/}" != "$GITHUB_REF" ]; then
-            V="${GITHUB_REF#refs/tags/}"
-          else
-            V="v0.2.0-dev"
-          fi
          FAST=""
          if [ "${NPM_CACHE_HIT:-}" = "true" ]; then FAST="--fast"; fi
-          echo "shaterd version: $V (npm cache hit: ${NPM_CACHE_HIT:-false})"
-          bash scripts/build-shaterd.sh "$V" $FAST
+          echo "shaterd version: $SHATER_VERSION / package ${SHATER_PKG_VERSION}-r${SHATER_PKG_RELEASE} (npm cache hit: ${NPM_CACHE_HIT:-false})"
+          bash scripts/build-shaterd.sh $FAST

      # Compile the 4 packages as .apk through the ImmortalWrt 25.12 SDK and
      # sign the per-arch packages.adb with the EC key (secret KEY_APK).
@@ -377,113 +285,21 @@ jobs:
          if-no-files-found: error

  # ---------------------------------------------------------------------------
-  # Publish once both arches are built. Rolling `latest` on dispatch, a versioned
-  # release on a `vX.Y.Z` tag. Self-contained (curl -> Gitea API).
-  release:
-    name: release
-    needs: build
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      - name: Download all arch feeds
-        uses: actions/download-artifact@v3
-        with:
-          path: artifacts
-
-      - name: Assemble release assets
-        id: assets
-        run: |
-          set -eu
-          mkdir -p release
-          # For each downloaded arch feed: one ready-to-serve tarball + loose ipks.
-          for d in artifacts/shater-*; do
-            [ -d "$d" ] || continue
-            arch="${d#artifacts/shater-}"
-            tar -C "$d" -czf "release/shater-feed-${arch}.tar.gz" .
-            # loose .ipk for direct `opkg install <url>` (dedupe shared _all ipks by name)
-            for ipk in "$d"/*.ipk; do
-              [ -e "$ipk" ] || continue
-              cp -n "$ipk" "release/$(basename "$ipk")"
-            done
-          done
-          # ship the feed's public key so routers can verify (see docs-shater/INSTALL.md)
-          cp -f dist/shater-feed.pub release/shater-feed.pub
-          ls -l release
-          echo "count=$(ls release | wc -l)" >> "$GITHUB_OUTPUT"
-
-      # restore the prebuilt usign binary (skips apt + cmake + clone + build)
-      - name: Cache CI tools (usign)
-        uses: actions/cache@v3.3.2
-        with:
-          path: .cache/tools
-          key: tools-usign-v1
-
-      - name: Install usign (feed signer)
-        run: bash ci/install-usign.sh
-
-      - name: Build & sign combined opkg feed index
-        # One Packages/Packages.gz over ALL loose .ipk (every arch + arch=all),
-        # with basename Filenames. opkg filters by Architecture, so a single
-        # release URL serves every device: BPI routers pick aarch64_cortex-a53 +
-        # all, the x86 testbed picks x86_64 + all. Signed with KEY_BUILD so
-        # routers keep check_signature on. This is what makes the release directly
-        # consumable as an `src/gz` feed (see docs-shater/INSTALL.md).
-        env:
-          KEY_BUILD: ${{ secrets.KEY_BUILD }}
-        run: bash ci/make-index.sh release
-
-      - name: Determine release identity
-        id: rel
-        run: |
-          set -eu
-          if [ "${GITHUB_REF#refs/tags/}" != "$GITHUB_REF" ]; then
-            echo "tag=${GITHUB_REF#refs/tags/}" >> "$GITHUB_OUTPUT"
-            echo "name=shater ${GITHUB_REF#refs/tags/}" >> "$GITHUB_OUTPUT"
-            echo "prerelease=false" >> "$GITHUB_OUTPUT"
-            echo "rolling=false" >> "$GITHUB_OUTPUT"
-          else
-            echo "tag=latest" >> "$GITHUB_OUTPUT"
-            echo "name=shater latest (main)" >> "$GITHUB_OUTPUT"
-            echo "prerelease=true" >> "$GITHUB_OUTPUT"
-            echo "rolling=true" >> "$GITHUB_OUTPUT"
-          fi
-
-      - name: Publish Gitea release
-        env:
-          TOKEN: ${{ secrets.RELEASE_TOKEN != '' && secrets.RELEASE_TOKEN || github.token }}
-          TAG: ${{ steps.rel.outputs.tag }}
-          NAME: ${{ steps.rel.outputs.name }}
-          PRERELEASE: ${{ steps.rel.outputs.prerelease }}
-          ROLLING: ${{ steps.rel.outputs.rolling }}
-          BODY: |
-            Automated build. Packages: shaterd + byedpi (per-arch), shater-core +
-            luci-app-shater (arch=all).
-            Targets: x86_64 (testbed) and aarch64_cortex-a53 (BPI-R3 + BPI-R4, mediatek/filogic).
-
-            ── Add as an opkg feed (recommended — then `opkg upgrade` just works) ──
-            This release is itself a SIGNED package feed; opkg filters by
-            architecture, so the same lines work on every device:
-              wget -O /etc/opkg/keys/5ac4b177689cb8e0 https://git.qomar.pw/omar/shater/releases/download/latest/shater-feed.pub
-              echo "src/gz shater https://git.qomar.pw/omar/shater/releases/download/latest" >> /etc/opkg/customfeeds.conf
-              opkg update
-              opkg install luci-app-shater        # pulls shater-core + shaterd too
-            The public-key install is one-time; after it, `opkg update/upgrade`
-            verify the signature with check_signature left on. Full guide: docs-shater/INSTALL.md.
-
-            ── Or install the loose .ipk directly / from the tarball feed ──
-              wget -O /tmp/f.tgz <this release>/shater-feed-aarch64_cortex-a53.tar.gz
-              mkdir -p /tmp/shater && tar -C /tmp/shater -xzf /tmp/f.tgz
-              opkg install /tmp/shater/luci-app-shater_*_all.ipk
-        run: bash ci/gitea-release.sh release/*
-
-  # ---------------------------------------------------------------------------
-  # Publish the apk lane: ONE release PER ARCH (apk package filenames carry no
-  # arch, and apk fetches `<name>-<ver>.apk` relative to the packages.adb URL —
-  # a flat multi-arch release would collide). Rolling `apk-latest-<arch>` on
-  # dispatch, `apk-<tag>-<arch>` on a version tag. The tags do NOT match the
-  # workflow's `v*` trigger, so publishing them cannot re-trigger the build.
+  # Publish: ONE release PER ARCH (apk package filenames carry no arch, and apk
+  # fetches `<name>-<ver>.apk` relative to the packages.adb URL — a flat
+  # multi-arch release would collide). Every run refreshes the ROLLING pointer
+  # `apk-latest-<arch>`; a `vX.Y.Z` tag run ALSO publishes the pinnable
+  # `apk-vX.Y.Z-<arch>`. The tags do NOT match the workflow's `v*` trigger, so
+  # publishing them cannot re-trigger the build.
+  #
+  # WHY THE ROLLING RELEASE IS PUBLISHED ON TAG RUNS TOO (fixed 2026-07-25):
+  # it used to be an either/or — `TAG=apk-latest-<arch>` on dispatch, ELSE
+  # `TAG=apk-<ver>-<arch>` — so once releases moved to tag pushes the rolling
+  # pointer was never written again. It froze at 0.2.0 (published 2026-07-24)
+  # while v0.2.9/v0.2.10 published fine, and every router whose
+  # /etc/apk/repositories.d/shater.list points at the rolling URL kept getting a
+  # successful, silent `apk update` with nothing new. Rolling is the whole point
+  # of that URL, so it is now written unconditionally and asserted afterwards.
  release-apk:
    name: release apk
    needs: build-apk
@@ -501,6 +317,9 @@ jobs:
        with:
          path: artifacts

+      # Identity of the VERSIONED release only. The rolling pointer is published
+      # on every run with fixed prerelease=true/rolling=true, so it needs nothing
+      # from here.
      - name: Determine release identity
        id: rel
        run: |
@@ -522,25 +341,95 @@ jobs:
          PRERELEASE: ${{ steps.rel.outputs.prerelease }}
          ROLLING: ${{ steps.rel.outputs.rolling }}
        run: |
-          set -eu
+          set -euo pipefail
          for d in artifacts/apkfeed-*; do
            [ -d "$d" ] || continue
            arch="${d#artifacts/apkfeed-}"
-            if [ "$VER" = latest ]; then TAG="apk-latest-$arch"; else TAG="apk-$VER-$arch"; fi
+            ROLL="apk-latest-$arch"
+
+            # The version we just built, read straight off the artifact
+            # (`shaterd-<ver>-r<rel>.apk`). NOT recomputed with ci/version.sh:
+            # this job checks out shallow, so it has no tags to describe from.
+            pkg=""
+            for a in "$d"/shaterd-*.apk; do
+              if [ -f "$a" ]; then pkg="$(basename "$a")"; fi
+            done
+            [ -n "$pkg" ] || { echo "[release-apk] ERROR: no shaterd-*.apk in $d"; exit 11; }
+            want="${pkg#shaterd-}"; want="${want%.apk}"
+            echo "[release-apk] arch=$arch built version=$want"
+
            BODY="Automated apk (OpenWrt/ImmortalWrt 25.12+) package repo for \`$arch\`.
          Packages: shaterd + byedpi (per-arch), shater-core + luci-app-shater (arch=all).
+          This build: \`$want\`.
          The index \`packages.adb\` is EC-signed; trust anchor \`shater-apk.pem\` (also in \`dist/\`).

-          ── Add as an apk repository (auto-updates via \`apk upgrade\`) ──
-            wget -O /etc/apk/keys/shater-apk.pem https://git.qomar.pw/omar/shater/releases/download/$TAG/shater-apk.pem
+          ── Add as an apk repository (rolling — install once, then just update) ──
+            wget -O /etc/apk/keys/shater-apk.pem https://git.qomar.pw/omar/shater/releases/download/$ROLL/shater-apk.pem
            echo \"https://git.qomar.pw/omar/shater/releases/download/apk-latest-\$(cat /etc/apk/arch)/packages.adb\" > /etc/apk/repositories.d/shater.list
            apk update
            apk add luci-app-shater      # pulls shater-core + shaterd too
            apk add byedpi               # optional: ByeDPI desync egress
-          Update: apk update && apk upgrade shaterd shater-core luci-app-shater byedpi
-          Full guide: docs-shater/INSTALL.md §6. The opkg/24.10 feed lives in the \`latest\` release."
-            echo "[release-apk] publishing $TAG from $d"
-            TAG="$TAG" NAME="shater apk $VER ($arch)" BODY="$BODY" \
-              PRERELEASE="$PRERELEASE" ROLLING="$ROLLING" \
+          \`apk-latest-<arch>\` is a MOVING pointer: every release run replaces its
+          assets, so the same repo line keeps serving the newest build. To pin a
+          version instead, point the repo line at
+          \`.../download/apk-vX.Y.Z-\$(cat /etc/apk/arch)/packages.adb\` — then the
+          file must be edited by hand for each upgrade.
+          ── Update — ALWAYS name the packages, NEVER a bare \`apk upgrade\` ──
+            apk update
+            apk upgrade shaterd shater-core luci-app-shater byedpi
+          A bare \`apk upgrade\` reconciles EVERY installed package against every
+          configured repo and can downgrade unrelated system packages; naming them
+          upgrades only those (apk-tools 3: \"If list of packages is provided, only
+          those packages are upgraded along with needed dependencies\").
+          Full guide: docs-shater/INSTALL.md §5."
+
+            # 1) the pinnable versioned release (tag runs only)
+            if [ "$VER" != latest ]; then
+              echo "[release-apk] publishing apk-$VER-$arch from $d"
+              TAG="apk-$VER-$arch" NAME="shater apk $VER ($arch)" BODY="$BODY" \
+                PRERELEASE="$PRERELEASE" ROLLING="$ROLLING" \
+                bash ci/gitea-release.sh "$d"/*
+            fi
+
+            # 2) the rolling pointer — ALWAYS, tag run included. ci/gitea-release.sh
+            #    deletes the existing release before recreating it, so the old
+            #    version's assets are REPLACED, never accumulated (two versions of
+            #    one package in one index would let apk choose, not us).
+            echo "[release-apk] publishing $ROLL from $d"
+            TAG="$ROLL" NAME="shater apk latest ($arch)" BODY="$BODY" \
+              PRERELEASE=true ROLLING=true \
              bash ci/gitea-release.sh "$d"/*
+
+            # 3) ASSERT the rolling release really serves THIS build — same class
+            #    of check as ci/sdk-build-apk.sh's package-version assert, and for
+            #    the same reason: the previous failure mode was silent. Reads the
+            #    published release back over the API and requires our three
+            #    tag-versioned packages at $want, the index, the key — and NO
+            #    left-over package asset at any other version.
+            api="$GITHUB_SERVER_URL/api/v1/repos/$GITHUB_REPOSITORY/releases/tags/$ROLL"
+            got="$(curl -fsS -H "Authorization: token $TOKEN" "$api" \
+                   | tr '{},' '\n\n\n' \
+                   | sed -n 's/.*"name"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p' | sort -u)" || {
+              echo "[release-apk] ERROR: cannot read back $ROLL from the API"; exit 12; }
+            echo "[release-apk] $ROLL assets: $(printf '%s ' $got)"
+            # here-string, NOT `printf | grep -q`: under `pipefail` the early
+            # exit of grep -q can SIGPIPE the writer and fail a passing check.
+            for f in "shaterd-$want.apk" "shater-core-$want.apk" \
+                     "luci-app-shater-$want.apk" packages.adb shater-apk.pem; do
+              grep -qxF "$f" <<<"$got" || {
+                echo "[release-apk] ERROR: $ROLL does not contain '$f' after publish."
+                echo "        A router pinned to the rolling URL would have silently"
+                echo "        stayed on its old version with a successful apk update."
+                exit 13; }
+            done
+            stale="$(grep -E '^(shaterd|shater-core|luci-app-shater)-.*\.apk$' <<<"$got" \
+                     | grep -vxF -e "shaterd-$want.apk" -e "shater-core-$want.apk" \
+                                 -e "luci-app-shater-$want.apk" || true)"
+            [ -z "$stale" ] || {
+              echo "[release-apk] ERROR: $ROLL still holds stale package assets:"
+              printf '          %s\n' $stale
+              echo "        Two versions of one package in one feed = apk picks by its"
+              echo "        own rules, not by our intent."
+              exit 14; }
+            echo "[release-apk] OK — $ROLL serves $want"
          done
@@ -63,7 +63,7 @@ nul
 /venv/
 /test/cache.db

-# feed artifacts (tracked public key dist/shater-feed.pub is force-added)
+# feed artifacts (the tracked apk trust anchor dist/shater-apk.pem is force-added)
 /dist/

 # local agent config (CLAUDE.md is deliberately tracked; .claude local settings are not)
@@ -49,29 +49,22 @@ Full list with MVP/T1/T2 tags — [`docs-shater/FEATURES.md`](docs-shater/FEATUR

 ## Install

-Two signed feeds. Pick by the router's OpenWrt version. Verbatim commands and the
-manual `.ipk`/`.apk` install are in [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md).
-
-**opkg (OpenWrt 24.10):**
+One signed **apk** feed (OpenWrt / ImmortalWrt / BananaWRT **25.12+**), one
+release per arch. Verbatim commands, the manual `.apk` install and the
+rolling-vs-pinned choice are in
+[`docs-shater/INSTALL.md`](docs-shater/INSTALL.md).

 ```sh
-wget -O /etc/opkg/keys/5ac4b177689cb8e0 \
-  https://git.qomar.pw/omar/shater/releases/download/latest/shater-feed.pub
-echo "src/gz shater https://git.qomar.pw/omar/shater/releases/download/latest" \
-  >> /etc/opkg/customfeeds.conf
-opkg update && opkg install luci-app-shater   # -> shater-core -> shaterd
-```
-
-**apk (OpenWrt / ImmortalWrt / BananaWRT 25.12+):**
-
-```sh
-wget -O /etc/apk/keys/shater-apk.pem \
-  "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/shater-apk.pem"
-echo "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/packages.adb" \
-  > /etc/apk/repositories.d/shater.list
+wget -O /etc/apk/keys/shater-apk.pem   "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/shater-apk.pem"
+echo "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/packages.adb"   > /etc/apk/repositories.d/shater.list
 apk update && apk add luci-app-shater         # -> shater-core -> shaterd
 ```

+`apk-latest-<arch>` is a moving pointer refreshed by every release run — install
+once and `apk update && apk upgrade shaterd shater-core luci-app-shater byedpi`
+keeps the router current. Point the repo line at `apk-vX.Y.Z-<arch>` instead to
+pin a build; that file then has to be edited by hand for every upgrade.
+
 shater ships **inert** (globals off) so install never breaks connectivity. After
 configuring nodes/rules: `uci set shater.globals.enabled=1 && uci commit shater`,
 then `shaterd apply` and `shaterd confirm`.
@@ -91,16 +84,17 @@ into `openwrt/shaterd/files/`. Details in
 | `panel/` | Admin SPA (Vite + React + TS) and its Go server |
 | `openwrt/` | Packages: `shaterd`, `shater-core`, `luci-app-shater`, `byedpi` |
 | `docs-shater/` | Product documentation |
-| `scripts/`, `ci/`, `.gitea/workflows/` | Build script, feed/release scripts, CI |
+| `scripts/`, `ci/`, `.gitea/workflows/` | Build script, apk feed/release scripts, CI |
 | `SPECS/`, `docs-lx/` | Engine-fork constitution/specs and feature-config reference |
 | `docs/`, `mkdocs.yml` | **Upstream** sing-box docs (mkdocs) — kept as-is |
 | `adapter/ cmd/ dns/ route/ option/ protocol/ transport/ …` | sing-box-lx engine tree |

 ## CI, upstream & license

-CI (`.gitea/workflows/release.yml`) builds all 4 packages and publishes signed
-feeds: opkg (usign, key `5ac4b177689cb8e0`) and apk (EC key `shater-apk.pem`). A
-`vX.Y.Z` tag → versioned release; `workflow_dispatch` → rolling `latest`.
+CI (`.gitea/workflows/release.yml`) builds all 4 packages and publishes a signed
+per-arch apk repo (EC key `shater-apk.pem`). A `vX.Y.Z` tag → the pinnable
+`apk-vX.Y.Z-<arch>`; every run also refreshes the rolling `apk-latest-<arch>` and
+asserts over the API that it really serves the version just built.

 The engine is the **sing-box-lx** fork — a thin downstream of upstream sing-box that
 lives by **rebase, never merge**; its constitution is
@@ -10,7 +10,7 @@

 [![License: GPL-3.0](https://img.shields.io/badge/license-GPL--3.0-blue.svg)](LICENSE)
 ![targets: x86_64 · aarch64_cortex-a53](https://img.shields.io/badge/targets-x86__64%20%C2%B7%20aarch64__cortex--a53-brightgreen.svg)
-![feeds: opkg 24.10 · apk 25.12](https://img.shields.io/badge/feeds-opkg%2024.10%20%C2%B7%20apk%2025.12-orange.svg)
+![feed: apk 25.12+](https://img.shields.io/badge/feed-apk%2025.12%2B-orange.svg)

 ---

@@ -128,36 +128,15 @@ data-plane, DNS-flow, apply-flow) — в [`docs-shater/ARCHITECTURE.md`](docs-sh

 ## Установка

-shater поставляется двумя подписанными фидами. Выберите по версии OpenWrt на роутере:
-
- **OpenWrt 24.10** → фид **opkg** (`.ipk`, `Packages.gz`, ключ usign).
- **OpenWrt / ImmortalWrt / BananaWRT 25.12+** → фид **apk** (`.apk`, `packages.adb`,
-  EC-ключ).
+shater поставляется одним подписанным **apk-фидом** (OpenWrt / ImmortalWrt /
+BananaWRT **25.12+**: `.apk`, индекс `packages.adb`, EC-ключ в `/etc/apk/keys/`).
+Старый opkg-фид (`.ipk`, 24.10) снят — оба наших роутера на 25.12 с apk-tools 3,
+бинаря `opkg` там просто нет (`docs-shater/DECISIONS.md` D22).

 Пакеты ставятся по зависимостям: `shaterd` → `shater-core` → `luci-app-shater`
 (+ опциональный `byedpi`). `shaterd` подтягивается автоматически как зависимость.

-### Путь A — фид opkg (OpenWrt 24.10)
-
-```sh
-# 1) доверяем ключу фида — ИМЯ файла обязано равняться отпечатку usign-ключа.
-wget -O /etc/opkg/keys/5ac4b177689cb8e0 \
-  https://git.qomar.pw/omar/shater/releases/download/latest/shater-feed.pub
-
-# 2) добавляем фид (один URL обслуживает все арки).
-echo "src/gz shater https://git.qomar.pw/omar/shater/releases/download/latest" \
-  >> /etc/opkg/customfeeds.conf
-
-# 3) обновляемся и ставим (shaterd подтянется как зависимость).
-opkg update
-opkg install luci-app-shater        # -> shater-core -> shaterd
-opkg install byedpi                 # опционально: ByeDPI desync-egress
-```
-
-Обновление: `opkg update && opkg upgrade shaterd shater-core luci-app-shater byedpi`
-(обновляйте только эти четыре пакета, не системные).
-
-### Путь B — фид apk (OpenWrt / ImmortalWrt / BananaWRT 25.12+)
+### Фид apk

 `/etc/apk/arch` сам выбирает нужный per-arch релиз (apk-релизы раздельны по арке):

@@ -176,11 +155,38 @@ apk add luci-app-shater        # -> shater-core -> shaterd
 apk add byedpi                 # опционально: ByeDPI desync-egress
 ```

-Обновление: `apk update && apk upgrade shaterd shater-core luci-app-shater byedpi`.
+Обновление — **перечисляйте пакеты явно, голый `apk upgrade` не запускайте**: без
+аргументов apk пересобирает состояние ВСЕХ установленных пакетов по ВСЕМ
+подключённым репозиториям и может задеть (в т.ч. откатить) посторонние системные
+пакеты.

-> Полные инструкции — раздельная установка из `.ipk`/`.apk` вручную, закрепление
-> версии (`vX.Y.Z` / `apk-vX.Y.Z-<arch>`), совместимость с BananaWRT
-> `25.12-mtk-vendor` — в [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md).
+```sh
+apk update
+apk upgrade shaterd shater-core luci-app-shater byedpi
+# эквивалент, дополнительно закрепляющий пакеты в world:
+# apk add -u shaterd shater-core luci-app-shater byedpi
+```
+
+Документация apk-tools 3 про `apk upgrade`: *«If list of packages is provided,
+only those packages are upgraded along with needed dependencies»*. Проверить
+установленные версии: `apk list -I shaterd shater-core luci-app-shater byedpi`.
+
+> **Роллинг или фиксация — это выбор URL в `shater.list`.** `apk-latest-<arch>`
+> — движущийся указатель: каждый релизный прогон заменяет его ассеты, поэтому
+> «поставил и забыл»: `apk update` сам видит новую сборку. `apk-vX.Y.Z-<arch>` —
+> фиксация на конкретной сборке: роутер не получит ничего нового, пока
+> `/etc/apk/repositories.d/shater.list` не отредактируют руками — на каждом
+> роутере и на каждый релиз. На `mini_router` сознательно прописан
+> версионированный URL, и ручная правка — его цена. Подробнее —
+> [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md) §5.1.
+
+> Версии пакетов CI берёт из git-тега (`vX.Y.Z` → `X.Y.Z-r1`, сборка вне тега →
+> `X.Y.Z-r<коммитов+1>`), поэтому каждая новая сборка действительно видна
+> менеджеру пакетов как новая. Подробности — `docs-shater/INSTALL.md` §2.1.
+
+> Полные инструкции — ручная установка из `.apk`, фиксация версии
+> (`apk-vX.Y.Z-<arch>`), совместимость с BananaWRT `25.12-mtk-vendor` — в
+> [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md).

 ### Включение

@@ -234,8 +240,8 @@ arm64}` с musl-static набором тегов (`CGO_ENABLED=0 GOOS=linux`), s
 | `openwrt/` | Пакеты: `shaterd`, `shater-core`, `luci-app-shater`, `byedpi` |
 | `docs-shater/` | Документация продукта (см. таблицу ниже) |
 | `scripts/` | `build-shaterd.sh` — сборка ship-артефакта |
-| `ci/` | Скрипты сборки фидов и релизов (SDK, usign/EC, Gitea API) |
-| `.gitea/workflows/` | `release.yml` — CI: сборка пакетов + подписанные фиды opkg/apk |
+| `ci/` | Скрипты сборки apk-фида и релизов (SDK, EC-подпись, Gitea API) |
+| `.gitea/workflows/` | `release.yml` — CI: сборка пакетов + подписанный apk-фид |
 | `SPECS/` | Конституция форка движка и спеки (Spec Kit) |
 | `docs-lx/` | Справочник конфигурации фич движка (`lx-config.md`, `.ru.md`) |
 | `lx-test/`, `submodules/` | Примеры конфигов движка и submodule AmneziaWG-рантайма |
@@ -249,16 +255,17 @@ arm64}` с musl-static набором тегов (`CGO_ENABLED=0 GOOS=linux`), s
 CI на **Gitea Actions** (`.gitea/workflows/release.yml`) собирает все 4 пакета и
 публикует **подписанные фиды**:

- **opkg (24.10):** один комбинированный релиз, подписан usign-ключом (публичный
-  `dist/shater-feed.pub`, отпечаток `5ac4b177689cb8e0`; секрет — в Gitea-secret
-  `KEY_BUILD`).
- **apk (25.12+):** параллельная линия, **по релизу на арку**, подписан EC-ключом
-  (`dist/shater-apk.pem`; секрет — `KEY_APK`).
+- **apk (25.12+)** — единственный формат: **по релизу на арку**, индекс
+  `packages.adb` подписан EC-ключом (публичный `dist/shater-apk.pem`; секрет — в
+  Gitea-secret `KEY_APK`).

-Триггеры: push тега **`vX.Y.Z`** → версионный релиз; `workflow_dispatch` →
-плавающий `latest`/`apk-latest-<arch>` (всегда свежий фид). Публикация — через
-Gitea API (`ci/gitea-release.sh`). Ключи **никогда не перегенерируются** — это
-инвалидировало бы доверие на всех развёрнутых роутерах.
+Триггеры: push тега **`vX.Y.Z`** → версионный релиз `apk-vX.Y.Z-<arch>`;
+`workflow_dispatch` → только роллинг. Роллинг `apk-latest-<arch>` обновляется
+**на каждом прогоне**, включая теговый, и после публикации проверяется через API:
+в нём обязаны лежать наши три пакета ровно собранной версии и ни одного ассета
+другой версии. Публикация — через Gitea API (`ci/gitea-release.sh`). Ключ
+**никогда не перегенерируется** — это инвалидировало бы доверие на всех
+развёрнутых роутерах.

 ---

@@ -283,7 +290,7 @@ build-тегами и живущий **ребейзом на каждый upstre
 | Документ | О чём |
 |----------|-------|
 | [`docs-shater/CONTEXT.md`](docs-shater/CONTEXT.md) | **Начните здесь** — контекст проекта, история v0.1→v0.2, testbed/инфра |
-| [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md) | Сборка ship-артефакта и установка обоих фидов (opkg/apk) |
+| [`docs-shater/INSTALL.md`](docs-shater/INSTALL.md) | Сборка ship-артефакта и установка apk-фида (роллинг/фиксация) |
 | [`docs-shater/ARCHITECTURE.md`](docs-shater/ARCHITECTURE.md) | One-binary дизайн, auth-handoff, data/DNS/apply-потоки (диаграммы) |
 | [`docs-shater/FEATURES.md`](docs-shater/FEATURES.md) | Полный список фич с тегами MVP/T1/T2 |
 | [`docs-shater/ROADMAP.md`](docs-shater/ROADMAP.md) | Фазовый план |
@@ -1,6 +1,6 @@
 #!/bin/sh
-# ci/build-feed-apk.sh — build the signed **apk** feed for ONE arch (the 25.12
-# lane — additive next to ci/build-feed.sh, which stays the opkg/24.10 lane).
+# ci/build-feed-apk.sh — build the signed **apk** feed for ONE arch (25.12+;
+# the only packaging lane shater has — see docs-shater/DECISIONS.md D22).
 #
 # Usage:  ci/build-feed-apk.sh <ARCH> <SDK_URL> <OUTDIR>
 #   e.g.  ci/build-feed-apk.sh aarch64_cortex-a53 \
@@ -10,14 +10,14 @@
 # This is the per-arch entrypoint the Gitea workflow's `build-apk` job calls.
 # It runs on the CI RUNNER and:
 #   1. asserts the prebuilt shaterd binary for this arch was already staged by
-#      scripts/build-shaterd.sh (same artifact-order contract as the opkg lane);
-#   2. drives a plain `debian:bookworm` container (workspace shared via
-#      `--volumes-from`, same trick as ci/build-feed.sh) that downloads the
-#      ImmortalWrt 25.12 apk-SDK tarball and runs ci/sdk-build-apk.sh in it:
-#      compile the 4 packages as .apk, then `apk mkndx --sign` the per-arch
-#      `packages.adb` index. Unlike the usign lane (index signed on the runner),
-#      apk indexing NEEDS the SDK's host `apk` tool, so index+sign happen inside
-#      the container.
+#      scripts/build-shaterd.sh (the artifact-order contract);
+#   2. drives a plain `debian:bookworm` container (the job's workspace volume is
+#      shared into it with `--volumes-from $(hostname)`; a bare `-v $PWD:...`
+#      points at a host path that does not exist under act_runner's DinD) that
+#      downloads the ImmortalWrt 25.12 apk-SDK tarball and runs
+#      ci/sdk-build-apk.sh in it: compile the 4 packages as .apk, then
+#      `apk mkndx --sign` the per-arch `packages.adb` index. Indexing NEEDS the
+#      SDK's host `apk` tool, so index+sign happen inside the container.
 #
 # Why the ImmortalWrt SDK (not openwrt/sdk images): the 25.12 fleet runs
 # BananaWRT 25.12-mtk-vendor = ImmortalWrt 25.12 base (target mediatek/filogic,
@@ -25,9 +25,9 @@
 # mediatek-filogic 25.12 tag — hence the official SDK tarball.
 #
 # Env:
-#   KEY_APK   EC (prime256v1) PRIVATE key PEM (Gitea repo secret — the apk analog
-#             of KEY_BUILD). If set, packages.adb carries an embedded signature
-#             verifiable by dist/shater-apk.pem (routers: /etc/apk/keys/).
+#   KEY_APK   EC (prime256v1) PRIVATE key PEM (Gitea repo secret). If set,
+#             packages.adb carries an embedded signature verifiable by
+#             dist/shater-apk.pem (routers: /etc/apk/keys/).
 #             If unset, an UNSIGNED index is produced (warning; not shippable —
 #             apk signatures are effectively mandatory).
 set -eu
@@ -55,6 +55,15 @@ fi

 chmod +x "$REPO"/ci/*.sh 2>/dev/null || true

+# --- 0.4) package version from the git tag ------------------------------------
+# The workflow puts these in the job env via `ci/version.sh --env >>
+# $GITHUB_ENV`; recompute here when run standalone. Passed into the container
+# below and re-exported to the unprivileged build user in ci/sdk-build-apk.sh.
+if [ -z "${SHATER_PKG_VERSION:-}" ] || [ -z "${SHATER_PKG_RELEASE:-}" ]; then
+  eval "$(sh "$REPO/ci/version.sh" --env)"
+fi
+echo "[apk-feed] package version: ${SHATER_PKG_VERSION}-r${SHATER_PKG_RELEASE}"
+
 # --- 0.5) runner-side caches --------------------------------------------------
 # All under $REPO/.cache so (a) actions/cache in the workflow can persist them
 # between runs and (b) the nested container sees them via --volumes-from.
@@ -63,7 +72,7 @@ chmod +x "$REPO"/ci/*.sh 2>/dev/null || true
 #         SDK; PKG_HASH still verifies every file, so stale = re-downloaded.
 #   apt/  debian:bookworm .deb archives for the host-deps install.
 # The nested container runs the build as an unprivileged user -> must be writable
-# (same reason as the chmod 0777 "$OUT" in ci/build-feed.sh).
+# (same reason as the chmod 0777 "$OUT" above).
 CACHE="$REPO/.cache"
 mkdir -p "$CACHE/sdk" "$CACHE/dl" "$CACHE/apt"
 chmod -R a+rwX "$CACHE/dl" "$CACHE/apt" 2>/dev/null || true
@@ -86,14 +95,16 @@ sh "$REPO/ci/fetch-sdk.sh" "$SDK_URL" "$SDK_TAR"

 # --- 1) SDK build + index + sign inside a debian container -------------------
 # `--volumes-from $(hostname)` shares THIS job container's workspace volume into
-# the nested container (see ci/build-feed.sh for why a bare -v does not work on
-# the act_runner DinD setup).
+# the nested container: a bare `-v $PWD:...` points at a host path that does not
+# exist under the act_runner DinD setup.
 echo "[apk-feed] SDK build arch=$ARCH (ImmortalWrt 25.12 apk-SDK)"
 docker pull -q debian:bookworm
 docker run --rm --volumes-from "$(hostname)" \
  -e ARCH="$ARCH" -e REPO="$REPO" -e OUT="$OUT" -e SDK_URL="$SDK_URL" \
  -e SDK_TAR="$SDK_TAR" -e DL_DIR="$CACHE/dl" -e APT_CACHE="$CACHE/apt" \
  -e FEEDS_CACHE="$FEEDS_CACHE" -e KEY_APK="${KEY_APK:-}" \
+  -e SHATER_PKG_VERSION="$SHATER_PKG_VERSION" \
+  -e SHATER_PKG_RELEASE="$SHATER_PKG_RELEASE" \
  debian:bookworm bash "$REPO/ci/sdk-build-apk.sh"

 # --- 2) sanity: the per-arch apk repo dir must be complete -------------------
@@ -1,92 +0,0 @@
-#!/bin/sh
-# ci/build-feed.sh — build the signed opkg feed for ONE arch.
-#
-# Usage:  ci/build-feed.sh <ARCH> <SDK_DOCKER_TAG> <OUTDIR>
-#   e.g.  ci/build-feed.sh x86_64             x86_64-24.10.4          out/x86_64
-#         ci/build-feed.sh aarch64_cortex-a53 mediatek-filogic-24.10.4 out/aarch64_cortex-a53
-#
-# This is the reusable per-arch entrypoint the Gitea workflow calls. It runs on
-# the CI RUNNER and:
-#   1. asserts the prebuilt shaterd binary for this arch was already staged by
-#      scripts/build-shaterd.sh (into openwrt/shaterd/files/) — proving artifact
-#      order: SPA+shaterd build BEFORE the SDK package build;
-#   2. drives the arch-matched `openwrt/sdk` docker image to compile all 4
-#      packages (ci/sdk-build.sh) and collect their .ipk into OUTDIR;
-#   3. builds + usign-signs the opkg `Packages` index over OUTDIR
-#      (ci/install-usign.sh + ci/make-index.sh; signs iff $KEY_BUILD is set).
-#
-# Env:
-#   KEY_BUILD   usign SECRET key (Gitea repo secret). If set, the feed index is
-#               signed and verifiable by dist/shater-feed.pub (fp 5ac4b177689cb8e0).
-#               If unset, an UNSIGNED feed is produced (make-index warns).
-set -eu
-
-ARCH="${1:?arch required (x86_64 | aarch64_cortex-a53)}"
-SDK_TAG="${2:?sdk docker tag required (e.g. x86_64-24.10.4)}"
-OUT="${3:?output dir required}"
-
-REPO="$(cd "$(dirname "$0")/.." && pwd)"
-mkdir -p "$OUT"; OUT="$(cd "$OUT" && pwd)"
-# $OUT is created here as ROOT on the runner, but the nested `openwrt/sdk`
-# container runs as the unprivileged `buildbot` (uid 1000) — so it must be able
-# to write the collected .ipk into $OUT. World-writable is set HERE (a chmod
-# from inside the container, as buildbot, cannot fix a root-owned dir).
-chmod 0777 "$OUT"
-
-# --- 0) the prebuilt shaterd binary must already be staged for this arch ------
-case "$ARCH" in
-  x86_64)             sfx=amd64 ;;
-  aarch64_cortex-a53) sfx=arm64 ;;
-  *) echo "[feed] ERROR: unsupported ARCH '$ARCH'"; exit 2 ;;
-esac
-if [ ! -f "$REPO/openwrt/shaterd/files/shaterd-$sfx.upx" ]; then
-  echo "[feed] ERROR: openwrt/shaterd/files/shaterd-$sfx.upx not staged."
-  echo "        Run scripts/build-shaterd.sh BEFORE ci/build-feed.sh." >&2
-  exit 3
-fi
-
-chmod +x "$REPO"/ci/*.sh 2>/dev/null || true
-
-# --- 0.5) persistent dl/ (package source tarballs) ----------------------------
-# Workspace dir restored/saved by actions/cache in the workflow and shared into
-# the nested SDK container via --volumes-from; becomes CONFIG_DOWNLOAD_FOLDER
-# there (ci/sdk-build.sh). PKG_HASH still verifies every file, so a stale cache
-# can never produce a wrong build. Must be writable by the container's
-# unprivileged buildbot user (same reason as the $OUT chmod above).
-DL_DIR="$REPO/.cache/dl"
-mkdir -p "$DL_DIR"
-chmod -R a+rwX "$DL_DIR" 2>/dev/null || true
-
-# --- 0.6) persistent feeds/ git checkouts -------------------------------------
-# Workspace dir restored/saved by actions/cache (key: feeds-opkg-<release>) and
-# symlinked over the SDK's feeds/ inside the container (ci/sdk-build.sh), so
-# `scripts/feeds update -a` fetches deltas instead of re-cloning base+packages+
-# luci from scratch (~7 min/run on this runner's slow github.com link).
-# Top-level chmod only: the contents are created by the container's uid-1000
-# build user and restored with the same ownership (tar-as-root preserves it).
-FEEDS_CACHE="$REPO/.cache/feeds/opkg"
-mkdir -p "$FEEDS_CACHE"
-chmod a+rwX "$REPO/.cache" "$REPO/.cache/feeds" "$FEEDS_CACHE" 2>/dev/null || true
-
-# --- 1) SDK package build (4 packages) in the arch-matched SDK image ----------
-# We drive the `openwrt/sdk` docker image directly (not openwrt/gh-action-sdk):
-# on a self-hosted Gitea act_runner the marketplace action fetch can be
-# unavailable, and we need a CLEAN single-feed layout. `--volumes-from
-# $(hostname)` shares THIS job container's workspace volume into the nested SDK
-# container — a bare `-v $PWD:...` points at a host path that does not exist
-# under the act_runner DinD setup. (Requires the job to run inside a container,
-# which Gitea Actions does by default.)
-echo "[feed] SDK build arch=$ARCH image=openwrt/sdk:$SDK_TAG"
-docker pull "openwrt/sdk:$SDK_TAG"
-docker run --rm --volumes-from "$(hostname)" \
-  -e ARCH="$ARCH" -e REPO="$REPO" -e OUT="$OUT" -e DL_DIR="$DL_DIR" \
-  -e FEEDS_CACHE="$FEEDS_CACHE" \
-  "openwrt/sdk:$SDK_TAG" \
-  sh "$REPO/ci/sdk-build.sh"
-
-# --- 2) index + sign the per-arch feed (usign, KEY_BUILD passed through) -------
-sh "$REPO/ci/install-usign.sh"
-KEY_BUILD="${KEY_BUILD:-}" bash "$REPO/ci/make-index.sh" "$OUT"
-
-echo "[feed] done arch=$ARCH -> $OUT"
-ls -l "$OUT"
@@ -2,24 +2,21 @@
 # ci/gen-apk-key.sh — generate the Shater **apk** feed signing keypair (25.12 lane).
 #
 # apk (OpenWrt/ImmortalWrt 25.12+) verifies package indexes with EC keys
-# (prime256v1 PEM), NOT usign — the existing usign identity
-# (dist/shater-feed.pub, fp 5ac4b177689cb8e0) keeps signing the opkg/24.10 feed
-# and is NOT touched by this script. This generates a SEPARATE, second identity:
+# (prime256v1 PEM). This is the ONLY feed identity shater has since the opkg
+# lane was removed (D22) — the old usign key is history, not a second lane.
 #
 #   dist/shater-apk.key   EC PRIVATE key.  NEVER commit (dist/ is gitignored).
 #                         Paste its full PEM contents into the Gitea repo secret
-#                         KEY_APK (the apk analog of the usign secret KEY_BUILD).
-#                         Then delete the local file (or keep it in a password
-#                         manager as the offline backup — losing it means every
-#                         deployed router must re-trust a new key).
-#   dist/shater-apk.pem   PUBLIC key. Commit it next to shater-feed.pub:
+#                         KEY_APK. Then delete the local file (or keep it in a
+#                         password manager as the offline backup — losing it
+#                         means every deployed router must re-trust a new key).
+#   dist/shater-apk.pem   PUBLIC key. Commit it:
 #                             git add -f dist/shater-apk.pem
 #                         (-f because /dist/ is gitignored). Routers install it
 #                         as /etc/apk/keys/shater-apk.pem.
 #
 # Run ONCE. Refuses to overwrite: regenerating the key invalidates the trust of
-# every router that already installed shater-apk.pem (same rule as D7 for the
-# usign key).
+# every router that already installed shater-apk.pem (see D22).
 set -eu

 REPO="$(cd "$(dirname "$0")/.." && pwd)"
@@ -1,60 +0,0 @@
-#!/bin/bash
-# Make `usign` available on the CI runner so ci/make-index.sh can sign the opkg
-# feed index. The OpenWrt SDK ships usign, but the index/signing step runs on the
-# bare runner (outside the SDK container), so we build the tiny standalone tool
-# from source (no libubox — it is intentionally dependency-free so it can
-# bootstrap a build system). No-op if usign is already on PATH.
-#
-# Ported unchanged from Shater v0.1 (ci/install-usign.sh): usign is
-# format-agnostic and the signing story is identical for the v0.2 4-package feed.
-#
-# CI cache: a previously-built binary is reused from $USIGN_CACHE (default:
-# <repo>/.cache/tools — a workspace dir the workflow persists via actions/cache),
-# skipping the apt + cmake + clone + build (~1 min). After a fresh build the
-# binary is copied there so the NEXT run hits the cache. usign is a tiny static
-# helper with no versioned protocol — a stale cached binary cannot mis-sign.
-set -eu
-
-REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
-TOOLS="${USIGN_CACHE:-$REPO_ROOT/.cache/tools}"
-
-# place <binary> — install onto PATH (system-wide if we can, else ~/bin)
-place() {
-  local SUDO=""; [ "$(id -u)" = 0 ] || SUDO="sudo"
-  if $SUDO install -m0755 "$1" /usr/local/bin/usign 2>/dev/null; then
-    :
-  else
-    mkdir -p "$HOME/bin"
-    install -m0755 "$1" "$HOME/bin/usign"
-    echo "$HOME/bin" >> "${GITHUB_PATH:-/dev/null}"
-    export PATH="$HOME/bin:$PATH"
-  fi
-}
-
-if command -v usign >/dev/null 2>&1; then
-  echo "[usign] already present: $(command -v usign)"
-  exit 0
-fi
-
-if [ -x "$TOOLS/usign" ]; then
-  place "$TOOLS/usign"
-  echo "[usign] restored from cache: $(command -v usign || echo "$HOME/bin/usign")"
-  exit 0
-fi
-
-SUDO=""; [ "$(id -u)" = 0 ] || SUDO="sudo"
-if ! command -v cmake >/dev/null 2>&1 || ! command -v cc >/dev/null 2>&1; then
-  $SUDO apt-get update -qq
-  $SUDO apt-get install -y -qq cmake gcc git
-fi
-
-tmp="$(mktemp -d)"
-# Canonical source; fall back to the GitHub mirror if git.openwrt.org is flaky.
-git clone --depth 1 https://git.openwrt.org/project/usign.git "$tmp/usign" \
-  || git clone --depth 1 https://github.com/openwrt/usign.git "$tmp/usign"
-( cd "$tmp/usign" && cmake -DCMAKE_BUILD_TYPE=Release . >/dev/null && make >/dev/null )
-
-place "$tmp/usign/usign"
-# seed the cache for the next run (best-effort)
-mkdir -p "$TOOLS" 2>/dev/null && install -m0755 "$tmp/usign/usign" "$TOOLS/usign" 2>/dev/null || true
-echo "[usign] built: $(command -v usign || echo "$HOME/bin/usign")"
@@ -1,39 +0,0 @@
-#!/bin/bash
-# Build the opkg feed index (Packages + Packages.gz) with SHA256 for a dir of
-# .ipk files, then optionally usign-sign it if $KEY_BUILD (the Gitea repo secret)
-# is set and usign is present. Arg $1 = feed dir.
-#
-# Ported from Shater v0.1 (ci/make-index.sh), unchanged. It is package-count and
-# package-name agnostic: it indexes whatever .ipk are in the dir, so it serves
-# BOTH the per-arch feed built by ci/build-feed.sh AND the combined release feed
-# assembled in the release job (shaterd + byedpi per-arch, shater-core +
-# luci-app-shater = _all). opkg filters by Architecture at install time, so one
-# combined URL serves every device.
-#
-# Feed format: opkg `src/gz` (.ipk + text Packages index, usign signature).
-# OpenWrt 24.10 (our SDK) still uses opkg; apk arrives at 25.12. The committed
-# trust anchor dist/shater-feed.pub is a usign (Ed25519) key, matching this.
-set -euo pipefail
-OUT="${1:?feed dir required}"; cd "$OUT"
-: > Packages
-for ipk in *.ipk; do
-  [ -e "$ipk" ] || continue
-  ctrl=$(tar -xzOf "$ipk" ./control.tar.gz | tar -xzO ./control)
-  sz=$(wc -c < "$ipk"); sha=$(sha256sum "$ipk" | cut -d' ' -f1)
-  printf '%s\n' "$ctrl" | sed '/^[[:space:]]*$/d' >> Packages
-  printf 'Filename: %s\nSize: %s\nSHA256sum: %s\n\n' "$ipk" "$sz" "$sha" >> Packages
-done
-gzip -kf Packages
-
-if [ -n "${KEY_BUILD:-}" ]; then
-  # Signing was requested — a missing/broken signer must FAIL the build, not
-  # silently ship an unsigned feed that routers with check_signature on reject.
-  command -v usign >/dev/null 2>&1 || { echo "[index] ERROR: KEY_BUILD set but usign not found" >&2; exit 1; }
-  umask 077; printf '%s\n' "$KEY_BUILD" > /tmp/usign.sec
-  usign -S -m Packages -s /tmp/usign.sec || { rm -f /tmp/usign.sec; echo "[index] ERROR: usign signing failed" >&2; exit 1; }
-  rm -f /tmp/usign.sec
-  echo "[index] signed -> Packages.sig ($(head -1 Packages.sig))"
-else
-  echo "[index] no KEY_BUILD -> UNSIGNED feed (opkg needs check_signature off, or set the secret)"
-fi
-echo "[index] contents:"; ls -l
@@ -10,7 +10,6 @@
 # the target fleet (BananaWRT 25.12-mtk-vendor = ImmortalWrt 25.12 base, its
 # distfeeds even point at downloads.immortalwrt.org/releases/25.12-SNAPSHOT) is
 # ImmortalWrt — so we extract the official ImmortalWrt SDK tarball ourselves.
-# Same --volumes-from workspace-sharing pattern as ci/sdk-build.sh (opkg lane).
 #
 # The OpenWrt buildsystem refuses to run as root, so the SDK build itself runs
 # as an unprivileged `build` user created here.
@@ -28,11 +27,16 @@ SDK_URL="${SDK_URL:?SDK_URL env required}"

 echo "[apk-sdk] arch=$ARCH repo=$REPO out=$OUT"
 echo "[apk-sdk] sdk=$SDK_URL"
+# Package version derived from the git tag by ci/version.sh (bug B4). Forwarded
+# to the unprivileged build user on the `su` line at the bottom of this file;
+# openwrt/{shaterd,shater-core,luci-app-shater}/Makefile pick it up from the
+# environment. byedpi keeps upstream ByeDPI's own version (see its Makefile).
+echo "[apk-sdk] package version: ${SHATER_PKG_VERSION:-<unset -> Makefile fallback>}-r${SHATER_PKG_RELEASE:-?}"
 test -f "$REPO/openwrt/shaterd/Makefile" || {
  echo "[apk-sdk] ERROR: feed not mounted ($REPO/openwrt/shaterd/Makefile missing)"; ls -la "$REPO" || true; exit 9; }

-# The prebuilt shaterd artifact must already be staged for this arch (same
-# contract as the opkg lane — scripts/build-shaterd.sh runs first).
+# The prebuilt shaterd artifact must already be staged for this arch
+# (artifact-order contract — scripts/build-shaterd.sh runs first).
 case "$ARCH" in
  x86_64)             sfx=amd64 ;;
  aarch64_cortex-a53) sfx=arm64 ;;
@@ -108,7 +112,7 @@ export HOME=/home/build
 cd "$SDKDIR"

 # Register this repo's openwrt/ as a src-link feed named `shater` (absolute
-# path required) — identical to the opkg lane (ci/sdk-build.sh).
+# path required).
 cp -f feeds.conf.default feeds.conf
 grep -q '^src-link shater ' feeds.conf || echo "src-link shater $REPO/openwrt" >> feeds.conf

@@ -392,6 +396,25 @@ done
 [ "$found" -ge 4 ] || { echo "[apk-sdk] ERROR: expected >=4 of OUR .apk, collected $found"; echo "[apk-sdk] (all .apk under bin/:)"; find bin -type f -name '*.apk' | head -20; exit 6; }
 echo "[apk-sdk] collected $found of our .apk"

+# --- assert the tag-derived version actually reached the packages -------------
+# B4's failure mode is a wrong-but-plausible version shipping silently, so the
+# env -> make hand-off is verified, not trusted: each of our three tag-versioned
+# packages must be named `<name>-<ver>-r<rel>.apk`. byedpi is excluded on purpose
+# (it carries upstream ByeDPI's own version). This runs BEFORE `apk mkndx`, so a
+# stale version can never even reach the index.
+if [ -n "${SHATER_PKG_VERSION:-}" ] && [ -n "${SHATER_PKG_RELEASE:-}" ]; then
+  want="${SHATER_PKG_VERSION}-r${SHATER_PKG_RELEASE}"
+  for p in shaterd shater-core luci-app-shater; do
+    [ -f "$OUT/${p}-${want}.apk" ] || {
+      echo "[apk-sdk] ERROR: $p was not built as version '$want'."
+      echo "        SHATER_PKG_VERSION/SHATER_PKG_RELEASE did not reach the package"
+      echo "        Makefile — the build would have shipped a stale version (bug B4)."
+      echo "[apk-sdk] collected:"; ls -1 "$OUT" | sed 's/^/          /'
+      exit 12; }
+  done
+  echo "[apk-sdk] version check OK — our 3 packages are $want"
+fi
+
 # --- index + sign: exactly how the OpenWrt 25.12 buildsystem does it ---------
 #   apk mkndx --root T --keys-dir T [--sign key] --allow-untrusted \
 #             --output packages.adb *.apk
@@ -423,7 +446,7 @@ INNER
 chmod 0644 /home/build/inner.sh

 su build -s /bin/bash -c \
-  "ARCH='$ARCH' REPO='$REPO' OUT='$OUT' SDKDIR='$SDKDIR' KEYFILE='${KEYFILE:-}' DL_DIR='${DL_DIR:-}' FEEDS_CACHE='${FEEDS_CACHE:-}' bash /home/build/inner.sh"
+  "ARCH='$ARCH' REPO='$REPO' OUT='$OUT' SDKDIR='$SDKDIR' KEYFILE='${KEYFILE:-}' DL_DIR='${DL_DIR:-}' FEEDS_CACHE='${FEEDS_CACHE:-}' SHATER_PKG_VERSION='${SHATER_PKG_VERSION:-}' SHATER_PKG_RELEASE='${SHATER_PKG_RELEASE:-}' bash /home/build/inner.sh"

 chmod -R a+rwX "$OUT" 2>/dev/null || true
 echo "[apk-sdk] OK arch=$ARCH — apk feed dir:"
@@ -1,116 +0,0 @@
-#!/bin/sh
-# Runs INSIDE an `openwrt/sdk:<target>-<ver>` container (CWD = SDK root
-# /builder). The job's workspace is shared into this container via
-# `docker run --volumes-from`, so the repo is visible at $REPO and output goes
-# to $OUT (a dir under the repo, hence also visible to the runner afterwards).
-#
-# Unlike Shater v0.1 (which compiled ONLY xrayctl in the SDK and hand-packed the
-# pure-data packages with tar), v0.2 builds ALL FOUR packages the canonical way,
-# via the SDK feed + `make package/<p>/compile`:
-#
-#   shaterd          prebuilt binary — Build/Compile only VALIDATES that
-#                    openwrt/shaterd/files/shaterd-<amd64|arm64>.upx was staged
-#                    by scripts/build-shaterd.sh on the runner BEFORE this ran.
-#                    (arch-specific .ipk: RSTRIP/STRIP disabled — packed ELF.)
-#   shater-core      PKGARCH=all data glue (procd init, sysctl, uci-defaults).
-#   luci-app-shater  PKGARCH=all LuCI thin launcher — its Makefile does
-#                    `include $(TOPDIR)/feeds/luci/luci.mk`, so the `luci` feed
-#                    MUST be updated first (that is what creates feeds/luci/luci.mk).
-#   byedpi           arch-specific C — the SDK cross-compiles ciadpi from the
-#                    upstream tarball (needs network for PKG_SOURCE_URL).
-#
-# Env (required): ARCH, REPO, OUT.
-set -eu
-ARCH="${ARCH:?ARCH env required}"
-REPO="${REPO:?REPO env required}"
-OUT="${OUT:?OUT env required}"
-mkdir -p "$OUT"
-
-echo "[sdk] arch=$ARCH repo=$REPO out=$OUT"
-test -f "$REPO/openwrt/shaterd/Makefile" || {
-  echo "[sdk] ERROR: feed not mounted ($REPO/openwrt/shaterd/Makefile missing)"; ls -la "$REPO" || true; exit 9; }
-
-# The prebuilt shaterd artifact must already be staged for this arch.
-case "$ARCH" in
-  x86_64)             sfx=amd64 ;;
-  aarch64_cortex-a53) sfx=arm64 ;;
-  *) echo "[sdk] ERROR: unsupported ARCH '$ARCH'"; exit 2 ;;
-esac
-test -f "$REPO/openwrt/shaterd/files/shaterd-$sfx.upx" || {
-  echo "[sdk] ERROR: openwrt/shaterd/files/shaterd-$sfx.upx not staged."
-  echo "        scripts/build-shaterd.sh must run on the runner before the SDK build."; exit 3; }
-
-# --- register this repo's openwrt/ as a src-link feed named `shater` ---------
-# src-link REQUIRES an absolute path; $REPO/openwrt is exactly a feed root (it
-# contains the 4 package dirs and nothing else that looks like a package).
-cp -f feeds.conf.default feeds.conf
-grep -q '^src-link shater ' feeds.conf || echo "src-link shater $REPO/openwrt" >> feeds.conf
-
-# Update metadata for ALL feeds: our `shater` feed + the SDK defaults (base,
-# luci, packages, routing, telephony). We need `luci` for feeds/luci/luci.mk and
-# `base`/`packages` for the runtime deps (kmod-nft-tproxy, kmod-nft-socket,
-# ip-full, rpcd, luci-base) to resolve.
-#
-# Persistent feeds checkouts: $FEEDS_CACHE (a workspace dir the runner restores
-# via actions/cache, shared into this container via --volumes-from) replaces
-# the SDK's ephemeral feeds/ dir, so `feeds update` git-fetches deltas instead
-# of re-cloning base+packages+luci every run (~7 min on the runner's slow
-# github.com link). Correctness-safe: update always checks out feeds.conf's
-# pinned revisions; if it ever fails on a cached checkout (e.g. a force-pushed
-# upstream), the cache is wiped and the update retried with fresh clones.
-if [ -n "${FEEDS_CACHE:-}" ] && mkdir -p "$FEEDS_CACHE" 2>/dev/null; then
-  rm -rf feeds
-  ln -s "$FEEDS_CACHE" feeds
-  echo "[sdk] feeds/ -> $FEEDS_CACHE (persistent cache)"
-fi
-echo "[sdk] feeds update -a"
-if ! ./scripts/feeds update -a; then
-  [ -L feeds ] || { echo "[sdk] ERROR: feeds update failed"; exit 8; }
-  echo "[sdk] WARNING: feeds update failed on cached checkouts — wiping cache, cloning fresh"
-  find "$FEEDS_CACHE" -mindepth 1 -maxdepth 1 -exec rm -rf {} + 2>/dev/null || true
-  ./scripts/feeds update -a
-fi
-
-echo "[sdk] feeds install (prefer shater feed)"
-./scripts/feeds install -p shater shaterd shater-core byedpi luci-app-shater
-
-# Select our packages, then defconfig. `make package/<p>/compile` builds the
-# explicit target regardless, but selecting first makes deps visible to defconfig.
-for p in shaterd shater-core byedpi luci-app-shater; do
-  echo "CONFIG_PACKAGE_$p=m" >> .config
-done
-# Route source downloads through OpenWrt's fast CDN mirror FIRST — sourceware.org
-# (elfutils) and other upstreams intermittently stall mid-transfer, and curl's
-# --connect-timeout doesn't cover a stalled stream, so the SDK download hangs the
-# build. LOCALMIRROR is tried before each package's own PKG_SOURCE_URL. (lx CI)
-echo 'CONFIG_LOCALMIRROR="https://sources.cdn.openwrt.org"' >> .config
-# Persistent dl/ across runs: $DL_DIR is a workspace dir the runner restores via
-# actions/cache (see ci/build-feed.sh). Correctness-safe: the buildroot verifies
-# PKG_HASH on every file already in dl/ and re-downloads on mismatch, so a stale
-# cache can never leak a wrong source into the build.
-if [ -n "${DL_DIR:-}" ]; then
-  echo "CONFIG_DOWNLOAD_FOLDER=\"$DL_DIR\"" >> .config
-fi
-echo "[sdk] defconfig"
-make defconfig >/dev/null
-
-# --- compile the 4 packages --------------------------------------------------
-for p in shaterd shater-core byedpi luci-app-shater; do
-  echo "[sdk] === build $p ==="
-  make "package/$p/compile" V=s -j"$(nproc)"
-done
-
-# --- collect ONLY our 4 packages' .ipk (per-arch shaterd/byedpi + _all core/luci)
-# NOT `find bin -name '*.ipk'`: the openwrt/sdk image ships HUNDREDS of prebuilt
-# kmod/base .ipk under bin/, which a blanket copy would pull into the feed and
-# get signed under OUR key. Match each package's own `<name>_<ver>_<arch>.ipk`.
-found=0
-for p in shaterd shater-core byedpi luci-app-shater; do
-  for ipk in $(find bin -type f -name "${p}_*.ipk"); do
-    cp -f "$ipk" "$OUT/"; found=$((found+1))
-  done
-done
-[ "$found" -ge 4 ] || { echo "[sdk] ERROR: expected >=4 of OUR .ipk, collected $found"; echo "[sdk] (all .ipk under bin/:)"; find bin -type f -name '*.ipk' | head -20; exit 4; }
-chmod -R a+rwX "$OUT" 2>/dev/null || true
-echo "[sdk] OK arch=$ARCH — collected $found of our .ipk:"
-ls -l "$OUT"
@@ -0,0 +1,131 @@
+#!/bin/sh
+# ci/version.sh — the SINGLE source of truth for "what version is this build?".
+#
+# WHY THIS EXISTS (bug B4)
+# -----------------------
+# PKG_VERSION/PKG_RELEASE used to be hand-written literals in the four package
+# Makefiles, and nobody remembered to bump them: v0.2.2 … v0.2.6 all shipped as
+# `shaterd 0.2.0-r3` with DIFFERENT binaries inside (v0.2.6's ELF is 5 491 616 B
+# vs r2's 5 488 336 B). Since apk offers an upgrade only when the feed's version
+# string differs from the installed one, `apk update` saw nothing new and the
+# routers could not be updated through the normal path at all.
+#
+# So the version is now DERIVED, in CI, from the git tag, and the package
+# Makefiles only carry a fallback for manual/offline builds.
+#
+# THE SCHEME
+# ----------
+#   tag push `vX.Y.Z`        -> PKG_VERSION=X.Y.Z   PKG_RELEASE=1
+#   any other build          -> PKG_VERSION=X.Y.Z of the NEAREST reachable tag,
+#   (workflow_dispatch,         PKG_RELEASE=<commits since that tag> + 1
+#    rolling `latest`)
+#   no tag / no git at all   -> PKG_VERSION=0.0.0   PKG_RELEASE=1  (+ warning)
+#
+# apk compares `<upstream>-r<rel>` as: the dotted upstream part first
+# (numerically, component by component), the `r<rel>` only as a tie-break.
+# Verified against the real tool, not from memory —
+# apk-tools 3.0.3 (`apk version -t`) and apk-tools 2.14.6:
+#       0.2.6-r1  > 0.2.0-r3      0.2.6-r12 > 0.2.6-r1
+#       0.2.7-r1  > 0.2.6-r12     0.0.0-r1  < 0.2.0-r3
+# That is exactly the ordering this scheme needs:
+#   * a release always outranks every rolling build that preceded it
+#     (0.2.7-r1 > 0.2.6-rN for any N — the dotted part decides), and
+#   * rolling builds between two releases grow monotonically (r2 < r10 < r11),
+# so a rolling build can never look newer than the next release, and the
+# `latest` feed still moves forward on every dispatch.
+#
+# +1 on the commit count (rather than the raw count) only avoids `-r0` and makes
+# a dispatch build of the tagged commit itself identical to the release build of
+# that same commit — which is the truth: same tree, same binary.
+#
+# `byedpi` is deliberately NOT versioned from our tag — see openwrt/byedpi/Makefile.
+#
+# USAGE
+#   ci/version.sh                  # or --env: eval-able / $GITHUB_ENV-able lines
+#   ci/version.sh --pkg-version    # X.Y.Z
+#   ci/version.sh --pkg-release    # R
+#   ci/version.sh --binary         # vX.Y.Z-rR[-g<sha>] for constant.Version
+#
+# Env:
+#   SHATER_REF / GITHUB_REF   when it is `refs/tags/<tag>` that tag wins and no
+#                             git history is needed (the tag-push path is exact
+#                             even on a shallow checkout).
+set -eu
+
+REPO="$(CDPATH='' cd -- "$(dirname -- "$0")/.." && pwd)"
+
+TAG=""
+EXACT=0
+N=0
+SHA=""
+
+# --- 1) an explicit tag ref is authoritative (and needs no git) --------------
+REF="${SHATER_REF:-${GITHUB_REF:-}}"
+case "$REF" in
+  refs/tags/*) TAG="${REF#refs/tags/}"; EXACT=1 ;;
+esac
+
+# --- 2) otherwise ask git for the nearest reachable release tag --------------
+# `--match 'v[0-9]*'` keeps non-release tags (latest, sdk-cache, apk-latest-*,
+# musl-toolchain-cache) out. This repo is a sing-box FORK and therefore also
+# carries upstream's v1.x tags — `git describe` picks the CLOSEST tag by commit
+# distance, so our own v0.2.x (a handful of commits back) always wins over
+# upstream's v1.x (thousands of commits back). The tag it picked is logged
+# below, so a surprise is visible in the CI log rather than silently shipped.
+if [ "$EXACT" -eq 0 ]; then
+  if D="$(git -C "$REPO" describe --tags --long --match 'v[0-9]*' 2>/dev/null)"; then
+    # `v0.2.6-1-g02c266188` -> TAG=v0.2.6 N=1 SHA=g02c266188.
+    # `%` strips the SHORTEST matching suffix, so a tag that itself contains a
+    # dash (`v0.2.0-healthplan`) survives intact.
+    TAG="${D%-*-g*}"
+    REST="${D#"$TAG"-}"
+    N="${REST%%-*}"
+    SHA="${REST#*-}"
+    if [ "$N" -eq 0 ]; then EXACT=1; fi
+  fi
+fi
+
+# --- 3) tag -> numeric PKG_VERSION ------------------------------------------
+# Keep the leading dotted-numeric run only: `v0.2.0-healthplan` -> `0.2.0`.
+VER=""
+if [ -n "$TAG" ]; then
+  VER="$(printf '%s' "${TAG#v}" | sed -n 's/^\([0-9][0-9.]*\).*/\1/p' | sed 's/\.*$//')"
+fi
+
+if [ -z "$VER" ]; then
+  # No release tag anywhere (shallow clone with no tags, a tarball export, a
+  # fresh fork). 0.0.0 is BELOW every version we have ever published, so such a
+  # build can never masquerade as an upgrade on a real router; the commit count
+  # still makes successive dev builds distinguishable.
+  VER="0.0.0"
+  EXACT=0
+  N="$(git -C "$REPO" rev-list --count HEAD 2>/dev/null || echo 0)"
+  SHA="$(git -C "$REPO" rev-parse --short HEAD 2>/dev/null || echo '')"
+  [ -z "$SHA" ] || SHA="g$SHA"
+  echo "[version] WARNING: no reachable vX.Y.Z tag (and/or no git) -> $VER" >&2
+fi
+
+# --- 4) PKG_RELEASE + the string stamped into the binary --------------------
+if [ "$EXACT" -eq 1 ]; then
+  REL=1
+  FULL="v${VER}-r${REL}"
+else
+  REL=$((N + 1))
+  FULL="v${VER}-r${REL}${SHA:+-$SHA}"
+fi
+
+echo "[version] tag='${TAG:-none}' commits_since=$N exact=$EXACT -> ${VER}-r${REL} (binary: $FULL)" >&2
+
+case "${1:---env}" in
+  --env|"")
+    printf 'SHATER_PKG_VERSION=%s\n' "$VER"
+    printf 'SHATER_PKG_RELEASE=%s\n' "$REL"
+    printf 'SHATER_VERSION=%s\n'     "$FULL"
+    ;;
+  --pkg-version) printf '%s\n' "$VER" ;;
+  --pkg-release) printf '%s\n' "$REL" ;;
+  --binary|--version) printf '%s\n' "$FULL" ;;
+  *)
+    echo "usage: $0 [--env|--pkg-version|--pkg-release|--binary]" >&2
+    exit 2 ;;
+esac
@@ -1,2 +0,0 @@
-untrusted comment: shater feed signing key
-RWRaxLF3aJy44JbcxSFujtrFFEQ8lIsnTkd1K5TdjIhdlC2c0wa0fv4V
@@ -31,12 +31,11 @@ Do not delete it — we port proven pieces from it. What v0.1 has:
 - **`luci-app-shater`** — a custom "instrument panel" LuCI app (client-side JS +
  ucode/rpcd ubus backend): Overview with a live Signal Path, Simple/Advanced
  toggle, quick-start wizard, Nodes/Subs/Rules/DNS/Live/Profiles/Settings pages.
- **CI + signed opkg feed** on Gitea: builds per-arch, signs the feed index with
-  usign, publishes a rolling `latest` Gitea release consumable as `src/gz`. **Feed
-  signing key fingerprint `5ac4b177689cb8e0`**; public key `dist/shater-feed.pub`,
-  secret in the Gitea repo secret `KEY_BUILD`.
+- **CI + a signed package feed** on Gitea: builds per-arch, signs the feed index,
+  publishes a rolling `latest` Gitea release the router consumes as a feed.
+  (v0.1 shipped `.ipk` signed with a usign key — that lane is retired, D22.)
 - Verified end-to-end on the VM: real LAN client proxied, DNS anti-leak, honest
-  fail-closed, opkg install/upgrade from the signed feed.
+  fail-closed, install/upgrade from the signed feed.

 v0.1 is engine-locked to **xray-core**; its generator, share-link parser and
 `run.json` are xray-shaped.
@@ -91,7 +90,7 @@ We are rebasing onto a new engine and a new UI architecture. Full rationale in
 - **`shater` branch `v0.1`** = the standalone xray-based version (frozen, ported
  from).
 - Until Phase 1 merges the engine in, `main` is the docs-first overlay seed you
-  are reading now (LICENSE, README, `docs-shater/`, `dist/shater-feed.pub`).
+  are reading now (LICENSE, README, `docs-shater/`, the feed signing key).

 ## What to port from v0.1 (don't rewrite these ideas)

@@ -105,8 +104,8 @@ overlay, don't redo:
 - **Subscription fetch** (HAPP emulation, fingerprint reconcile, per-sub cache)
  and the flexible **ruleset/list** model — though sing-box has its own share-link
  parser and config schema we now target.
- **CI feed build + usign signing + Gitea release** (adapt to the single forked
-  binary; keep key `5ac4b177689cb8e0`).
+- **CI feed build + index signing + Gitea release** (adapted to the single forked
+  binary; the format is apk, signed with the EC key — D22).
 - The LuCI **design system** (the "instrument panel" identity) — reused for the
  mini-dashboard and as the panel's visual language.

@@ -122,9 +121,9 @@ filter/stats engine wired into sing-box's DNS.
  `https://github.com/SagerNet/sing-box`).
 - **CI:** Gitea Actions (act_runner + Docker). v0.1's workflow was removed from
  `main`; new CI is added when the v0.2 build exists.
- **Feed signing:** usign key `5ac4b177689cb8e0`; secret in repo secret
-  `KEY_BUILD`; public key `dist/shater-feed.pub` (kept so existing installs keep
-  verifying).
+- **Feed signing:** EC (prime256v1) key for the apk index; secret in the repo
+  secret `KEY_APK`; public key `dist/shater-apk.pem`, installed on routers as
+  `/etc/apk/keys/shater-apk.pem`. Never regenerate it (D22).
 - **Test VM:** OpenWrt 24.10.3 x86_64 in Docker (`docker ps --filter
  name=openwrt-vm`). SSH via the ssh-manager MCP server `local_openwrt`
  (localhost:2222, root/openwrt). LuCI at `http://127.0.0.1:8080` (root/openwrt),
@@ -64,11 +64,17 @@ sing-box is GPL-3.0; linking it makes the combined work GPL-3.0. Our own files m
 stay GPL-2.0-or-later (which permits the upgrade), but the project LICENSE is
 GPL-3.0 for clarity.

-## D7 — Keep the v0.1 feed signing identity
+## D7 — Keep the v0.1 feed signing identity  *(SUPERSEDED by D22)*
 The usign feed key `5ac4b177689cb8e0` (public key in `dist/shater-feed.pub`,
 secret in Gitea secret `KEY_BUILD`) carries over, so routers that already trust it
 keep verifying v0.2 packages. Do not regenerate it without a documented rotation.

+> **Superseded 2026-07-25 (D22).** The opkg feed this identity signed no longer
+> exists, so there is nothing left for the key to verify. It was never rotated or
+> compromised — it is simply unused. `dist/shater-feed.pub` was deleted from the
+> tree; the reasoning, and how to resurrect the identity if it is ever needed
+> again, is in D22.
+
 ## D8 — Preserve, don't destroy: v0.1 lives on its branch
 The reset moved the full working xray-based project to the `v0.1` branch and
 cleaned `main`. Nothing is lost; reusable logic (reliability layer, nft/routing,
@@ -95,6 +101,10 @@ runtime, forcing an ELF with `PT_INTERP=/lib64/ld-linux-x86-64.so.2` + `PT_DYNAM
  plane is tproxy/redirect (netplane); generate never emits a tun inbound, so
  the userspace gvisor netstack (~3.6 MB) is unreachable. If a tun inbound ever
  appears it falls back to the system stack — re-add the tag then.
+  **REVERTED 2026-07-25 — that reasoning was wrong and shipped a dead feature.**
+  gVisor is not only the tun stack: it is the netstack of the **WireGuard
+  endpoint**, which we do emit and do declare [MVP]. See D23; the tag is back and
+  is now held there by a test.
 - 2026-07-23: `with_clash_api` also dropped. The admin panel is shater's own
  web server and generate never emits a `clash_api` service; the desktop/CLI
  `LX_TAGS` keeps the tag for external dashboards.
@@ -424,3 +434,235 @@ on the next render.
 Consequence: all delay numbers are comparable (least ping ranks apples against
 apples), and group settings lose two footgun fields while Settings keeps the
 two that actually govern every check.
+
+## D21 — A rule's destination is a rule-set, and nothing else
+Decided 2026-07-25 (product owner). `config rule` carried THREE ways to say
+where traffic is going: `dst_domain` (an inline domain list), `dst_ip` (an inline
+CIDR list) and `dst_ruleset` (a reference to a `config ruleset`). Three
+mechanisms meant three sets of semantics to learn and keep straight, and the
+inline ones were the worse half of the trade: they are re-parsed per rule instead
+of being compiled once into a `.srs`, they cannot be shared between rules, and
+their matcher vocabulary had drifted from the rule-set one in a way nobody could
+see (below).
+
+**Decision: `dst_domain` and `dst_ip` are removed (schema v2). `dst_ruleset` is
+the only destination matcher.** `Src`, `dst_port` and `proto` are untouched —
+they are not lists of destinations and have no rule-set form.
+
+- **Rejected: keep the inline lists as a shorthand.** "One obvious way" is the
+  whole point; a shorthand that quietly means something different from the long
+  form (see the bare-entry trap) is worse than no shorthand.
+- **Rejected: promote inline lists to rule-sets lazily at generate time.** The
+  config on disk would then not say what the router does, and the panel would
+  have to render a list the user cannot find or edit.
+
+### The bare-entry trap, and how the migration handles it
+The two contexts already disagreed about exactly one spelling, silently:
+
+| entry              | in a rule (`dst_domain`) | in a rule-set (`entry`) | migrated to  |
+|--------------------|--------------------------|-------------------------|--------------|
+| `example.com`      | **exact host**           | **host + subdomains**   | `full:example.com` |
+| `full:example.com` | exact host               | exact host              | unchanged    |
+| `suffix:example.com` / `.example.com` | host + subdomains | host + subdomains | unchanged |
+| `keyword:ads`      | substring                | substring               | unchanged    |
+| `regexp:^ads\.`    | pattern                  | pattern *(added here)*  | unchanged    |
+| `geosite:x` / `geoip:x` | inert (engine field removed) | inert (unknown prefix) | unchanged |
+
+`shaterd migrate` (schema v1→v2, `shater/model/migrate.go`) creates one inline
+`config ruleset` per rule that still carries a legacy list — `rule-<rule name>`
+for domains, `rule-<rule name>-ip` for addresses — moves the entries across with
+the conversion above, appends the new name to `dst_ruleset`, and deletes the old
+option. It is idempotent, it resumes an interrupted run, and it never overwrites
+a hand-written rule-set that already owns the generated name (it picks
+`rule-<name>-2`). `regexp:` support was added to inline rule-sets in the same
+change precisely so the move can be lossless.
+
+`geosite:`/`geoip:` entries are copied VERBATIM rather than promoted to a
+`source=geosite` rule-set: those matchers have been inert since the engine
+dropped the route-rule geosite/geoip fields, and turning a dead matcher live
+during an upgrade would be a behaviour change, not a migration. The text is kept
+so the operator can see it and convert it deliberately.
+
+**One deliberate semantic change, called out:** a rule that used BOTH lists
+matched them with AND (an engine route rule ANDs its matcher fields), which is
+almost never what "these sites and these networks" meant. The two generated
+rule-sets are ORed, because `rule_set: [a, b]` matches when either matches. Such
+a rule matches more after the migration than before; it affects only configs that
+used both fields at once.
+
+That AND→OR change is about the ENGINE's TCP/UDP path, and it deliberately does
+**not** extend to the untunnelable-protocol plane (`shater/apply/untunnelable.go`,
+the ping / IPTV / VPN-passthrough policy in nftables). There, a v1
+`dst_domain + dst_ip` rule could never claim a packet that carries no domain, so
+the plan skipped it; reading the migrated form as OR would have made the address
+half suddenly decisive, and with `target=direct` that means an upgrade quietly
+sending previously-tunnelled ICMP out with the client's real source address. A
+rule whose rule-sets are known to match by NAME is therefore still skipped by that
+plan, and the skip is reported ("a routing rule matches by name … as well as by
+address"). Split the rule in two if you want the addresses decided there.
+
+### The vocabulary is about ENTRIES YOU TYPE, not about every list body
+The table above is the vocabulary of an **inline** rule-set's `entry` values (and
+of the DNS-filter/device lists, which share the classifier). The other two rule-set
+sources are not other spellings of it:
+
+| source                       | what it is                       | vocabulary |
+|------------------------------|----------------------------------|------------|
+| `inline`                     | entries you type                 | the table above |
+| `url` → `.srs` / `.json`     | a compiled rule-set, engine-owned | the engine's, not ours |
+| `url` → anything else        | a hosts / one-domain-per-line / AdBlock TEXT FILE | **none** — every line is a domain plus its subdomains |
+| `file`                       | a local `.srs` / `.json`          | the engine's, not ours |
+
+**Rejected: run text lists through the entry classifier too.** A published
+AdGuard/OISD list is full of colon-bearing tokens that are ordinary filter syntax
+(`##…:has(…)`, `$domain=`, absolute URLs); classifying them would either mis-import
+them or bury the operator under hundreds of "unrecognised prefix" warnings per
+list. The formats also disagree structurally — a hosts line carries several names,
+so the text parser works per token, while an entry is a whole line. And `regexp:`
+arriving from a third-party URL is a pattern compiled into the router's matcher and
+evaluated per query, which is a very different proposition from one the operator
+typed.
+
+So the difference stands and is paid for in diagnostics instead: a text list
+containing `full:` / `suffix:` / `keyword:` / `regexp:` is reported per list, on
+every generate, naming the entries and pointing at `source=inline` where they work
+(`warnListEntryVocabulary`, `shater/generate/ruleset.go`). The check tests only
+those four markers, never the general `word:` shape, so it fires on a human's
+mistake and stays quiet on published filter syntax.
+
+Consequence: one destination mechanism, one vocabulary, one place a list is
+edited; every list is compiled once and reused. The panel's rule editor drops its
+Domain(s) and IP/CIDR(s) fields; its destination control is a checkbox list of
+the rulesets that already exist, and nothing more. Creating and filling a list
+stays in the Rulesets panel — **rejected: a "create a list from here" shortcut in
+the rule editor**, because a second place to author a list is a second place for
+its semantics and its duplicate-name rules to drift, and the whole point of this
+decision was to stop having two.
+
+## D22 — One packaging lane: apk. The opkg/`.ipk` lane is deleted, not disabled
+Decided 2026-07-25 (product owner). CI built and published TWO signed feeds from
+every run: opkg/usign (`.ipk` + `Packages.gz`, OpenWrt 24.10) and apk/EC (`.apk` +
+`packages.adb`, OpenWrt/ImmortalWrt 25.12). The opkg half served nobody. Checked
+on the actual hardware, not inferred:
+
+| Device | Firmware | pkg arch | package manager |
+|---|---|---|---|
+| `mini_router` (BPi-R3 Mini) | ImmortalWrt 25.12.1 | `aarch64_cortex-a53` | apk-tools 3.0.5 |
+| `main_router` (BPi-R4)      | OpenWrt 25.12.0     | `aarch64_cortex-a53` | apk-tools 3.0.5 — **no `opkg` binary on the system at all** |
+
+**Decision: delete the opkg lane outright.** Removed: the `build` + `release`
+jobs from `.gitea/workflows/release.yml`; `ci/build-feed.sh`, `ci/sdk-build.sh`,
+`ci/make-index.sh`, `ci/install-usign.sh`; and the trust anchor
+`dist/shater-feed.pub`. The Gitea secret `KEY_BUILD` is now referenced by
+nothing and can be deleted from the repo settings. `ci/version.sh`,
+`ci/gitea-release.sh` and `ci/fetch-sdk.sh` are shared or apk-only and stay.
+
+- **Rejected: keep the lane but stop triggering it** (comment it out / gate it on
+  a dispatch input). Dead code in CI is worse than no code: it keeps a second SDK
+  matrix, a second signing key and a second feed layout alive in everyone's head
+  and in every future edit, and it silently rots because nothing runs it. The
+  24.10 SDK images it pins are themselves a frozen dependency.
+- **Rejected: keep `dist/shater-feed.pub` as a historical artifact.** A committed
+  trust anchor is an instruction — it invites someone to follow the old install
+  path for a feed that is no longer produced. Nothing is lost by removing it:
+  git history still holds the file, the SECRET half is untouched in `KEY_BUILD`,
+  and a usign secret key blob contains its own public half, so the identity can
+  be reconstructed if a 24.10 device ever has to be served again. Deleting the
+  file is reversible; a stale trust anchor pointing at an unmaintained feed is
+  the thing that quietly misleads.
+- **Not done: revoking or rotating the usign key.** There is no incident. It is
+  retired, not burned (D7).
+
+Consequence: one SDK, one key, one feed layout, one set of install instructions.
+It also makes the rolling release `apk-latest-<arch>` the *only* install path
+that does not require hand-editing a file per release — which is why the same
+change fixed it: publishing was an either/or (`apk-latest-<arch>` on dispatch,
+ELSE `apk-vX.Y.Z-<arch>` on a tag), so once releases moved to tag pushes the
+rolling pointer stopped being written and froze at `0.2.0` while v0.2.9/v0.2.10
+shipped — routers on the rolling URL got a successful, silent `apk update` with
+nothing new. `release-apk` now writes the rolling pointer on every run and
+asserts, by reading the published release back over the Gitea API, that it holds
+our three tag-versioned packages at exactly the version just built and no asset
+at any other version.
+
+
+## D23 — The router tag set is a checked contract, not a string literal
+`with_gvisor` was trimmed from the router set on 2026-07-23 (D9) as "unreachable
+code: we never emit a tun inbound". True about tun — and irrelevant, because
+gVisor is also the netstack of the **WireGuard endpoint**, which shater emits and
+FEATURES.md declares [MVP] (AmneziaWG is called *"a driving requirement"*). Every
+binary shipped between then and 2026-07-25 answered a configured WireGuard node
+with:
+
+```
+create instance: initialize endpoint[0]: create WireGuard device:
+gVisor is not included in this build, rebuild with -tags with_gvisor
+```
+
+`transport/wireguard/device_stack_stub.go` (`//go:build !with_gvisor`) returns
+`tun.ErrGVisorNotIncluded` from **both** device constructors, so
+`system_interface: true` is not an escape hatch either: WireGuard was 100% dead
+in the shipped artifact while the panel offered it, the parser accepted `wg://`,
+`awg://` and wg-quick `.conf` imports, and the owner had 7 WireGuard sections in
+UCI on a production router.
+
+- **Decision:** `with_gvisor` is part of the router tag set and stays there for
+  as long as we ship WireGuard. It costs **~2.8 MB raw / ~0.65 MB UPX per arch**
+  (measured 2026-07-25, both arches; `/overlay` on the production router is
+  6.9 GB with 205 MB used). A tag whose absence turns a declared feature into a
+  runtime error is not "dead weight" — it is the feature.
+
+### Why the bug was invisible, and what now makes it visible
+The defect was not a typo in a tag list. It was that **nothing connected the tag
+list to the feature list**, and the shipped tag combination was the one build
+configuration nothing exercised: the whole test suite compiles with the FULL
+upstream set (`with_gvisor` included), so `TestAmneziaWGEndpoint` passed happily
+while the artifact it was supposed to vouch for could not create a WireGuard
+device. Tests proved the code was right; they never proved the *build* was.
+
+Three pieces now hold it together:
+
+1. **One definition of the set** — `scripts/router-tags.sh` (`SHATER_ROUTER_TAGS`
+   + `SHATER_ROUTER_LDFLAGS`), sourced by `scripts/build-shaterd.sh` and by the
+   checker. The tag list used to live as a literal inside the build script, i.e.
+   in a file no test reads. A second copy is a second truth.
+2. **A declared-feature table** — `shater/buildtags`: every tag-gated capability
+   we promise, with the exact tags it needs *to run* and why (the code anchor).
+   `TestRouterTagSetCoversDeclaredFeatures` parses the shell file and fails if a
+   declared feature lost a tag. It needs no build tags, no Linux, no network and
+   no privileges, so it runs in every plain `go test ./...` — including on the
+   Windows dev host, where nothing else can see the shipped configuration.
+3. **A construction test under the shipped tags** —
+   `shater/generate.TestShippedTagSetConstructsDeclaredProtocols` drives one node
+   of every declared protocol (ss/vmess/trojan/vless ws-grpc-httpupgrade-quic-
+   xhttp/REALITY/uTLS-fp/hysteria2/tuic/**wg**/**awg**) through `box.New`+`Start`.
+   `scripts/check-router-tags.sh` runs it **with `SHATER_ROUTER_TAGS`**, and CI
+   runs that script (`.gitea/workflows/release.yml`) *before* the artifact is
+   built. In a router-tag-set run nothing may be skipped: a protocol that is not
+   compiled in fails the run instead of quietly disappearing from it.
+
+(2) catches a trim the moment it is made and names the feature it kills; (3)
+catches what a list comparison cannot — a tag that is present but insufficient.
+Neither is a substitute for the other. A new protocol in `shater/parse` +
+`shater/generate` means a new row in `buildtags.Features` and a new probe case;
+`TestEveryTagGatedFeatureIsProbed` fails until both exist.
+
+- **Rejected: "just add the tag".** The one-line fix restores WireGuard and
+  leaves the mechanism that hid it fully intact — the next size-driven trim is
+  equally invisible. The tag is the smallest part of this decision.
+- **Rejected: run the WHOLE test suite with the router tag set in CI.** It is the
+  obvious move and it does not work: parts of the suite legitimately depend on
+  upstream-only tags, and the run costs a second full compile of a 25 MB binary's
+  worth of packages on every release. A focused, unprivileged construction test
+  buys the same evidence for ~10 s and, unlike a full run, can be *required* to
+  skip nothing.
+- **Rejected: assert the tag set against upstream's `DEFAULT_BUILD_TAGS`.** That
+  makes any trim a failure, which turns the check into noise and re-litigates D9
+  on every upstream rebase. The contract is with our own feature list, not with
+  upstream's.
+- **Not done: dropping `with_lx_command`.** It is inert for `shaterd` — nothing
+  under `shater/` imports `sing-box/daemon` or `experimental/libbox`, and
+  `go list -deps ./shater/cmd/shaterd` links neither, so it costs zero bytes. It
+  stays only so the router set remains a subset of the lx desktop set. Noted
+  because "a tag that buys nothing" is the mirror image of this bug and should be
+  removed deliberately, not silently.
@@ -13,8 +13,12 @@ usable release, **[T1]** next, **[T2]** later. Phases refer to `ROADMAP.md`.
 - **[MVP]** TPROXY transparent proxy for multiple LAN interfaces (TCP + UDP), SNI/
  Host/QUIC sniffing.
 - **[MVP]** First-match routing rules by source (IP/CIDR/MAC/interface/zone),
-  destination (domain/suffix/keyword/geosite), reusable domain/IP lists, port,
-  proto → target (outbound/selector/chain/direct/block) + egress.
+  destination, port, proto → target (outbound/selector/chain/direct/block) + egress.
+  A rule names its **destination through a rule-set only** — a reusable named list
+  (inline domains/CIDRs, a local or remote file, or a geosite/geoip category) that is
+  compiled once into a `.srs` and shared by every rule that references it. Domain
+  entries take `full:` (exact), `suffix:` / a leading dot (host + subdomains),
+  `keyword:` (substring) and `regexp:`; a bare entry means host + subdomains.
 - **[MVP]** Node groups with balancer/observatory (least-ping/failover/round-robin).
 - **[T1]** Multi-hop chains (L1→Ln); per-rule egress selection; egress via any
  interface/tunnel (e.g. an AmneziaWG tunnel).
@@ -89,8 +93,8 @@ usable release, **[T1]** next, **[T2]** later. Phases refer to `ROADMAP.md`.
  SIM uplink → different egress); backup/restore; i18n (EN + RU).

 ## Ops & distribution
- **[MVP]** Single signed binary; signed opkg feed on Gitea (reuse key
-  `5ac4b177689cb8e0`); one-line install; `opkg upgrade`.
+- **[MVP]** Single signed binary; signed apk feed on Gitea (EC key
+  `dist/shater-apk.pem`); one-line install; named-package `apk upgrade`.
 - **[T1]** Upstream-rebase cadence (track sing-box-lx tags) with a smoke suite.
- **[T2]** apk (OpenWrt 25.x) packaging; multi-router fleet management; REST/gRPC
-  external API; Telegram bot.
+- **[T2]** Multi-router fleet management; REST/gRPC external API; Telegram bot.
+  (apk packaging landed and is now the only lane — D22.)
@@ -1,7 +1,7 @@
 # Shater v0.2 — Build & Install

 How to build the ship artifact (the SPA-embedded `shaterd` binary) and install
-the OpenWrt feed onto a router.
+the signed apk repo onto a router.

 ## 1. Build the `shaterd` binary

@@ -26,7 +26,9 @@ What it does:
 Arg / env:

 - `VERSION` — stamped into `constant.Version`. Resolution: positional arg →
-  `$SHATER_VERSION` → `git describe --tags` → `v0.2.0-dev`.
+  `$SHATER_VERSION` → `ci/version.sh --binary` → `v0.2.0-dev`. `ci/version.sh` is
+  the **same** computation the package version comes from (§2.1), so the string
+  the panel shows always matches what `apk list -I shaterd` reports.
 - `--fast` — skip `npm ci` when `panel/node_modules` already exists.
 - `UPX=/path/to/upx` — override the UPX binary (default `upx` on `PATH`). UPX is
  cross-arch, so one host packs both the amd64 and aarch64 ELFs. (Note: UPX also
@@ -41,22 +43,44 @@ UPX="…/scratchpad/upx-4.2.4-win64/upx.exe" scripts/build-shaterd.sh v0.2.0 --f
 The `dist/*` and `openwrt/shaterd/files/shaterd-*.upx` outputs are gitignored —
 they are release artifacts, not source.

-Tag set (D9 — keep in sync with `docs-shater/DECISIONS.md`):
+Tag set (D9/D23) — defined in **one** place, `scripts/router-tags.sh`, which
+documents every tag and is sourced by the build:

 ```
-with_quic,with_wireguard,with_utls,
+with_gvisor,with_quic,with_wireguard,with_utls,
 badlinkname,tfogo_checklinkname0,with_xhttp,with_awg,with_lx_command
 ```

 We drop `with_purego,with_naive_outbound`: they pull cronet-go, which forces a
 glibc `PT_INTERP` even under `CGO_ENABLED=0`, making the binary unusable on musl.
-We drop `with_gvisor`: the shater data plane is tproxy/redirect and generate
-never emits a tun inbound, so the userspace gvisor netstack is unreachable code.
 We drop `with_clash_api`: the admin panel is shater's own web server and the
 generator never emits a `clash_api` service, so the Clash server is dead code.
 We drop `with_dhcp`: shater resolver types are `udp/tcp/doh/dot/local/fakeip`;
 a `dhcp://` DNS transport is never generated or registered.

+`with_gvisor` was dropped in 2026-07 as "unreachable — we emit no tun inbound"
+and **put back on 2026-07-25**: gVisor is also the netstack of the WireGuard
+endpoint, so without it every `wg://`/`awg://` node died at apply time with
+*"gVisor is not included in this build"* while the panel still offered the
+feature. It costs ~2.8 MB raw / ~0.65 MB UPX per arch. Full story: `DECISIONS.md`
+D23.
+
+### Changing the tag set
+
+Run the guard — it is what stands between a size trim and a silently dead
+feature, and CI runs it before the artifact is built:
+
+```sh
+scripts/check-router-tags.sh     # from Windows/macOS it re-execs itself in golang:1.26
+```
+
+It (1) fails if a feature declared in `FEATURES.md` lost a build tag it needs to
+run (`shater/buildtags`, no tags/OS/network required) and (2) constructs one node
+of every declared protocol through `box.New` **compiled with the shipped tag
+set** — nothing may be skipped in that run. Adding a protocol to
+`shater/parse`+`shater/generate` means adding a row to `buildtags.Features` and a
+probe case in `shater/generate/shipped_tags_linux_test.go`.
+
 ## 2. Packages

 Four OpenWrt packages live under `openwrt/`:
@@ -84,22 +108,65 @@ it. Because the binary is UPX-packed, the package disables the SDK's default str
 feed installed and run `make package/shaterd/compile` (and the others) per target.
 See `openwrt-package-build-ci` for SDK/feed mechanics.

+### 2.1 Package versions come from the git tag
+
+`PKG_VERSION`/`PKG_RELEASE` are **not** maintained by hand. They used to be, and
+nobody bumped them: **v0.2.2 … v0.2.6 all shipped as `shaterd 0.2.0-r3`** with
+different binaries inside (v0.2.6's ELF is 5 491 616 B against r2's 5 488 336 B).
+apk offers an upgrade only when the feed's version string differs from the
+installed one, so `apk update` saw nothing new and the routers could not be
+updated through the normal path at all.
+
+`ci/version.sh` now derives them from `git describe`, once per CI job:
+
+| Build | `PKG_VERSION` | `PKG_RELEASE` | `constant.Version` |
+|---|---|---|---|
+| tag push `v0.2.7`             | `0.2.7` | `1` | `v0.2.7-r1` |
+| dispatch, 3 commits past `v0.2.7` | `0.2.7` | `4` | `v0.2.7-r4-g<sha>` |
+| no reachable tag / no git     | `0.0.0` | `1` | `v0.0.0-r1` |
+
+Ordering is what makes this safe (checked with `apk version -t` on apk-tools
+3.0.3): the dotted part decides first, `-rN` only breaks ties — so
+`0.2.7-r1 > 0.2.6-r12 > 0.2.6-r1 > 0.2.0-r3`. A release therefore always
+outranks every rolling build before it, rolling builds between two releases grow
+monotonically, and an untagged build (`0.0.0`) can never masquerade as an
+upgrade.
+
+The value travels as `SHATER_PKG_VERSION`/`SHATER_PKG_RELEASE` in the SDK build
+environment; the Makefiles read it with a literal fallback for manual/offline
+builds. `ci/sdk-build-apk.sh` then **asserts** the produced `.apk` really carries
+it, so a lost variable fails the build instead of shipping a stale version. The
+release job asserts the same version again on the published rolling repo (§5.1).
+
+`byedpi` is deliberately excluded — `PKG_VERSION:=0.17.3` is *upstream ByeDPI's*
+version, which is what `PKG_HASH` pins and what tells you which ByeDPI is
+installed. Stamping our tag on it would also be a downgrade: every comparator
+reads `0.2.7 < 0.17.3` (component-wise, `2 < 17`). Bump its `PKG_RELEASE` by hand
+when our packaging of it changes.
+
 ## 3. Install on a router

-Install order follows the deps (`shaterd` → `shater-core` → `luci-app-shater`):
+**The normal path is the signed apk repo — §5.** This section is the manual
+fallback (a router with no route to the Gitea host, or a hand-carried build).
+
+Install order follows the deps (`shaterd` → `shater-core` → `luci-app-shater`).
+apk filenames carry no architecture, so make sure you copied the `.apk` built for
+*this* router's arch (`cat /etc/apk/arch`):

 ```sh
-opkg install shaterd_0.2.0-1_<arch>.ipk        # or: apk add shaterd   (25.12+)
-opkg install shater-core_0.2.0-1_all.ipk
-opkg install luci-app-shater_0.2.0-1_all.ipk
-opkg install byedpi_0.17.3-1_<arch>.ipk        # optional: ByeDPI egress
+# <ver> = the release version, e.g. 0.2.7-r1 (§2.1 — it comes from the git tag)
+# --allow-untrusted: our member .apk are unsigned by design — trust lives in the
+# signed packages.adb index (§5), which a loose file install does not consult.
+apk add --allow-untrusted ./shaterd-<ver>.apk
+apk add --allow-untrusted ./shater-core-<ver>.apk
+apk add --allow-untrusted ./luci-app-shater-<ver>.apk
+apk add --allow-untrusted ./byedpi-0.17.3-r1.apk       # optional: ByeDPI egress
 ```

-Installing from a signed feed instead:
+From the repo instead (§5 sets it up once), deps pull the rest in:

 ```sh
-# add the feed (customfeeds.conf / apk repositories), then:
-opkg update && opkg install shater-core luci-app-shater   # shaterd pulled in as a dep
+apk update && apk add luci-app-shater      # -> shater-core -> shaterd
 ```

 ## 4. Enable
@@ -119,83 +186,55 @@ daemon (`shaterd run`), which owns the engine, the `inet shater` data plane, pol
 routing, in-process DNS, and the admin panel (default `:8088`). The LuCI app's
 "Open panel" button mints a single-use token and hands the browser off to the panel.

-## 5. Add the signed feed (recommended — then `opkg upgrade` just works)
+## 5. The signed apk repo (the normal install path)

-CI (`.gitea/workflows/release.yml`) publishes every build as a **rolling `latest`
-Gitea release** that is itself a signed opkg `src/gz` feed: the release holds the
-`.ipk` for all arches, a `Packages`/`Packages.gz` index, a usign `Packages.sig`,
-and the public key `shater-feed.pub`. opkg filters by `Architecture`, so the **same
-two lines work on every device** (x86 testbed picks `x86_64 + all`; the BPI routers
-pick `aarch64_cortex-a53 + all`).
+OpenWrt/ImmortalWrt **25.12** packages with Alpine's **apk**: `.apk` files, a
+binary `packages.adb` index, EC (prime256v1) keys in `/etc/apk/keys/`, and
+effectively mandatory signatures (unsigned needs `--allow-untrusted`). This is
+the only format shater publishes — the `.ipk`/opkg lane was removed in 2026-07
+(`DECISIONS.md` D22); every device we serve is on 25.12 with apk-tools 3.

-> **Format:** OpenWrt 24.10 (our SDK) uses **opkg** (`.ipk`, `Packages.gz`, usign),
-> so the feed is `src/gz` and the trust anchor is the usign key
-> `dist/shater-feed.pub` (fingerprint **`5ac4b177689cb8e0`**). apk only replaces
-> opkg at OpenWrt **25.12** — see §6.
-
-One-time setup on the router:
-
-```sh
-# 1) trust the feed key — the FILENAME must equal the usign key fingerprint.
-wget -O /etc/opkg/keys/5ac4b177689cb8e0 \
-  https://git.qomar.pw/omar/shater/releases/download/latest/shater-feed.pub
-
-# 2) add the feed (one URL serves every arch).
-echo "src/gz shater https://git.qomar.pw/omar/shater/releases/download/latest" \
-  >> /etc/opkg/customfeeds.conf
-
-# 3) refresh + install (shaterd is pulled in as a dependency).
-opkg update
-opkg install luci-app-shater        # -> shater-core -> shaterd
-opkg install byedpi                 # optional: ByeDPI desync egress
-```
-
-With the key installed, opkg's default `check_signature 1` verifies the feed on
-every `opkg update`; no `--nocheck-signature` needed. A **tagged** release
-(`vX.Y.Z`) publishes the identical layout at
-`.../releases/download/vX.Y.Z` if you prefer to pin a version instead of tracking
-`latest`.
-
-### Updating
-
-```sh
-opkg update
-opkg upgrade shaterd shater-core luci-app-shater byedpi   # only our own packages
-```
-
-Updates are only offered when the feed's `Version` differs from the installed one,
-so **bump `PKG_RELEASE`** (or `PKG_VERSION`) in the package Makefile on every
-shipped change — otherwise `opkg upgrade` sees the same version and does nothing.
-Do **not** `opkg upgrade` base/system packages from this feed; upgrade only the
-four shater packages above.
-
-## 6. apk feed (OpenWrt/ImmortalWrt 25.12+ — incl. BananaWRT 25.12-mtk-vendor)
-
-OpenWrt/ImmortalWrt **25.12** replaces opkg with Alpine's **apk**: `.apk` files,
-a binary `packages.adb` index, EC (prime256v1) keys in `/etc/apk/keys/`, and
-effectively mandatory signatures (unsigned needs `--allow-untrusted`). The
-package **Makefiles are unchanged** — the SDK release decides the format.
-
-CI builds this lane **in parallel** with the opkg feed (same manual triggers:
-`v*` tag push or `workflow_dispatch`): the `build-apk` jobs in
-`.gitea/workflows/release.yml` compile the same 4 packages through the official
-**ImmortalWrt 25.12 SDK** (tarballs from
+CI (`v*` tag push or `workflow_dispatch`) compiles the 4 packages through the
+official **ImmortalWrt 25.12 SDK** (tarballs from
 `downloads.immortalwrt.org/releases/25.12.1/targets/{x86/64,mediatek/filogic}/`)
-and publish **one release per arch** — rolling `apk-latest-x86_64` /
-`apk-latest-aarch64_cortex-a53`, or `apk-vX.Y.Z-<arch>` for a tagged version.
-Per-arch (unlike the combined opkg release) because apk filenames carry no
-architecture and packages are fetched relative to the `packages.adb` URL.
+and publishes **one release per arch**: the rolling `apk-latest-x86_64` /
+`apk-latest-aarch64_cortex-a53`, plus `apk-vX.Y.Z-<arch>` on a tag. Per-arch
+because apk filenames carry no architecture and packages are fetched *relative to
+the `packages.adb` URL*, so one flat multi-arch release would collide.

-> **Key:** apk cannot use the usign key. The apk trust anchor is the separate EC
-> public key **`dist/shater-apk.pem`** (generated once by `ci/gen-apk-key.sh`;
-> private half lives ONLY in the Gitea secret **`KEY_APK`**, the apk analog of
-> `KEY_BUILD`). Never regenerate either key — that invalidates every deployed
-> router's trust. The usign identity `shater-feed.pub` keeps signing the
-> opkg/24.10 feed, untouched.
+> **Key:** the trust anchor is the EC public key **`dist/shater-apk.pem`**
+> (generated once by `ci/gen-apk-key.sh`; the private half lives ONLY in the
+> Gitea secret **`KEY_APK`**). Never regenerate it — that invalidates every
+> deployed router's trust.

-One-time setup on a 25.12 router (BananaWRT `25.12-mtk-vendor` on the BPI-R3
-mini, BPI-R4 on 25.12, or the future 25.12 VM — `/etc/apk/arch` picks the right
-per-arch release automatically):
+### 5.1 Rolling or pinned — pick the repo URL deliberately
+
+The repo line names an **index file**, and which one you name is the whole
+update policy:
+
+| Repo line points at | Behaviour | Cost |
+|---|---|---|
+| `apk-latest-<arch>/packages.adb` (**rolling**) | Every release run REPLACES this release's assets, so `apk update && apk upgrade <our packages>` always sees the newest build. Install once, never touch the file again. | You get whatever CI published last; there is no per-router pin. |
+| `apk-vX.Y.Z-<arch>/packages.adb` (**pinned**) | The router stays on exactly that build. `apk update` will never offer a newer shater. | `/etc/apk/repositories.d/shater.list` must be edited **by hand on every upgrade**, on every router. |
+
+`mini_router` is deliberately on a **pinned** URL — a considered choice, and the
+hand-edit per release is its price. Use rolling unless you specifically want to
+freeze a device.
+
+> The rolling release used to go stale silently: publishing was an either/or, so
+> tag runs wrote only `apk-vX.Y.Z-<arch>` and `apk-latest-<arch>` was last
+> refreshed on 2026-07-24 at `0.2.0` while v0.2.9/v0.2.10 shipped. A router on
+> the rolling URL kept getting a successful `apk update` with nothing new. Fixed
+> 2026-07-25: `release-apk` writes the rolling pointer on **every** run and then
+> reads the release back over the Gitea API, asserting it holds our three
+> tag-versioned packages at exactly the version just built and **no** leftover
+> asset at another version (two versions of one package in one index would let
+> apk choose instead of us).
+
+### 5.2 One-time setup on the router
+
+BananaWRT `25.12-mtk-vendor` on the BPI-R3 mini, OpenWrt 25.12 on the BPI-R4, or
+the testbed VM — `/etc/apk/arch` picks the right per-arch release automatically:

 ```sh
 # 1) trust the apk feed key (any *.pem filename under /etc/apk/keys works).
@@ -203,6 +242,7 @@ wget -O /etc/apk/keys/shater-apk.pem \
  "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/shater-apk.pem"

 # 2) add the repo — the line points at the packages.adb INDEX FILE itself.
+#    (rolling; for a pinned router put apk-vX.Y.Z-$(cat /etc/apk/arch) here — §5.1)
 echo "https://git.qomar.pw/omar/shater/releases/download/apk-latest-$(cat /etc/apk/arch)/packages.adb" \
  > /etc/apk/repositories.d/shater.list

@@ -212,17 +252,34 @@ apk add luci-app-shater        # -> shater-core -> shaterd
 apk add byedpi                 # optional: ByeDPI desync egress
 ```

-### Updating
+### 5.3 Updating
+
+**Never run a bare `apk upgrade`.** With no arguments apk reconciles *every*
+installed package against *every* configured repository at once; on a router
+whose distfeeds point at a moving snapshot that can pull in — or roll back —
+unrelated system packages. Always name ours:

 ```sh
 apk update
-apk upgrade shaterd shater-core luci-app-shater byedpi   # only our own packages
+apk upgrade shaterd shater-core luci-app-shater byedpi
 ```

-Same rule as opkg: an upgrade is only offered when the feed version differs, so
-bump `PKG_RELEASE`/`PKG_VERSION` on every shipped change (apk shows it as
-`0.2.0-r1`). Pin a version instead of tracking rolling by pointing the repo line
-at `.../download/apk-vX.Y.Z-$(cat /etc/apk/arch)/packages.adb`.
+apk-tools 3 documents exactly this behaviour for `apk upgrade`: *"When no
+packages are specified, all packages are upgraded if possible. If list of
+packages is provided, only those packages are upgraded along with needed
+dependencies."* The equivalent form, which additionally re-pins the packages in
+`world`, is:
+
+```sh
+apk add -u shaterd shater-core luci-app-shater byedpi     # -u = --upgrade
+```
+
+Drop `byedpi` from either list if you never installed it. Check what you are on
+with `apk list -I shaterd shater-core luci-app-shater byedpi` — the version reads
+`0.2.7-r1` (§2.1: `PKG_VERSION-rPKG_RELEASE`, derived from the git tag by CI, so
+every build really is a new version; before that fix v0.2.2…v0.2.6 all published
+as `0.2.0-r3` and `apk update` offered nothing). Rolling vs pinned repo URL —
+§5.1.

 ### BananaWRT `25.12-mtk-vendor` compatibility

@@ -195,7 +195,7 @@ type Chain struct { Name string; Hops []string } // "group:<n>" | "node:<n>", L1
 type Egress struct { Name,Type,Interface,Target string } // interface|proxy|direct|block
 type Rule struct {
 	Name string; Enabled bool; Order int
-	Src []string; DstDomain,DstRuleset,DstIP []string; DstPort,Proto string
+	Src []string; DstRuleset []string; DstPort,Proto string // dst = ruleset only (v0.2 schema v2)
 	Target string // chain:|group:|node:|direct|block
 	Egress,Kill string
 	SchedEnabled bool; SchedDays []string; SchedStart,SchedEnd string; SchedUTCOffset int
@@ -257,8 +257,12 @@ Apply/rollback: `apSnapshot` (run→last-good, nft→last-good.nft, route marks)
 - `config node`: name, enabled, uri, mux, mux_concurrency, xudp_concurrency, xudp_udp443, sockopt_mark, tcp_fast_open, tcp_keepalive_idle.
 - `config group`: name, source, subscription, list node, strategy, include/exclude/filter_proto/filter_country, dedup, probe_url, probe_interval.
 - `config chain`: name, list hop.  `config egress`: name, type, interface, target.
- `config ruleset`: name, type(domain|ipcidr), source(inline|file|url), url, path, format, update_interval, list entry.
- `config rule`: name, enabled, order, list src/dst_domain/dst_ruleset/dst_ip, dst_port, proto, target, egress, kill, sched_enabled, list sched_day, sched_start/end/tz.
+- `config ruleset`: name, type(domain|ipcidr), source(inline|file|url|geosite|geoip), url, path, format, update_interval, list category, list entry.
+- `config rule`: name, enabled, order, list src, list dst_ruleset, dst_port, proto, target, egress, kill, sched_enabled, list sched_day, sched_start/end, sched_utc_offset.
+  v0.1 carried `dst_domain`/`dst_ip` on the rule itself; **schema v2 removed both** — a
+  destination is a `config ruleset` and nothing else. `shaterd migrate` folds each legacy
+  list into a generated `rule-<name>` (and `rule-<name>-ip`) inline ruleset; see
+  `DECISIONS.md` D21 for the entry-by-entry conversion table.
 - `config preset`: name, enabled, order, target.  `config profile`: name, enabled, priority, list match_iface, probe_url, probe_mode, sched_*, list enable_rule/disable_rule, default_target, default_egress.
 - `config resolver`: name, type, address, detour, pool.  `config dns_rule`: order, list match_domain/match_src, resolver.

@@ -6,7 +6,7 @@ OpenWrt). Лицо репозитория и быстрый старт — в к
 | Документ | О чём |
 |----------|-------|
 | [CONTEXT.md](CONTEXT.md) | **Начните здесь** — контекст проекта, история v0.1→v0.2, решения в кратце, testbed/инфра |
-| [INSTALL.md](INSTALL.md) | Сборка ship-артефакта (`shaterd`) и установка обоих фидов — opkg (24.10) и apk (25.12+) |
+| [INSTALL.md](INSTALL.md) | Сборка ship-артефакта (`shaterd`) и установка apk-фида (25.12+): роллинг или фиксация версии |
 | [ARCHITECTURE.md](ARCHITECTURE.md) | One-binary дизайн, auth-handoff LuCI→панель, data/DNS/apply-потоки (диаграммы) |
 | [FEATURES.md](FEATURES.md) | Полный список фич с тегами MVP/T1/T2 |
 | [ROADMAP.md](ROADMAP.md) | Фазовый план |
@@ -104,7 +104,7 @@ build new logic in the `shater/`, `panel/`, `openwrt/` overlay.

 ## Phase 8 — Ship it  ✅ DONE
 - Adapt CI to build/sign the single forked binary for both arches; publish the
-  signed opkg feed (reuse key `5ac4b177689cb8e0`); install/upgrade docs.
+  signed feed (apk since D22, EC key `dist/shater-apk.pem`); install/upgrade docs.
 - Set an upstream-rebase cadence (merge new sing-box-lx tags, run the smoke suite).

 ## Cross-cutting (every phase)
@@ -0,0 +1,170 @@
+# Живое тестирование shater v0.2.6 на mini_router
+
+**Дата:** 2026-07-25
+**Устройство:** Bananapi BPi-R3 Mini · ImmortalWrt **25.12-linkup** · `aarch64_cortex-a53`
+**Установка:** из подписанного apk-фида `apk-v0.2.6-aarch64_cortex-a53`
+**Пакеты:** `shaterd 0.2.0-r3`, `shater-core 0.2.0-r3`, `luci-app-shater 0.2.0-r2`, `byedpi 0.17.3-r1`
+**Сборка:** CI run 61, коммит `024e9308c` (вершина `main`)
+
+Сценарий: полное удаление предыдущей установки → чистая установка из фида →
+проверка дефолтного состояния → восстановление рабочего конфига с подписками
+(315 узлов) → функциональная проверка.
+
+**Итог: 79 проверок, 74 PASS, 5 находок** (детали и разбор — в
+`shater-bugs-2026-07-25.md` на рабочем столе).
+
+---
+
+## 1. Релиз и фид
+
+| # | Проверка | Результат |
+|---|---|---|
+| T1 | Публикация `apk-v0.2.6-<arch>` для обеих архитектур | PASS |
+| T2 | Ассеты: 4 `.apk` + `packages.adb` + `shater-apk.pem` | PASS |
+| T3 | `apk update` принимает индекс (проверка EC-подписи) | PASS |
+| T4 | Пакеты видны в нужных версиях (r3/r3/r2) | PASS |
+| T5 | Диагностика сборки: `kmod packages selected (=m): 0` (было 1078) | PASS |
+| T6 | Собраны ровно наши 4 пакета | PASS |
+| T7 | opkg-лейн v0.2.6 (24.10) тоже зелёный | PASS |
+
+## 2. Установка
+
+| # | Проверка | Результат |
+|---|---|---|
+| T8 | `apk add luci-app-shater byedpi` — 4 пакета | PASS |
+| T9 | Зависимости `kmod-nft-tproxy`/`kmod-nft-socket` из базового фида | PASS |
+| T10 | Целостность: `apk manifest` = sha256 файла на диске | PASS |
+| T11 | Установлен именно бинарь v0.2.6 (5 491 616 Б vs 5 488 336 Б в r2) | PASS |
+| T12 | init-скрипты `shater`, `shater-cron` | PASS |
+| T13 | `sysctl.d/99-shater.conf`, `hotplug.d/iface/99-shater` | PASS |
+| T14 | boot-линки `S99shater`, `K10shater`, `S96shater-cron` | PASS |
+
+## 3. Дефолтное состояние (чистая установка)
+
+| # | Проверка | Результат |
+|---|---|---|
+| T15 | Дефолтный конфиг создан uci-defaults (27 строк) | PASS |
+| T16 | `enabled='0'` — плоскость не ставится без согласия | PASS |
+| T17 | Заготовлен tproxy-inbound на LAN, пресеты выключены | PASS |
+| T18 | Демон стартует, `plane=none`, `table=false` | PASS |
+| T19 | Права конфига `-rw-------` (0600) | PASS |
+
+## 4. Восстановление рабочего конфига
+
+| # | Проверка | Результат |
+|---|---|---|
+| T20 | Восстановление из бэкапа (3213 UCI-строк) | PASS |
+| T21 | Кэш подписок цел: 315 узлов в 4 файлах | PASS |
+| T22 | `shaterd migrate` → `ok`, схема v1 | PASS |
+| T23 | Старт с реальным конфигом: `active`, `engine_running`, `plane=full` | PASS |
+
+## 5. Data plane
+
+| # | Проверка | Результат |
+|---|---|---|
+| T24 | Таблица `inet shater` создана (9 цепочек/сетов) | PASS |
+| T25 | 16 tproxy-правил | PASS |
+| T26 | `ip rule from all fwmark 0x2000 lookup shater` | PASS |
+| T27 | `accept_local=1` на `br-lan` | PASS |
+| T28 | DNS-divert: `dport 53 → tproxy :12345` для LAN-интерфейсов | PASS |
+| T29 | DoT заблокирован: `dport 853 reject` | PASS |
+| T30 | `block_doh=1`, правила присутствуют | PASS |
+| T31 | **Kill-switch fail-closed**: цепочка `forward` завершается `drop` для LAN (v4+v6) | PASS |
+| T32 | fw4 и dnsmasq не тронуты (свои таблицы целы) | PASS |
+
+## 6. Панель и API
+
+| # | Проверка | Результат |
+|---|---|---|
+| T33 | SPA отдаётся на `:8088` | PASS |
+| T34 | `shaterd mint-token` выдаёт одноразовый токен | PASS |
+| T35 | `/api/status` без сессии → **401** | PASS |
+| T36 | `/api/session` (POST, JSON) → 200 + cookie `HttpOnly; SameSite=Strict; Max-Age=28800` | PASS |
+| T37 | `/api/status` по cookie отдаёт данные, совпадающие с CLI | PASS |
+| T38 | `/api/config` — 340 записей узлов | PASS |
+| T39 | `/api/groups/health` — 103 протестировано, 13 живых, выбран `FR-vless-8` | PASS |
+| T40 | `/api/devices` — устройства с IPv4/IPv6/MAC | PASS |
+| T41 | `/api/interfaces` — `ewan/eth1 10.0.0.125/24 zone=wan` | PASS |
+| T42 | `/api/ruleset/status` — remote-ruleset обновлён сегодня | PASS |
+| T43 | `/api/stats` — memory backend, счётчики и top-domains | PASS |
+| T44 | `/api/stats/log` — query-log с доменом, qtype, rcode, сервером | PASS |
+| T45 | `/api/log?range=100` — пусто (следствие `log_file='0'`, не дефект) | OK |
+
+## 7. Жизненный цикл конфигурации
+
+| # | Проверка | Результат |
+|---|---|---|
+| T46 | `shaterd apply` → `{"changed":false}`, `can_rollback=true` | PASS |
+| T47 | `shaterd confirm` снимает авто-откат (`can_rollback=false`) | PASS |
+| T48 | `shaterd rollback` после confirm корректно сообщает об отсутствии last-good | PASS |
+| T49 | `shaterd reconcile` (SIGHUP) не роняет движок | PASS |
+| T50 | `shaterd sub update all-qomar` — реально обновил 143 узла | PASS |
+| T51 | `shaterd blocklist update` → reconcile signalled | PASS |
+| T52 | `shaterd schedule due` → reconcile signalled | PASS |
+
+## 8. Устойчивость
+
+| # | Проверка | Результат |
+|---|---|---|
+| T53 | `kill -9` демона → procd поднимает новый PID | PASS |
+| T54 | После respawn: `engine_running=true`, `plane=full` | PASS |
+| T55 | `stop` снимает таблицу `inet shater` полностью | PASS |
+| T56 | `stop` → пауза → `start`: плоскость восстанавливается | PASS |
+| T57 | Сеть при остановленном shater не деградирует | PASS |
+| T58 | Память: 253 МБ занято из 2 ГБ при работающем движке | PASS |
+
+## 9. DNS
+
+| # | Проверка | Результат |
+|---|---|---|
+| T59 | Резолв через `127.0.0.1` | PASS |
+| T60 | LAN-клиенты резолвят через движок (query-log растёт) | PASS |
+| T61 | `.lan`-домены остаются за dnsmasq | PASS |
+| T62 | dnsmasq жив и слушает на всех адресах | PASS |
+| T63 | **Резолв через LAN-адрес `10.67.0.1` после `restart`** | **FAIL — B3** |
+| T64 | Тот же резолв после `stop` → пауза → `start` | PASS |
+
+## 10. Конфигурация и логи
+
+| # | Проверка | Результат |
+|---|---|---|
+| T65 | 5 правил маршрутизации, 2 профиля, активен `ethernet-uplink` | PASS |
+| T66 | **Два правила `default`, оба catch-all — нижнее живое, верхнее мертво** | **FAIL — B1** |
+| T67 | **`shaterd nodes` всегда возвращает `[]`** | **FAIL — B2** |
+| T68 | Логи уходят в syslog (`log_syslog=1`, 22 записи) | PASS |
+| T69 | **ANSI-escape коды в syslog** | **FAIL — B5** |
+| T70 | `loglevel=warning` соблюдается | PASS |
+| T71–T79 | Прочие проверки состояния (статус-поля, права, uptime, счётчики, целостность таблиц) | PASS |
+
+---
+
+## Находки
+
+| ID | Суть | Важность |
+|---|---|---|
+| **B1** | Два catch-all правила `default`; одно из них не работает никогда. **Поправка к первоначальному диагнозу:** правило без условий задаёт `route.Final`, а не выпускается как match-all, поэтому выигрывает ПОСЛЕДНЕЕ (`order=100 → group:auto`) — трафик идёт через прокси, а мёртвая настройка это `order=20 → direct` | средняя |
+| **B2** | `shaterd nodes` — заглушка, всегда `[]`, хотя usage обещает список узлов (в кэше 315, в `/api/config` 340) | средняя |
+| **B3** | После `service shater restart` резолв к LAN-адресу роутера не работает и не восстанавливается; `stop`+пауза+`start` — работает (гонка) | средняя |
+| **B4** | `PKG_RELEASE` не менялся с v0.2.1 → v0.2.2…v0.2.6 выходят как `r3` при разном содержимом; `apk upgrade` не увидит обновления | средняя |
+| **B5** | ANSI-раскраска попадает в syslog | низкая |
+
+Разбор с воспроизведением — в `shater-bugs-2026-07-25.md`.
+
+## История CI по этому релизу
+
+Путь до зелёной сборки apk-лейна занял четыре итерации, каждая вскрывала
+следующий слой одной причины:
+
+| Тег | Что чинили | Итог |
+|---|---|---|
+| v0.2.2 | — (первый прогон с фиксами аудита) | `Disk quota exceeded`, 3593 `apk mkpkg kmod-*` |
+| v0.2.3 | `.config` строится с нуля, а не дописывается | 1078 kmod — SDK вообще не везёт `.config` |
+| v0.2.4 | Выключены `ALL`/`ALL_KMODS`/`ALL_NONSHARED` | 1078 kmod — они выбираются не через `ALL_KMODS` |
+| v0.2.5 | Второй проход: явное `is not set` для каждого kmod | 1078 kmod — kconfig игнорирует user-значение у беспромптовых символов |
+| **v0.2.6** | Удаление сгенерированных блоков `config PACKAGE_*` (`default m`) из `Config-build.in` | **0 kmod, сборка зелёная** |
+
+Корень: `target/sdk/Makefile` генерирует `Config-build.in` прогоном
+`convert-config.pl` по конфигу бильдбота, где `ALL_KMODS=y` уже развернулся в
+`CONFIG_PACKAGE_kmod-*=m` на каждый модуль. Фильтр `next if /^(# )?CONFIG_PACKAGE/`
+в скрипте стоит в ветке `else`, куда строка со знаком `=` не попадает, поэтому
+каждый kmod приезжает в SDK как безусловный `default m`.
@@ -15,6 +15,17 @@
 include $(TOPDIR)/rules.mk

 PKG_NAME:=byedpi
+
+# DELIBERATELY NOT auto-versioned from our git tag (unlike shaterd/shater-core/
+# luci-app-shater, which take SHATER_PKG_VERSION/SHATER_PKG_RELEASE from
+# ci/version.sh). PKG_VERSION here is THIRD-PARTY UPSTREAM's version — it is what
+# PKG_SOURCE_URL/PKG_HASH pin, and what tells an operator which ByeDPI is
+# actually installed. Stamping our tag on it would be both a lie and a
+# regression: our tags are 0.2.x, and the version comparator (apk-tools 3,
+# verified) reads 0.2.7 < 0.17.3 — component-wise numerically, 2 < 17
+# — so the "new" package would be a DOWNGRADE and routers would refuse it.
+# Bump PKG_RELEASE BY HAND when *our packaging* of it changes (init script, uci
+# defaults, build flags); bump PKG_VERSION+PKG_HASH when upstream releases.
 PKG_VERSION:=0.17.3
 PKG_RELEASE:=1

@@ -24,8 +24,13 @@ LUCI_TITLE:=LuCI thin launcher for Shater (mini dashboard + panel handoff)
 LUCI_DEPENDS:=+shater-core +rpcd
 LUCI_PKGARCH:=all

-PKG_VERSION:=0.2.0
-PKG_RELEASE:=2
+# Version comes from the git tag via ci/version.sh -> SHATER_PKG_VERSION /
+# SHATER_PKG_RELEASE in the SDK build env (see openwrt/shaterd/Makefile for the
+# full rationale — bug B4). The literals are the manual/offline fallback only.
+# These MUST stay above the luci.mk include: luci.mk only defaults PKG_VERSION/
+# PKG_RELEASE when they are still unset, and the i18n subpackages inherit them.
+PKG_VERSION:=$(if $(SHATER_PKG_VERSION),$(SHATER_PKG_VERSION),0.2.0)
+PKG_RELEASE:=$(if $(SHATER_PKG_RELEASE),$(SHATER_PKG_RELEASE),1)

 PKG_MAINTAINER:=Shater <maqrota@icloud.com>
 PKG_LICENSE:=GPL-3.0-or-later
@@ -13,8 +13,13 @@
 include $(TOPDIR)/rules.mk

 PKG_NAME:=shater-core
-PKG_VERSION:=0.2.0
-PKG_RELEASE:=3
+
+# Version comes from the git tag via ci/version.sh -> SHATER_PKG_VERSION /
+# SHATER_PKG_RELEASE in the SDK build env (see openwrt/shaterd/Makefile for the
+# full rationale — bug B4: v0.2.2…v0.2.6 all shipped as 0.2.0-r3). The literals
+# are the manual/offline fallback only.
+PKG_VERSION:=$(if $(SHATER_PKG_VERSION),$(SHATER_PKG_VERSION),0.2.0)
+PKG_RELEASE:=$(if $(SHATER_PKG_RELEASE),$(SHATER_PKG_RELEASE),1)

 PKG_MAINTAINER:=Shater <maqrota@icloud.com>
 PKG_LICENSE:=GPL-2.0-or-later
@@ -62,7 +62,29 @@ config inbound
 #	list node 'my-node'
 #
 # A routing rule. target: chain:<n>|group:<n>|node:<n>|egress:<n>|direct|block.
-# Match on src / dst_domain / dst_ruleset / dst_ip / dst_port / proto.
+# Match on src / dst_ruleset / dst_port / proto. A rule with NO matcher at all is
+# the default route for everything that reached it.
+#
+# WHERE the traffic is going is named ONLY by dst_ruleset — one or more
+# `config ruleset` names; the rule matches when ANY of them matches. There is no
+# inline domain or address list on a rule (`dst_domain`/`dst_ip` were removed in
+# schema v2): a destination list is written once as a ruleset, compiled into a
+# .srs and shared by every rule that references it. `shaterd migrate` converts
+# older configs automatically, creating a `rule-<name>` ruleset per rule.
+#config ruleset
+#	option name 'blocked-video'
+#	option type 'domain'
+#	option source 'inline'
+#	list entry 'youtube.com'
+#	list entry 'suffix:googlevideo.com'
+#
+#config rule
+#	option name 'video-via-main'
+#	option enabled '1'
+#	option order '50'
+#	list dst_ruleset 'blocked-video'
+#	option target 'group:main'
+#
 #config rule
 #	option name 'all-via-main'
 #	option enabled '1'
@@ -83,11 +105,16 @@ config inbound
 #	option type 'direct'
 #	option dpi 'fragment'
 #
+#config ruleset
+#	option name 'youtube'
+#	option source 'geosite'
+#	list category 'youtube'
+#
 #config rule
 #	option name 'youtube-fragment'
 #	option enabled '1'
 #	option order '50'
-#	list dst_domain 'geosite:youtube'
+#	list dst_ruleset 'youtube'
 #	option target 'egress:frag'
 #
 # A DNS resolver (type: doh|dot|plain|local|fakeip). `detour` routes its queries
@@ -47,6 +47,13 @@ PROG=/usr/bin/shaterd
 # hotplug/shater-cron touch the data plane. tmpfs => cleared by reboot, so
 # nothing reconciles before this init has run at boot.
 ACTIVE_FLAG=/var/run/shater.active
+# Written by `shaterd run`; the single-owner token this init waits on so a
+# restart never overlaps a new data plane with the previous one's teardown.
+PIDFILE=/var/run/shaterd.pid
+# Seconds `start` will wait for a predecessor to finish its teardown. Must be
+# >= term_timeout below (procd's hard cap on a predecessor's life after SIGTERM)
+# so we never give up while procd is still letting it shut down cleanly.
+STOP_WAIT_SECS=40

 # --- helpers ---------------------------------------------------------------

@@ -66,6 +73,54 @@ _slog() {
 	[ "$(uci -q get shater.globals.log_syslog)" = "0" ] || logger -t shater "$@"
 }

+# Echo the pid of a LIVE `shaterd run`, or fail. The pidfile is written by the
+# daemon itself and removed only by the daemon that owns it, AFTER its teardown
+# has completed — so "pidfile names a live process" is precisely "the previous
+# data plane has not been dismantled yet".
+shater_daemon_pid() {
+	local pid
+	pid=$(cat "$PIDFILE" 2>/dev/null) || return 1
+	[ -n "$pid" ] || return 1
+	kill -0 "$pid" 2>/dev/null || return 1
+	echo "$pid"
+}
+
+# Block until no predecessor daemon is left, bounded by STOP_WAIT_SECS.
+#
+# WHY THIS EXISTS. procd's `stop` is ASYNCHRONOUS: rc.common's `restart` is
+# literally `stop; start`, and the `service delete` ubus call returns the moment
+# procd has SENT SIGTERM — not when the instance is gone. `start` therefore
+# re-adds the instance while the outgoing `shaterd run` is still executing its
+# honest teardown (engine close, then `nft delete table`, `ip rule`/`ip route`
+# removal and the per-iface sysctl restore). The result is that `restart` is NOT
+# equivalent to `stop` + pause + `start`: the new plane is stood up on top of
+# kernel state the old one has not finished removing, which is what B3 (DNS to
+# the router's own LAN address dead after a restart, and never recovering) came
+# out of. Waiting here restores the equivalence, and costs literally nothing when
+# there is no predecessor — the check runs before the first sleep.
+#
+# Returning non-zero does NOT abort the start: the daemon carries its own
+# single-owner guard and will refuse (or wait) on its side. Better to hand the
+# decision to the process that can actually see the plane than to leave the box
+# with no service at all.
+shater_wait_stopped() {
+	local i=0 pid
+	pid=$(shater_daemon_pid) || return 0
+	_slog -p daemon.info \
+		"restart: waiting for the previous shaterd (pid $pid) to finish tearing the data plane down"
+	while [ "$i" -lt "$STOP_WAIT_SECS" ]; do
+		sleep 1
+		i=$((i + 1))
+		shater_daemon_pid >/dev/null || {
+			_slog -p daemon.info "restart: previous shaterd exited after ${i}s; starting a fresh one"
+			return 0
+		}
+	done
+	_slog -p daemon.warn \
+		"restart: previous shaterd (pid $pid) still alive after ${STOP_WAIT_SECS}s — starting anyway"
+	return 1
+}
+
 # --- procd lifecycle -------------------------------------------------------

 start_service() {
@@ -87,9 +142,31 @@ start_service() {
 		return 0
 	fi

+	# Do not stand a new data plane up on top of one that is still being taken
+	# down. On `restart` procd has only just SIGTERMed the previous instance and
+	# returned; this is the handshake that makes `restart` == `stop` + pause +
+	# `start`. It also keeps `migrate` below from rewriting UCI underneath a
+	# daemon that is still reading it. No-op (and no delay) when nothing is
+	# running, which is the boot case.
+	shater_wait_stopped
+
 	# Bring the UCI schema forward before the daemon reads it (idempotent;
 	# refuses a newer schema) so an upgraded package never applies a stale config.
-	"$PROG" migrate >/dev/null 2>&1
+	#
+	# THE FAILURE IS LOGGED, NOT SWALLOWED. This is the only place the schema
+	# migration runs at boot (`shaterd run`, the SIGHUP reconcile and the panel's
+	# config write all read UCI directly), so if it fails here it does not get
+	# retried until the next start. And it CAN fail for a mundane reason — a full
+	# /overlay makes `uci commit` fail — after which the config still carries the
+	# schema-v1 `dst_domain`/`dst_ip` options. The daemon holds every rule that
+	# still has them DISABLED and reports it, so nothing is silently misrouted, but
+	# rules the operator wrote are then not in force and the reason has to be
+	# visible somewhere. Hence: log the binary's own stderr, and start anyway —
+	# refusing to start would take the admin panel down with it, and the panel is
+	# the only way to fix the box.
+	local migrate_out
+	migrate_out=$("$PROG" migrate 2>&1) || _slog -p daemon.err \
+		"UCI schema migration FAILED: ${migrate_out:-no output from $PROG migrate}. Starting anyway; routing rules that still carry the removed dst_domain/dst_ip options stay DISABLED until this succeeds. Free space on /overlay and re-run '$PROG migrate', or restart the service."

 	procd_open_instance shater
 	# shaterd runs in the FOREGROUND under procd (must never daemonize). `run` is
@@ -111,7 +188,16 @@ start_service() {
 	procd_set_param stderr 1
 	# Give the daemon room to run its honest teardown (engine.Close + netplane
 	# restore) before procd SIGKILLs it.
-	procd_set_param term_timeout 10
+	#
+	# 30s, not 10s: an engine holding a few hundred outbounds closes its
+	# urltest/observatory goroutines and flushes experimental.cache_file to FLASH
+	# before the netplane teardown even starts, and on eMMC/NAND that alone can
+	# outlast 10s. A SIGKILL there aborts the teardown at an arbitrary point and
+	# leaves the plane HALF removed — the nft table gone but the policy routing
+	# still installed, or vice versa — which is precisely the class of leftover
+	# state the successor's idempotent fast-path cannot see and never repairs.
+	# Shutdown is bounded by procd either way; we are only choosing where.
+	procd_set_param term_timeout 30
 	procd_close_instance

 	# Mark the stack live for hotplug/cron — but ONLY when interception is
@@ -141,10 +227,12 @@ stop_service() {
 reload_service() {
 	# Fired by the `shater` config.change reload-trigger (LuCI Save & Apply /
 	# reload_config). Simplest correct behaviour: stop + start. `stop` clears the
-	# flag and SIGTERMs the daemon (honest teardown); `start` re-guards on
-	# enabled and, if still enabled, launches a fresh `shaterd run` that reads
-	# the new UCI and applies it. When the stack is disabled, `start` is a no-op,
-	# so a disable+apply cleanly tears everything down.
+	# flag and SIGTERMs the daemon (honest teardown); `start` WAITS for that
+	# teardown to actually finish (shater_wait_stopped) and then launches a fresh
+	# `shaterd run` that reads the new UCI and applies it. When the stack is
+	# disabled, `start` is a no-op, so a disable+apply cleanly tears everything
+	# down. Because the wait lives in start_service, this path gets the same
+	# stop-then-start ordering guarantee as `restart`.
 	stop
 	start
 }
@@ -34,8 +34,17 @@
 include $(TOPDIR)/rules.mk

 PKG_NAME:=shaterd
-PKG_VERSION:=0.2.0
-PKG_RELEASE:=3
+
+# VERSIONING — derived from the git tag, NOT hand-maintained here (bug B4).
+# ci/version.sh turns `git describe` into SHATER_PKG_VERSION/SHATER_PKG_RELEASE
+# (tag vX.Y.Z -> X.Y.Z + r1; off-tag -> last tag + r<commits+1>), and
+# ci/build-feed-apk.sh exports them into the SDK build env. ci/sdk-build-apk.sh
+# then ASSERTS that the produced .apk really carries that version, so a lost env
+# can never silently ship a stale one again.
+# The literals below are ONLY the manual/offline fallback (no CI, no git) — they
+# are not "the release version"; releases are named by the tag.
+PKG_VERSION:=$(if $(SHATER_PKG_VERSION),$(SHATER_PKG_VERSION),0.2.0)
+PKG_RELEASE:=$(if $(SHATER_PKG_RELEASE),$(SHATER_PKG_RELEASE),1)

 PKG_MAINTAINER:=Shater <maqrota@icloud.com>
 PKG_LICENSE:=GPL-3.0-or-later
@@ -95,8 +104,8 @@ define Package/shaterd/install
 	$(INSTALL_BIN) $(CURDIR)/files/$(SHATERD_BIN) $(1)/usr/bin/shaterd
 endef

-# This package ships ONLY the binary — no init script — so opkg's default
-# postinst never touches the running service. On `opkg upgrade shaterd` the new
+# This package ships ONLY the binary — no init script — so the package manager's
+# postinst never touches the running service. On `apk upgrade shaterd` the new
 # ELF lands at /usr/bin/shaterd while the OLD image keeps running from its
 # unlinked inode: the upgrade silently has no effect until the next reboot, and
 # meanwhile the new CLI (`shaterd reconcile`, `status`, `mint-token` — invoked by
@@ -8,7 +8,8 @@
    "dev": "vite",
    "build": "tsc --noEmit && vite build",
    "preview": "vite preview",
-    "typecheck": "tsc --noEmit"
+    "typecheck": "tsc --noEmit",
+    "test": "node --test src/*.test.ts"
  },
  "dependencies": {
    "react": "^18.3.1",
@@ -405,3 +405,73 @@
  color: var(--dim);
  max-width: 74ch;
 }
+
+/* ---- inline rename (shared) ----
+   The pencil-in-the-row interaction: click the ✎ beside a name, type over it,
+   Enter commits / Esc cancels / blur commits. Lifted out of Devices.css when
+   Nodes grew the same affordance — one interaction, one set of rules, so the two
+   pages can never drift apart. `--locked` is the same control with the action
+   withheld: it stays visible and focusable-looking so a missing rename reads as
+   a stated rule, not a dead button. */
+.inline-rename {
+  flex: none;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  width: 22px;
+  height: 22px;
+  padding: 0;
+  border: 1px solid transparent;
+  border-radius: 5px;
+  background: none;
+  color: var(--faint);
+  font-size: 12px;
+  line-height: 1;
+  cursor: pointer;
+  transition: color 0.15s, background 0.15s, border-color 0.15s;
+}
+.inline-rename:hover:not(:disabled) {
+  color: var(--accent);
+  background: color-mix(in srgb, var(--accent) 12%, transparent);
+}
+.inline-rename:focus-visible {
+  color: var(--accent);
+  border-color: var(--accent);
+  outline: 2px solid var(--accent);
+  outline-offset: 1px;
+}
+.inline-rename:disabled {
+  opacity: 0.5;
+  cursor: default;
+}
+/* Withheld, not broken: keep the glyph readable and let the cursor say "there is
+   a reason" rather than dimming it into invisibility. */
+.inline-rename--locked {
+  opacity: 0.75;
+  cursor: help;
+}
+.inline-rename--locked:hover {
+  color: var(--dim);
+  background: none;
+}
+
+.inline-rename-input {
+  min-width: 0;
+  max-width: 24ch;
+  padding: 4px 8px;
+  border: 1px solid var(--accent);
+  border-radius: 6px;
+  background: var(--sink);
+  color: var(--ink);
+  font-size: 13px;
+  font-weight: 600;
+  letter-spacing: 0.01em;
+  box-shadow: 0 1px 2px var(--shadow) inset;
+}
+.inline-rename-input:focus-visible {
+  outline: 2px solid var(--accent);
+  outline-offset: 1px;
+}
+.inline-rename-input:disabled {
+  opacity: 0.55;
+}
@@ -51,6 +51,38 @@ export class ApiError extends Error {
 */
 export type Plane = 'full' | 'hold' | 'none'

+/**
+ * Where the router's traffic actually ENDS UP, decided by the daemon from the
+ * engine config it is running (apply.Status.traffic ← generate.TrafficOf).
+ *
+ *   tunnel  — the default route goes into a tunnel: everything not matched by a
+ *             more specific rule is proxied.
+ *   split   — the default leaves directly, but some rules do tunnel their traffic.
+ *   direct  — the default leaves directly and nothing is tunnelled at all.
+ *   blocked — the default is the fail-closed backstop: unmatched traffic is
+ *             dropped, not let out. Nothing leaks.
+ *
+ * `plane` DOES NOT ANSWER THIS and must never be read as if it did. `plane` says
+ * how much of the data plane is installed (nft table, policy routing, engine up);
+ * a router whose only rule is `default → direct` has all of it and sends the whole
+ * LAN out the plain WAN with its real address. That combination — plane "full",
+ * traffic "direct" — was live on a user's router under a green "Protected" LED.
+ */
+export type TrafficVerdict = 'tunnel' | 'split' | 'direct' | 'blocked'
+
+export interface Traffic {
+  // '' or absent ⇒ not known (daemon that predates this field, nothing applied
+  // yet, or the plane is on hold). NEVER treat unknown as 'tunnel'.
+  verdict?: TrafficVerdict | ''
+  // The outbound tag the engine's default route names, in the engine's own
+  // vocabulary ("direct", "block", a node/group tag). Diagnostic — wording is
+  // driven by `verdict`, never by parsing this.
+  default?: string
+  // How many of the engine's route rules send their matched traffic into a tunnel.
+  // Separates "some of your traffic is protected" from "none of it is".
+  tunnel_rules?: number
+}
+
 /**
 * One thing the last apply could not do. Deliberately fail-OPEN with a warning
 * rather than refusing the whole config (the alternative was taking the network
@@ -88,6 +120,10 @@ export interface Status {
  // How much of the data plane is installed. Absent on older daemons ⇒ unknown,
  // in which case the UI shows nothing rather than guessing "full".
  plane?: Plane
+  // Where the traffic actually goes under the running config. Absent on older
+  // daemons ⇒ unknown; see TrafficVerdict for why this is a separate question
+  // from `plane`.
+  traffic?: Traffic
  // Findings from the last apply. ALWAYS an array from the daemon (never null);
  // empty means the last apply was clean. Pre-sorted critical-first and capped at
  // 50, where a truncated list ends with an `info` entry saying "suppressed".
@@ -809,9 +845,17 @@ export interface Rule {
  Enabled: boolean
  Order: number
  Src?: string[] | null
-  DstDomain?: string[] | null
+  /**
+   * WHERE the traffic is going — the rule's only destination matcher. Each entry
+   * names a {@link Ruleset}; the rule matches when ANY of them matches.
+   *
+   * There is no inline domain or address list on a rule. `dst_domain`/`dst_ip`
+   * were removed in schema v2, and `shaterd migrate` folds every existing one
+   * into a generated `rule-<name>` ruleset, so a destination list is written and
+   * edited in exactly one place and compiled once into a .srs that every rule
+   * referencing it shares.
+   */
  DstRuleset?: string[] | null
-  DstIP?: string[] | null
  DstPort?: string
  /**
   * Narrow the rule to one transport or one sniffed application protocol. A
@@ -1187,6 +1231,61 @@ export function getStatsConns(q: number | StatsLogQuery = {}): Promise<ConnLogEn
  return MOCK ? mock.getStatsConns(o) : req<ConnLogEntry[]>(`api/stats/conns${statsLogQS(o)}`)
 }

+/**
+ * One routing rule's reachability verdict — the rule analogue of
+ * {@link ChainHealth}.used: a quiet note about the ROUTING CONFIG, never a health
+ * signal.
+ *
+ * `unreachable` means the rule can NEVER take effect, whatever the traffic. Today
+ * the daemon reports exactly one certain case, and it is a subtle one: a rule with
+ * no conditions at all is not matched in sequence — it becomes the router's
+ * default. Two such rules therefore retire each other, and the LAST one by Order
+ * wins, so an earlier "default → direct" is dead even though it sorts first. A
+ * condition-less rule never retires a rule that HAS conditions: those are matched
+ * ahead of the default whatever their Order.
+ *
+ * `index` is the rule's position in GET /api/config's `Rules`, which is how a
+ * verdict is matched to a row — rule names are not unique, and the config that
+ * prompted this had two rules both called `default`. `name`/`order` are echoed so
+ * a page holding a verdict fetched before an edit can check it still describes the
+ * row it is about to badge, and drop it silently otherwise.
+ */
+export interface RuleReach {
+  index: number
+  name: string
+  order: number
+  unreachable: boolean
+  /** The rule that supersedes this one; absent when `unreachable` is false. */
+  shadowed_by?: string
+  /** Its index in `Rules`, or -1 when there is none. */
+  shadowed_by_index: number
+  shadowed_by_order?: number
+  /** Operator-facing sentence; absent when `unreachable` is false. */
+  reason?: string
+  /**
+   * Whether the rule is IN FORCE right now — `Rule.Enabled` after the active WAN
+   * profile's overrides. This is NOT `GET /api/config`'s `Enabled`: that one is
+   * the desired state the page PUTs back, and on a router with profiles the two
+   * legitimately disagree. Draw rows from this; keep the switch on the other.
+   */
+  effective_enabled: boolean
+  /** The active profile that CHANGED this rule's state; absent when none did. */
+  overridden_by?: string
+  /** Which way it went. Absent together with `overridden_by`. */
+  override?: 'enabled' | 'disabled'
+}
+
+/** GET /api/rules/reachability. `rules` is ALWAYS an array, one entry per rule in
+ *  the same order as GET /api/config's `Rules`. */
+export interface RulesReachability {
+  rules: RuleReach[]
+}
+
+/** GET /api/rules/reachability — which routing rules can never fire, and why. */
+export function getRulesReachability(): Promise<RulesReachability> {
+  return MOCK ? mock.getRulesReachability() : req<RulesReachability>('api/rules/reachability')
+}
+
 /** GET /api/ruleset/status — remote rule-set / blocklist freshness + rule counts. */
 export function getRulesetStatus(): Promise<RulesetStatus[]> {
  return MOCK ? mock.getRulesetStatus() : req<RulesetStatus[]>('api/ruleset/status')
@@ -1,4 +1,5 @@
-/* Buttons — mono, uppercase. .btn is ghost; .btn.primary is solid orange. */
+/* Buttons — mono, uppercase. .btn is ghost; .btn.primary is solid orange;
+ * .btn.crit is the solid-red destructive commit. */
 .btn {
  display: inline-block;
  padding: 7px 12px;
@@ -30,3 +31,16 @@
  color: #fff;
  filter: brightness(1.05);
 }
+
+/* Destructive commit. The fill is crit stepped a little toward black so white
+ * label text clears 4.5:1 in BOTH themes — the raw --crit is bright enough in
+ * dark mode to fall under it. Red here always means "this removes something". */
+.btn.crit {
+  border-color: transparent;
+  background: color-mix(in srgb, var(--crit) 88%, #000);
+  color: #fff;
+}
+.btn.crit:hover {
+  color: #fff;
+  filter: brightness(1.08);
+}
@@ -1,19 +1,27 @@
 import './Button.css'
+import { forwardRef } from 'react'
 import type { ButtonHTMLAttributes } from 'react'

 export interface ButtonProps extends ButtonHTMLAttributes<HTMLButtonElement> {
-  /** `primary` is the solid-orange call to action; `ghost` is the default. */
-  variant?: 'ghost' | 'primary'
+  /**
+   * `primary` is the solid-orange call to action; `crit` is the solid-red
+   * destructive commit (delete, remove) — semantic crit, never the accent;
+   * `ghost` is the default.
+   */
+  variant?: 'ghost' | 'primary' | 'crit'
 }

-export function Button({ variant = 'ghost', className, type, ...rest }: ButtonProps) {
+/** Ref-forwarding so a dialog can park focus on a specific button. */
+export const Button = forwardRef<HTMLButtonElement, ButtonProps>(function Button(
+  { variant = 'ghost', className, type, ...rest },
+  ref,
+) {
  return (
    <button
+      ref={ref}
      type={type ?? 'button'}
-      className={['btn', variant === 'primary' ? 'primary' : '', className]
-        .filter(Boolean)
-        .join(' ')}
+      className={['btn', variant === 'ghost' ? '' : variant, className].filter(Boolean).join(' ')}
      {...rest}
    />
  )
-}
+})
@@ -0,0 +1,162 @@
+/* <ConfirmDialog> — the safety interlock plate.
+ *
+ * This replaces the browser's native confirm dialog, which a browser can mute for
+ * good ("prevent this page from creating additional dialogs"): after that it
+ * returns false with no dialog at all, so every delete button in the panel goes
+ * dead and silent with no way to recover short of a page reload. We draw the
+ * plate ourselves, so nothing can suppress it.
+ *
+ * Faceplate language: a small rack module lifted off the panel — corner screws
+ * (reused from Faceplate.css), an engraved label, a groove above the actions.
+ * Destructive intent is carried by the crit semantic, never by the orange accent:
+ * accent means "this control is active", crit means "this destroys something".
+ */
+
+/* The veil is a fixed dark wash in both themes — a light scrim over a light
+ * panel would not read as "the panel is out of reach". Follows the tokens.css
+ * pattern: light base, dark via media query, data-theme overrides win both ways. */
+.cfm-scrim {
+  --cfm-veil: rgba(33, 29, 21, 0.52);
+}
+@media (prefers-color-scheme: dark) {
+  .cfm-scrim {
+    --cfm-veil: rgba(0, 0, 0, 0.66);
+  }
+}
+:root[data-theme='light'] .cfm-scrim {
+  --cfm-veil: rgba(33, 29, 21, 0.52);
+}
+:root[data-theme='dark'] .cfm-scrim {
+  --cfm-veil: rgba(0, 0, 0, 0.66);
+}
+
+.cfm-scrim {
+  position: fixed;
+  inset: 0;
+  z-index: 200;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  /* Short viewports: the plate scrolls with the veil instead of being clipped. */
+  overflow-y: auto;
+  padding: calc(var(--u, 8px) * 2);
+  background: var(--cfm-veil);
+  animation: cfm-veil-in 0.14s ease-out;
+}
+
+.cfm-card {
+  position: relative;
+  width: min(32rem, 100%);
+  max-height: calc(100dvh - var(--u, 8px) * 4);
+  overflow-y: auto;
+  padding: calc(var(--u, 8px) * 3.25);
+  border: 1px solid var(--groove);
+  border-radius: 12px;
+  /* same brushed plate as <Faceplate>, one step brighter so it reads as lifted */
+  background:
+    repeating-linear-gradient(
+      90deg,
+      transparent 0 2px,
+      color-mix(in srgb, var(--edge) 30%, transparent) 2px 3px
+    ),
+    linear-gradient(180deg, var(--raised), color-mix(in srgb, var(--raised) 82%, var(--panel)));
+  box-shadow:
+    0 1px 0 var(--edge) inset,
+    0 30px 60px -22px var(--shadow),
+    0 4px 12px var(--shadow);
+  animation: cfm-card-in 0.18s cubic-bezier(0.2, 0.7, 0.3, 1);
+}
+.cfm-card:focus {
+  outline: none;
+}
+
+/* `still` is set from usePrefersReducedMotion — the plate appears, it never
+ * travels. (The global reduced-motion rule in tokens.css also neutralises the
+ * duration; this keeps the intent explicit at the component.) */
+.cfm-scrim.still,
+.cfm-scrim.still .cfm-card {
+  animation: none;
+}
+
+@keyframes cfm-veil-in {
+  from {
+    opacity: 0;
+  }
+  to {
+    opacity: 1;
+  }
+}
+@keyframes cfm-card-in {
+  from {
+    opacity: 0;
+    transform: translateY(6px) scale(0.99);
+  }
+  to {
+    opacity: 1;
+    transform: none;
+  }
+}
+
+/* ---- header: engraved label + state LED ---- */
+.cfm-hd {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  margin-bottom: calc(var(--u, 8px) * 1.5);
+}
+.cfm-label {
+  flex: 1;
+  font-family: var(--font-mono);
+  font-size: 10px;
+  letter-spacing: var(--track-label-wide, 0.24em);
+  color: var(--dim);
+  text-transform: uppercase;
+}
+
+/* ---- copy ---- */
+.cfm-title {
+  margin: 0;
+  font-family: var(--font-mono);
+  font-weight: 700;
+  font-size: 17px;
+  line-height: 1.35;
+  color: var(--ink);
+  /* names can be long and unbroken — wrap rather than push the plate wide */
+  overflow-wrap: anywhere;
+}
+.cfm-body {
+  margin: calc(var(--u, 8px) * 1.5) 0 0;
+  max-width: 52ch;
+  font-family: var(--font-sans);
+  font-size: 13.5px;
+  line-height: 1.6;
+  color: var(--dim);
+  overflow-wrap: anywhere;
+}
+
+/* ---- action bar ---- */
+.cfm-actions {
+  display: flex;
+  justify-content: flex-end;
+  gap: calc(var(--u, 8px));
+  margin-top: calc(var(--u, 8px) * 3);
+  padding-top: calc(var(--u, 8px) * 2);
+  border-top: 1px solid var(--groove);
+}
+
+@media (max-width: 420px) {
+  .cfm-card {
+    padding: calc(var(--u, 8px) * 2.5);
+  }
+  .cfm-actions {
+    flex-wrap: wrap;
+  }
+  .cfm-actions .btn {
+    flex: 1 1 auto;
+    text-align: center;
+  }
+  /* screws crowd a small plate — drop them rather than collide with the copy */
+  .cfm-card > .screw {
+    display: none;
+  }
+}
@@ -0,0 +1,281 @@
+import './ConfirmDialog.css'
+import {
+  createContext,
+  useCallback,
+  useContext,
+  useEffect,
+  useId,
+  useRef,
+  useState,
+} from 'react'
+import type { ReactNode } from 'react'
+import { createPortal } from 'react-dom'
+import { Button } from './Button'
+import { Led } from './Led'
+import { usePrefersReducedMotion } from './usePrefersReducedMotion'
+
+/**
+ * How the confirming button is painted.
+ *
+ *   crit    — the action destroys something. Semantic crit, never the accent.
+ *   neutral — the action is a normal commit the operator should read first
+ *             (a warning before saving); the accent's call-to-action is correct.
+ */
+export type ConfirmTone = 'crit' | 'neutral'
+
+export interface ConfirmOptions {
+  /** Engraved eyebrow, e.g. "DELETE RULE". Names the operation, not the object. */
+  label?: string
+  /** The question. One line, ends in "?". */
+  title: string
+  /** The consequence — what changes on the router if this goes through. */
+  body?: ReactNode
+  /** Verb on the confirming button. Defaults to "Delete". */
+  confirmLabel?: string
+  /** Verb on the dismissing button. Defaults to "Cancel". */
+  cancelLabel?: string
+  /** Defaults to `crit` — the overwhelmingly common case is a delete. */
+  tone?: ConfirmTone
+}
+
+export interface ConfirmDialogProps extends ConfirmOptions {
+  open: boolean
+  /** Called exactly once per dialog, with the operator's answer. */
+  onResolve: (confirmed: boolean) => void
+}
+
+const FOCUSABLE =
+  'button:not([disabled]), [href], input:not([disabled]), select:not([disabled]), textarea:not([disabled]), [tabindex]:not([tabindex="-1"])'
+
+/**
+ * The modal plate itself. Normally reached through `useConfirm()`; exported so a
+ * page that wants to own the open state can render it directly.
+ *
+ * Keyboard contract:
+ *   - focus moves to Cancel on open, so a reflex Enter dismisses, never deletes;
+ *   - Tab / Shift+Tab cycle inside the plate and cannot reach the page behind it;
+ *   - Esc answers "no";
+ *   - on close, focus returns to whatever opened the dialog.
+ */
+export function ConfirmDialog({
+  open,
+  onResolve,
+  label,
+  title,
+  body,
+  confirmLabel = 'Delete',
+  cancelLabel = 'Cancel',
+  tone = 'crit',
+}: ConfirmDialogProps) {
+  const titleId = useId()
+  const bodyId = useId()
+  const cardRef = useRef<HTMLDivElement>(null)
+  const cancelRef = useRef<HTMLButtonElement>(null)
+  const openerRef = useRef<HTMLElement | null>(null)
+  const reduced = usePrefersReducedMotion()
+
+  // Take the page out of the tab order, park focus on Cancel, and hand focus
+  // back to the opener when the plate goes away.
+  useEffect(() => {
+    if (!open) return
+    const opener = document.activeElement
+    openerRef.current = opener instanceof HTMLElement ? opener : null
+
+    const prevOverflow = document.body.style.overflow
+    document.body.style.overflow = 'hidden'
+
+    // Cancel is the resting place: an Enter or a Space meant for the page lands
+    // on "no". The destructive button is one Tab away, deliberately.
+    ;(cancelRef.current ?? cardRef.current)?.focus()
+
+    return () => {
+      document.body.style.overflow = prevOverflow
+      const back = openerRef.current
+      openerRef.current = null
+      if (back && document.contains(back)) back.focus()
+    }
+  }, [open])
+
+  // Esc answers no; Tab is caged. Capture phase so a page-level key handler
+  // never sees keys aimed at the dialog.
+  useEffect(() => {
+    if (!open) return
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === 'Escape') {
+        e.preventDefault()
+        e.stopPropagation()
+        onResolve(false)
+        return
+      }
+      if (e.key !== 'Tab') return
+      const card = cardRef.current
+      if (!card) return
+      const list = Array.from(card.querySelectorAll<HTMLElement>(FOCUSABLE))
+      if (list.length === 0) {
+        e.preventDefault()
+        card.focus()
+        return
+      }
+      const first = list[0]
+      const last = list[list.length - 1]
+      const active = document.activeElement as HTMLElement | null
+      if (!active || !card.contains(active)) {
+        e.preventDefault()
+        ;(e.shiftKey ? last : first).focus()
+      } else if (e.shiftKey && active === first) {
+        e.preventDefault()
+        last.focus()
+      } else if (!e.shiftKey && active === last) {
+        e.preventDefault()
+        first.focus()
+      }
+    }
+    document.addEventListener('keydown', onKey, true)
+    return () => document.removeEventListener('keydown', onKey, true)
+  }, [open, onResolve])
+
+  if (!open) return null
+
+  return createPortal(
+    <div
+      className={['cfm-scrim', reduced ? 'still' : ''].filter(Boolean).join(' ')}
+      // A click on the field around the plate means "not now". Mousedown (not
+      // click) so a text selection dragged out of the plate can't dismiss it.
+      onMouseDown={(e) => {
+        if (e.target === e.currentTarget) onResolve(false)
+      }}
+    >
+      <div
+        className={`cfm-card tone-${tone}`}
+        ref={cardRef}
+        tabIndex={-1}
+        role="alertdialog"
+        aria-modal="true"
+        aria-labelledby={titleId}
+        aria-describedby={body != null ? bodyId : undefined}
+      >
+        <i className="screw tl" aria-hidden="true" />
+        <i className="screw tr" aria-hidden="true" />
+        <i className="screw bl" aria-hidden="true" />
+        <i className="screw br" aria-hidden="true" />
+
+        {/* Lamp first, then the engraved label — the way a real panel reads, and
+            it keeps the LED off the corner screw. */}
+        <div className="cfm-hd">
+          <Led variant={tone === 'crit' ? 'crit' : 'amber'} />
+          <span className="cfm-label">{label ?? (tone === 'crit' ? 'Confirm delete' : 'Confirm')}</span>
+        </div>
+
+        <h2 className="cfm-title" id={titleId}>
+          {title}
+        </h2>
+        {body != null && (
+          <p className="cfm-body" id={bodyId}>
+            {body}
+          </p>
+        )}
+
+        <div className="cfm-actions">
+          <Button ref={cancelRef} onClick={() => onResolve(false)}>
+            {cancelLabel}
+          </Button>
+          <Button variant={tone === 'crit' ? 'crit' : 'primary'} onClick={() => onResolve(true)}>
+            {confirmLabel}
+          </Button>
+        </div>
+      </div>
+    </div>,
+    document.body,
+  )
+}
+
+// ---- provider + hook --------------------------------------------------------
+
+interface Request extends ConfirmOptions {
+  id: number
+  resolve: (v: boolean) => void
+}
+
+const ConfirmCtx = createContext<((o: ConfirmOptions) => Promise<boolean>) | null>(null)
+
+/**
+ * Mount once at the app root. Everything below can then ask a question and await
+ * the answer.
+ */
+export function ConfirmProvider({ children }: { children: ReactNode }) {
+  const [req, setReq] = useState<Request | null>(null)
+  const pending = useRef<Request | null>(null)
+  const seq = useRef(0)
+
+  const confirm = useCallback(
+    (opts: ConfirmOptions) =>
+      new Promise<boolean>((resolve) => {
+        // A second question while one is open answers the first with "no" rather
+        // than leaving its promise — and its caller — hanging forever.
+        pending.current?.resolve(false)
+        seq.current += 1
+        const next: Request = { ...opts, id: seq.current, resolve }
+        pending.current = next
+        setReq(next)
+      }),
+    [],
+  )
+
+  const settle = useCallback((confirmed: boolean) => {
+    const open = pending.current
+    pending.current = null
+    setReq(null)
+    open?.resolve(confirmed)
+  }, [])
+
+  // Teardown must not strand a caller mid-await.
+  useEffect(
+    () => () => {
+      pending.current?.resolve(false)
+      pending.current = null
+    },
+    [],
+  )
+
+  // A question belongs to the page that asked it. The provider outlives the
+  // hash router, so a navigation would otherwise leave a stale plate floating
+  // over a page it has nothing to do with — answer it "no" and clear it.
+  useEffect(() => {
+    const onNav = () => {
+      if (pending.current) settle(false)
+    }
+    window.addEventListener('hashchange', onNav)
+    return () => window.removeEventListener('hashchange', onNav)
+  }, [settle])
+
+  return (
+    <ConfirmCtx.Provider value={confirm}>
+      {children}
+      {req !== null && <ConfirmDialog key={req.id} open onResolve={settle} {...req} />}
+    </ConfirmCtx.Provider>
+  )
+}
+
+/**
+ * Ask the operator, get a definite answer:
+ *
+ *   const confirm = useConfirm()
+ *   if (!(await confirm({ title: 'Delete rule "x"?', body: '…' }))) return
+ *
+ * The returned function is stable, so it is safe in a useCallback dep list. It
+ * always settles — cancel, Esc, click-outside and teardown all resolve `false`;
+ * only the confirming button resolves `true`.
+ *
+ * Name it `confirm` at the call site on purpose: the local binding shadows the
+ * global one inside that component, so an accidental bare `confirm(...)` cannot
+ * reach the suppressible native dialog.
+ */
+export function useConfirm(): (o: ConfirmOptions) => Promise<boolean> {
+  const ctx = useContext(ConfirmCtx)
+  if (!ctx) {
+    // Loud on purpose. A fallback that quietly resolved false would rebuild the
+    // exact bug this component exists to kill.
+    throw new Error('useConfirm() needs <ConfirmProvider> above it (mounted in main.tsx)')
+  }
+  return ctx
+}
@@ -17,6 +17,8 @@ export { Button } from './Button'
 export type { ButtonProps } from './Button'
 export { Select } from './Select'
 export type { SelectProps, SelectOption } from './Select'
+export { ConfirmDialog, ConfirmProvider, useConfirm } from './ConfirmDialog'
+export type { ConfirmDialogProps, ConfirmOptions, ConfirmTone } from './ConfirmDialog'
 export { Clock } from './Clock'
 export { CatSuggest } from './CatSuggest'
 export { SrcPicker } from './SrcPicker'
@@ -2,12 +2,17 @@ import { StrictMode } from 'react'
 import { createRoot } from 'react-dom/client'
 import './tokens.css'
 import { App } from './App'
+import { ConfirmProvider } from './components'

 const rootEl = document.getElementById('root')
 if (!rootEl) throw new Error('#root not found')

+// ConfirmProvider sits ABOVE <App> so it survives App's early returns (the
+// unauth / no-link plates) — useConfirm() can never find itself without a host.
 createRoot(rootEl).render(
  <StrictMode>
-    <App />
+    <ConfirmProvider>
+      <App />
+    </ConfirmProvider>
  </StrictMode>,
 )
@@ -6,7 +6,7 @@
 // state mutates in-memory so the Apply / Confirm / Rollback flow is exercisable.
 //
 // Type-only imports from api.ts (erased at build) keep this free of a runtime cycle.
-import type { ApplyResult, ChainHealth, ConnLogEntry, DiscoveredDevice, GroupHealth, GroupMemberHealth, GroupsHealth, GroupTestResult, GroupTestStart, GroupTestStatus, Interface, Model, QueryLogEntry, RulesetCategories, RulesetCheck, RulesetStatus, Stats, StatsLogPage, StatsLogQuery, Status, StatusWarning } from './api'
+import type { ApplyResult, ChainHealth, ConnLogEntry, DiscoveredDevice, GroupHealth, GroupMemberHealth, GroupsHealth, GroupTestResult, GroupTestStart, GroupTestStatus, Interface, Model, Profile, QueryLogEntry, RuleReach, RulesReachability, RulesetCategories, RulesetCheck, RulesetStatus, Stats, StatsLogPage, StatsLogQuery, Status, StatusWarning, Traffic } from './api'

 let armed = false // a pending commit-confirm auto-rollback
 let hasLastGood = false // a predecessor config exists to roll back to (post-apply)
@@ -146,6 +146,12 @@ const CONFIG: Model = {
    { Name: 'block-ads', Enabled: true, Order: 10, DstRuleset: ['ad-hosts'], Target: 'block' },
    { Name: 'ru-bypass', Enabled: true, Order: 20, DstRuleset: ['ru-inside'], Target: 'direct' },
    { Name: 'private-direct', Enabled: true, Order: 30, DstRuleset: ['private-nets'], Target: 'direct' },
+    // A SECOND condition-less rule, above the real default. It reads like a working
+    // rule and does nothing: a rule with no conditions becomes the router's default,
+    // and the last such rule by Order wins — so this one never applies. It is in the
+    // fixture on purpose, to exercise the "never applies" badge; the field config
+    // that prompted it had two rules BOTH named `default` (orders 20 and 100).
+    { Name: 'default-bypass', Enabled: true, Order: 40, Target: 'direct' },
    { Name: 'default-tunnel', Enabled: true, Order: 900, Target: 'group:auto' },
  ],
  // Named match-lists a rule points DstRuleset at. url + geosite + geoip are remote
@@ -296,6 +302,96 @@ const RULESET_STATUS: RulesetStatus[] = [
  { tag: 'rs-ru-geoip-ru', name: 'ru-geoip', category: 'ru', kind: 'ruleset', remote: true, last_updated: '', interval_seconds: 86_400, rule_count: 0 },
 ]

+/** GET /api/rules/reachability. Mirrors the daemon's analysis over CONFIG.Rules:
+ *  a rule with no conditions is the router's default, and the LAST such rule by
+ *  Order wins — every earlier one can never apply. It reads the live CONFIG so
+ *  edits made in `?mock` keep the badge honest.
+ *
+ *  It also mirrors model.ResolveActiveProfile + ApplyProfileRuleOverrides, because
+ *  `effective_enabled` is the whole point of the endpoint: CONFIG's `mobile-uplink`
+ *  is active and both enables and disables rules, so `?mock` shows the same
+ *  desired-vs-effective split the field config does. */
+export async function getRulesReachability(): Promise<RulesReachability> {
+  await wait(60)
+  const rules = CONFIG.Rules ?? []
+
+  // A pin naming an existing, ENABLED profile wins outright. Otherwise auto-select:
+  // highest Priority among enabled profiles, ties by Name, skipping any with an
+  // iface condition (the WAN watcher owns those and expresses its verdict as the pin).
+  const profiles = CONFIG.Profiles ?? []
+  const pinned = String(CONFIG.Globals?.ActiveProfile ?? '').trim()
+  let prof: Profile | null = profiles.find((p) => p.Enabled && p.Name === pinned) ?? null
+  if (!prof) {
+    for (const p of profiles) {
+      if (!p.Enabled || (p.MatchIface ?? []).length > 0) continue
+      const pp = p.Priority ?? 0
+      const bp = prof?.Priority ?? 0
+      if (!prof || pp > bp || (pp === bp && p.Name < prof.Name)) prof = p
+    }
+  }
+  // Enable first, then Disable, so a name in both ends up disabled (Disable wins).
+  const effective = rules.map((r) => Boolean(r.Enabled))
+  if (prof) {
+    const force = (names: string[] | null | undefined, on: boolean) => {
+      for (const raw of names ?? []) {
+        const n = raw.trim()
+        rules.forEach((r, i) => {
+          if (r.Name === n) effective[i] = on
+        })
+      }
+    }
+    force(prof.EnableRules, true)
+    force(prof.DisableRules, false)
+  }
+  const activeProfile = prof
+
+  const out: RuleReach[] = rules.map((r, index) => ({
+    index,
+    name: String(r.Name ?? ''),
+    order: Number(r.Order ?? 0),
+    unreachable: false,
+    shadowed_by_index: -1,
+    effective_enabled: effective[index],
+    // Annotate only where the profile actually FLIPPED the outcome — a profile that
+    // disables an already-off rule has overridden nothing the operator can see.
+    ...(activeProfile && effective[index] !== Boolean(r.Enabled)
+      ? {
+          overridden_by: activeProfile.Name,
+          override: effective[index] ? ('enabled' as const) : ('disabled' as const),
+        }
+      : {}),
+  }))
+  const conditionless = (r: (typeof rules)[number]): boolean =>
+    !(r.Src ?? []).length &&
+    !(r.DstRuleset ?? []).length &&
+    !String(r.DstPort ?? '').trim() &&
+    !String(r.Proto ?? '').trim()
+  const target = (r: (typeof rules)[number]): string =>
+    String(r.Target ?? '').trim() || (r.Egress ? `egress:${String(r.Egress).trim()}` : '')
+  const defaults = rules
+    .map((r, index) => ({ r, index }))
+    // The EFFECTIVE flag, not the configured one: a rule the active profile
+    // switched off is not in force and cannot retire anything (model's
+    // RuleReachability runs over the effective set for the same reason).
+    .filter(({ r, index }) => effective[index] && conditionless(r) && target(r))
+    .sort((a, b) => Number(a.r.Order ?? 0) - Number(b.r.Order ?? 0) || a.index - b.index)
+  const winner = defaults[defaults.length - 1]
+  if (winner) {
+    for (const { index } of defaults.slice(0, -1)) {
+      out[index].unreachable = true
+      out[index].shadowed_by = String(winner.r.Name ?? '')
+      out[index].shadowed_by_index = winner.index
+      out[index].shadowed_by_order = Number(winner.r.Order ?? 0)
+      out[index].reason =
+        `this rule has no conditions, so it sets the default for all traffic — but rule ` +
+        `"${winner.r.Name}" (order ${winner.r.Order}) has none either and comes after it, so ` +
+        `"${target(winner.r)}" is the default the router uses and this rule's target ` +
+        `"${target(rules[index])}" is never applied`
+    }
+  }
+  return { rules: out }
+}
+
 export async function getRulesetStatus(): Promise<RulesetStatus[]> {
  await wait(90)
  return RULESET_STATUS.map((r) => ({ ...r }))
@@ -357,6 +453,11 @@ export async function getRulesetCategories(source: string): Promise<RulesetCateg
 //   ?mock&warn=1     → a full warning set (critical + warning + info) on top
 //   ?mock&ks=open    → healthy plane but a FAIL-OPEN kill-switch, which is what
 //                      makes the untunnelable policy inert (F8 case 4)
+//   ?mock&traffic=…  → with the plane FULL, where the traffic actually ends up:
+//                      split | direct | blocked | blackout | unknown. `direct` is
+//                      the field case the readout used to call "Protected" (one
+//                      rule, `default → direct`); `unknown` is a daemon too old to
+//                      report. Default: tunnel.
 function mockPlane(): { plane: 'full' | 'hold' | 'none'; engine: boolean; killSwitch: string } {
  const q = typeof location === 'undefined' ? '' : location.search
  const params = new URLSearchParams(q)
@@ -368,6 +469,30 @@ function mockPlane(): { plane: 'full' | 'hold' | 'none'; engine: boolean; killSw
  return { plane: 'full', engine: true, killSwitch }
 }

+// The daemon's verdict on where traffic goes (apply.Status.traffic). Only
+// meaningful with the plane installed: with the engine down there is no running
+// config to judge, and the daemon reports the unknown/zero value — so do the same
+// here rather than leaving a stale "tunnel" behind a dead engine.
+function mockTraffic(plane: 'full' | 'hold' | 'none'): Traffic | undefined {
+  if (plane !== 'full') return { verdict: '', default: '', tunnel_rules: 0 }
+  const params = new URLSearchParams(typeof location === 'undefined' ? '' : location.search)
+  switch (params.get('traffic')) {
+    case 'split':
+      return { verdict: 'split', default: 'direct', tunnel_rules: 3 }
+    case 'direct':
+      return { verdict: 'direct', default: 'direct', tunnel_rules: 0 }
+    case 'blocked':
+      return { verdict: 'blocked', default: 'block', tunnel_rules: 2 }
+    case 'blackout':
+      return { verdict: 'blocked', default: 'block', tunnel_rules: 0 }
+    case 'unknown':
+      // A daemon that predates the field sends no `traffic` at all.
+      return undefined
+    default:
+      return { verdict: 'tunnel', default: 'auto', tunnel_rules: 1 }
+  }
+}
+
 const MOCK_WARNINGS: StatusWarning[] = [
  {
    severity: 'critical',
@@ -470,6 +595,7 @@ export async function getStatus(): Promise<Status> {
    can_rollback: armed || hasLastGood,
    engine_running: engine,
    plane,
+    traffic: mockTraffic(plane),
    warnings: mockWarnings(killSwitch),
    // Process uptime. Anchored to when this tab loaded plus a fixed head start, so
    // the reading ticks forward across polls exactly like the real daemon's does.
@@ -1,6 +1,6 @@
 import './DNS.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { Button, CatSuggest, Led, SrcPicker, Toggle } from '../components'
+import { Button, CatSuggest, Led, SrcPicker, Toggle, useConfirm } from '../components'
 import { apply as apiApply, getConfig, putConfig, ApiError } from '../api'
 import type { Alert, DNSRule, Model, Resolver } from '../api'

@@ -220,6 +220,7 @@ function describeDetour(
 // ---- page ------------------------------------------------------------------

 export default function DNS() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<DNSModel | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)

@@ -422,15 +423,19 @@ export default function DNS() {
  )

  const removeBlocklist = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = blocklists[idx]
-      if (!window.confirm(`Delete blocklist “${target.Name}”? This removes it from the config.`))
-        return
+      const ok = await confirm({
+        label: 'Delete blocklist',
+        title: `Delete blocklist “${target.Name}”?`,
+        body: 'This removes it from the config.',
+      })
+      if (!ok) return
      const next = blocklists.filter((_, i) => i !== idx)
      void save({ ...config, Blocklists: next }, `Deleted ${target.Name}`)
    },
-    [config, blocklists, save],
+    [config, blocklists, save, confirm],
  )

  // ---- allowlist mutations --------------------------------------------------
@@ -457,15 +462,19 @@ export default function DNS() {
  )

  const removeAllowlist = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = allowlists[idx]
-      if (!window.confirm(`Delete allowlist “${target.Name}”? This removes it from the config.`))
-        return
+      const ok = await confirm({
+        label: 'Delete allowlist',
+        title: `Delete allowlist “${target.Name}”?`,
+        body: 'This removes it from the config.',
+      })
+      if (!ok) return
      const next = allowlists.filter((_, i) => i !== idx)
      void save({ ...config, Allowlists: next }, `Deleted ${target.Name}`)
    },
-    [config, allowlists, save],
+    [config, allowlists, save, confirm],
  )

  // ---- resolver mutations ---------------------------------------------------
@@ -492,11 +501,15 @@ export default function DNS() {
  )

  const removeResolver = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = resolvers[idx]
-      if (!window.confirm(`Delete resolver “${target.Name}”? This removes it from the config.`))
-        return
+      const ok = await confirm({
+        label: 'Delete resolver',
+        title: `Delete resolver “${target.Name}”?`,
+        body: 'This removes it from the config.',
+      })
+      if (!ok) return
      const next = resolvers.filter((_, i) => i !== idx)
      // Don't leave default/fallback pointing at a resolver that no longer exists.
      const g = { ...config.Globals }
@@ -578,15 +591,19 @@ export default function DNS() {
  )

  const removeDNSRule = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = dnsRules[idx]
-      if (!window.confirm(`Delete this DNS rule? Matching queries fall back to the default resolver.`))
-        return
+      const ok = await confirm({
+        label: 'Delete DNS rule',
+        title: 'Delete this DNS rule?',
+        body: 'Matching queries fall back to the default resolver.',
+      })
+      if (!ok) return
      const next = dnsRules.filter((_, i) => i !== idx)
      void save({ ...config, DNSRules: next }, `Deleted DNS rule → ${target.Resolver}`)
    },
-    [config, dnsRules, save],
+    [config, dnsRules, save, confirm],
  )

  // ---- alert mutations ------------------------------------------------------
@@ -610,14 +627,19 @@ export default function DNS() {
  )

  const removeAlert = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = alerts[idx]
-      if (!window.confirm(`Delete alert “${target.Name}”? This removes it from the config.`)) return
+      const ok = await confirm({
+        label: 'Delete alert',
+        title: `Delete alert “${target.Name}”?`,
+        body: 'This removes it from the config.',
+      })
+      if (!ok) return
      const next = alerts.filter((_, i) => i !== idx)
      void save({ ...config, Alerts: next }, `Deleted ${target.Name}`)
    },
-    [config, alerts, save],
+    [config, alerts, save, confirm],
  )

  const setAlertVia = useCallback(
@@ -138,57 +138,9 @@

 /* inline rename: a quiet pencil affordance beside the name, and the mono input
   it swaps to — in the same sink/groove tone as the domain editors. */
-.dev-rename {
-  flex: none;
-  display: inline-flex;
-  align-items: center;
-  justify-content: center;
-  width: 22px;
-  height: 22px;
-  padding: 0;
-  border: 1px solid transparent;
-  border-radius: 5px;
-  background: none;
-  color: var(--faint);
-  font-size: 12px;
-  line-height: 1;
-  cursor: pointer;
-  transition: color 0.15s, background 0.15s, border-color 0.15s;
-}
-.dev-rename:hover:not(:disabled) {
-  color: var(--accent);
-  background: color-mix(in srgb, var(--accent) 12%, transparent);
-}
-.dev-rename:focus-visible {
-  color: var(--accent);
-  border-color: var(--accent);
-  outline: 2px solid var(--accent);
-  outline-offset: 1px;
-}
-.dev-rename:disabled {
-  opacity: 0.5;
-  cursor: default;
-}
-.dev-name-input {
-  min-width: 0;
-  max-width: 24ch;
-  padding: 4px 8px;
-  border: 1px solid var(--accent);
-  border-radius: 6px;
-  background: var(--sink);
-  color: var(--ink);
-  font-size: 13px;
-  font-weight: 600;
-  letter-spacing: 0.01em;
-  box-shadow: 0 1px 2px var(--shadow) inset;
-}
-.dev-name-input:focus-visible {
-  outline: 2px solid var(--accent);
-  outline-offset: 1px;
-}
-.dev-name-input:disabled {
-  opacity: 0.55;
-}
+/* The pencil button and the name input now live in App.css as .inline-rename /
+   .inline-rename-input — Nodes grew the same affordance and the two pages must
+   not drift. */
 .dev-id-l2 {
  display: flex;
  align-items: center;
@@ -1,6 +1,6 @@
 import './Devices.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { Button, Led, Module, Toggle } from '../components'
+import { Button, Led, Module, Toggle, useConfirm } from '../components'
 import type { LedVariant } from '../components'
 import { apply as apiApply, getConfig, getDevices, putConfig, ApiError } from '../api'
 import type { Device, DiscoveredDevice, Model } from '../api'
@@ -81,6 +81,7 @@ function networkLabel(row: DeviceRow): string {
 // ---- page ------------------------------------------------------------------

 export default function Devices() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)

@@ -251,17 +252,22 @@ export default function Devices() {
  const nameOf = (row: DeviceRow) => row.cfg?.Name || row.hostname || row.ip || 'device'

  const removeControl = useCallback(
-    (row: DeviceRow) => {
+    async (row: DeviceRow) => {
      if (!config) return
      const devs = asArray(config.Devices)
      const idx = matchDevice(devs, row.mac, row.ip)
      if (idx < 0) return
      const nm = devs[idx].Name || nameOf(row)
-      if (!window.confirm(`Stop managing “${nm}”? Its per-device rules are removed; it falls back to network defaults.`))
-        return
+      const ok = await confirm({
+        label: 'Stop managing device',
+        title: `Stop managing “${nm}”?`,
+        body: 'Its per-device rules are removed; it falls back to network defaults.',
+        confirmLabel: 'Stop managing',
+      })
+      if (!ok) return
      void save({ ...config, Devices: devs.filter((_, i) => i !== idx) }, `Removed control for ${nm}`)
    },
-    [config, save],
+    [config, save, confirm],
  )

  const loading = config === null && loadError === null && devices === null && devError === null
@@ -471,7 +477,7 @@ function DeviceCard({
            {renaming ? (
              <input
                ref={nameInput}
-                className="dev-name-input mono"
+                className="inline-rename-input mono"
                type="text"
                spellCheck={false}
                autoComplete="off"
@@ -497,7 +503,7 @@ function DeviceCard({
                </span>
                <button
                  type="button"
-                  className="dev-rename"
+                  className="inline-rename"
                  onClick={beginRename}
                  disabled={busy}
                  aria-label={`Rename ${name}`}
@@ -1,6 +1,6 @@
 import './Networks.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { Button, Led, Select, Toggle } from '../components'
+import { Button, Led, Select, Toggle, useConfirm } from '../components'
 import { apply as apiApply, getConfig, putConfig, ApiError } from '../api'
 import type { Inbound, Interface, Model, Status } from '../api'
 import { isLanNetwork, isWanNetwork, useInterfaces } from '../srcOptions'
@@ -242,6 +242,7 @@ function computeWarnings(inbounds: Inbound[], ifaces: Interface[]): Warning[] {
 // ---- page ------------------------------------------------------------------

 export default function Networks({ status }: { status?: Status | null }) {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)
  const ifaces = useInterfaces()
@@ -392,23 +393,21 @@ export default function Networks({ status }: { status?: Status | null }) {
  )

  const removeInbound = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = inbounds[idx]
-      if (
-        !window.confirm(
-          `Delete inbound “${target.Name}”?${
-            intercepts(target)
-              ? ` ${target.Network || 'Its network'} stops going through the tunnel.`
-              : ''
-          }`,
-        )
-      )
-        return
+      const ok = await confirm({
+        label: 'Delete inbound',
+        title: `Delete inbound “${target.Name}”?`,
+        body: intercepts(target)
+          ? `${target.Network || 'Its network'} stops going through the tunnel.`
+          : undefined,
+      })
+      if (!ok) return
      const next = inbounds.filter((_, i) => i !== idx)
      void save({ ...config, Inbounds: next }, `Deleted ${target.Name}`)
    },
-    [config, inbounds, save],
+    [config, inbounds, save, confirm],
  )

  return (
@@ -727,3 +727,53 @@ select.fp-input {
    width: 9rem;
  }
 }
+
+/* ---- inline node rename ----
+   The pencil / input pair itself is shared (.inline-rename[-input] in App.css);
+   only the row-local sizing and the refusal message live here. A node name is
+   longer than a device name (it carries a protocol and a host), so the field is
+   given more room than the shared 24ch default. */
+.node-name-input {
+  max-width: 32ch;
+  font-family: var(--font-mono);
+  font-size: 12.5px;
+}
+/* Why a rename was refused, pinned under the row it was typed in. Semantic crit:
+   the name did not change, and that must not be mistaken for a saved edit. */
+.row-err {
+  margin: 2px 0 0;
+  font-size: 11.5px;
+  line-height: 1.45;
+  color: var(--crit);
+  max-width: 68ch;
+}
+/* Stated once per subscription bucket: the same rule the locked control in every
+   row carries, so the absent rename is explained before it is looked for. */
+.group-note {
+  margin: 0;
+  padding: 8px 12px;
+  border: 1px solid var(--groove);
+  border-top: 0;
+  background: color-mix(in srgb, var(--sink) 25%, transparent);
+  font-size: 11.5px;
+  line-height: 1.5;
+  color: var(--faint);
+}
+/* The optional name sits beside the link input on a wide row and drops onto its
+   own line when the row can no longer hold both. */
+.add-name {
+  flex: 0 1 22ch;
+  min-width: 12ch;
+}
+.add-row--conf .add-name {
+  flex: none;
+  align-self: stretch;
+}
+@media (max-width: 640px) {
+  .add-row {
+    flex-wrap: wrap;
+  }
+  .add-name {
+    flex: 1 1 100%;
+  }
+}
@@ -2,7 +2,7 @@ import './Nodes.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
 import type { ReactNode } from 'react'
 import type { LedVariant } from '../components'
-import { Button, Led, Toggle } from '../components'
+import { Button, Led, Toggle, useConfirm } from '../components'
 import {
  apply as apiApply,
  getConfig,
@@ -158,6 +158,219 @@ function uniqueName(base: string, taken: Set<string>): string {
  return `${seed}-${i}`
 }

+// ---- node names are identity, not a caption --------------------------------
+//
+// A node's Name IS its sing-box outbound tag and the only thing every reference
+// to it spells: a rule target `node:<name>`, a chain hop, a manual group's member
+// list, a resolver detour, an alert delivery, a subscription fetch detour. Rename
+// the node alone and every one of those points at nothing — and an unresolved
+// target does NOT fall back to the default route, the daemon BLOCKS that traffic.
+// So the rename either carries every reference with it, or it is refused.
+
+/** Reserved outbound tags. A node called this is skipped by the generator entirely. */
+const RESERVED_TAGS = ['direct', 'block']
+
+/**
+ * Prefixes that `model.SplitTarget` reads as a KIND, not as part of a name. A
+ * name starting with one of them makes every bare reference to it ambiguous with
+ * a real `kind:name` reference, so it is refused rather than half-supported.
+ */
+const KIND_PREFIXES = ['node', 'group', 'egress', 'chain', 'direct', 'block']
+
+/** Names are rendered into a line-oriented `uci export`; control chars are stripped there. */
+function hasControlChar(s: string): boolean {
+  for (let i = 0; i < s.length; i++) {
+    const c = s.charCodeAt(i)
+    if (c < 0x20 || c === 0x7f) return true
+  }
+  return false
+}
+
+/**
+ * Why `name` cannot be a node name here, or null if it can.
+ *
+ * Every rule mirrors something the daemon actually does with the name, not a
+ * house style: reserved tags make generate skip the node; a duplicate makes two
+ * outbounds share a tag and the manager silently keeps the last one; a group of
+ * the same name is dropped by buildGroups ("rename the group"); an
+ * `egress-<name>` collision takes over a real egress outbound; and a control
+ * character is rewritten to a space by sanitizeUCIValue on write, so the saved
+ * name would not be the one you typed.
+ */
+function nodeNameError(
+  raw: string,
+  m: Model | null,
+  self: string | null,
+): string | null {
+  const name = raw.trim()
+  if (!name) return 'A node needs a name.'
+  if (hasControlChar(name))
+    return 'Names can’t contain line breaks or control characters — they’re stripped when the config is written.'
+  if (RESERVED_TAGS.some((t) => t.toLowerCase() === name.toLowerCase()))
+    return `“${name}” is a reserved target name — a node called that is skipped by the engine. Pick another.`
+  const head = name.includes(':') ? name.slice(0, name.indexOf(':')).toLowerCase() : ''
+  if (head && KIND_PREFIXES.includes(head))
+    return `A name starting with “${head}:” reads as a ${head} reference everywhere it’s used. Pick another.`
+  if (!m) return null
+
+  const clash = asArray(m.Nodes).find((n) => n.Name === name && n.Name !== self)
+  if (clash)
+    return clash.FromSub
+      ? `“${name}” is already a node from subscription “${clash.FromSub}”. Two nodes with one name share a single outbound — pick another.`
+      : `“${name}” is already another node. Pick another.`
+  if (asArray(m.Groups).some((g) => g.Name === name))
+    return `A group is already named “${name}”. The engine drops the group when a node takes its name — pick another.`
+  const egressClash = asArray(m.Egresses).find((e) => `egress-${e.Name}` === name)
+  if (egressClash)
+    return `“${name}” is the outbound tag of egress “${egressClash.Name}”. Pick another.`
+  return null
+}
+
+/** One place a node name is written, as a short label for the rename summary. */
+interface NodeRefSite {
+  /** Which section — drives the "N rules, M groups" count. */
+  kind: 'rule' | 'group' | 'chain' | 'resolver' | 'alert' | 'subscription' | 'egress'
+  label: string
+}
+
+/** A target/detour string naming this node in its prefixed form (`node:<name>`). */
+const isNodeRef = (v: string | undefined | null, name: string): boolean =>
+  (v ?? '') === `node:${name}`
+
+/** …or in the bare form the engine also resolves (a group member, a bare hop/target). */
+const isBareRef = (v: string | undefined | null, name: string): boolean => (v ?? '') === name
+
+/**
+ * Every place `name` is written outside the node itself. Both spellings count:
+ * `resolveTarget` falls through to a bare node lookup, and a manual group's
+ * member list is bare by contract.
+ */
+function findNodeReferences(m: Model, name: string): NodeRefSite[] {
+  const out: NodeRefSite[] = []
+  for (const r of asArray(m.Rules)) {
+    if (isNodeRef(r.Target, name) || isBareRef(r.Target, name))
+      out.push({ kind: 'rule', label: `rule “${r.Name}” target` })
+  }
+  for (const g of asArray(m.Groups)) {
+    if (asArray(g.Nodes).some((n) => n === name))
+      out.push({ kind: 'group', label: `group “${g.Name}” member` })
+  }
+  for (const c of asArray(m.Chains)) {
+    if (asArray(c.Hops).some((h) => isNodeRef(h, name) || isBareRef(h, name)))
+      out.push({ kind: 'chain', label: `chain “${c.Name}” hop` })
+  }
+  for (const r of asArray(m.Resolvers)) {
+    if (isNodeRef(r.Detour, name)) out.push({ kind: 'resolver', label: `resolver “${r.Name}” DNS path` })
+  }
+  for (const a of asArray(m.Alerts)) {
+    if (isNodeRef(a.Via, name)) out.push({ kind: 'alert', label: `alert “${a.Name}” delivery` })
+  }
+  for (const s of asArray(m.Subscriptions)) {
+    if (isNodeRef(s.FetchDetour, name))
+      out.push({ kind: 'subscription', label: `subscription “${s.Name}” fetch` })
+  }
+  for (const e of asArray(m.Egresses)) {
+    if (isNodeRef(e.Target, name)) out.push({ kind: 'egress', label: `egress “${e.Name}” target` })
+  }
+  return out
+}
+
+/**
+ * What makes a rename impossible to carry rather than merely wide.
+ *
+ * A BARE reference is just a name; the engine resolves it node-first, then group.
+ * If something else already answers to the old name, we cannot tell which object
+ * a bare reference meant, and rewriting it would move a reference the operator
+ * never pointed at this node. That is a half-done cascade, so the rename is
+ * refused instead — with the collision named, so it can be fixed.
+ */
+function bareAmbiguity(m: Model, name: string): string | null {
+  const group = asArray(m.Groups).find((g) => g.Name === name)
+  if (!group) return null
+  const bare = [
+    ...asArray(m.Rules)
+      .filter((r) => isBareRef(r.Target, name))
+      .map((r) => `rule “${r.Name}”`),
+    ...asArray(m.Chains)
+      .filter((c) => asArray(c.Hops).some((h) => isBareRef(h, name)))
+      .map((c) => `chain “${c.Name}”`),
+  ]
+  if (bare.length === 0) return null
+  return `A group is also named “${name}”, and ${bare.join(', ')} point${bare.length === 1 ? 's' : ''} at that bare name — there is no way to tell which of the two is meant. Rename the group first, then this node.`
+}
+
+/**
+ * Rewrite every reference from `from` to `to`. Returns a NEW Model with only the
+ * touched sections replaced; the Nodes section is the caller's business.
+ *
+ * Bare references are rewritten too — that is the whole point for a manual
+ * group's member list — which is safe only because `bareAmbiguity` has already
+ * refused the one case where a bare name could mean something else.
+ */
+function renameNodeReferences(m: Model, from: string, to: string): Model {
+  if (from === to) return m
+  /** Prefixed-only sites (a detour is never spelled bare). */
+  const pfx = (v: string | undefined) => (isNodeRef(v, from) ? `node:${to}` : v)
+  /** Sites that accept either spelling — each is rewritten in the spelling it already uses. */
+  const either = (v: string | undefined) => {
+    if (isNodeRef(v, from)) return `node:${to}`
+    if (isBareRef(v, from)) return to
+    return v
+  }
+
+  const next: Model = { ...m }
+  if (m.Rules) next.Rules = m.Rules.map((r) => ({ ...r, Target: either(r.Target) }))
+  if (m.Groups)
+    next.Groups = m.Groups.map((g) => ({
+      ...g,
+      Nodes: g.Nodes ? g.Nodes.map((n) => (n === from ? to : n)) : g.Nodes,
+    }))
+  if (m.Chains)
+    next.Chains = m.Chains.map((c) => ({
+      ...c,
+      Hops: c.Hops ? c.Hops.map((h) => either(h) ?? h) : c.Hops,
+    }))
+  if (m.Resolvers) next.Resolvers = m.Resolvers.map((r) => ({ ...r, Detour: pfx(r.Detour) }))
+  if (m.Alerts) next.Alerts = m.Alerts.map((a) => ({ ...a, Via: pfx(a.Via) }))
+  if (m.Subscriptions)
+    next.Subscriptions = m.Subscriptions.map((s) => ({ ...s, FetchDetour: pfx(s.FetchDetour) }))
+  if (m.Egresses) next.Egresses = m.Egresses.map((e) => ({ ...e, Target: pfx(e.Target) }))
+  return next
+}
+
+/** "3 rules, 1 group and 2 chains" — what the rename is about to rewrite. */
+function refSummary(refs: NodeRefSite[]): string {
+  const plural: Record<NodeRefSite['kind'], [string, string]> = {
+    rule: ['rule', 'rules'],
+    group: ['group', 'groups'],
+    chain: ['chain', 'chains'],
+    resolver: ['resolver', 'resolvers'],
+    alert: ['alert', 'alerts'],
+    subscription: ['subscription', 'subscriptions'],
+    egress: ['egress', 'egresses'],
+  }
+  const order: NodeRefSite['kind'][] = [
+    'rule', 'group', 'chain', 'resolver', 'alert', 'subscription', 'egress',
+  ]
+  const parts = order
+    .map((k) => [k, refs.filter((r) => r.kind === k).length] as const)
+    .filter(([, n]) => n > 0)
+    .map(([k, n]) => `${n} ${plural[k][n === 1 ? 0 : 1]}`)
+  if (parts.length === 1) return parts[0]
+  return `${parts.slice(0, -1).join(', ')} and ${parts[parts.length - 1]}`
+}
+
+/**
+ * The name a rename just committed to, waiting for its row to come back.
+ *
+ * A row is keyed by the node's NAME, so committing a rename unmounts the row and
+ * mounts a different one — carrying the focused element away with it. This baton
+ * survives that remount: the row that reappears under the new name claims it and
+ * puts the keyboard back on its own rename button, instead of dropping the user
+ * on <body> halfway down a list of 300 nodes.
+ */
+let pendingRenameFocus: string | null = null
+
 // A subscription with more than this many nodes starts collapsed so the list
 // doesn't become one endless scroll; an active search overrides it.
 const LARGE_GROUP = 20
@@ -289,6 +502,7 @@ function DetourSelect({
 // ---- page ------------------------------------------------------------------

 export default function Nodes() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)

@@ -385,6 +599,9 @@ export default function Nodes() {
  const [nodeInput, setNodeInput] = useState('')
  const [nodeErr, setNodeErr] = useState<string | null>(null)
  const [addMode, setAddMode] = useState<'link' | 'conf'>('link')
+  // Optional. Empty keeps the old behaviour (a name derived from the server
+  // address), so "paste a link, press Add" stays a two-step path.
+  const [nodeName, setNodeName] = useState('')
  const [importing, setImporting] = useState(false)

  // ---- node search + collapsible grouping -----------------------------------
@@ -445,8 +662,12 @@ export default function Nodes() {
      try {
        const { uri, name } = await importWg(conf)
        const taken = new Set(nodes.map((n) => n.Name))
+        // A typed name is used AS TYPED — uniqueName would silently turn a
+        // collision into "name-2", which is the confusion this field exists to
+        // end. It is validated instead, and a clash is refused out loud above.
+        const wanted = nodeName.trim()
        const node: NodeCfg = {
-          Name: uniqueName(name || 'wireguard', taken),
+          Name: wanted || uniqueName(name || 'wireguard', taken),
          Enabled: true,
          URI: uri,
          FromSub: '',
@@ -455,6 +676,7 @@ export default function Nodes() {
        const ok = await save({ ...config, Nodes: [...nodes, node] }, `Added ${node.Name}`)
        if (ok) {
          setNodeInput('')
+          setNodeName('')
          setAddMode('link')
        }
      } catch (e) {
@@ -463,11 +685,20 @@ export default function Nodes() {
        setImporting(false)
      }
    },
-    [config, nodes, save, flash],
+    [config, nodes, nodeName, save, flash],
  )

  const addNode = useCallback(async () => {
    if (!config) return
+    // The name is checked BEFORE the import round-trip, so a bad name costs
+    // nothing and the message lands in the form next to the field.
+    if (nodeName.trim()) {
+      const bad = nodeNameError(nodeName, config, null)
+      if (bad) {
+        setNodeErr(bad)
+        return
+      }
+    }
    // Auto-detect a pasted config, whichever input it landed in.
    if (nodeInput.includes(WG_MARKER)) {
      await addWgConf(nodeInput)
@@ -486,10 +717,20 @@ export default function Nodes() {
    const parsed = parseShareLink(uri)
    const taken = new Set(nodes.map((n) => n.Name))
    const base = parsed.suggested || `${parsed.proto.toLowerCase()}-${parsed.host}`.replace(/[^\w.:-]+/g, '-')
-    const node: NodeCfg = { Name: uniqueName(base, taken), Enabled: true, URI: uri, FromSub: '', Egress: '' }
+    const wanted = nodeName.trim()
+    const node: NodeCfg = {
+      Name: wanted || uniqueName(base, taken),
+      Enabled: true,
+      URI: uri,
+      FromSub: '',
+      Egress: '',
+    }
    const ok = await save({ ...config, Nodes: [...nodes, node] }, `Added ${node.Name}`)
-    if (ok) setNodeInput('')
-  }, [config, nodeInput, nodes, save, addMode, addWgConf])
+    if (ok) {
+      setNodeInput('')
+      setNodeName('')
+    }
+  }, [config, nodeInput, nodeName, nodes, save, addMode, addWgConf])

  const toggleNode = useCallback(
    (idx: number, on: boolean) => {
@@ -501,14 +742,88 @@ export default function Nodes() {
  )

  const removeNode = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = nodes[idx]
-      if (!window.confirm(`Delete node “${target.Name}”? This removes it from the config.`)) return
+      const ok = await confirm({
+        label: 'Delete node',
+        title: `Delete node “${target.Name}”?`,
+        body: 'This removes it from the config.',
+      })
+      if (!ok) return
      const next = nodes.filter((_, i) => i !== idx)
      void save({ ...config, Nodes: next }, `Deleted ${target.Name}`)
    },
-    [config, nodes, save],
+    [config, nodes, save, confirm],
+  )
+
+  /**
+   * Rename a manual node, carrying every reference with it.
+   *
+   * The name is this node's identity: its outbound tag, and the exact string a
+   * rule target, a chain hop, a manual group's member list, a resolver detour, an
+   * alert delivery and a subscription fetch detour all spell. So the rename is one
+   * atomic save of the Nodes section AND every referencing section, or it does not
+   * happen at all:
+   *
+   *   - an invalid or colliding name is refused with the reason (`nodeNameError`);
+   *   - a name a GROUP also answers to, with bare references pointing at it, is
+   *     refused too — there is no way to know which object those meant, and
+   *     guessing would move a reference the operator never pointed here;
+   *   - anything else is shown exactly what it will rewrite, and only then saved.
+   *
+   * Errors surface through `onError` so they land in the row that was edited.
+   */
+  const renameNode = useCallback(
+    async (idx: number, raw: string, onError: (msg: string) => void): Promise<boolean> => {
+      if (!config) return false
+      const target = nodes[idx]
+      const from = target.Name
+      const to = raw.trim()
+      if (to === from) return true
+      // Subscription names come back from the feed on the next update; renaming
+      // one would be undone without warning, so this path is manual-only.
+      if (target.FromSub) {
+        onError(`“${from}” is named by subscription “${target.FromSub}” — the feed rewrites it on the next update.`)
+        return false
+      }
+      const bad = nodeNameError(to, config, from)
+      if (bad) {
+        onError(bad)
+        return false
+      }
+      const blocked = bareAmbiguity(config, from)
+      if (blocked) {
+        onError(blocked)
+        return false
+      }
+
+      const refs = findNodeReferences(config, from)
+      if (refs.length > 0) {
+        const shown = refs.slice(0, 4).map((r) => r.label)
+        const more = refs.length - shown.length
+        const ok = await confirm({
+          tone: 'neutral',
+          label: 'Rename node',
+          title: `Rename “${from}” to “${to}”?`,
+          body: `This also updates ${refSummary(refs)} that point at it — ${shown.join(', ')}${more > 0 ? `, and ${more} more` : ''}. They are saved together, so nothing is left pointing at the old name.`,
+          confirmLabel: 'Rename',
+        })
+        if (!ok) return false
+      }
+
+      // One PUT: the node and every reference move in the same write, so no
+      // intermediate state exists where a reference dangles.
+      const carried = renameNodeReferences(config, from, to)
+      const next = asArray(carried.Nodes).map((n, i) => (i === idx ? { ...n, Name: to } : n))
+      return save(
+        { ...carried, Nodes: next },
+        refs.length > 0
+          ? `Renamed to ${to} — updated ${refs.length} reference${refs.length === 1 ? '' : 's'}`
+          : `Renamed to ${to}`,
+      )
+    },
+    [config, nodes, save, confirm],
  )

  // Pin (or clear) one node's dial egress. Same optimistic save→apply path as
@@ -563,16 +878,20 @@ export default function Nodes() {
  )

  const removeSub = useCallback(
-    (idx: number) => {
+    async (idx: number) => {
      if (!config) return
      const target = subs[idx]
      const hasCache = nodes.some((n) => n.FromSub === target.Name)
-      const extra = hasCache ? ' Its cached nodes stay until you next apply.' : ''
-      if (!window.confirm(`Delete subscription “${target.Name}”?${extra}`)) return
+      const ok = await confirm({
+        label: 'Delete subscription',
+        title: `Delete subscription “${target.Name}”?`,
+        body: hasCache ? 'Its cached nodes stay until you next apply.' : undefined,
+      })
+      if (!ok) return
      const next = subs.filter((_, i) => i !== idx)
      void save({ ...config, Subscriptions: next }, `Deleted ${target.Name}`)
    },
-    [config, subs, nodes, save],
+    [config, subs, nodes, save, confirm],
  )

  // Commit an options edit for one subscription. The editor hands back a fully
@@ -736,6 +1055,20 @@ export default function Nodes() {
                disabled={busy || importing || !config}
              />
            )}
+            <input
+              className="fp-input add-name"
+              type="text"
+              spellCheck={false}
+              autoComplete="off"
+              placeholder="Name (optional)"
+              aria-label="Node name — optional"
+              value={nodeName}
+              onChange={(e) => {
+                setNodeName(e.target.value)
+                if (nodeErr) setNodeErr(null)
+              }}
+              disabled={busy || importing || !config}
+            />
            <Button type="submit" variant="primary" disabled={busy || importing || !config}>
              {importing ? 'Importing…' : saving ? 'Saving…' : addMode === 'conf' ? 'Import' : 'Add node'}
            </Button>
@@ -743,7 +1076,8 @@ export default function Nodes() {
          <p className="add-hint">
            {addMode === 'conf'
              ? 'Paste a wg-quick / AmneziaWG .conf — it starts with [Interface].'
-              : 'vless://, ss://, trojan://, hysteria2://… A pasted [Interface] config is imported automatically.'}
+              : 'vless://, ss://, trojan://, hysteria2://… A pasted [Interface] config is imported automatically.'}{' '}
+            Leave the name empty and it’s taken from the server address; you can rename it later.
          </p>
        </div>
        {nodeErr && (
@@ -800,6 +1134,7 @@ export default function Nodes() {
                onToggle={() => toggleGroup(g)}
                onToggleNode={toggleNode}
                onRemoveNode={removeNode}
+                onRenameNode={renameNode}
                onSetEgress={setNodeEgress}
              />
            ))}
@@ -911,6 +1246,7 @@ function NodeGroup({
  onToggle,
  onToggleNode,
  onRemoveNode,
+  onRenameNode,
  onSetEgress,
 }: {
  group: NodeGroupData
@@ -920,6 +1256,7 @@ function NodeGroup({
  onToggle: () => void
  onToggleNode: (idx: number, on: boolean) => void
  onRemoveNode: (idx: number) => void
+  onRenameNode: (idx: number, name: string, onError: (msg: string) => void) => Promise<boolean>
  onSetEgress: (idx: number, egress: string) => Promise<boolean>
 }) {
  const panelId = `node-group-${group.key || 'manual'}`
@@ -940,6 +1277,12 @@ function NodeGroup({
          <span className="group-count mono">{count}</span>
        </button>
      </h3>
+      {open && group.key !== '' && (
+        <p className="group-note">
+          Names come from the subscription feed and are rewritten on every update, so nodes in this
+          list can’t be renamed here.
+        </p>
+      )}
      {open && (
        <ul id={panelId} className="rows-list group-rows">
          {group.items.map(({ node, idx }) => (
@@ -950,6 +1293,7 @@ function NodeGroup({
              egressNames={egressNames}
              onToggle={(on) => onToggleNode(idx, on)}
              onDelete={() => onRemoveNode(idx)}
+              onRename={(name, onError) => onRenameNode(idx, name, onError)}
              onSetEgress={(egress) => onSetEgress(idx, egress)}
            />
          ))}
@@ -965,6 +1309,7 @@ function NodeRow({
  egressNames,
  onToggle,
  onDelete,
+  onRename,
  onSetEgress,
 }: {
  node: NodeCfg
@@ -972,6 +1317,7 @@ function NodeRow({
  egressNames: string[]
  onToggle: (on: boolean) => void
  onDelete: () => void
+  onRename: (name: string, onError: (msg: string) => void) => Promise<boolean>
  onSetEgress: (egress: string) => Promise<boolean>
 }) {
  const { proto, host, hasCreds } = useMemo(() => parseShareLink(node.URI), [node.URI])
@@ -982,6 +1328,68 @@ function NodeRow({
  const [open, setOpen] = useState(false)
  const panelId = `node-egress-${node.FromSub || 'manual'}-${node.Name}`

+  // ---- inline rename (same interaction as a device row) ---------------------
+  // Enter commits, Esc cancels, blur commits; a ref-guard keeps Esc-then-blur
+  // from committing twice. Unlike a device, the commit can be REFUSED (a name
+  // collision, or references that can't be carried), so the input stays open
+  // with the reason under it instead of closing on a change that never happened.
+  const [renaming, setRenaming] = useState(false)
+  const [draft, setDraft] = useState(node.Name)
+  const [renameErr, setRenameErr] = useState<string | null>(null)
+  const nameInput = useRef<HTMLInputElement>(null)
+  const renameBtn = useRef<HTMLButtonElement>(null)
+  const finished = useRef(false)
+
+  const beginRename = () => {
+    setDraft(node.Name)
+    setRenameErr(null)
+    finished.current = false
+    setRenaming(true)
+  }
+  const finishRename = async (commit: boolean) => {
+    if (finished.current) return
+    finished.current = true
+    const nm = draft.trim()
+    if (!commit || !nm || nm === node.Name) {
+      setRenaming(false)
+      setRenameErr(null)
+      return
+    }
+    // Armed BEFORE the save: the renamed row remounts the moment the config
+    // state lands, which is before this await resolves. Arming afterwards would
+    // always miss it.
+    pendingRenameFocus = nm
+    const ok = await onRename(nm, (msg) => setRenameErr(msg))
+    if (ok) {
+      setRenaming(false)
+      setRenameErr(null)
+    } else {
+      if (pendingRenameFocus === nm) pendingRenameFocus = null
+      // Refused — hold the field open on the rejected text so it can be fixed.
+      finished.current = false
+      nameInput.current?.focus()
+    }
+  }
+
+  useEffect(() => {
+    if (renaming) {
+      nameInput.current?.focus()
+      nameInput.current?.select()
+    }
+  }, [renaming])
+
+  // Claim the baton if this row is the one the rename produced. The row remounts
+  // while the PUT is still in flight, so on that first pass the button is still
+  // disabled and focus() would be a silent no-op — the baton is held until the
+  // save settles and this effect re-runs with a focusable button.
+  useEffect(() => {
+    if (pendingRenameFocus !== node.Name) return
+    const btn = renameBtn.current
+    if (!btn || btn.disabled) return
+    pendingRenameFocus = null
+    btn.focus()
+  }, [node.Name, busy])
+
  return (
    <li className={`row-item node-row${open ? ' node-row--open' : ''}`}>
      <div className="row-head">
@@ -993,10 +1401,73 @@ function NodeRow({
        />
        <div className="row-main">
          <div className="row-line1">
-            <span className="row-name">{node.Name}</span>
+            {renaming ? (
+              <input
+                ref={nameInput}
+                className="inline-rename-input node-name-input mono"
+                type="text"
+                spellCheck={false}
+                autoComplete="off"
+                value={draft}
+                aria-label={`Rename node ${node.Name}`}
+                aria-invalid={renameErr ? true : undefined}
+                onChange={(e) => {
+                  setDraft(e.target.value)
+                  if (renameErr) setRenameErr(null)
+                }}
+                onBlur={() => void finishRename(true)}
+                onKeyDown={(e) => {
+                  if (e.key === 'Enter') {
+                    e.preventDefault()
+                    void finishRename(true)
+                  } else if (e.key === 'Escape') {
+                    e.preventDefault()
+                    void finishRename(false)
+                  }
+                }}
+                disabled={busy}
+              />
+            ) : (
+              <>
+                <span className="row-name" title={node.Name}>
+                  {node.Name}
+                </span>
+                {managed ? (
+                  // Not hidden — withheld, with the reason attached. A control
+                  // that quietly isn't there reads as a bug; this one states the
+                  // rule, and the same sentence is on the group header above.
+                  <button
+                    type="button"
+                    className="inline-rename inline-rename--locked"
+                    disabled
+                    aria-label={`Can’t rename ${node.Name} — its name comes from subscription “${node.FromSub}” and is rewritten on the next update`}
+                    title={`Named by subscription “${node.FromSub}” — the feed rewrites this name on the next update. Rename it in the subscription, or add the node manually.`}
+                  >
+                    🔒
+                  </button>
+                ) : (
+                  <button
+                    ref={renameBtn}
+                    type="button"
+                    className="inline-rename"
+                    onClick={beginRename}
+                    disabled={busy}
+                    aria-label={`Rename node ${node.Name}`}
+                    title="Rename"
+                  >
+                    ✎
+                  </button>
+                )}
+              </>
+            )}
            <span className="badge">{proto}</span>
            {node.Stale && <span className="badge badge--warn">stale</span>}
          </div>
+          {renameErr && (
+            <p className="row-err" role="alert">
+              {renameErr}
+            </p>
+          )}
          <div className="row-line2 mono">
            <span className="row-host">{host}</span>
            {hasCreds && (
@@ -341,7 +341,7 @@ export function Overview({
          led={{ variant: len(config?.Rules) ? 'on' : 'amber' }}
          rows={[
            { k: 'egresses', v: String(len(config?.Egresses)) },
-            { k: 'default', v: defaultTarget(config), hot: true },
+            { k: 'default', v: defaultTarget(status, config), hot: true },
          ]}
        />

@@ -563,12 +563,20 @@ const NAV_LABEL: Record<Route, string> = {
 // for the apply/rollback flow, where the individual flags are the actual
 // subject of the page.)

-function defaultTarget(config: Model | null): string {
-  const rules = config?.Rules ?? []
-  if (rules.length === 0) return '—'
-  // The highest Order enabled rule is the effective catch-all.
-  const enabled = rules.filter((r) => r.Enabled)
-  if (enabled.length === 0) return 'none'
-  const last = enabled.reduce((a, b) => (b.Order >= a.Order ? b : a))
-  return last.Target || last.Egress || last.Name
+/** Where everything not matched by a rule goes — the engine's route `final`.
+ *
+ *  Taken from the daemon (status.traffic.default), which reads it off the config
+ *  it is running. The guess this replaced was "the highest-Order enabled rule",
+ *  and that is not what the default is: a rule only becomes the default by having
+ *  NO conditions at all, whatever its Order (model.IsCatchAll), so a specific
+ *  high-Order rule was routinely printed here as the router's default. It also
+ *  described the config on disk rather than the one running, and could not see a
+ *  target that failed to resolve and fell back.
+ *
+ *  Falls back to the rule count only when the daemon has not reported — never to
+ *  a guess about where traffic goes. */
+function defaultTarget(status: Status | null, config: Model | null): string {
+  const d = status?.traffic?.default
+  if (d) return d
+  return len(config?.Rules) === 0 ? '—' : 'not reported'
 }
@@ -1,6 +1,6 @@
 import './Profiles.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { Button, Led, Toggle } from '../components'
+import { Button, Led, Toggle, useConfirm } from '../components'
 import { apply as apiApply, getConfig, getInterfaces, putConfig, ApiError } from '../api'
 import type { Interface, Model, Profile } from '../api'

@@ -36,6 +36,7 @@ function namesOf(v: unknown): string[] {
 // ---- page ------------------------------------------------------------------

 export default function Profiles() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)

@@ -192,9 +193,14 @@ export default function Profiles() {
  )

  const deleteProfile = useCallback(
-    (name: string) => {
+    async (name: string) => {
      if (!config) return
-      if (!window.confirm(`Delete profile “${name}”? Its overrides stop applying.`)) return
+      const ok = await confirm({
+        label: 'Delete profile',
+        title: `Delete profile “${name}”?`,
+        body: 'Its overrides stop applying.',
+      })
+      if (!ok) return
      const next = profiles.filter((p) => p.Name !== name)
      const g =
        config.Globals.ActiveProfile === name
@@ -202,7 +208,7 @@ export default function Profiles() {
          : config.Globals
      void save({ ...config, Profiles: next, Globals: g }, `Deleted ${name}`)
    },
-    [config, profiles, save],
+    [config, profiles, save, confirm],
  )

  // ---- expansion (only one profile editor open at a time) -------------------
@@ -58,6 +58,45 @@
  color: var(--ink);
 }

+/* ---- active-profile banner ----
+ *
+ * Deliberately NOT the accent plate the save→apply bar wears above. Orange is
+ * "there is something for you to do" on this faceplate, and an active WAN profile
+ * is a standing condition, not a pending action. A quiet plate with an amber tag
+ * reads as "note the state" — and it is the SAME amber the overridden rows below
+ * carry, so the banner and its rows are visibly one story rather than two
+ * unrelated oddities. */
+.rt-prof-banner {
+  display: flex;
+  align-items: flex-start;
+  gap: 10px;
+  margin: 0 0 calc(var(--u, 8px) * 2.5);
+  padding: 10px 14px;
+  border: 1px solid color-mix(in srgb, var(--amber) 35%, var(--groove));
+  border-radius: 8px;
+  background: color-mix(in srgb, var(--amber) 7%, transparent);
+  font-family: var(--font-sans);
+  font-size: 12px;
+  line-height: 1.55;
+  color: var(--dim);
+}
+.rt-prof-banner strong {
+  color: var(--ink);
+  font-weight: 600;
+}
+.rt-prof-tag {
+  flex: none;
+  margin-top: 1px;
+  padding: 2px 7px;
+  border: 1px solid color-mix(in srgb, var(--amber) 55%, var(--groove));
+  border-radius: 999px;
+  background: color-mix(in srgb, var(--amber) 12%, transparent);
+  font-size: 9px;
+  letter-spacing: var(--track-label);
+  text-transform: uppercase;
+  color: var(--amber);
+}
+
 /* ---- empty state ---- */
 .rt-empty {
  padding: calc(var(--u, 8px) * 4) 0 calc(var(--u, 8px) * 3);
@@ -252,6 +291,88 @@
  color: var(--faint);
 }

+/* ---- a rule that can never fire (superseded by a later condition-less rule) ----
+ *
+ * Warn semantics only: --amber, never --accent. Orange is the ACTIVE state on this
+ * faceplate, and a rule the router ignores is the opposite of active — painting it
+ * orange is what made two `default` rows look equally live. It is a dashed amber
+ * frame, an amber order chip, and a dimmed target, so the row reads as "wired but
+ * not connected" without shouting: nothing is broken, one setting is just inert. */
+.rt-rule.dead {
+  border-style: dashed;
+  border-color: color-mix(in srgb, var(--amber) 55%, var(--groove));
+  background: var(--panel);
+  box-shadow: none;
+}
+.rt-ord.dead {
+  color: var(--amber);
+  border-color: color-mix(in srgb, var(--amber) 45%, var(--groove));
+}
+.rt-badge.dead {
+  padding: 1px 7px;
+  border: 1px solid color-mix(in srgb, var(--amber) 55%, var(--groove));
+  border-radius: 999px;
+  background: color-mix(in srgb, var(--amber) 12%, transparent);
+  color: var(--amber);
+}
+.rt-dead-note {
+  font-family: var(--font-sans);
+  font-size: 11.5px;
+  line-height: 1.45;
+  color: var(--dim);
+}
+/* The target is still what the operator asked for, so it stays readable — just
+ * quiet, because the router is not using it. */
+.rt-rule.dead .rt-target {
+  border-style: dashed;
+  opacity: 0.62;
+}
+
+/* ---- a rule the active WAN profile overrides ----
+ *
+ * The row itself needs no new paint: an overridden-off rule already wears `.off`
+ * (it is off, whatever its switch says) and an overridden-on rule wears nothing
+ * (it is on). What was missing was never colour — it was the sentence naming who
+ * decided. So this is the per-row twin of the banner and borrows .rt-dead-note's
+ * type wholesale: same voice, same size, one <p> margin to reset. */
+/* The same pill as .rt-badge.dead, so the two override states read as one pair,
+ * but in accent — a rule the profile forces ON is active, and active is orange on
+ * this faceplate. The pill is also what keeps it from running into the plain
+ * "default route · final" badge beside it, where "final on · by profile" read as
+ * one phrase. */
+.rt-badge.prof-on {
+  padding: 1px 7px;
+  border: 1px solid var(--accent-soft);
+  border-radius: 999px;
+  background: color-mix(in srgb, var(--accent) 10%, transparent);
+}
+
+.rt-prof-note {
+  margin: 0;
+}
+.rt-prof-note strong {
+  color: var(--ink);
+  font-weight: 600;
+}
+
+/* Switch + its legend. The caption shows ONLY while a profile overrides the rule,
+ * and it is what keeps the control honest: the plate says what the router is
+ * doing, this says the switch is about the saved setting. A legend under the
+ * control it names is the faceplate's own idiom. */
+.rt-switch {
+  display: inline-flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 3px;
+}
+.rt-switch-note {
+  font-family: var(--font-mono);
+  font-size: 8.5px;
+  letter-spacing: var(--track-label);
+  text-transform: uppercase;
+  color: var(--faint);
+}
+
 /* ---- target chip (styled like the artifact's group:auto mono chips) ---- */
 .rt-target {
  display: inline-flex;
@@ -359,9 +480,6 @@
  gap: 5px;
  min-width: 0;
 }
-.rt-field-wide {
-  grid-column: span 2;
-}
 .rt-flabel {
  font-family: var(--font-mono);
  font-size: 9px;
@@ -478,6 +596,8 @@ select.rt-input {
  border-color: var(--accent);
  box-shadow: 0 1px 0 var(--edge) inset, 0 0 0 1px var(--accent-soft);
 }
+/* "no matchers" flag in the plate foot — shared by BOTH rule forms (add and
+ * edit), so the same non-blocking warning reads identically in either. */
 .rt-edit-warn {
  font-family: var(--font-mono);
  font-size: 11.5px;
@@ -800,9 +920,6 @@ select.rt-input {
    justify-content: flex-start;
    align-self: start;
  }
-  .rt-field-wide {
-    grid-column: auto;
-  }
  .rt-rs-row {
    grid-template-columns: 1fr;
    row-gap: 10px;
@@ -1,16 +1,17 @@
 import './Routing.css'
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
 import type { FormEvent, ReactNode } from 'react'
-import { Button, CatSuggest, SrcPicker, Toggle } from '../components'
+import { Button, CatSuggest, SrcPicker, Toggle, useConfirm } from '../components'
 import {
  apply as apiApply,
  getConfig,
  putConfig,
+  getRulesReachability,
  getRulesetStatus,
  updateRuleset as apiUpdateRuleset,
  ApiError,
 } from '../api'
-import type { Model, Rule, Ruleset, RulesetStatus } from '../api'
+import type { Model, Rule, RuleReach, Ruleset, RulesetStatus } from '../api'

 // ---------------------------------------------------------------------------
 // The api.ts `Rule` is a deliberately thin subset (Name/Enabled/Order/Target/
@@ -21,9 +22,7 @@ import type { Model, Rule, Ruleset, RulesetStatus } from '../api'
 // ---------------------------------------------------------------------------
 type RRule = Rule & {
  Src?: string[] | null
-  DstDomain?: string[] | null
  DstRuleset?: string[] | null
-  DstIP?: string[] | null
  DstPort?: string
  Proto?: string
  Kill?: string
@@ -34,6 +33,29 @@ type RRule = Rule & {
  // Minutes east of UTC anchoring the schedule's wall-clock times; captured
  // from the editing browser on save (the router has no tzdata). 0 ⇒ UTC.
  SchedUTCOffset?: number
+  // The daemon's unmigrated-rule tripwire (model.Rule.LegacyDst), read-only here.
+  // Non-empty ⇒ the config STILL carries the schema-v1 `dst_domain`/`dst_ip` that
+  // schema v2 removed, i.e. `shaterd migrate` never ran or could not commit. Each
+  // element is the raw `<option>=<value>` text so the panel can quote what was
+  // found. The daemon holds such a rule disabled; the panel only reports it (it
+  // is never rendered back to UCI, so a config write from here drains it out).
+  // Field name is the Go one: model.Rule has no json tags.
+  LegacyDst?: string[] | null
+}
+
+/**
+ * Whether a rule is in force, kept strictly apart from whether it is switched on.
+ *
+ * `on` is the EFFECTIVE state — what the router is actually doing — and every mark
+ * on the row is drawn from it. `profile`/`dir` are set only when the active WAN
+ * profile is the reason the two differ, so the row can name who overrode the
+ * saved setting instead of leaving the operator to guess why a switch that reads
+ * "on" routes nothing.
+ */
+type RuleForce = {
+  on: boolean
+  profile: string | null
+  dir: 'enabled' | 'disabled' | null
 }

 /**
@@ -96,7 +118,6 @@ function ProtoOptions({ value }: { value: string }) {

 const len = (a: unknown[] | null | undefined): number => (a ? a.length : 0)
 const byOrder = (a: RRule, b: RRule): number => a.Order - b.Order
-const csv = (s: string): string[] => s.split(',').map((x) => x.trim()).filter(Boolean)

 // --- ruleset helpers --------------------------------------------------------
 // A `config ruleset` (api.ts Ruleset) is a named domain/ipcidr list a rule
@@ -212,18 +233,59 @@ function everyLabel(sec: number): string {
  return `every ${sec}s`
 }

-/** A rule with no matcher of any kind is the effective catch-all (route Final). */
+/** A rule with no matcher of any kind is the effective catch-all (route Final).
+ *  Mirrors model.IsCatchAll on the daemon side — the two must agree or the
+ *  "never applies" badge lands on a different row than the apply warning.
+ *
+ *  AN UNMIGRATED RULE IS NEVER A CATCH-ALL, and that is the first thing checked
+ *  here, exactly as on the Go side (model/reachability.go). When LegacyDst is
+ *  non-empty the rule's destination is still written in the schema-v1 options
+ *  the parser no longer reads, so its lack of matchers means "the destination is
+ *  unreadable", not "matches everything" — reading it the other way is precisely
+ *  what turned an uncommitted `shaterd migrate` into route Final for the whole
+ *  router. The daemon holds such a rule disabled and reports false here; if this
+ *  copy disagreed, the panel would paint the row "default route · final" while
+ *  the daemon routes nothing through it. */
 function isCatchAll(r: RRule): boolean {
+  if (len(r.LegacyDst) > 0) return false
  return (
    len(r.Src) === 0 &&
-    len(r.DstDomain) === 0 &&
    len(r.DstRuleset) === 0 &&
-    len(r.DstIP) === 0 &&
    !(r.DstPort && r.DstPort.trim()) &&
    !(r.Proto && r.Proto.trim())
  )
 }

+/** isCatchAll's twin for a form still being edited: the live fields of either
+ *  rule form with no matcher left in them.
+ *
+ *  Such a rule is not "matches all" in the ordinary sense — the engine emits it
+ *  as route.Final, and among several the LAST one in rule order owns it. So what
+ *  saving one actually does depends on what is last right now, and there are
+ *  three cases:
+ *    - no rules at all, or the last rule is a conditional one → the new rule
+ *      lands last and TAKES the default, silently retargeting every otherwise
+ *      unmatched flow (e.g. the whole LAN to `direct`, past the tunnel);
+ *    - the last rule is already a catch-all → nextOrder() deliberately inserts
+ *      the new one BEFORE it (and bumps the old one up), so the existing default
+ *      keeps route.Final and the new rule is dead on arrival — the daemon
+ *      reports it as shadowed and the row renders as such.
+ *  Both outcomes are worth a warning, and neither form knows which it will be
+ *  (the add form has no rule list), so the shared text says only what is certain:
+ *  the rule has no matchers. It is a legal configuration either way, so neither
+ *  form blocks it — they warn, from this one predicate, so the flag cannot drift
+ *  out of sync between add and edit. */
+function formHasNoMatchers(f: {
+  src: string[]
+  port: string
+  rulesets: string[]
+  proto: string
+}): boolean {
+  return (
+    f.src.length === 0 && f.port.trim() === '' && f.rulesets.length === 0 && f.proto.trim() === ''
+  )
+}
+
 /** Effective routing target for a rule (Target wins; a bare Egress is a target too). */
 function effectiveTarget(r: RRule): string {
  if (r.Target && r.Target.trim()) return r.Target.trim()
@@ -262,16 +324,19 @@ interface TargetGroups {
  nodes: TargetOpt[] // node:<n> (huge — rendered last)
 }

-// Free-text destination matchers offered by the ADD form. Domains are NOT one of
-// them (the user's call): domain matching goes through named rulesets — that's
-// what they exist for. 'none' = the rule matches by rulesets/source/proto alone.
-// (Legacy rules that already carry DstDomain stay editable in the edit form.)
-type MatchKind = 'none' | 'ip' | 'port'
+// The add form's fields. WHERE traffic is going is a ruleset choice and nothing
+// else — a rule has no inline domain or address list any more, so the old
+// Match-kind picker (rulesets / ip / port) collapsed into a plain Port field
+// beside the ruleset picker. The cost is real and accepted: routing a single
+// domain is no longer done here — you leave for the Rulesets panel, create the
+// list, fill it, and come back to check it. A "create a list from here" shortcut
+// was proposed and rejected (DECISIONS.md D21): a second place to author a list
+// is a second place for its entry semantics and duplicate-name rules to drift,
+// which is the exact thing D21 removed.
 interface AddForm {
  name: string
  src: string[]
-  matchKind: MatchKind
-  matchValue: string
+  port: string
  rulesets: string[]
  proto: string
  target: string
@@ -283,8 +348,7 @@ interface AddForm {
 const EMPTY_FORM: AddForm = {
  name: '',
  src: [],
-  matchKind: 'none',
-  matchValue: '',
+  port: '',
  rulesets: [],
  proto: '',
  target: 'direct',
@@ -317,6 +381,7 @@ const browserTZName = (): string => {
 }

 export default function Routing() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)
  const [actionError, setActionError] = useState<string | null>(null)
@@ -334,6 +399,22 @@ export default function Routing() {
    toastTimer.current = window.setTimeout(() => setToast(null), 2600)
  }, [])

+  // Which rules can never fire, keyed by their position in the model's Rules array
+  // — NOT by name. The config that made this necessary had two rules both called
+  // `default`, which is exactly when a name-keyed verdict badges the wrong row.
+  const [reach, setReach] = useState<Map<number, RuleReach>>(new Map())
+  const loadReach = useCallback(async () => {
+    try {
+      const { rules } = await getRulesReachability()
+      setReach(new Map(rules.map((r) => [r.index, r])))
+    } catch {
+      // An older daemon has no such endpoint, and a stopped one answers nothing.
+      // Drop the verdicts rather than keep stale ones: no badge is honest, a badge
+      // about the previous config is not.
+      setReach(new Map())
+    }
+  }, [])
+
  const load = useCallback(async () => {
    try {
      setLoadError(null)
@@ -341,7 +422,8 @@ export default function Routing() {
    } catch (e) {
      setLoadError(errMsg(e))
    }
-  }, [])
+    void loadReach()
+  }, [loadReach])
  useEffect(() => {
    void load()
  }, [load])
@@ -407,6 +489,71 @@ export default function Routing() {
  // Every rule name, for the edit form's duplicate-name guard (it excludes self).
  const ruleNames = useMemo(() => new Set(rules.map((r) => r.Name)), [rules])

+  // Each rule's position in the model's Rules array — the key the daemon's
+  // reachability verdicts use. `rules` above is a sorted COPY of the same object
+  // references, so identity survives the sort and this map stays valid.
+  const modelIndex = useMemo(() => {
+    const m = new Map<RRule, number>()
+    ;((config?.Rules as RRule[] | null | undefined) ?? []).forEach((r, i) => m.set(r, i))
+    return m
+  }, [config])
+
+  /**
+   * The daemon's verdict for one rule, or null when we have none that describes it.
+   *
+   * Verdicts are fetched separately from the config, so between an optimistic edit
+   * and the refetch they can describe the PREVIOUS rule list. Re-checking the
+   * echoed name and order is what stops that window from putting a "never applies"
+   * badge on a working rule: a mismatch means the verdict is not about this row,
+   * and no badge is the honest answer.
+   */
+  const verdictOf = useCallback(
+    (r: RRule): RuleReach | null => {
+      const i = modelIndex.get(r)
+      if (i === undefined) return null
+      const v = reach.get(i)
+      if (!v || v.name !== r.Name || v.order !== r.Order) return null
+      return v
+    },
+    [modelIndex, reach],
+  )
+
+  const shadowOf = useCallback(
+    (r: RRule): { by: string; byOrder: number; reason: string } | null => {
+      const v = verdictOf(r)
+      if (!v || !v.unreachable || !v.shadowed_by) return null
+      return { by: v.shadowed_by, byOrder: v.shadowed_by_order ?? 0, reason: v.reason ?? '' }
+    },
+    [verdictOf],
+  )
+
+  /**
+   * Whether a rule is IN FORCE, and who decided that — the two states this page
+   * used to conflate.
+   *
+   * `Rule.Enabled` from /api/config is the DESIRED state: what the operator saved,
+   * what the switch edits, what gets PUT back. The active WAN profile can override
+   * it in either direction, and then the desired state is no longer what the router
+   * is doing. Drawing the row from `Enabled` is what let a config with two rules
+   * `enabled '1'` show two live switches while the engine ran one chain.
+   *
+   * With no verdict — an older daemon, a stopped one, or one still describing the
+   * previous config — the desired state is all we know, so the row falls back to it
+   * and claims no profile rather than inventing one. The `typeof` guard is for the
+   * older daemon specifically: it answers without `effective_enabled` at all, and
+   * reading `undefined` as false would gray out every rule on the page.
+   */
+  const forceOf = useCallback(
+    (r: RRule): RuleForce => {
+      const v = verdictOf(r)
+      if (!v || typeof v.effective_enabled !== 'boolean') {
+        return { on: !!r.Enabled, profile: null, dir: null }
+      }
+      return { on: v.effective_enabled, profile: v.overridden_by ?? null, dir: v.override ?? null }
+    },
+    [verdictOf],
+  )
+
  // Rulesets are named domain/IP lists rules match against (rule.DstRuleset).
  const rulesets = useMemo<Ruleset[]>(
    () => [...((config?.Rulesets as Ruleset[] | null | undefined) ?? [])],
@@ -458,6 +605,10 @@ export default function Routing() {
        await putConfig(next)
        setSavedPending(true)
        flash(okMsg)
+        // The verdicts describe the config on disk, which just changed — re-ask.
+        // Adding or moving a rule is precisely what turns a working default into a
+        // superseded one, and vice versa.
+        void loadReach()
      } catch (e) {
        setConfig(prev)
        setActionError(errMsg(e))
@@ -466,7 +617,7 @@ export default function Routing() {
        setSaving(false)
      }
    },
-    [config, flash],
+    [config, flash, loadReach],
  )

  const commitRules = useCallback(
@@ -503,13 +654,17 @@ export default function Routing() {
  // Delete the ruleset AND strip its name from any rule that referenced it, so no
  // rule is left pointing at a matcher that no longer exists (one atomic persist).
  const deleteRuleset = useCallback(
-    (name: string) => {
+    async (name: string) => {
      if (!config) return
      const used = rulesetUsage.get(name) ?? 0
-      const warn = used
-        ? `Delete ruleset "${name}"? It'll be removed from ${used} rule${used === 1 ? '' : 's'} that match it.`
-        : `Delete ruleset "${name}"?`
-      if (!window.confirm(warn)) return
+      const ok = await confirm({
+        label: 'Delete ruleset',
+        title: `Delete ruleset "${name}"?`,
+        body: used
+          ? `It'll be removed from ${used} rule${used === 1 ? '' : 's'} that match it.`
+          : undefined,
+      })
+      if (!ok) return
      const nextRulesets = rulesets.filter((r) => r.Name !== name)
      const nextRules = rules.map((r) => {
        const cur = r.DstRuleset ?? []
@@ -520,7 +675,7 @@ export default function Routing() {
        `ruleset ${name} deleted`,
      )
    },
-    [config, rules, rulesets, rulesetUsage, persist],
+    [config, rules, rulesets, rulesetUsage, persist, confirm],
  )

  const onToggle = useCallback(
@@ -560,17 +715,26 @@ export default function Routing() {
  )

  const onDelete = useCallback(
-    (name: string) => {
-      if (!window.confirm(`Delete rule "${name}"? Traffic it matched will fall through to the next rule.`)) return
+    async (name: string) => {
+      const ok = await confirm({
+        label: 'Delete rule',
+        title: `Delete rule "${name}"?`,
+        body: 'Traffic it matched will fall through to the next rule.',
+      })
+      if (!ok) return
      commitRules(
        rules.filter((r) => r.Name !== name),
        `${name} deleted`,
      )
    },
-    [rules, commitRules],
+    [rules, commitRules, confirm],
  )

  // Insert a new rule just above the catch-all (so a specific rule can actually match).
+  // isCatchAll() is false for an unmigrated rule, which is the right answer here too:
+  // such a rule is held disabled and owns no default route, so there is nothing to
+  // insert ahead of — the new rule simply goes last, where a rule with no matchers
+  // does become the default.
  const nextOrder = useCallback((): { order: number; bumpCatchAll?: { name: string; order: number } } => {
    if (rules.length === 0) return { order: 10 }
    const last = rules[rules.length - 1]
@@ -622,18 +786,15 @@ export default function Routing() {
        return
      }
      setFormError(null)
-      const mv = form.matchValue.trim()
      const rule: RRule = {
        Name: name,
        Enabled: true,
        Order: 0,
        Src: form.src,
-        // Domains are matched via rulesets only — the add form has no free-text
-        // domain matcher by design.
-        DstDomain: [],
+        // Destination = rulesets, always. Domains and addresses live in a
+        // `config ruleset` so one list serves every rule that needs it.
        DstRuleset: form.rulesets,
-        DstIP: form.matchKind === 'ip' ? csv(mv) : [],
-        DstPort: form.matchKind === 'port' ? mv : '',
+        DstPort: form.port.trim(),
        Proto: form.proto,
        Target: form.target,
        Egress: '',
@@ -704,7 +865,14 @@ export default function Routing() {
    )
  }

-  const enabledCount = rules.filter((r) => r.Enabled).length
+  // One force verdict per displayed rule, computed once and handed down — the row,
+  // the counter and the banner must all be reading the SAME answer.
+  const force = rules.map((r) => forceOf(r))
+  // EFFECTIVE, not configured. A counter that added up saved switches said "2 / 2
+  // active" for a config the router was running one rule of.
+  const enabledCount = force.filter((f) => f.on).length
+  const overridden = force.filter((f) => f.profile !== null)
+  const overrideProfile = overridden[0]?.profile ?? null

  return (
    <section className="page" aria-label="Routing rules">
@@ -713,12 +881,28 @@ export default function Routing() {
          Rules run top to bottom on the bus — the <strong>first match wins</strong>. Traffic that
          reaches the bottom follows the default route.
        </p>
-        <span className="rt-count mono" aria-label={`${enabledCount} of ${rules.length} rules active`}>
+        <span className="rt-count mono" aria-label={`${enabledCount} of ${rules.length} rules in force`}>
          {enabledCount}
-          <small> / {rules.length} active</small>
+          <small> / {rules.length} in force</small>
        </span>
      </div>

+      {/* Said once at the top, so the per-row badges below read as consequences of
+          one thing rather than as N unrelated oddities. Only shown when a profile
+          actually changed something: a router that uses no profiles, or one whose
+          profile agrees with every saved switch, gets no banner at all. */}
+      {overrideProfile && (
+        <p className="rt-prof-banner" role="status">
+          <span className="rt-prof-tag mono">profile</span>
+          <span>
+            <strong className="mono">{overrideProfile}</strong> is the active WAN profile and is
+            overriding {overridden.length === 1 ? '1 rule' : `${overridden.length} rules`} below. The
+            switches keep showing what you saved; the rows show what the router is running. Change
+            which rules a profile forces on the <strong>Profiles</strong> page.
+          </span>
+        </p>
+      )}
+
      {actionError && (
        <p className="page-error" role="alert">
          {actionError}
@@ -737,14 +921,18 @@ export default function Routing() {
      {rules.length === 0 ? (
        <div className="rt-empty">
          <p>No rules — all traffic follows the default route.</p>
-          <p className="rt-empty-sub">Add a rule below to steer a domain, address, or port.</p>
+          <p className="rt-empty-sub">Add a rule below to steer a destination list, source, or port.</p>
        </div>
      ) : (
        <ol className="rt-list" aria-label="Routing rules in first-match order">
          {rules.map((r, i) =>
            editingRule === r.Name ? (
              <RuleEditForm
-                key={r.Name}
+                // Rule names are NOT unique in the wild — the config that prompted
+                // the never-applies badge had two rules called `default`, and a
+                // duplicate React key makes the second row shadow the first. The
+                // model index disambiguates without changing row identity.
+                key={`${modelIndex.get(r) ?? i}:${r.Name}`}
                initial={r}
                names={ruleNames}
                targets={targets}
@@ -755,12 +943,14 @@ export default function Routing() {
              />
            ) : (
              <RuleRow
-                key={r.Name}
+                key={`${modelIndex.get(r) ?? i}:${r.Name}`}
                rule={r}
                index={i}
                total={rules.length}
                busy={saving}
                editingOther={editingRule !== null}
+                shadow={shadowOf(r)}
+                force={force[i]}
                onEdit={onEditRule}
                onToggle={onToggle}
                onMove={onMove}
@@ -810,6 +1000,8 @@ function RuleRow({
  total,
  busy,
  editingOther,
+  shadow,
+  force,
  onEdit,
  onToggle,
  onMove,
@@ -820,6 +1012,11 @@ function RuleRow({
  total: number
  busy: boolean
  editingOther: boolean
+  /** Set when the daemon reports this rule can never fire; null when it can. */
+  shadow: { by: string; byOrder: number; reason: string } | null
+  /** Whether the rule is IN FORCE, and which profile decided that (see RuleForce).
+   *  Every mark on this row comes from here; `rule.Enabled` drives only the switch. */
+  force: RuleForce
  onEdit: (name: string) => void
  onToggle: (name: string) => void
  onMove: (name: string, dir: 'up' | 'down') => void
@@ -829,16 +1026,32 @@ function RuleRow({
  // the first-match order can't shift under the open form. Edit itself stays live —
  // clicking it just swaps which row is being edited.
  const frozen = busy || editingOther
-  const isDefault = isCatchAll(rule)
+  // A rule with no conditions is the router's default — but only ONE of them can
+  // be, and the daemon says which. A superseded one must not wear the default's
+  // marks (the dashed accent frame, the "· final" order chip, the "everything not
+  // matched above" line): those are the claim that made two `default` rules
+  // indistinguishable in the first place.
+  const dead = shadow !== null
+  // The unmigrated-rule tripwire (see RRule.LegacyDst / model.IsCatchAll): this
+  // rule's destination is still in the removed schema-v1 options, so the daemon
+  // holds it disabled. It is NOT an ordinary disabled rule — nobody switched it
+  // off — so it gets the same "wired but not connected" amber treatment as a
+  // shadowed rule, plus a badge and a line saying what to run. isCatchAll()
+  // already refuses to call it the default, so `final` marks cannot land here.
+  const legacyDst = (rule.LegacyDst ?? []).filter(Boolean)
+  const unmigrated = legacyDst.length > 0
+  const inert = dead || unmigrated
+  const isDefault = isCatchAll(rule) && !dead
  const target = effectiveTarget(rule)
  const tone = targetTone(target)
-  const cls = [
-    'rt-rule',
-    rule.Enabled ? '' : 'off',
-    isDefault ? 'final' : '',
-  ]
+  // Dimmed by the EFFECTIVE state, never by the saved one. A rule the active
+  // profile switched off is not in force, and the row has to read that way even
+  // though its switch — which edits the saved setting — is still on.
+  const cls = ['rt-rule', force.on ? '' : 'off', isDefault ? 'final' : '', inert ? 'dead' : '']
    .filter(Boolean)
    .join(' ')
+  // What the switch says, spelled out, for the moment the two disagree.
+  const savedState = rule.Enabled ? 'on' : 'off'

  return (
    <li className={cls}>
@@ -852,7 +1065,7 @@ function RuleRow({
        >
          ▲
        </button>
-        <span className={isDefault ? 'rt-ord final' : 'rt-ord'}>
+        <span className={isDefault ? 'rt-ord final' : inert ? 'rt-ord dead' : 'rt-ord'}>
          {isDefault ? '·' : rule.Order}
        </span>
        <button
@@ -870,14 +1083,56 @@ function RuleRow({
        <div className="rt-head">
          <span className="rt-name">{rule.Name}</span>
          {isDefault && <span className="rt-badge">default route · final</span>}
+          {unmigrated && <span className="rt-badge dead">held off · not migrated</span>}
+          {dead && <span className="rt-badge dead">never applies</span>}
+          {/* Amber for the rule the profile switched OFF (warn semantics: wired but
+              not connected), accent for the one it switched ON — orange is the
+              faceplate's active state, and a force-enabled rule is exactly that. */}
+          {force.dir === 'disabled' && <span className="rt-badge dead">off · by profile</span>}
+          {force.dir === 'enabled' && <span className="rt-badge prof-on">on · by profile</span>}
        </div>
        <div className="rt-match">
-          {isDefault ? (
+          {unmigrated ? (
+            // Why the rule is off and what fixes it. Same voice as the shadow note:
+            // state, cause, one command. The daemon says the same thing through the
+            // apply warnings (model.ValidateRules); this puts it on the row it is about.
+            <span className="rt-dead-note">
+              Destination still written the old way (<span className="mono">{legacyDst.join(', ')}</span>
+              ) — this config was never migrated, so shaterd cannot read where this rule sends
+              traffic and holds it disabled. Run <span className="mono">shaterd migrate</span> on the
+              router to turn those entries into a ruleset, then enable the rule again.
+            </span>
+          ) : dead ? (
+            // The badge says it never fires; this line says what beat it and what to
+            // do. Visible text, not a tooltip — the operator has to be able to find
+            // the other rule, and two rows can carry the same name.
+            <span className="rt-dead-note" title={shadow.reason}>
+              “{shadow.by}” (order {shadow.byOrder}) has no conditions either and runs after this
+              one, so it is the default the router uses. Give this rule a condition, or delete one
+              of the two.
+            </span>
+          ) : isDefault ? (
            <span className="rt-nomatch">everything not matched above</span>
          ) : (
            <Matchers rule={rule} />
          )}
        </div>
+        {/* Added BELOW the matchers, not instead of them: the rule's conditions are
+            still worth reading — the operator is deciding whether to change the
+            profile or the rule.
+            Two clauses only. The banner at the top of the page already carries the
+            general explanation and the way to change it, and a profile that
+            overrides several rules would otherwise repeat that paragraph on every
+            one of them. What is left is the part only this row can say: whether it
+            is in force, and what its own switch is showing instead. */}
+        {force.profile && (
+          <p className="rt-dead-note rt-prof-note">
+            {force.dir === 'disabled' ? 'Not in force' : 'In force'} — profile{' '}
+            <strong className="mono">{force.profile}</strong> switches this rule{' '}
+            {force.dir === 'disabled' ? 'off' : 'on'}. The switch still reads{' '}
+            <strong>{savedState}</strong>: that is the saved setting.
+          </p>
+        )}
      </div>

      <div className={`rt-target ${tone}`} title={`target: ${target}`}>
@@ -897,12 +1152,40 @@ function RuleRow({
        >
          Edit
        </button>
-        <Toggle
-          pressed={rule.Enabled}
-          onChange={() => onToggle(rule.Name)}
-          label={`${rule.Enabled ? 'Disable' : 'Enable'} rule ${rule.Name}`}
-          disabled={frozen}
-        />
+        {/* An unmigrated rule cannot be switched on from here, and the switch says
+            so rather than pretending: the daemon holds it disabled, but a config
+            write from the panel DROPS the unreadable legacy options (render.go
+            emits neither), so enabling it here would save a live rule with no
+            destination left at all — the catch-all this tripwire exists to
+            prevent. `shaterd migrate` clears LegacyDst and the switch comes back. */}
+        {/* The switch edits the SAVED setting and nothing else, so it keeps showing
+            rule.Enabled even while the active profile forces the opposite. Mirroring
+            the effective state here would be worse than the bug it replaces: the
+            operator would flip a switch that was never theirs, and the PUT would
+            write the profile's decision into UCI as if they had chosen it. The row
+            above says what the router is doing; the "saved" caption says what this
+            control is for. */}
+        <span className="rt-switch">
+          <Toggle
+            pressed={rule.Enabled}
+            onChange={() => onToggle(rule.Name)}
+            label={
+              unmigrated
+                ? `Rule ${rule.Name} is held disabled until the config is migrated`
+                : force.profile
+                  ? `Saved setting for rule ${rule.Name} is ${savedState}; profile ${force.profile} is forcing it ${
+                      force.dir === 'disabled' ? 'off' : 'on'
+                    }. This switch changes the saved setting only.`
+                  : `${rule.Enabled ? 'Disable' : 'Enable'} rule ${rule.Name}`
+            }
+            disabled={frozen || unmigrated}
+          />
+          {force.profile && (
+            <span className="rt-switch-note" aria-hidden="true">
+              saved
+            </span>
+          )}
+        </span>
        <button
          type="button"
          className="rt-del"
@@ -935,9 +1218,7 @@ function Matchers({ rule }: { rule: RRule }): ReactNode {
    )
  }
  listChip('src', rule.Src, 'src')
-  listChip('dns', rule.DstDomain, 'dom')
  listChip('ruleset', rule.DstRuleset, 'rs')
-  listChip('ip', rule.DstIP, 'ip')
  if (rule.DstPort && rule.DstPort.trim()) {
    chips.push(
      <span className="rt-chip" key="port">
@@ -1053,7 +1334,16 @@ function TargetOptions({ targets, current }: { targets: TargetGroups; current?:
  )
 }

-/** The dst_ruleset checkbox group. Renders nothing when no rulesets exist. */
+/**
+ * The destination picker: which rulesets this rule matches (dst_ruleset).
+ *
+ * Checkboxes and nothing else. This is the ONLY way a rule names a destination,
+ * so it renders even when the config has no lists yet — an empty picker that says
+ * where lists come from is the honest answer, and hiding it would leave the rule
+ * form with no destination control at all. Building and filling a list is the
+ * Rulesets panel's job, deliberately kept out of the rule editor so a list is
+ * created in exactly one place.
+ */
 function RulesetPicker({
  options,
  selected,
@@ -1065,24 +1355,26 @@ function RulesetPicker({
  busy: boolean
  onToggle: (name: string) => void
 }): ReactNode {
-  if (options.length === 0) return null
  return (
    <div className="rt-rsel">
-      <span className="rt-flabel">Match rulesets — dst_ruleset</span>
-      <div className="rt-rsel-opts" role="group" aria-label="Match these rulesets">
-        {options.map((n) => {
-          const on = selected.includes(n)
-          return (
-            <label key={n} className={on ? 'rt-rsel-opt on' : 'rt-rsel-opt'}>
-              <input type="checkbox" checked={on} onChange={() => onToggle(n)} disabled={busy} />
-              <span className="mono">{n}</span>
-            </label>
-          )
-        })}
-      </div>
+      <span className="rt-flabel">Destination — dst_ruleset</span>
+      {options.length > 0 && (
+        <div className="rt-rsel-opts" role="group" aria-label="Match these rulesets">
+          {options.map((n) => {
+            const on = selected.includes(n)
+            return (
+              <label key={n} className={on ? 'rt-rsel-opt on' : 'rt-rsel-opt'}>
+                <input type="checkbox" checked={on} onChange={() => onToggle(n)} disabled={busy} />
+                <span className="mono">{n}</span>
+              </label>
+            )
+          })}
+        </div>
+      )}
      <p className="rt-rsel-hint">
-        The rule also matches any traffic in the checked list(s). Combine with a domain, address, or
-        port, or use a ruleset on its own.
+        {options.length === 0
+          ? 'No rulesets yet. Add one under Rulesets below, then come back and check it here — a rule matches a destination through a ruleset only.'
+          : 'The rule matches traffic in ANY checked list. Narrow it further with a source, port or protocol.'}
      </p>
    </div>
  )
@@ -1208,7 +1500,14 @@ function AddRule({
    set('rulesets', form.rulesets.includes(n) ? form.rulesets.filter((x) => x !== n) : [...form.rulesets, n])
  const toggleDay = (d: string) =>
    set('schedDays', form.schedDays.includes(d) ? form.schedDays.filter((x) => x !== d) : [...form.schedDays, d])
-  const matchPlaceholder = form.matchKind === 'ip' ? '10.0.0.0/8, 100.64.0.0/10' : '443, 8080-8090'
+
+  // Same flag as the edit form, from the same predicate: no matcher = this rule
+  // asks to be route.Final. Whether it wins depends on what is last — it takes
+  // the default when there is no catch-all yet (or the last rule is conditional),
+  // and lands dead when there is one, because nextOrder() inserts ahead of it.
+  // Both are worth flagging, so the text states the fact and not the outcome.
+  // Shown, never blocking: a default route is a legal thing to write.
+  const noMatchers = formHasNoMatchers(form)

  return (
    <form className="rt-add" onSubmit={onSubmit} aria-label="Add a routing rule">
@@ -1241,32 +1540,17 @@ function AddRule({
        </label>

        <label className="rt-field">
-          <span className="rt-flabel">Match</span>
-          <select
+          <span className="rt-flabel">Port(s)</span>
+          <input
            className="rt-input mono"
-            value={form.matchKind}
-            onChange={(e) => set('matchKind', e.target.value as MatchKind)}
-          >
-            <option value="none">rulesets only</option>
-            <option value="ip">ip / cidr</option>
-            <option value="port">port</option>
-          </select>
+            value={form.port}
+            onChange={(e) => set('port', e.target.value)}
+            placeholder="443, 8080-8090"
+            autoComplete="off"
+            spellCheck={false}
+          />
        </label>

-        {form.matchKind !== 'none' && (
-          <label className="rt-field rt-field-wide">
-            <span className="rt-flabel">{form.matchKind === 'port' ? 'Port(s)' : 'Address(es)'}</span>
-            <input
-              className="rt-input mono"
-              value={form.matchValue}
-              onChange={(e) => set('matchValue', e.target.value)}
-              placeholder={matchPlaceholder}
-              autoComplete="off"
-              spellCheck={false}
-            />
-          </label>
-        )}
-
        <label className="rt-field">
          <span className="rt-flabel">Proto</span>
          <select
@@ -1313,6 +1597,9 @@ function AddRule({
        <Button type="submit" variant="primary" disabled={busy}>
          {busy ? 'Saving…' : 'Add rule'}
        </Button>
+        {noMatchers && !error && (
+          <span className="rt-edit-warn">no matchers — matches everything</span>
+        )}
        {error && (
          <span className="rt-add-error" role="alert">
            {error}
@@ -1324,11 +1611,11 @@ function AddRule({
 }

 // --- edit-a-rule plate (inline, replaces the row it edits) ------------------
-// Unlike AddRule, editing exposes all three destination matchers at once
-// (Domain(s) / IP-CIDR(s) / Port) rather than a single Match picker — a real rule
-// can carry several matcher kinds simultaneously and none may be silently dropped.
-// The full original rule is spread into the result on save, so Order / Enabled /
-// Kill / Egress (and anything else off-form) survive untouched.
+// Same fields as AddRule, on purpose: a rule carries exactly one destination
+// mechanism (rulesets) plus port/proto/source, so there is nothing an edit can
+// reveal that the add form hides. The full original rule is spread into the
+// result on save, so Order / Enabled / Kill / Egress (and anything else off-form)
+// survive untouched.
 function RuleEditForm({
  initial,
  names,
@@ -1348,8 +1635,6 @@ function RuleEditForm({
 }) {
  const [name, setName] = useState(initial.Name)
  const [src, setSrc] = useState<string[]>([...(initial.Src ?? [])])
-  const [domain, setDomain] = useState((initial.DstDomain ?? []).join(', '))
-  const [ip, setIp] = useState((initial.DstIP ?? []).join(', '))
  const [port, setPort] = useState(initial.DstPort ?? '')
  const [proto, setProto] = useState(initial.Proto ?? '')
  const [target, setTarget] = useState(effectiveTarget(initial))
@@ -1365,14 +1650,15 @@ function RuleEditForm({
  const toggleDay = (d: string) =>
    setSchedDays((cur) => (cur.includes(d) ? cur.filter((x) => x !== d) : [...cur, d]))

+  // This rule's destination may still be in the schema-v1 options (RRule.LegacyDst).
+  // The form cannot show or edit them — they are not fields any more — and saving
+  // writes the model back through render.go, which does not emit them. So a save
+  // here silently DISCARDS that destination. Say so before it happens; the fix is
+  // `shaterd migrate`, which converts them into a ruleset this form can check.
+  const legacyDst = (initial.LegacyDst ?? []).filter(Boolean)
+
  // A rule with no matcher of any kind is a catch-all — legal, but worth flagging.
-  const noMatchers =
-    src.length === 0 &&
-    csv(domain).length === 0 &&
-    csv(ip).length === 0 &&
-    port.trim() === '' &&
-    rulesets.length === 0 &&
-    proto.trim() === ''
+  const noMatchers = formHasNoMatchers({ src, port, rulesets, proto })

  const submit = (e: FormEvent) => {
    e.preventDefault()
@@ -1394,8 +1680,6 @@ function RuleEditForm({
      ...initial,
      Name: nm,
      Src: src,
-      DstDomain: csv(domain),
-      DstIP: csv(ip),
      DstPort: port.trim(),
      DstRuleset: rulesets,
      Proto: proto,
@@ -1414,7 +1698,15 @@ function RuleEditForm({
      <form className="rt-add rt-edit-form" onSubmit={submit} aria-label={`Edit rule ${initial.Name}`}>
        <div className="rt-add-hd">
          <span className="rt-add-title">Edit {initial.Name}</span>
-          <span className="rt-add-sub">Empty a field to drop that matcher. Save, then Apply.</span>
+          <span className="rt-add-sub">
+            Uncheck a list or clear a field to drop that matcher. Save, then Apply.
+          </span>
+          {legacyDst.length > 0 && (
+            <span className="rt-edit-warn">
+              not migrated — saving discards its old destination ({legacyDst.join(', ')}); run
+              shaterd migrate first
+            </span>
+          )}
        </div>

        <div className="rt-fields">
@@ -1444,37 +1736,6 @@ function RuleEditForm({
            />
          </label>

-          {/* Domains are matched via rulesets by design — this legacy field only
-              appears when the rule ALREADY carries free-text domains, so they
-              stay visible and clearable rather than silently preserved. */}
-          {(initial.DstDomain ?? []).length > 0 && (
-            <label className="rt-field rt-field-wide">
-              <span className="rt-flabel">Domain(s) — legacy</span>
-              <input
-                className="rt-input mono"
-                value={domain}
-                onChange={(e) => setDomain(e.target.value)}
-                placeholder="youtube.com, *.googlevideo.com"
-                autoComplete="off"
-                spellCheck={false}
-                disabled={busy}
-              />
-            </label>
-          )}
-
-          <label className="rt-field rt-field-wide">
-            <span className="rt-flabel">IP / CIDR(s)</span>
-            <input
-              className="rt-input mono"
-              value={ip}
-              onChange={(e) => setIp(e.target.value)}
-              placeholder="10.0.0.0/8, 100.64.0.0/10"
-              autoComplete="off"
-              spellCheck={false}
-              disabled={busy}
-            />
-          </label>
-
          <label className="rt-field">
            <span className="rt-flabel">Port(s)</span>
            <input
@@ -1,6 +1,6 @@
 import './Targets.css'
 import { Fragment, useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { Button, Led, Toggle } from '../components'
+import { Button, Led, Toggle, useConfirm } from '../components'
 import type { LedVariant } from '../components'
 import {
  apply as apiApply,
@@ -161,18 +161,18 @@ function renameReferences(m: Model, kind: RefKind, from: string, to: string): Mo
 }

 /**
- * The sentence a delete confirmation appends: what still points at this target,
- * and what happens to it. Empty list ⇒ an explicit "nothing references it", so
- * the operator can delete a stray with confidence instead of guessing.
+ * The body of a delete confirmation: what still points at this target, and what
+ * happens to it. Empty list ⇒ an explicit "nothing references it", so the
+ * operator can delete a stray with confidence instead of guessing.
 */
 function refWarning(refs: RefSite[]): string {
-  if (refs.length === 0) return ' Nothing references it.'
+  if (refs.length === 0) return 'Nothing references it.'
  const shown = refs.slice(0, 4).map((r) => r.label)
  const more = refs.length - shown.length
  const list = `${shown.join(', ')}${more > 0 ? `, and ${more} more` : ''}`
  return refs.length === 1
-    ? ` It is referenced by ${list}, whose traffic will be blocked (an unresolved target never falls through to the default route).`
-    : ` It is referenced by ${refs.length} places — ${list} — whose traffic will be blocked (an unresolved target never falls through to the default route).`
+    ? `It is referenced by ${list}, whose traffic will be blocked (an unresolved target never falls through to the default route).`
+    : `It is referenced by ${refs.length} places — ${list} — whose traffic will be blocked (an unresolved target never falls through to the default route).`
 }

 /**
@@ -457,6 +457,7 @@ interface Opt {
 // ---- page ------------------------------------------------------------------

 export default function Targets() {
+  const confirm = useConfirm()
  const [config, setConfig] = useState<Model | null>(null)
  const [loadError, setLoadError] = useState<string | null>(null)

@@ -787,13 +788,18 @@ export default function Targets() {
  )

  const removeGroup = useCallback(
-    (name: string) => {
+    async (name: string) => {
      if (!config) return
      const refs = findReferences(config, 'group', name)
-      if (!window.confirm(`Delete group “${name}”?${refWarning(refs)}`)) return
+      const ok = await confirm({
+        label: 'Delete group',
+        title: `Delete group “${name}”?`,
+        body: refWarning(refs),
+      })
+      if (!ok) return
      void save({ ...config, Groups: groups.filter((g) => g.Name !== name) }, `Deleted ${name}`)
    },
-    [config, groups, save],
+    [config, groups, save, confirm],
  )

  // ---- chain mutations ------------------------------------------------------
@@ -820,13 +826,18 @@ export default function Targets() {
  )

  const removeChain = useCallback(
-    (name: string) => {
+    async (name: string) => {
      if (!config) return
      const refs = findReferences(config, 'chain', name)
-      if (!window.confirm(`Delete chain “${name}”?${refWarning(refs)}`)) return
+      const ok = await confirm({
+        label: 'Delete chain',
+        title: `Delete chain “${name}”?`,
+        body: refWarning(refs),
+      })
+      if (!ok) return
      void save({ ...config, Chains: chains.filter((c) => c.Name !== name) }, `Deleted ${name}`)
    },
-    [config, chains, save],
+    [config, chains, save, confirm],
  )

  // ---- egress mutations -----------------------------------------------------
@@ -853,13 +864,18 @@ export default function Targets() {
  )

  const removeEgress = useCallback(
-    (name: string) => {
+    async (name: string) => {
      if (!config) return
      const refs = findReferences(config, 'egress', name)
-      if (!window.confirm(`Delete egress “${name}”?${refWarning(refs)}`)) return
+      const ok = await confirm({
+        label: 'Delete egress',
+        title: `Delete egress “${name}”?`,
+        body: refWarning(refs),
+      })
+      if (!ok) return
      void save({ ...config, Egresses: egresses.filter((e) => e.Name !== name) }, `Deleted ${name}`)
    },
-    [config, egresses, save],
+    [config, egresses, save, confirm],
  )

  const busy = saving || applying
@@ -0,0 +1,150 @@
+// protectionState — the one sentence the whole panel shows about "am I protected".
+//
+// Run with `npm test` (node's built-in test runner + native TypeScript stripping;
+// no test dependency is added to the SPA, which ships inside the daemon binary).
+//
+// The case this file was written for is "plane full, traffic direct": the exact
+// state of a live router — one enabled rule, `default → direct`, no groups, no
+// rule-sets — where every part of the data plane was installed and the readout
+// therefore said "Protected — traffic from your network is going through the
+// tunnel", under a green LED, while the whole LAN went out the plain WAN.
+//
+// planeState.ts has no runtime imports (both of its imports are `import type`),
+// so this runs against the real module with nothing stubbed.
+
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+
+import { protectionState } from './planeState.ts'
+import type { Status, Traffic } from './api.ts'
+
+/** A healthy, fully-installed router; `traffic` is what each case varies. */
+function status(over: Partial<Status> = {}): Status {
+  return {
+    running: true,
+    enabled: true,
+    active: true,
+    table: true,
+    hash: 'abc',
+    version: '1.11.0-shater',
+    kill_switch: 'closed',
+    engine_running: true,
+    plane: 'full',
+    warnings: [],
+    ...over,
+  }
+}
+
+function withTraffic(traffic: Traffic | undefined): Status {
+  return status({ traffic })
+}
+
+// --- the field case ---------------------------------------------------------
+
+test('plane full + default direct is NOT reported as protected', () => {
+  const s = protectionState(withTraffic({ verdict: 'direct', default: 'direct', tunnel_rules: 0 }))
+  assert.notEqual(s.headline, 'Protected')
+  assert.equal(s.variant, 'crit')
+  assert.equal(s.alarm, true)
+  // The claim that was false must not survive anywhere in the copy.
+  assert.doesNotMatch(s.detail, /going through the tunnel/)
+  // ...and the honest consequence must be stated, not implied.
+  assert.match(s.detail, /real address/)
+})
+
+// --- the other verdicts under a full plane ----------------------------------
+
+test('plane full + default into a tunnel is protected', () => {
+  const s = protectionState(withTraffic({ verdict: 'tunnel', default: 'auto', tunnel_rules: 1 }))
+  assert.equal(s.variant, 'on')
+  assert.equal(s.headline, 'Protected')
+  assert.equal(s.alarm, false)
+})
+
+test('plane full + direct default with tunnelling rules is split, not protected', () => {
+  const s = protectionState(withTraffic({ verdict: 'split', default: 'direct', tunnel_rules: 3 }))
+  assert.equal(s.variant, 'amber')
+  assert.notEqual(s.headline, 'Protected')
+  // Says how much is protected, and that the default is not.
+  assert.match(s.detail, /3 rules/)
+  assert.match(s.detail, /normal internet connection/)
+  // A working selective setup must not raise a banner on every other page.
+  assert.equal(s.alarm, false)
+})
+
+test('split names a single rule in the singular', () => {
+  const s = protectionState(withTraffic({ verdict: 'split', default: 'direct', tunnel_rules: 1 }))
+  assert.match(s.detail, /^One rule sends traffic/)
+})
+
+test('plane full + blocked default with rules leaks nothing and is never crit', () => {
+  const s = protectionState(withTraffic({ verdict: 'blocked', default: 'block', tunnel_rules: 2 }))
+  assert.equal(s.variant, 'amber')
+  assert.equal(s.alarm, false)
+  assert.match(s.detail, /nothing is leaving unprotected/)
+})
+
+test('plane full + blocked default with no rules says the network has no way out', () => {
+  const s = protectionState(withTraffic({ verdict: 'blocked', default: 'block', tunnel_rules: 0 }))
+  assert.equal(s.variant, 'amber')
+  assert.equal(s.alarm, true)
+  assert.doesNotMatch(s.detail, /going through the tunnel/)
+})
+
+test('plane full with no verdict claims nothing either way', () => {
+  for (const t of [undefined, { verdict: '' as const }]) {
+    const s = protectionState(withTraffic(t))
+    assert.notEqual(s.headline, 'Protected')
+    assert.equal(s.variant, 'amber')
+    assert.equal(s.alarm, false)
+  }
+})
+
+// --- the branches that were already correct ---------------------------------
+
+test('no status yet', () => {
+  const s = protectionState(null)
+  assert.equal(s.variant, 'off')
+  assert.equal(s.alarm, false)
+})
+
+test('service switched off is a deliberate state, not a fault', () => {
+  const s = protectionState(status({ enabled: false }))
+  assert.equal(s.variant, 'amber')
+  assert.equal(s.headline, 'Turned off')
+  assert.equal(s.alarm, false)
+})
+
+test('hold: the kill-switch caught it — protected, offline', () => {
+  const s = protectionState(status({ plane: 'hold', engine_running: false, active: false }))
+  assert.equal(s.variant, 'amber')
+  assert.equal(s.alarm, true)
+  assert.match(s.headline, /blocked/)
+})
+
+test('none + fail-closed is the leak, and it is crit', () => {
+  const s = protectionState(status({ plane: 'none', table: false, engine_running: false }))
+  assert.equal(s.variant, 'crit')
+  assert.equal(s.alarm, true)
+})
+
+test('none + fail-open is the operator’s documented choice, stated not alarmed at', () => {
+  const s = protectionState(
+    status({ plane: 'none', table: false, engine_running: false, kill_switch: 'open' }),
+  )
+  assert.equal(s.variant, 'amber')
+  assert.equal(s.alarm, true)
+})
+
+test('daemon too old to send `plane` keeps its own fallback', () => {
+  // Nothing here may depend on `traffic`: a daemon with no `plane` has no
+  // `traffic` either, and this branch reads what it can observe instead.
+  const { plane, ...noPlane } = status()
+  void plane
+  assert.equal(protectionState(noPlane as Status).headline, 'Protected')
+  assert.equal(protectionState({ ...noPlane, running: false } as Status).headline, 'Service stopped')
+  assert.equal(
+    protectionState({ ...noPlane, active: false } as Status).headline,
+    'Starting up',
+  )
+})
@@ -11,7 +11,7 @@
 // same router differently.

 import type { LedVariant } from './components'
-import type { Status } from './api'
+import type { Status, Traffic } from './api'

 export interface ProtectionState {
  variant: LedVariant
@@ -31,6 +31,24 @@ export interface ProtectionState {
 *   hold — the kill-switch caught it. Protected, but offline.
 *   none (fail-closed) — there is no protection at all. Online, and exposed.
 * Collapsing them would erase the only difference that matters.
+ *
+ * PLANE IS NOT THE WHOLE ANSWER, AND THAT USED TO BE A LIE. `plane: 'full'`
+ * returned "Protected — traffic from your network is going through the tunnel",
+ * which is a claim `plane` cannot support: it only says the nft table, the policy
+ * routing and the engine are all installed. Where the diverted packets go once the
+ * engine has them is decided by the engine's default route, and a router in the
+ * field ran with one rule — `default → direct`, no groups, no rule-sets. Fully
+ * installed plane, zero tunnel, whole LAN out the plain WAN with its real address,
+ * green LED, "Protected".
+ *
+ * So `full` now branches on `status.traffic`, the daemon's verdict on the config
+ * it is actually running (see api.ts TrafficVerdict). It is computed on the daemon
+ * because only the daemon knows what was GENERATED and STARTED: the panel's
+ * /api/config is desired state, which diverges from the running one whenever edits
+ * are unapplied or a rollback is pending, and re-deriving the default route from it
+ * would mean a second implementation of the generator's rule loop — schedules,
+ * shadowed catch-alls, targets that failed to resolve and fell back to direct —
+ * drifting against the first.
 */
 export function protectionState(status: Status | null): ProtectionState {
  if (!status) {
@@ -56,12 +74,7 @@ export function protectionState(status: Status | null): ProtectionState {

  switch (status.plane) {
    case 'full':
-      return {
-        variant: 'on',
-        headline: 'Protected',
-        detail: 'Traffic from your network is going through the tunnel.',
-        alarm: false,
-      }
+      return fullPlaneState(status.traffic)
    case 'hold':
      return {
        variant: 'amber',
@@ -112,3 +125,87 @@ export function protectionState(status: Status | null): ProtectionState {
    alarm: false,
  }
 }
+
+/**
+ * The plane is fully installed — now say where the traffic it carries ends up.
+ *
+ * Only `tunnel` earns "Protected". The other verdicts each describe a real router
+ * someone can be sitting in front of, and they are kept apart because the thing to
+ * DO about them differs:
+ *
+ *   split   — deliberate for most people who reach it, accidental for the rest
+ *             (a default rule that was never pointed anywhere). Amber, but no
+ *             alarm: raising a banner on every page of a working selective setup
+ *             is how a banner stops being read.
+ *   direct  — the engine is running and forwarding every connection out the plain
+ *             WAN. The traffic outcome is identical to `plane: 'none'` under a
+ *             closed kill-switch, so it gets the same weight: crit, and it
+ *             interrupts. The wording differs because the fix does — nothing
+ *             failed here, the routing simply says "direct".
+ *   blocked — the fail-closed default. Nothing is leaking, so this is never crit;
+ *             with no tunnelling rules at all it means the network has no way out
+ *             and someone should be told why.
+ *   unknown — an older daemon, or the seconds between this daemon starting and its
+ *             first apply. We do not know, so we do not claim. Saying "Protected"
+ *             here is the exact bug being removed.
+ */
+function fullPlaneState(traffic: Traffic | undefined): ProtectionState {
+  const tunnelRules = traffic?.tunnel_rules ?? 0
+
+  switch (traffic?.verdict) {
+    case 'tunnel':
+      return {
+        variant: 'on',
+        headline: 'Protected',
+        detail: 'Traffic from your network is going through the tunnel.',
+        alarm: false,
+      }
+    case 'split':
+      return {
+        variant: 'amber',
+        headline: 'Partly protected — the rest goes out directly',
+        detail: `${ruleCount(tunnelRules)} through the tunnel. Everything they don’t match leaves through your normal internet connection, with your real address.`,
+        alarm: false,
+      }
+    case 'direct':
+      return {
+        variant: 'crit',
+        headline: 'Not protected — nothing is going through the tunnel',
+        detail:
+          'The service is running, but your routing sends every connection straight out your normal internet connection, with your real address. On the Routing page, point the default rule at a group or a node.',
+        alarm: true,
+      }
+    case 'blocked':
+      return tunnelRules > 0
+        ? {
+            variant: 'amber',
+            headline: 'Partly protected — everything else is blocked',
+            detail: `${ruleCount(tunnelRules)} through the tunnel. Anything they don’t match is blocked instead of being let out, so nothing is leaving unprotected.`,
+            alarm: false,
+          }
+        : {
+            variant: 'amber',
+            headline: 'Nothing is getting out',
+            detail:
+              'No rule sends traffic anywhere, so every connection from your network is being blocked rather than let out unprotected. Add a default rule on the Routing page.',
+            alarm: true,
+          }
+    default:
+      return {
+        variant: 'amber',
+        headline: 'Checking where traffic goes',
+        detail:
+          'The router is up and handling your traffic. It hasn’t reported yet whether that traffic is going through the tunnel.',
+        alarm: false,
+      }
+  }
+}
+
+/** "One rule sends traffic" / "4 rules send traffic", so the detail lines above
+ *  can name a number the operator can go and count on the Routing page. Falls
+ *  back to the vague form only if the daemon sent a verdict without a count. */
+function ruleCount(n: number): string {
+  if (n <= 0) return 'Some traffic goes'
+  if (n === 1) return 'One rule sends traffic'
+  return `${n} rules send traffic`
+}
@@ -18,5 +18,10 @@
    "noFallthroughCasesInSwitch": true,
    "forceConsistentCasingInFileNames": true
  },
-  "include": ["src", "vite.config.ts"]
+  "include": ["src", "vite.config.ts"],
+  // *.test.ts runs under node's built-in test runner (`npm test`), which strips
+  // types rather than checking them. They are excluded here because they import
+  // node:test / node:assert, and the SPA deliberately carries no @types/node — it
+  // is embedded in the daemon binary, so every devDependency is weight on a router.
+  "exclude": ["src/**/*.test.ts"]
 }
@@ -16,6 +16,7 @@ import (
 	"github.com/sagernet/sing/common"
 	"github.com/sagernet/sing/common/buf"
 	"github.com/sagernet/sing/common/control"
+	E "github.com/sagernet/sing/common/exceptions" // lx:tproxy_writeback_connect
 	M "github.com/sagernet/sing/common/metadata"
 	N "github.com/sagernet/sing/common/network"
 	"github.com/sagernet/sing/common/udpnat2"
@@ -68,7 +69,26 @@ func (t *TProxy) Start(stage adapter.StartStage) error {
 }

 func (t *TProxy) Close() error {
-	return t.listener.Close()
+	err := t.listener.Close()
+	// lx:begin tproxy_writeback_connect
+	// Closing the listener stops INGRESS but leaves every live UDP NAT session in
+	// the cache, and each session holds a write-back socket bound to its original
+	// destination. Nothing else ever wakes those sessions: the cache evicts
+	// lazily (on the next Get/Add), and after this inbound is gone there is no
+	// next Get. For a DNS session answered in-engine by hijack-dns there is not
+	// even an outbound connection whose failure could unwind it, so its socket
+	// survives until a GC finalizer happens to reach it.
+	//
+	// That is invisible upstream, where an inbound is closed once at shutdown. On
+	// this fork the engine is REBUILT on every config apply, so each apply would
+	// strand another generation of sockets on the router's own LAN :53. Purge
+	// evicts every session; the cache's OnEvict closes the conn, which unblocks
+	// the session's routing goroutine and runs its onClose — the one place the
+	// write-back socket is actually closed. Purge AFTER the listener so a packet
+	// arriving mid-teardown cannot re-create a session behind us.
+	t.udpNat.Purge()
+	// lx:end tproxy_writeback_connect
+	return err
 }

 func (t *TProxy) NewConnection(ctx context.Context, conn net.Conn, metadata adapter.InboundContext, onClose N.CloseHandlerFunc) {
@@ -121,18 +141,92 @@ type tproxyPacketWriter struct {
 	conn        *net.UDPConn
 }

+// lx:begin tproxy_writeback_connect
+//
+// The TPROXY UDP write-back socket is bound to the ORIGINAL DESTINATION, so the
+// client sees the reply coming from the address it addressed. Upstream leaves
+// that socket UNCONNECTED (net.ListenPacket + WriteToUDPAddrPort), and that is
+// the bug this block exists for.
+//
+// An unconnected socket bound to <addr>:<port> is, as far as the kernel is
+// concerned, a RECEIVER for that address:port — and because the socket also
+// carries SO_REUSEADDR it silently joins the UDP demultiplex set of whatever
+// else is bound there. Nothing ever reads from it: this writer only sends. So
+// every datagram the kernel happens to hand it is lost.
+//
+// On a router that transparently intercepts LAN DNS towards its OWN address
+// (`nft ... udp dport 53 tproxy ...`) the original destination IS the router's
+// LAN address, so each intercepted DNS session parks another silent receiver on
+// <router-lan-ip>:53 right next to dnsmasq's socket — observed on the stand:
+// 33 such sockets against dnsmasq's one, several with a growing Recv-Q. The
+// host's own queries to that address take the loopback path, are never diverted,
+// and are therefore demultiplexed among all of them: they land in one of the
+// silent sockets at random and time out, permanently and unpredictably, while
+// every other DNS path on the box keeps working.
+//
+// Connecting the socket fixes it at the root. compute_score() in the kernel's
+// UDP lookup REJECTS a connected socket for any peer other than the connected
+// one, so a write-back socket can no longer be handed a datagram it will not
+// read. Nothing about the reply changes — same spoofed source address, same
+// single peer, one Write instead of one WriteTo.
+//
+// The unconnected path is kept verbatim for a destination that cannot be bound
+// (a domain socksaddr), so no existing case regresses.
+
+// newTProxyWriteBack creates the connected write-back socket: local address =
+// the original destination (transparent bind), peer = the client. A package
+// variable so the behaviour can be tested without CAP_NET_ADMIN.
+var newTProxyWriteBack = func(w *tproxyPacketWriter, destination M.Socksaddr) (*net.UDPConn, error) {
+	var dialer net.Dialer
+	dialer.LocalAddr = destination.UDPAddr()
+	dialer.Control = control.Append(dialer.Control, control.ReuseAddr())
+	dialer.Control = control.Append(dialer.Control, redir.TProxyWriteBack())
+	conn, err := w.listener.DialContext(dialer, w.ctx, "udp", w.source.String())
+	if err != nil {
+		return nil, err
+	}
+	udpConn, loaded := conn.(*net.UDPConn)
+	if !loaded {
+		conn.Close()
+		return nil, E.New("tproxy write back: unexpected connection type ", conn)
+	}
+	return udpConn, nil
+}
+
+// lx:end tproxy_writeback_connect
+
 func (w *tproxyPacketWriter) WritePacket(buffer *buf.Buffer, destination M.Socksaddr) error {
 	defer buffer.Release()
 	if w.listener.ListenOptions().NetNs == "" {
 		conn := w.conn
 		if w.destination == destination && conn != nil {
-			_, err := conn.WriteToUDPAddrPort(buffer.Bytes(), w.source)
+			// lx:begin tproxy_writeback_connect
+			// The cached socket is CONNECTED to the client, so this is a plain
+			// Write. A failed write also closes it: upstream only dropped the
+			// reference, leaving the fd to the GC finalizer.
+			_, err := conn.Write(buffer.Bytes())
 			if err != nil {
+				conn.Close()
 				w.conn = nil
 			}
 			return err
+			// lx:end tproxy_writeback_connect
 		}
 	}
+	// lx:begin tproxy_writeback_connect
+	if destination.IsIP() {
+		udpConn, err := newTProxyWriteBack(w, destination)
+		if err != nil {
+			return err
+		}
+		if w.listener.ListenOptions().NetNs == "" && w.destination == destination {
+			w.conn = udpConn
+		} else {
+			defer udpConn.Close()
+		}
+		return common.Error(udpConn.Write(buffer.Bytes()))
+	}
+	// lx:end tproxy_writeback_connect
 	var listenConfig net.ListenConfig
 	listenConfig.Control = control.Append(listenConfig.Control, control.ReuseAddr())
 	listenConfig.Control = control.Append(listenConfig.Control, redir.TProxyWriteBack())
@@ -0,0 +1,234 @@
+package redirect
+
+// Regression cover for the lx:tproxy_writeback_connect block in tproxy.go.
+//
+// The failure it guards against, seen on a live BPi-R3 Mini: `shaterd` held 33
+// UNCONNECTED UDP sockets on the router's own LAN address :53 — the write-back
+// sockets of intercepted DNS sessions — alongside dnsmasq's single socket on the
+// same address:port. Nothing reads a write-back socket, so every host-originated
+// query that the kernel demultiplexed into one of them was silently dropped
+// (Recv-Q climbing, sender timing out). Connecting the socket to the one peer it
+// ever talks to removes it from the demultiplex set for everybody else.
+//
+// The real socket needs CAP_NET_ADMIN (IP_TRANSPARENT) and a foreign bind, so
+// these tests drive the seam (newTProxyWriteBack) with an ordinary connected
+// loopback socket. That is enough to pin both properties that actually broke:
+//
+//   - the write path uses the CONNECTED form. If WritePacket ever goes back to
+//     WriteToUDPAddrPort, Go returns ErrWriteToConnected on a connected socket
+//     and these tests fail — i.e. the test cannot pass with an unconnected
+//     write-back socket, which is exactly the regression.
+//   - repeated writes to the same destination REUSE one socket instead of
+//     accumulating a new one per packet.
+
+import (
+	"context"
+	"net"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/sagernet/sing-box/common/listener"
+	"github.com/sagernet/sing-box/option"
+
+	"github.com/sagernet/sing/common/buf"
+	M "github.com/sagernet/sing/common/metadata"
+	N "github.com/sagernet/sing/common/network"
+	"github.com/sagernet/sing/common/udpnat2"
+)
+
+// writeBackHarness stands up a loopback "client" socket and a tproxyPacketWriter
+// whose socket factory returns a plain connected UDP socket aimed at it. It
+// returns the writer, the client socket and a pointer to the factory call count.
+func writeBackHarness(t *testing.T) (*tproxyPacketWriter, *net.UDPConn, *int) {
+	t.Helper()
+
+	client, err := net.ListenUDP("udp", &net.UDPAddr{IP: net.IPv4(127, 0, 0, 1)})
+	if err != nil {
+		t.Fatalf("client socket: %v", err)
+	}
+	t.Cleanup(func() { client.Close() })
+
+	source := netip.MustParseAddrPort(client.LocalAddr().String())
+	w := &tproxyPacketWriter{
+		ctx:    context.Background(),
+		source: source,
+		// A listener with empty options is enough: WritePacket only reads
+		// ListenOptions().NetNs, and the factory is stubbed below.
+		listener: listener.New(listener.Options{
+			Context: context.Background(),
+			Listen:  option.ListenOptions{},
+		}),
+		destination: M.SocksaddrFrom(netip.MustParseAddr("127.0.0.1"), 53),
+	}
+
+	calls := 0
+	orig := newTProxyWriteBack
+	t.Cleanup(func() { newTProxyWriteBack = orig })
+	newTProxyWriteBack = func(w *tproxyPacketWriter, destination M.Socksaddr) (*net.UDPConn, error) {
+		calls++
+		// The real implementation binds the ORIGINAL DESTINATION transparently
+		// and connects to w.source; here we only reproduce the connect half,
+		// which is the property under test.
+		conn, err := net.DialUDP("udp", nil, net.UDPAddrFromAddrPort(w.source))
+		if err != nil {
+			return nil, err
+		}
+		return conn, nil
+	}
+	return w, client, &calls
+}
+
+// readOne reads one datagram from the client socket with a short deadline.
+func readOne(t *testing.T, client *net.UDPConn) string {
+	t.Helper()
+	_ = client.SetReadDeadline(time.Now().Add(2 * time.Second))
+	b := make([]byte, 512)
+	n, _, err := client.ReadFrom(b)
+	if err != nil {
+		t.Fatalf("client read: %v", err)
+	}
+	return string(b[:n])
+}
+
+// TestWriteBackUsesConnectedSocket: the reply must go out over a CONNECTED
+// socket. On the pre-fix code the write is WriteToUDPAddrPort, which Go refuses
+// on a connected socket ("use of WriteTo with pre-connected connection"), so
+// this test is red for exactly the shape that caused the outage.
+func TestWriteBackUsesConnectedSocket(t *testing.T) {
+	w, client, calls := writeBackHarness(t)
+
+	if err := w.WritePacket(buf.As([]byte("first")).ToOwned(), w.destination); err != nil {
+		t.Fatalf("WritePacket: %v", err)
+	}
+	if got := readOne(t, client); got != "first" {
+		t.Fatalf("client got %q, want %q", got, "first")
+	}
+	if *calls != 1 {
+		t.Fatalf("expected one write-back socket, got %d", *calls)
+	}
+}
+
+// TestWriteBackReusesOneSocket: a session that keeps answering the same client
+// must keep ONE socket, not open a fresh one per datagram. The live box carried
+// one socket per intercepted DNS session already; one per PACKET would turn a
+// nuisance into an fd exhaustion.
+func TestWriteBackReusesOneSocket(t *testing.T) {
+	w, client, calls := writeBackHarness(t)
+
+	for i, payload := range []string{"a", "b", "c", "d"} {
+		if err := w.WritePacket(buf.As([]byte(payload)).ToOwned(), w.destination); err != nil {
+			t.Fatalf("WritePacket %d: %v", i, err)
+		}
+		if got := readOne(t, client); got != payload {
+			t.Fatalf("packet %d: client got %q, want %q", i, got, payload)
+		}
+	}
+	if *calls != 1 {
+		t.Fatalf("four packets to one destination must share one socket, got %d sockets", *calls)
+	}
+	if w.conn == nil {
+		t.Fatalf("the write-back socket must be cached on the writer for reuse")
+	}
+}
+
+// TestWriteBackClosesSocketOnWriteFailure: upstream dropped the reference to a
+// failed socket without closing it, leaving the fd to the GC finalizer. On a box
+// that already parks one socket per DNS session that is the wrong direction.
+func TestWriteBackClosesSocketOnWriteFailure(t *testing.T) {
+	w, client, _ := writeBackHarness(t)
+
+	if err := w.WritePacket(buf.As([]byte("warm")).ToOwned(), w.destination); err != nil {
+		t.Fatalf("WritePacket: %v", err)
+	}
+	readOne(t, client)
+
+	cached := w.conn
+	if cached == nil {
+		t.Fatalf("expected a cached socket after the first write")
+	}
+	// Close it behind WritePacket's back so the next write fails, exactly as a
+	// dead peer or a torn-down plane would make it fail.
+	cached.Close()
+
+	if err := w.WritePacket(buf.As([]byte("boom")).ToOwned(), w.destination); err == nil {
+		t.Fatalf("a write on a closed socket must report the failure")
+	}
+	if w.conn != nil {
+		t.Fatalf("a failed write must drop the cached socket")
+	}
+	// A second Close on an already-closed conn is an error, which is how we know
+	// WritePacket closed it rather than merely forgetting it.
+	if err := cached.Close(); err == nil {
+		t.Fatalf("WritePacket must CLOSE the failed socket, not just nil the field")
+	}
+}
+
+// --- Close() must release the UDP NAT sessions --------------------------------
+
+// natSessionHandler stands in for the router: it drains the session conn until
+// it errors (which is what Close does to it) and then reports onClose, exactly
+// as the real routing goroutine does. onClose is where the write-back socket is
+// closed, so "onClose fired" is the observable proof the socket was released.
+type natSessionHandler struct {
+	released chan error
+}
+
+func (h *natSessionHandler) NewPacketConnectionEx(ctx context.Context, conn N.PacketConn, source M.Socksaddr, destination M.Socksaddr, onClose N.CloseHandlerFunc) {
+	go func() {
+		for {
+			buffer := buf.NewSize(1024)
+			_, err := conn.ReadPacket(buffer)
+			buffer.Release()
+			if err != nil {
+				if onClose != nil {
+					onClose(err)
+				}
+				h.released <- err
+				return
+			}
+		}
+	}()
+}
+
+// TestTProxyCloseReleasesNatSessions pins the apply-level half of the leak.
+//
+// Closing the inbound used to close the listener only. The NAT cache evicts
+// lazily, so after the inbound is gone nothing ever touches it again and every
+// live session — with the write-back socket it holds on the router's own
+// LAN :53 — was left to a GC finalizer. On this fork the engine is rebuilt on
+// every config apply, so that is one stranded generation of sockets per apply.
+func TestTProxyCloseReleasesNatSessions(t *testing.T) {
+	handler := &natSessionHandler{released: make(chan error, 1)}
+	tp := &TProxy{
+		ctx: context.Background(),
+		listener: listener.New(listener.Options{
+			Context: context.Background(),
+			Listen:  option.ListenOptions{},
+		}),
+	}
+	prepared := 0
+	tp.udpNat = udpnat.New(handler, func(source M.Socksaddr, destination M.Socksaddr, userData any) (bool, context.Context, N.PacketWriter, N.CloseHandlerFunc) {
+		prepared++
+		return true, context.Background(), nil, func(error) {}
+	}, time.Minute, false)
+
+	tp.udpNat.NewPacket(
+		[][]byte{{0x00}},
+		M.SocksaddrFrom(netip.MustParseAddr("10.67.0.2"), 40000),
+		M.SocksaddrFrom(netip.MustParseAddr("10.67.0.1"), 53),
+		nil,
+	)
+	if prepared != 1 {
+		t.Fatalf("expected one NAT session, got %d", prepared)
+	}
+
+	if err := tp.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	select {
+	case <-handler.released:
+	case <-time.After(5 * time.Second):
+		t.Fatal("Close must release every live NAT session (and with it the write-back socket it holds)")
+	}
+}
@@ -18,17 +18,21 @@
 #      OpenWrt package can $(INSTALL_BIN) the arch-matched artifact.
 #   5. Prints a size table + a per-arch static check (ELF type / no PT_INTERP).
 #
-# Router build tag set = D9 (musl-static). We deliberately DROP with_purego and
-# with_naive_outbound: they pull cronet-go, which forces a glibc PT_INTERP even
-# with CGO_ENABLED=0, making the binary unusable on musl OpenWrt.
+# Router build tag set = D9/D23 (musl-static). It is DEFINED IN, and only in,
+# scripts/router-tags.sh (sourced below) — that file documents every tag and is
+# machine-checked against the declared feature list by shater/buildtags's test.
+# Run scripts/check-router-tags.sh after touching it.
 #
 # Usage:
 #   scripts/build-shaterd.sh [VERSION] [--fast]
 #
 #   VERSION   Version string stamped into constant.Version. Resolution order:
 #               1) this positional arg, if given
-#               2) $SHATER_VERSION, if set
-#               3) `git describe --tags` (nearest tag + commit)
+#               2) $SHATER_VERSION, if set (CI sets it from ci/version.sh)
+#               3) `ci/version.sh --binary` — THE single source of truth shared
+#                  with the package version (vX.Y.Z-rR[-g<sha>], derived from
+#                  the git tag exactly like PKG_VERSION/PKG_RELEASE), so the
+#                  string the panel shows always matches `apk info shaterd`
 #               4) fallback: v0.2.0-dev
 #   --fast    Skip `npm ci` when panel/node_modules already exists (dev speed-up).
 #
@@ -69,24 +73,25 @@ if [ -n "$VERSION_ARG" ]; then
 	VERSION="$VERSION_ARG"
 elif [ -n "${SHATER_VERSION:-}" ]; then
 	VERSION="$SHATER_VERSION"
-elif VERSION="$(git -C "$REPO" describe --tags 2>/dev/null)"; then
-	: # git describe succeeded
+elif VERSION="$(sh "$REPO/ci/version.sh" --binary)"; then
+	# Same computation the PACKAGE version comes from (ci/version.sh), so the
+	# binary's constant.Version and the .ipk/.apk version can never drift apart.
+	: # ci/version.sh always succeeds (it falls back to 0.0.0 without git)
 else
 	VERSION="v0.2.0-dev"
 fi
 [ -n "$VERSION" ] || VERSION="v0.2.0-dev"

 # --- config -----------------------------------------------------------------
-# D9 router tag set (musl-static). Keep in sync with docs-shater/DECISIONS.md D9.
-# No with_gvisor: the data plane is tproxy/redirect (netplane), generate never
-# emits a tun inbound, so the userspace gvisor stack was 3.6 MB of dead weight
-# (tun would fall back to the system stack anyway).
-# No with_clash_api: the panel is shater's own; generate never emits a clash_api
-# service ("the shater generator emits none of those" — shater/engine/engine.go).
-# No with_dhcp: shater resolvers are udp/tcp/doh/dot/local/fakeip — no "dhcp://"
-# DNS transport is ever generated, and the slim registry never registers it.
-ROUTER_TAGS="with_quic,with_wireguard,with_utls,badlinkname,tfogo_checklinkname0,with_xhttp,with_awg,with_lx_command"
-LDFLAGS="-X github.com/sagernet/sing-box/constant.Version=${VERSION} -checklinkname=0 -s -w -buildid="
+# D9/D23 router tag set (musl-static). The set itself lives in ONE place —
+# scripts/router-tags.sh — because it is also parsed by shater/buildtags's test,
+# which proves it still covers every feature docs-shater/FEATURES.md declares.
+# Do not re-inline it here: that split is exactly how `with_gvisor` went missing
+# while `with_wireguard` stayed (D23).
+# shellcheck source=router-tags.sh
+. "$SCRIPT_DIR/router-tags.sh"
+ROUTER_TAGS="$SHATER_ROUTER_TAGS"
+LDFLAGS="-X github.com/sagernet/sing-box/constant.Version=${VERSION} ${SHATER_ROUTER_LDFLAGS} -s -w -buildid="

 UPX_BIN="${UPX:-upx}"
 # UPX itself treats the environment variable UPX as extra command-line options, so
@@ -106,6 +111,20 @@ echo "  version : $VERSION"
 echo "  tags    : $ROUTER_TAGS"
 echo "  upx     : $UPX_BIN"
 echo "  go      : $(go version)"
+
+# --- gate: does this tag set still support what we declare? (D23) ------------
+# Cheap (one tiny tag-less package, no network, ~1 s) and it travels with the
+# BUILD rather than with a CI config, so an artifact produced by hand on a
+# developer's machine gets the same guarantee. The heavier half — actually
+# constructing every declared protocol under these tags — is
+# scripts/check-router-tags.sh, which CI runs before this script.
+if ! (cd "$REPO" && go test -count=1 ./shater/buildtags/ >/dev/null); then
+	echo >&2
+	echo "  ABORT: the router tag set no longer covers a declared feature." >&2
+	echo "         Details: go test ./shater/buildtags/" >&2
+	echo "         Full check: scripts/check-router-tags.sh" >&2
+	exit 1
+fi
 echo

 # --- step 1: build the SPA --------------------------------------------------
@@ -0,0 +1,118 @@
+#!/usr/bin/env bash
+#
+# check-router-tags.sh — prove the SHIPPED build-tag set still supports every
+# feature shater declares (D23).
+#
+# WHY (2026-07-25): the router tag set is a trimmed subset of upstream's, but the
+# test suite builds with the FULL upstream set — so the one combination we
+# actually ship was never exercised. `with_gvisor` got trimmed while
+# `with_wireguard` stayed, and every shipped binary answered a WireGuard node
+# with "gVisor is not included in this build". Compiling is not evidence.
+#
+# WHAT IT RUNS
+#   1. shater/buildtags, TAG-LESS — reads scripts/router-tags.sh and fails if a
+#      declared feature (buildtags.Features) lost a build tag it needs. Cheap,
+#      hostable anywhere, catches the trim at the moment it happens.
+#   2. shater/generate + shater/buildtags, WITH THE SHIPPED TAG SET on linux —
+#      constructs one node of every declared protocol through box.New+Start, and
+#      cross-checks that the tag detectors match the set the compiler was given.
+#      This is the half that catches "the tag is there but insufficient".
+#
+# The run is unprivileged (no tproxy inbound is built) and offline apart from Go
+# module downloads.
+#
+# Usage:
+#   scripts/check-router-tags.sh
+#
+# Env:
+#   SHATER_GO_IMAGE   docker image used to reach linux from a non-linux host
+#                     (default golang:1.26 — keep it >= go.mod's toolchain).
+#   SHATER_NO_DOCKER=1  fail instead of falling back to docker.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO="$(cd "$SCRIPT_DIR/.." && pwd)"
+cd "$REPO"
+
+# shellcheck source=router-tags.sh
+. "$SCRIPT_DIR/router-tags.sh"
+
+echo "== router build-tag check =="
+echo "  tags   : $SHATER_ROUTER_TAGS"
+echo "  ldflags: $SHATER_ROUTER_LDFLAGS"
+echo
+
+# --- 1. static: does the set still cover the declared features? --------------
+# No tags, no OS constraint: this is the check that would have caught the outage
+# on the developer's own machine.
+echo "== [1/2] declared features vs. the shipped tag set (no tags needed) =="
+go test -count=1 ./shater/buildtags/
+echo
+
+# --- 2. behavioural: does the shipped combination actually construct? --------
+# The protocol-construction test is linux-only (box.New validates the loop-guard
+# routing_mark only there). From a non-linux host, re-exec this script inside a
+# golang container rather than silently skipping — a skipped guard is no guard.
+if [ "$(go env GOOS)" != "linux" ] && [ "${SHATER_TAGCHECK_IN_DOCKER:-0}" != "1" ]; then
+	if [ "${SHATER_NO_DOCKER:-0}" = "1" ] || ! command -v docker >/dev/null 2>&1; then
+		echo "  ERROR: step 2 needs linux (GOOS=$(go env GOOS)) and docker is unavailable/disabled." >&2
+		echo "         Run this script on the linux CI runner or the OpenWrt VM." >&2
+		exit 1
+	fi
+	image="${SHATER_GO_IMAGE:-golang:1.26}"
+	echo "== [2/2] re-exec on linux via docker ($image) =="
+	host_repo="$REPO"
+	command -v cygpath >/dev/null 2>&1 && host_repo="$(cygpath -w "$REPO")"
+	# Named volumes keep the module/build cache warm between runs; MSYS2_ARG_CONV_EXCL
+	# stops Git Bash from rewriting the container-side paths into windows ones.
+	MSYS2_ARG_CONV_EXCL='*' MSYS_NO_PATHCONV=1 docker run --rm \
+		-v "$host_repo":/src \
+		-v shater-tagcheck-gomod:/go/pkg/mod \
+		-v shater-tagcheck-gocache:/root/.cache/go-build \
+		-w /src \
+		-e SHATER_TAGCHECK_IN_DOCKER=1 \
+		"$image" bash scripts/check-router-tags.sh
+	exit $?
+fi
+
+echo "== [2/2] every declared protocol constructs under the SHIPPED tags =="
+out="$(mktemp)"
+trap 'rm -f "$out"' EXIT
+set +e
+SHATER_ROUTER_TAG_CHECK=1 go test -count=1 -v \
+	-tags "$SHATER_ROUTER_TAGS" -ldflags "$SHATER_ROUTER_LDFLAGS" \
+	-run 'TestShippedTagSetConstructsDeclaredProtocols|TestEveryTagGatedFeatureIsProbed' \
+	./shater/generate/ >"$out" 2>&1
+rc_gen=$?
+set -e
+sed 's/^/    /' "$out"
+
+set +e
+SHATER_ROUTER_TAG_CHECK=1 go test -count=1 -v \
+	-tags "$SHATER_ROUTER_TAGS" -ldflags "$SHATER_ROUTER_LDFLAGS" \
+	-run 'TestCompiledTagsMatchTheShippedSet' \
+	./shater/buildtags/ >"$out" 2>&1
+rc_tags=$?
+set -e
+sed 's/^/    /' "$out"
+
+if [ "$rc_gen" -ne 0 ] || [ "$rc_tags" -ne 0 ]; then
+	echo
+	echo "  FAILED: the tag set we SHIP cannot do what we declare." >&2
+	exit 1
+fi
+
+# A guard that silently runs nothing is worse than no guard: prove the tests were
+# actually compiled in and executed (build tags / file renames could exclude them).
+for want in TestShippedTagSetConstructsDeclaredProtocols TestCompiledTagsMatchTheShippedSet; do
+	if ! SHATER_ROUTER_TAG_CHECK=1 go test -count=1 -v \
+		-tags "$SHATER_ROUTER_TAGS" -ldflags "$SHATER_ROUTER_LDFLAGS" \
+		-run "$want" ./shater/generate/ ./shater/buildtags/ 2>&1 | grep -q -- "--- PASS: $want"; then
+		echo "  FAILED: $want did not run (build-tag/file-name drift?)" >&2
+		exit 1
+	fi
+done
+
+echo
+echo "== OK: the shipped tag set covers every declared feature, and every =="
+echo "==     declared protocol constructs through box.New under it.      =="
@@ -0,0 +1,78 @@
+# shellcheck shell=sh
+#
+# router-tags.sh — THE build-tag set of the shipped `shaterd` router binary.
+#
+# This file is DATA, not a program: it is `.`-sourced by
+#   - scripts/build-shaterd.sh     (the ship build)
+#   - scripts/check-router-tags.sh (the guard that proves the set is complete)
+# and it is PARSED by shater/buildtags/buildtags_test.go, which asserts that
+# every feature docs-shater/FEATURES.md declares supported has its build tags
+# present here. Change the set here and nowhere else.
+#
+# WHY A TRIMMED SET AT ALL (D9): upstream's DEFAULT_BUILD_TAGS registers the
+# whole sing-box zoo. We drop what shater/generate can never emit, because a
+# router binary pays for every tag twice — flash and (UPX unpacks into anonymous
+# pages) resident RAM. We do NOT drop what a declared feature needs to run.
+#
+# WHY THIS FILE EXISTS (the 2026-07-25 WireGuard outage): the set used to be a
+# string literal inside build-shaterd.sh, with nothing connecting it to the
+# feature list. `with_gvisor` was trimmed as "unreachable code" while
+# `with_wireguard` stayed — so every shipped binary answered a WireGuard node
+# with "gVisor is not included in this build". No test caught it: the test suite
+# builds with the FULL upstream tag set, so the SHIPPED combination was never
+# exercised. One file + one test now hold the two halves together (D23).
+#
+# ---- the set -----------------------------------------------------------------
+#   with_gvisor            userspace netstack. REQUIRED BY with_wireguard: both
+#                          transport/wireguard device constructors
+#                          (newStackDevice AND newSystemStackDevice) are stubs
+#                          returning tun.ErrGVisorNotIncluded without it, so
+#                          box.New dies at "create WireGuard device". Not
+#                          optional as long as we ship WireGuard/AmneziaWG.
+#   with_quic              hysteria2 + tuic outbounds, QUIC/HTTP3 DNS
+#                          transports, and the vless/vmess `type=quic` transport
+#                          (shater/registry/registry_quic.go).
+#   with_wireguard         registers the wireguard endpoint
+#                          (shater/registry/registry_wireguard.go).
+#   with_awg               AmneziaWG obfuscation params (jc/jmin/jmax, s1-s4,
+#                          h1-h4, i1-i5) actually reach the device
+#                          (transport/wireguard/device_awg.go). A driving
+#                          product requirement — FEATURES.md marks it [MVP].
+#   with_utls              uTLS fingerprints AND REALITY: common/tls/
+#                          reality_client.go is itself `//go:build with_utls`.
+#   with_xhttp             the XHTTP/SplitHTTP v2ray transport
+#                          (shater/parse emits type=xhttp).
+#   badlinkname            badtls fast path (common/badtls/*.go are
+#                          `go1.25 && badlinkname`). Needs -checklinkname=0 in
+#                          SHATER_ROUTER_LDFLAGS below or the LINK step fails.
+#   tfogo_checklinkname0   same deal for tfo-go's linkname use.
+#   with_lx_command        lx daemon command server. Inert for shaterd (nothing
+#                          under shater/ imports sing-box/daemon or libbox —
+#                          `go list -deps ./shater/cmd/shaterd` links neither),
+#                          kept only so the router set stays a subset of the lx
+#                          desktop set. Costs nothing; safe to drop later.
+#
+# Deliberately NOT here (each is unreachable for shater, not merely unused):
+#   with_purego,with_naive_outbound  cronet-go forces a glibc PT_INTERP even at
+#                                    CGO_ENABLED=0 -> will not run on musl (D9).
+#   with_clash_api                   the panel is shater's own web server;
+#                                    generate emits no clash_api service.
+#   with_dhcp                        resolver types are udp/tcp/doh/dot/local/
+#                                    fakeip; no dhcp:// transport is generated.
+#   with_tailscale,with_acme,with_ech,with_usbip,with_cloudflared,with_ocm,
+#   with_ccm,with_v2ray_api,with_reality_server
+#                                    nothing in shater/parse or shater/generate
+#                                    can produce them; hysteria2 `ech=` is
+#                                    refused with a warning in the parser.
+#
+# Adding a tag here is cheap. REMOVING one is a product decision: run
+# `scripts/check-router-tags.sh` — it fails if the set no longer covers a
+# declared feature, and it fails if the shipped combination cannot construct
+# every protocol through box.New.
+
+SHATER_ROUTER_TAGS="with_gvisor,with_quic,with_wireguard,with_utls,badlinkname,tfogo_checklinkname0,with_xhttp,with_awg,with_lx_command"
+
+# Linker flags the tag set REQUIRES (they are not optional trimming: `badlinkname`
+# without -checklinkname=0 fails at link time with
+# "invalid reference to crypto/tls.(*Conn).handlePostHandshakeMessage").
+SHATER_ROUTER_LDFLAGS="-checklinkname=0"
@@ -74,6 +74,19 @@ type Applier struct {
 	stateMu sync.RWMutex
 	holding bool

+	// traffic is WHERE THE TRAFFIC GOES under the config that is currently running:
+	// tunnelled, split, straight out, or blocked (see generate.TrafficOf). It is
+	// computed from the generated option.Options at the moment they are handed to
+	// the engine, so it describes what runs rather than what is on disk.
+	//
+	// Deliberately separate from `holding` and from Plane, which answer "how much of
+	// the data plane is installed". The panel used to read Plane == "full" as
+	// "protected" and said so under a green LED on a router whose only rule was
+	// `default -> direct`; the plane really was fully installed, and the whole LAN
+	// really was going out the plain WAN. Zero value = not known (no successful
+	// apply in this process yet), which is NOT the same as "tunnel".
+	traffic generate.Traffic // guarded by stateMu
+
 	// lastWarnings is the normalised warning set from the last SUCCESSFUL apply,
 	// published through Status so the panel can show fail-open degradations
 	// (a blocklist that did not load, a DoH host left reachable) instead of
@@ -488,6 +501,23 @@ func (a *Applier) applyLocked(m *model.Model) (bool, error) {
 		return changed, err
 	}

+	// The plane is COMPLETE only here: table + policy routing + sysctls. ApplyNft
+	// already flushed the DNS conntrack when it loaded the ruleset, but that is
+	// one step too early — ApplyRouting is idempotent BY del-then-add, so it opens
+	// a window in which the fwmark rule is momentarily absent, and any DNS flow
+	// that crosses that window is tracked against a plane that is still being
+	// assembled. Flushing once more now that every piece is in place is what makes
+	// "no entry survives the transition" actually true. Only on a real change (the
+	// fast path assembled nothing), best-effort, and cheap: the :53 entry count is
+	// bounded by the number of clients.
+	if !nftCurrent {
+		if n, ferr := netplane.FlushDNSConntrack(); ferr != nil {
+			a.log.Debug("flush DNS conntrack after plane change: ", ferr)
+		} else if n > 0 {
+			a.log.Debug("plane changed: dropped ", n, " stale DNS conntrack entries")
+		}
+	}
+
 	// (4) success. Bump the effective-state generation ONLY when something really
 	// moved: a no-op reconcile must not invalidate an armed commit-confirm window
 	// (cron reconciles every minute — counting those would cancel every rollback
@@ -497,6 +527,12 @@ func (a *Applier) applyLocked(m *model.Model) (bool, error) {
 	}
 	a.setHolding(false)
 	a.lastGood = m
+	// Publish where this config actually sends traffic, read off the very options
+	// the engine was just handed (a.eng.Apply above). The engine's hash fast-path
+	// may have skipped a swap, in which case these options are hash-equal to what is
+	// already running — either way they are the running config, which is the only
+	// config this verdict may describe.
+	a.setTraffic(generate.TrafficOf(opts))
 	// Publish the warnings of THIS successful apply, in one normalised set, and log
 	// them in one consistent format. Status carries them to the panel so a
 	// fail-open degradation is visible in the UI instead of only in logread.
@@ -540,6 +576,11 @@ func (a *Applier) applyLocked(m *model.Model) (bool, error) {
 // With the kill-switch OPEN nothing is installed — fail-open is the operator's
 // documented choice and this must not quietly override it.
 func (a *Applier) holdLocked(m *model.Model, cause error) {
+	// The engine is not carrying anything, so whatever the last running config did
+	// with traffic is no longer true of this router. Forget it either way — a stale
+	// "tunnel" verdict left behind by a config that is no longer running is the same
+	// reassuring lie in a different place.
+	a.setTraffic(generate.Traffic{})
 	if !killSwitchClosed(m.Globals) {
 		a.log.Warn("engine is down and kill_switch=open: LAN traffic is NOT protected (documented fail-open): ", cause)
 		return
@@ -623,24 +664,52 @@ func (a *Applier) setHolding(v bool) {
 	a.stateMu.Unlock()
 }

+// Traffic returns where the traffic of the CURRENTLY RUNNING config goes. The
+// zero value means no config of this process's is running (nothing applied yet,
+// or the plane was torn down / put on hold), and callers must render that as
+// "unknown", never as protected.
+func (a *Applier) Traffic() generate.Traffic {
+	a.stateMu.RLock()
+	defer a.stateMu.RUnlock()
+	return a.traffic
+}
+
+func (a *Applier) setTraffic(t generate.Traffic) {
+	a.stateMu.Lock()
+	a.traffic = t
+	a.stateMu.Unlock()
+}
+
 func (a *Applier) setWarnings(ws []Warning) {
 	a.stateMu.Lock()
 	a.lastWarnings = ws
 	a.stateMu.Unlock()
 }

-// Warnings returns the normalised warning set from the last successful apply.
-// Never nil: an empty slice means "the last apply was clean", which the panel
-// must render differently from "no apply has run yet" (Active/Plane cover that).
+// Warnings returns the normalised warning set from the last successful apply,
+// PLUS whatever is wrong right now that no apply can describe. Never nil: an
+// empty slice means "the last apply was clean", which the panel must render
+// differently from "no apply has run yet" (Active/Plane cover that).
+//
+// The live half is currently the engine's abandoned generations
+// (engineTeardownWarnings). It is computed at READ time rather than folded into
+// lastWarnings on purpose: a superseded box that will not shut down is a
+// condition of the process, not a property of a config. Folding it in would make
+// it appear only after the NEXT successful apply and then stay published long
+// after the shutdown finally completed — reporting a leak that is over, and
+// staying silent about one that is not. Read-time means it shows up the instant
+// it happens and clears itself the instant it resolves.
 func (a *Applier) Warnings() []Warning {
+	var out []Warning
+	if a.eng != nil {
+		out = engineTeardownWarnings(a.eng.PendingCloses())
+	}
 	a.stateMu.RLock()
 	defer a.stateMu.RUnlock()
-	if a.lastWarnings == nil {
+	if out == nil && a.lastWarnings == nil {
 		return []Warning{}
 	}
-	out := make([]Warning, len(a.lastWarnings))
-	copy(out, a.lastWarnings)
-	return out
+	return append(out, a.lastWarnings...)
 }

 // Reconcile re-reads UCI and either tears down (disabled) or re-applies (enabled).
@@ -706,6 +775,7 @@ func (a *Applier) Teardown() error {
 	a.lastGood = nil
 	a.lastNft = ""
 	a.setHolding(false)
+	a.setTraffic(generate.Traffic{})
 	a.setWarnings(nil)
 	// The plane is gone, so the logged set no longer describes anything. Forget it,
 	// and the next apply re-announces its warnings in full rather than staying
@@ -986,6 +1056,20 @@ type Status struct {
 	//	         say so rather than looking healthy.
 	Plane string `json:"plane"`

+	// Traffic is WHERE THE TRAFFIC GOES under the running config — tunnelled, split,
+	// straight out, or blocked (see generate.Traffic).
+	//
+	// Plane does NOT answer this, and reading it as if it did is the defect this
+	// field exists for. Plane == "full" only means the table, the policy routing and
+	// the engine are all in place; a router whose one rule is `default -> direct` has
+	// all three and sends every packet out the plain WAN with its real address. The
+	// panel showed that as "Protected — traffic is going through the tunnel".
+	//
+	// Zero value (Verdict == "") means unknown: no successful apply has run in this
+	// daemon process yet, or the plane is on hold / torn down. A consumer must render
+	// that as unknown and never as protected.
+	Traffic generate.Traffic `json:"traffic"`
+
 	// Warnings is the normalised warning set from the last successful apply:
 	// everything that was skipped, degraded or left un-applied while the apply
 	// still succeeded. Always non-nil so the panel can map over it unconditionally.
@@ -1065,6 +1149,7 @@ func (a *Applier) Status() Status {
 		Hash:          a.eng.Hash(),
 		CanRollback:   a.canRollback(),
 		EngineRunning: a.eng.Running(),
+		Traffic:       a.Traffic(),
 		Warnings:      a.Warnings(),
 	}
 	s.StartedUnix, s.UptimeSeconds = processUptime(time.Now())
@@ -0,0 +1,283 @@
+package apply
+
+// Regression cover for the leaked-engine-generation defect.
+//
+// Observed on the router: one shaterd process was carrying up to FOUR sing-box
+// instances at once. sing-box stamps every log line with the elapsed seconds of
+// ITS OWN instance, so the same process printed `ERROR[2015]` and `ERROR[0129]`
+// in the same second — two engines half an hour apart in age, both alive, both
+// dialling, both holding WireGuard devices built from the same private keys. A
+// full daemon stop+start collapsed it back to one generation, which places the
+// leak squarely on the config re-apply path rather than on startup.
+//
+// The tests below pin the two halves of the fix:
+//
+//	TestApplySwapsLeaveExactlyOneEngineGeneration — the healthy path really
+//	  retires the old instance (its listener is provably gone), N times in a row.
+//	TestStuckEngineCloseDoesNotBlockTheApply     — a shutdown that never returns
+//	  is bounded, does not stall the apply, and is REPORTED as a critical warning
+//	  for exactly as long as it is true.
+
+import (
+	"io"
+	"net"
+	"net/netip"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	C "github.com/sagernet/sing-box/constant"
+	"github.com/sagernet/sing-box/option"
+	"github.com/sagernet/sing-box/shater/engine"
+	"github.com/sagernet/sing/common/json/badoption"
+)
+
+// mixedOn builds a minimal but REAL engine config: a mixed inbound bound to
+// 127.0.0.1:port plus a direct outbound. Two configs with different ports hash
+// differently, so each Apply is a genuine swap rather than a hash-gate no-op —
+// and the bound port is the observable that proves whether the old instance
+// actually died.
+func mixedOn(port uint16) option.Options {
+	listen := badoption.Addr(netip.MustParseAddr("127.0.0.1"))
+	return option.Options{
+		Log: &option.LogOptions{Level: "error"},
+		Inbounds: []option.Inbound{{
+			Type: C.TypeMixed,
+			Tag:  "mixed-in",
+			Options: &option.HTTPMixedInboundOptions{
+				ListenOptions: option.ListenOptions{Listen: &listen, ListenPort: port},
+			},
+		}},
+		Outbounds: []option.Outbound{{
+			Type:    C.TypeDirect,
+			Tag:     "direct-out",
+			Options: &option.DirectOutboundOptions{},
+		}},
+	}
+}
+
+// portFree reports whether 127.0.0.1:port can be bound right now — i.e. whether
+// the instance that used to listen there is really gone. Retried briefly because
+// a listener is released by Close, not by the return of Close's caller.
+func portFree(port uint16) bool {
+	deadline := time.Now().Add(3 * time.Second)
+	for {
+		ln, err := net.Listen("tcp", net.JoinHostPort("127.0.0.1", strconv.Itoa(int(port))))
+		if err == nil {
+			_ = ln.Close()
+			return true
+		}
+		if time.Now().After(deadline) {
+			return false
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+}
+
+// TestApplySwapsLeaveExactlyOneEngineGeneration is the core regression: after N
+// sequential applies the process must be carrying ONE engine, not N.
+//
+// "Carrying" is checked two ways on purpose. Generations() is the engine's own
+// accounting (running instance + every retirement still in flight) and would
+// catch a retirement that silently never completes. The port check is
+// independent of that accounting: if the superseded instance were still alive it
+// would still hold its listener, and the bind would fail. A fix that only
+// reset a pointer would pass the first check and fail the second.
+func TestApplySwapsLeaveExactlyOneEngineGeneration(t *testing.T) {
+	const (
+		firstPort = 18801
+		applies   = 5
+	)
+
+	a := New(engine.New(), nil)
+	t.Cleanup(func() { _ = a.eng.Close() })
+
+	for i := 0; i < applies; i++ {
+		port := uint16(firstPort + i)
+		changed, err := a.eng.Apply(mixedOn(port))
+		if err != nil {
+			t.Fatalf("apply #%d (port %d): %v", i+1, port, err)
+		}
+		if !changed {
+			t.Fatalf("apply #%d: every config here differs, so the swap must be real", i+1)
+		}
+		if got := a.eng.Generations(); got != 1 {
+			t.Fatalf("after apply #%d the process carries %d engine instances, want exactly 1 — "+
+				"a superseded generation is still alive (this is the four-generations-in-one-process leak)",
+				i+1, got)
+		}
+		if stuck := a.eng.PendingCloses(); len(stuck) != 0 {
+			t.Fatalf("after apply #%d: %d generation(s) abandoned, want none: %+v", i+1, len(stuck), stuck)
+		}
+		if i > 0 {
+			prev := uint16(firstPort + i - 1)
+			if !portFree(prev) {
+				t.Fatalf("after apply #%d the PREVIOUS generation still holds 127.0.0.1:%d — "+
+					"the old engine was replaced in the field but never actually stopped", i+1, prev)
+			}
+		}
+		// The warning set must stay clean while teardown is healthy: a critical
+		// warning that cries wolf on every apply is worse than none.
+		for _, w := range a.Warnings() {
+			if w.Section == "engine" {
+				t.Fatalf("after apply #%d a healthy swap produced an engine warning: %+v", i+1, w)
+			}
+		}
+	}
+
+	if err := a.eng.Close(); err != nil {
+		t.Fatalf("close: %v", err)
+	}
+	if got := a.eng.Generations(); got != 0 {
+		t.Fatalf("after Close the process carries %d engine instances, want 0", got)
+	}
+	if !portFree(firstPort + applies - 1) {
+		t.Fatalf("after Close the last generation still holds its listener")
+	}
+}
+
+// hangingCloser returns a box closer that BLOCKS the first close it is handed
+// until release() is called, and performs every later close normally. That is the
+// shape of the real fault: one subsystem of one generation (a WireGuard endpoint)
+// refuses to come down, while the rest of the process is fine.
+func hangingCloser() (closer func(io.Closer) error, release func()) {
+	gate := make(chan struct{})
+	first := make(chan struct{}, 1)
+	first <- struct{}{}
+	return func(c io.Closer) error {
+			select {
+			case <-first:
+				<-gate      // the stuck generation: never returns until released
+				return c.Close() // ...and then really does close, so the port frees
+			default:
+				return c.Close()
+			}
+		}, func() {
+			close(gate)
+		}
+}
+
+// TestStuckEngineCloseDoesNotBlockTheApply pins all four requirements of the
+// bounded teardown at once:
+//
+//	1. the apply COMPLETES — a shutdown that never returns must not hold the
+//	   control plane (and therefore the panel) hostage;
+//	2. the new engine is running afterwards — fail-closed semantics are unchanged,
+//	   the swap succeeded;
+//	3. the abandoned generation is REPORTED as a critical warning, by name, for as
+//	   long as it is still running — it is not silently swallowed;
+//	4. generations do not stack: a further apply while the leak persists leaves
+//	   one live instance plus the one abandoned one, not three.
+func TestStuckEngineCloseDoesNotBlockTheApply(t *testing.T) {
+	const budget = 200 * time.Millisecond
+	restoreBudget := engine.SetCloseBudget(budget)
+	defer restoreBudget()
+
+	a := New(engine.New(), nil)
+
+	// Generation 1 comes up with the REAL closer still installed.
+	if _, err := a.eng.Apply(mixedOn(18811)); err != nil {
+		t.Fatalf("apply #1: %v", err)
+	}
+
+	closer, release := hangingCloser()
+	restoreCloser := engine.SetBoxCloser(closer)
+	released := false
+	defer func() {
+		if !released {
+			release()
+		}
+		restoreCloser()
+		_ = a.eng.Close()
+	}()
+
+	// (1) Generation 2: the retirement of generation 1 will never return.
+	start := time.Now()
+	changed, err := a.eng.Apply(mixedOn(18812))
+	elapsed := time.Since(start)
+	if err != nil {
+		t.Fatalf("apply #2 must SUCCEED despite the stuck teardown: %v", err)
+	}
+	if !changed {
+		t.Fatalf("apply #2: expected a real swap")
+	}
+	// Generously bounded: the budget plus box.New+Start. The point is that it
+	// returned at all — before the fix this waited on Box.Close forever.
+	if elapsed > budget+20*time.Second {
+		t.Fatalf("apply #2 took %s: a stuck teardown must not stall the apply", elapsed)
+	}
+
+	// (2) fail-closed semantics unchanged: the new engine really is up.
+	if !a.eng.Running() {
+		t.Fatalf("apply #2: the new engine must be running")
+	}
+
+	// (3) the leak is visible, named, and critical.
+	stuck := a.eng.PendingCloses()
+	if len(stuck) != 1 {
+		t.Fatalf("PendingCloses() = %+v, want exactly the one abandoned generation", stuck)
+	}
+	if stuck[0].Generation != 1 {
+		t.Errorf("abandoned generation = %d, want 1", stuck[0].Generation)
+	}
+	ws := a.Status().Warnings // the exact set `shaterd status` and the panel read
+	var found *Warning
+	for i := range ws {
+		if ws[i].Section == "engine" {
+			found = &ws[i]
+			break
+		}
+	}
+	if found == nil {
+		t.Fatalf("a superseded engine that will not shut down produced NO warning; "+
+			"Status would show a healthy router: %+v", ws)
+	}
+	if found.Severity != SeverityCritical {
+		t.Errorf("stuck-teardown warning severity = %q, want %q", found.Severity, SeverityCritical)
+	}
+	if !strings.Contains(found.Name, "generation 1") {
+		t.Errorf("stuck-teardown warning must name the generation, got Name=%q", found.Name)
+	}
+	if !strings.Contains(found.Message, "STILL RUNNING") {
+		t.Errorf("stuck-teardown warning must say the instance is still running, got %q", found.Message)
+	}
+	if got := a.eng.Generations(); got != 2 {
+		t.Fatalf("Generations() = %d, want 2 (one live + one abandoned)", got)
+	}
+
+	// (4) another apply while the leak persists must not add a THIRD generation:
+	// with a generation abandoned the swap goes close-old-then-start-new, so the
+	// process still holds one live instance plus the one that will not die.
+	if _, err := a.eng.Apply(mixedOn(18813)); err != nil {
+		t.Fatalf("apply #3: %v", err)
+	}
+	if got := a.eng.Generations(); got != 2 {
+		t.Fatalf("Generations() = %d after a third apply, want 2 — generations are stacking, "+
+			"which is exactly the four-live-engines fault", got)
+	}
+	if stuck := a.eng.PendingCloses(); len(stuck) != 1 || stuck[0].Generation != 1 {
+		t.Fatalf("PendingCloses() = %+v, want only the original abandoned generation 1", stuck)
+	}
+
+	// (5) and it CLEARS: when the shutdown finally completes the warning goes away
+	// on its own. A leak report that outlives the leak trains the operator to
+	// ignore the panel.
+	release()
+	released = true
+	deadline := time.Now().Add(5 * time.Second)
+	for len(a.eng.PendingCloses()) > 0 && time.Now().Before(deadline) {
+		time.Sleep(10 * time.Millisecond)
+	}
+	if got := a.eng.PendingCloses(); len(got) != 0 {
+		t.Fatalf("the finished shutdown is still reported as abandoned: %+v", got)
+	}
+	for _, w := range a.Warnings() {
+		if w.Section == "engine" {
+			t.Fatalf("the engine warning outlived the leak it describes: %+v", w)
+		}
+	}
+	if got := a.eng.Generations(); got != 1 {
+		t.Fatalf("Generations() = %d after the stuck shutdown completed, want 1", got)
+	}
+}
@@ -0,0 +1,47 @@
+package apply
+
+import (
+	"errors"
+	"testing"
+
+	"github.com/sagernet/sing-box/shater/engine"
+	"github.com/sagernet/sing-box/shater/generate"
+	"github.com/sagernet/sing-box/shater/model"
+)
+
+// TestStatusReportsTraffic pins the wiring the panel's headline depends on.
+//
+// Plane says how much of the data plane is installed; it does NOT say where the
+// traffic goes, and reading it as if it did put "Protected — traffic is going
+// through the tunnel" on a router whose only rule was `default -> direct`. The
+// verdict that answers the real question travels in Status.Traffic, so it must
+// (a) start unknown, (b) surface what the last successful apply published, and
+// (c) go back to unknown the moment the engine stops carrying that config.
+func TestStatusReportsTraffic(t *testing.T) {
+	a := New(engine.New(), nil)
+
+	// A fresh applier has applied nothing, so it knows nothing. The zero value must
+	// NOT read as any verdict — least of all "tunnel".
+	if got := a.Status().Traffic; got.Verdict != "" {
+		t.Fatalf("fresh applier: Traffic.Verdict = %q, want \"\" (unknown)", got.Verdict)
+	}
+
+	a.setTraffic(generate.Traffic{Verdict: generate.VerdictTunnel, Default: "auto", TunnelRules: 2})
+	got := a.Status().Traffic
+	if got.Verdict != generate.VerdictTunnel || got.Default != "auto" || got.TunnelRules != 2 {
+		t.Fatalf("Status().Traffic = %+v, want the verdict the last apply published", got)
+	}
+
+	// The engine is down and the config it was running is no longer in force. A
+	// verdict left over from it would be the same reassuring lie, one layer down.
+	// (kill_switch=open takes holdLocked's early return, which is precisely the path
+	// that must still forget the verdict.)
+	g := model.DefaultGlobals()
+	g.KillSwitch = "open"
+	a.mu.Lock()
+	a.holdLocked(&model.Model{Globals: g}, errors.New("engine start failed"))
+	a.mu.Unlock()
+	if got := a.Status().Traffic; got.Verdict != "" {
+		t.Fatalf("after hold: Traffic.Verdict = %q, want \"\" (unknown) — the engine is not carrying that config", got.Verdict)
+	}
+}
@@ -20,6 +20,7 @@ import (
 	"net/netip"
 	"strings"

+	C "github.com/sagernet/sing-box/constant"
 	"github.com/sagernet/sing-box/option"
 	"github.com/sagernet/sing-box/shater/model"
 	"github.com/sagernet/sing-box/shater/netplane"
@@ -30,6 +31,110 @@ import (
 // engine.RuleSetIPCIDRs).
 type ruleSetCIDRs func(tag string) ([]netip.Prefix, bool)

+// SCHEMA v2 MOVED A RULE'S DESTINATION OUT OF THE RULE, AND THIS FILE HAD TO FOLLOW.
+//
+// Until D21 a routing rule spelled its destination inline: `dst_domain` became
+// DefaultRule.Domain*, `dst_ip` became DefaultRule.IPCIDR. Both were visible right
+// on the rule, so matchesUntunnelable could read them off it and the ADDRESS half
+// could only ever be resolved through the engine for the rare geosite/geoip case.
+//
+// Since v2 a rule names a `config ruleset` and generate materialises it into
+// Route.RuleSet, so BOTH halves arrive by reference. That broke two things at once,
+// in opposite directions:
+//
+//   - DOMAINS WENT INVISIBLE. A v1 rule `dst_domain=[bank.ru] dst_ip=[10/8]` ANDed
+//     its matchers, so it could not match an ICMP packet (no domain in one) and this
+//     file skipped it. After migration the same rule reads `rule_set: [rule-x,
+//     rule-x-ip]`, whose two sets are ORed — so the address set alone would make the
+//     rule look applicable, and the plan would start emitting an ALLOW (target
+//     direct) or a DENY (target block) for 10/8 that the operator never asked for.
+//     The ALLOW is the dangerous half: it sends previously-tunnelled ICMP out with
+//     the client's real address, i.e. an upgrade silently opening a leak. So a rule
+//     whose rule-sets are KNOWN to carry domain matchers is treated exactly as its
+//     v1 form was: inapplicable to untunnelable traffic. (The OR is deliberate and
+//     documented in D21 for the ENGINE's TCP/UDP path; it does not follow that a
+//     leak-guard should widen itself during an upgrade without a word.)
+//   - ADDRESSES WENT ENGINE-ONLY. `rule-x-ip` is an INLINE rule-set: its prefixes
+//     are sitting right there in the generated options. Asking the engine for them
+//     — and, when the engine is not up, truncating the whole walk and denying
+//     everything — was needless. Inline sets are now read from the options, so an
+//     engine that has not started yet no longer costs the operator their ping.
+//
+// Both come from the same index, built once per plan: what the CONFIG ITSELF can
+// say about each rule-set. Remote/local sets stay the engine's business.
+
+// ruleSetFacts is what the generated options alone reveal about one rule-set.
+//
+// opaque means "not knowable from here" — a local/remote set (only the engine has
+// its contents) or an inline set with a rule shape this code does not model. An
+// opaque set is never used to conclude anything; it falls through to the engine
+// lookup exactly as before.
+type ruleSetFacts struct {
+	opaque    bool
+	hasDomain bool           // carries a matcher only a NAMED destination can satisfy
+	cidrs     []netip.Prefix // addresses, when they are knowable here
+}
+
+// indexRuleSets summarises every rule-set in the generated options by tag.
+func indexRuleSets(sets []option.RuleSet) map[string]ruleSetFacts {
+	out := make(map[string]ruleSetFacts, len(sets))
+	for _, rs := range sets {
+		var f ruleSetFacts
+		// option.RuleSet leaves Type empty for inline in some marshalled forms; the
+		// generator always sets it, and both spellings mean the same thing here.
+		if rs.Type != C.RuleSetTypeInline && rs.Type != "" {
+			f.opaque = true
+			out[rs.Tag] = f
+			continue
+		}
+		for _, hr := range rs.InlineOptions.Rules {
+			if hr.Type != C.RuleTypeDefault && hr.Type != "" {
+				// A logical headless rule, or a shape sing-box adds later. Never guessed
+				// at — the same rule the walk itself follows for a logical route rule.
+				f = ruleSetFacts{opaque: true}
+				break
+			}
+			d := hr.DefaultOptions
+			if headlessNeedsUnavailableMatcher(d) {
+				f.hasDomain = true
+			}
+			f.cidrs = append(f.cidrs, parsePrefixes(d.IPCIDR)...)
+		}
+		out[rs.Tag] = f
+	}
+	return out
+}
+
+// headlessNeedsUnavailableMatcher is matchesUntunnelable's predicate for the
+// HEADLESS rule shape a rule-set carries. It reports whether the rule demands
+// something a packet with no domain, no ports, no sniffed protocol and no process
+// identity cannot supply — in which case that rule-set cannot claim such a packet.
+//
+// The generator only ever emits domain matchers or ip_cidr into an inline routing
+// rule-set, so in practice this is the domain test; the rest is there so a future
+// matcher is refused rather than silently ignored.
+func headlessNeedsUnavailableMatcher(d option.DefaultHeadlessRule) bool {
+	switch {
+	case len(d.Domain) > 0, len(d.DomainSuffix) > 0, len(d.DomainKeyword) > 0,
+		len(d.DomainRegex) > 0, len(d.AdGuardDomain) > 0:
+		return true // no domain in an ICMP/ESP/GRE packet
+	case len(d.Port) > 0, len(d.PortRange) > 0, len(d.SourcePort) > 0, len(d.SourcePortRange) > 0:
+		return true
+	case len(d.Network) > 0, len(d.QueryType) > 0:
+		return true
+	case len(d.ProcessName) > 0, len(d.ProcessPath) > 0, len(d.ProcessPathRegex) > 0,
+		len(d.PackageName) > 0, len(d.PackageNameRegex) > 0:
+		return true
+	case len(d.WIFISSID) > 0, len(d.WIFIBSSID) > 0:
+		return true
+	case d.Invert:
+		// Same reasoning as the route-rule case: inverting flips every conservative
+		// assumption, so the set may only ever withhold, never grant.
+		return true
+	}
+	return false
+}
+
 // buildUntunnelablePlan reduces the resolved routing rules to the first-match
 // walk the nft forward chain can evaluate for a packet that has no domain, no
 // ports and no sniffable protocol.
@@ -57,6 +162,9 @@ func buildUntunnelablePlan(opts option.Options, lookup ruleSetCIDRs) *netplane.U
 		// No routing at all: nothing is provably direct.
 		return plan
 	}
+	// Since schema v2 a rule's destination lives in a rule-set, so the rule-sets
+	// have to be read alongside the rules. See the block above ruleSetFacts.
+	sets := indexRuleSets(opts.Route.RuleSet)

 	for _, r := range opts.Route.Rules {
 		if r.Type == "logical" {
@@ -78,7 +186,10 @@ func buildUntunnelablePlan(opts option.Options, lookup ruleSetCIDRs) *netplane.U
 		// and the destination logic never allowed anything. Its `protocol: dns`
 		// matcher makes it inapplicable to a packet with no stream to sniff, which
 		// is the correct and obvious reading once the checks are in this order.
-		if !matchesUntunnelable(d) {
+		if !matchesUntunnelable(d, sets) {
+			if note, ok := namedDestinationNote(d, sets); ok {
+				plan.Warnings = append(plan.Warnings, note)
+			}
 			continue // needs a domain/port/protocol: cannot apply to this traffic
 		}

@@ -96,21 +207,28 @@ func buildUntunnelablePlan(opts option.Options, lookup ruleSetCIDRs) *netplane.U
 		mt := netplane.UntunnelableMatch{Allow: verdict}

 		// Destination predicate: explicit ip_cidr plus every rule-set's addresses.
+		// An INLINE rule-set is answered straight out of the config — it is fully
+		// present there — so a rule whose lists are all inline no longer depends on
+		// the engine being up. Only local/remote sets still have to be asked for.
 		dst := parsePrefixes(d.IPCIDR)
-		resolvable := true
+		var unresolved string
 		for _, tag := range d.RuleSet {
+			if f, known := sets[tag]; known && !f.opaque {
+				dst = append(dst, f.cidrs...)
+				continue
+			}
 			prefixes, ok := lookup(tag)
 			if !ok {
-				resolvable = false
+				unresolved = tag
 				break
 			}
 			dst = append(dst, prefixes...)
 		}
-		if !resolvable {
+		if unresolved != "" {
 			plan.Warnings = append(plan.Warnings, fmt.Sprintf(
 				"list %q is not loaded yet, so ping/IPTV/VPN passthrough stays blocked for the "+
 					"addresses it covers (it resolves itself once the list downloads)",
-				strings.Join(d.RuleSet, ", ")))
+				unresolved))
 			return plan
 		}
 		hasDstMatcher := len(d.IPCIDR) > 0 || len(d.RuleSet) > 0
@@ -198,7 +316,20 @@ func classifyAction(a option.RuleAction) (isDirect bool, class int) {
 // Every matcher listed here is one the packet cannot satisfy, so a rule carrying
 // any of them is inapplicable rather than universally matching. Matchers ANDed
 // within a rule mean one unsatisfiable matcher makes the whole rule unsatisfiable.
-func matchesUntunnelable(d option.DefaultRule) bool {
+//
+// sets carries the same test for the matchers that are no longer written on the
+// rule: since schema v2 a destination is a rule-set reference, so a rule's domains
+// live in Route.RuleSet rather than in d.Domain*. A rule referencing a set that is
+// KNOWN to match by name is inapplicable here for exactly the reason a d.Domain
+// rule is — the packet has no name to match. Sets whose contents this code cannot
+// see (local/remote) say nothing either way; the address walk below already refuses
+// to conclude anything from them.
+func matchesUntunnelable(d option.DefaultRule, sets map[string]ruleSetFacts) bool {
+	for _, tag := range d.RuleSet {
+		if sets[tag].hasDomain {
+			return false
+		}
+	}
 	switch {
 	case len(d.Domain) > 0, len(d.DomainSuffix) > 0, len(d.DomainKeyword) > 0,
 		len(d.DomainRegex) > 0, len(d.Geosite) > 0:
@@ -229,6 +360,41 @@ func matchesUntunnelable(d option.DefaultRule) bool {
 	return true
 }

+// namedDestinationNote explains the one skip an operator can actually be
+// surprised by: a rule that names BOTH a domain list and an address list.
+//
+// A domain-only rule contributes nothing here and always has, so saying so would be
+// noise. A mixed rule is different — its addresses look like something this plan
+// could act on, and it is precisely the shape `shaterd migrate` produces from a v1
+// rule that carried dst_domain and dst_ip together. The rule is skipped so the
+// upgrade cannot quietly change what ping does (see the block above ruleSetFacts);
+// that decision is worth one line in the operator's warning list rather than none.
+//
+// ok=false when there is nothing surprising to report.
+func namedDestinationNote(d option.DefaultRule, sets map[string]ruleSetFacts) (string, bool) {
+	var named []string
+	addressed := len(d.IPCIDR) > 0
+	for _, tag := range d.RuleSet {
+		f, known := sets[tag]
+		if known && f.hasDomain {
+			named = append(named, tag)
+			continue
+		}
+		// An unknown/opaque set may well hold addresses; so may a knowable one.
+		if !known || f.opaque || len(f.cidrs) > 0 {
+			addressed = true
+		}
+	}
+	if len(named) == 0 || !addressed {
+		return "", false
+	}
+	return fmt.Sprintf(
+		"a routing rule matches by name (%s) as well as by address, and ping/IPTV/VPN-passthrough traffic "+
+			"carries no name — so that rule is left out of the ping/IPTV decision entirely and later rules "+
+			"decide those addresses. Split it into a name rule and an address rule if you want the addresses "+
+			"decided here", strings.Join(named, ", ")), true
+}
+
 // parsePrefixes converts CIDR/bare-address strings to prefixes, dropping anything
 // unparseable (it cannot be matched on, so it must not silently widen a set).
 func parsePrefixes(in []string) []netip.Prefix {
@@ -22,6 +22,16 @@ func tunnelModel() *model.Model {
 	return m
 }

+// pinnedIPSet declares the inline `type=ipcidr` rule-set a rule pins its
+// destination addresses with. Since schema v2 a routing rule has no dst_ip of its
+// own: addresses are a `config ruleset`, so the generated rule carries a rule_set
+// reference and the plan resolves the actual prefixes through the lookup below —
+// i.e. through the RUNNING engine in production (engine.RuleSetIPCIDRs), not out
+// of the config text. planFor's table stands in for that.
+func pinnedIPSet(name string, cidrs ...string) model.Ruleset {
+	return model.Ruleset{Name: name, Type: "ipcidr", Source: "inline", Entries: cidrs}
+}
+
 // planFor generates m and builds the untunnelable plan, resolving rule-set tags
 // from the supplied table. A tag absent from the table reports "not loaded".
 func planFor(t *testing.T, m *model.Model, sets map[string][]string) *netplane.UntunnelablePlan {
@@ -58,11 +68,12 @@ func renderPlan(t *testing.T, m *model.Model, plan *netplane.UntunnelablePlan) s
 // only 8.8.8.8 may be un-pingable and the rest of the internet must answer.
 func TestOnlyPinnedAddressIsTunnelled(t *testing.T) {
 	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{pinnedIPSet("pin", "8.8.8.8/32")}
 	m.Rules = []model.Rule{
-		{Name: "pin", Enabled: true, Order: 10, DstIP: []string{"8.8.8.8/32"}, Target: "group:auto"},
+		{Name: "pin", Enabled: true, Order: 10, DstRuleset: []string{"pin"}, Target: "group:auto"},
 		{Name: "rest", Enabled: true, Order: 99, Target: "direct"},
 	}
-	plan := planFor(t, m, nil)
+	plan := planFor(t, m, map[string][]string{"rs-pin": {"8.8.8.8/32"}})

 	if !plan.DefaultAllow {
 		t.Fatalf("a catch-all `direct` rule must make the default ALLOW; plan=%+v", plan)
@@ -221,14 +232,148 @@ func TestDomainRuleDoesNotAffectUntunnelable(t *testing.T) {
 	}
 }

-// TestBlockedRuleDenies: an explicitly blocked destination stays dropped.
-func TestBlockedRuleDenies(t *testing.T) {
+// --- schema v2: the destination moved into rule-sets, and so did the analysis ---
+
+// migratedRule reproduces exactly what `shaterd migrate` makes of a v1 rule that
+// carried BOTH lists: `dst_domain=[bank.ru] dst_ip=[203.0.113.0/24]` becomes two
+// inline rule-sets, `rule-x` and `rule-x-ip`, and the rule references both.
+func migratedRule(target string) *model.Model {
 	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{
+		{Name: "rule-x", Type: "domain", Source: "inline", Entries: []string{"full:bank.ru"}},
+		pinnedIPSet("rule-x-ip", "203.0.113.0/24"),
+	}
 	m.Rules = []model.Rule{
-		{Name: "bad", Enabled: true, Order: 10, DstIP: []string{"203.0.113.0/24"}, Target: "block"},
+		{Name: "x", Enabled: true, Order: 10,
+			DstRuleset: []string{"rule-x", "rule-x-ip"}, Target: target},
+		{Name: "dflt", Enabled: true, Order: 99, Target: "group:auto"},
+	}
+	return m
+}
+
+// TestMigratedDomainAndIPRuleStaysOutOfTheUntunnelablePlan is the upgrade
+// regression. In v1 the rule ANDed its domain and its addresses, so it could never
+// claim a packet that carries no domain and this plan skipped it. After the
+// migration the same rule reads `rule_set: [rule-x, rule-x-ip]`, and rule_set
+// entries are ORed — so the address set alone made the rule look applicable and the
+// plan started emitting a verdict for 203.0.113.0/24 that nobody asked for.
+//
+// With target=direct that verdict is an ALLOW, i.e. an upgrade quietly sending
+// previously-tunnelled ICMP out with the client's real source address. That is the
+// half that makes this a leak and not just a surprise, so it is checked first.
+func TestMigratedDomainAndIPRuleStaysOutOfTheUntunnelablePlan(t *testing.T) {
+	// Both engine states, because they fail differently: with the engine UP the old
+	// code emitted a live verdict for the addresses, and with it DOWN the same
+	// mistake hid behind the unresolvable-list truncation. Neither may happen.
+	for _, target := range []string{"direct", "block", "group:auto"} {
+		for _, engine := range []string{"up", "down"} {
+			t.Run(target+"/engine-"+engine, func(t *testing.T) {
+				m := migratedRule(target)
+				var loaded map[string][]string
+				if engine == "up" {
+					loaded = map[string][]string{"rs-rule-x": {}, "rs-rule-x-ip": {"203.0.113.0/24"}}
+				}
+				plan := planFor(t, m, loaded)
+
+				if len(plan.Matches) != 0 {
+					t.Fatalf("a rule that matches by NAME as well as by address must contribute no "+
+						"address step (it cannot match a packet that carries no name): %+v", plan.Matches)
+				}
+				if plan.DefaultAllow {
+					t.Errorf("the catch-all routes into the tunnel, so the default must still deny")
+				}
+				fwd := renderPlan(t, m, plan)
+				if strings.Contains(fwd, "203.0.113.0/24") {
+					t.Errorf("the migrated address list must not reach the data plane through the "+
+						"untunnelable policy:\n%s", fwd)
+				}
+			})
+		}
+	}
+}
+
+// TestMigratedDomainAndIPRuleSaysWhyItWasSkipped: skipping is the safe answer, but
+// a silently different ping after an upgrade is the complaint this whole change
+// exists to answer. The mixed rule — the exact shape migration produces — is named.
+func TestMigratedDomainAndIPRuleSaysWhyItWasSkipped(t *testing.T) {
+	plan := planFor(t, migratedRule("direct"), nil)
+	var seen bool
+	for _, w := range plan.Warnings {
+		if strings.Contains(w, "matches by name") && strings.Contains(w, "rs-rule-x") {
+			seen = true
+		}
+	}
+	if !seen {
+		t.Fatalf("the skip must be explained and the list named; warnings=%v", plan.Warnings)
+	}
+}
+
+// TestDomainOnlyRuleSetIsQuiet: a rule whose only destination is a NAME list has
+// always contributed nothing here and is not a surprise, so it must not produce a
+// note. Only the mixed shape is worth a line.
+func TestDomainOnlyRuleSetIsQuiet(t *testing.T) {
+	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{
+		{Name: "names", Type: "domain", Source: "inline", Entries: []string{"full:bank.ru"}},
+	}
+	m.Rules = []model.Rule{
+		{Name: "names", Enabled: true, Order: 10, DstRuleset: []string{"names"}, Target: "group:auto"},
 		{Name: "rest", Enabled: true, Order: 99, Target: "direct"},
 	}
 	plan := planFor(t, m, nil)
+	if !plan.DefaultAllow {
+		t.Errorf("a name-only rule must not withhold the catch-all allow; plan=%+v", plan)
+	}
+	for _, w := range plan.Warnings {
+		if strings.Contains(w, "matches by name") {
+			t.Errorf("a name-only rule is not a surprise and needs no note: %q", w)
+		}
+	}
+}
+
+// TestInlineRuleSetNeedsNoEngine: an inline rule-set's addresses are sitting in the
+// generated config. Asking the engine for them — and, while it is down, truncating
+// the whole walk so ping/IPTV/VPN passthrough dies everywhere — was a conservative
+// answer to a question that did not have to be asked. The verdict must now be the
+// same whether or not the engine is up.
+func TestInlineRuleSetNeedsNoEngine(t *testing.T) {
+	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{pinnedIPSet("pin", "8.8.8.8/32")}
+	m.Rules = []model.Rule{
+		{Name: "pin", Enabled: true, Order: 10, DstRuleset: []string{"pin"}, Target: "group:auto"},
+		{Name: "rest", Enabled: true, Order: 99, Target: "direct"},
+	}
+
+	down := planFor(t, m, nil)                                         // engine not up
+	up := planFor(t, m, map[string][]string{"rs-pin": {"8.8.8.8/32"}}) // engine up
+
+	for name, plan := range map[string]*netplane.UntunnelablePlan{"engine-down": down, "engine-up": up} {
+		if !plan.DefaultAllow {
+			t.Fatalf("%s: the catch-all direct must allow the rest of the internet; plan=%+v", name, plan)
+		}
+		if len(plan.Matches) != 1 || plan.Matches[0].Allow {
+			t.Fatalf("%s: the tunnelled address must produce one DENY step, got %+v", name, plan.Matches)
+		}
+		if got := plan.Matches[0].Dst4; len(got) != 1 || got[0] != "8.8.8.8/32" {
+			t.Fatalf("%s: step destinations = %v, want [8.8.8.8/32]", name, got)
+		}
+	}
+	for _, w := range down.Warnings {
+		if strings.Contains(w, "not loaded yet") {
+			t.Errorf("an inline list is never 'not loaded': %q", w)
+		}
+	}
+}
+
+// TestBlockedRuleDenies: an explicitly blocked destination stays dropped.
+func TestBlockedRuleDenies(t *testing.T) {
+	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{pinnedIPSet("bad", "203.0.113.0/24")}
+	m.Rules = []model.Rule{
+		{Name: "bad", Enabled: true, Order: 10, DstRuleset: []string{"bad"}, Target: "block"},
+		{Name: "rest", Enabled: true, Order: 99, Target: "direct"},
+	}
+	plan := planFor(t, m, map[string][]string{"rs-bad": {"203.0.113.0/24"}})
 	if len(plan.Matches) != 1 || plan.Matches[0].Allow {
 		t.Fatalf("a blocked destination must deny untunnelable traffic too: %+v", plan.Matches)
 	}
@@ -361,11 +506,12 @@ func TestPlanNeverAcceptsTCPOrUDP(t *testing.T) {
 			m := tunnelModel()
 			m.Globals.IPv6 = true
 			m.Globals.Untunnelable = policy
+			m.Rulesets = []model.Ruleset{pinnedIPSet("pin", "8.8.8.8/32")}
 			m.Rules = []model.Rule{
-				{Name: "pin", Enabled: true, Order: 10, DstIP: []string{"8.8.8.8/32"}, Target: "group:auto"},
+				{Name: "pin", Enabled: true, Order: 10, DstRuleset: []string{"pin"}, Target: "group:auto"},
 				{Name: "rest", Enabled: true, Order: 99, Target: "direct"},
 			}
-			fwd := renderPlan(t, m, planFor(t, m, nil))
+			fwd := renderPlan(t, m, planFor(t, m, map[string][]string{"rs-pin": {"8.8.8.8/32"}}))

 			// Only the lines the untunnelable policy emits are in scope: the tproxy
 			// diverts in prerouting legitimately match TCP/UDP, which is their job.
@@ -424,12 +570,13 @@ func TestLocalPlaneSurvivesEveryPlan(t *testing.T) {
 // only grant those clients, not everyone.
 func TestSourceScopedRuleNarrowsTheAllow(t *testing.T) {
 	m := tunnelModel()
+	m.Rulesets = []model.Ruleset{pinnedIPSet("lab", "198.51.100.0/24")}
 	m.Rules = []model.Rule{
 		{Name: "lab", Enabled: true, Order: 10, Src: []string{"192.168.9.0/24"},
-			DstIP: []string{"198.51.100.0/24"}, Target: "direct"},
+			DstRuleset: []string{"lab"}, Target: "direct"},
 		{Name: "dflt", Enabled: true, Order: 99, Target: "group:auto"},
 	}
-	plan := planFor(t, m, nil)
+	plan := planFor(t, m, map[string][]string{"rs-lab": {"198.51.100.0/24"}})
 	if len(plan.Matches) != 1 {
 		t.Fatalf("expected one step, got %+v", plan.Matches)
 	}
@@ -24,7 +24,9 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"time"

+	"github.com/sagernet/sing-box/shater/engine"
 	"github.com/sagernet/sing-box/shater/model"
 	"github.com/sagernet/sing-box/shater/netplane"
 )
@@ -91,6 +93,15 @@ var criticalMarkers = []string{
 	"not covered",     // an interface outside the fail-closed guard
 	"fail-closed",     //  ''
 	"REJECTED",        // an unusable interface name
+	// A condition-less rule retired by a later condition-less rule whose target is
+	// `direct` (generate/route.go warnUnreachableRules): the operator's default
+	// policy — a tunnel, or a block — is not the one the router uses, so everything
+	// unmatched leaves on the plain WAN. The marker is the full clause, not the
+	// shorter "leaves over the plain WAN" that ruleKillFallback's kill=open note
+	// also contains: that one is a DELIBERATE per-rule bypass the operator asked
+	// for, and grading it critical here would be a different decision made by
+	// accident.
+	"leaves over the plain WAN with your real IP address",
 }

 // protectionSections are entity kinds whose whole purpose is to block or divert
@@ -260,6 +271,50 @@ func untunnelablePolicyWarnings(g model.Globals, planNotes []string) []Warning {
 	}
 }

+// engineTeardownWarnings turns the engine's ABANDONED generations — superseded
+// sing-box instances whose shutdown overran the hard close budget and are still
+// running inside this process — into operator-facing warnings.
+//
+// Critical, without hesitation. A leaked generation is not untidiness:
+//
+//   - it still holds its WireGuard devices, and two devices built from the same
+//     private key evict each other at the peer (one session per public key), so
+//     the leak reproduces BETWEEN generations exactly the fault
+//     generate/wgdedup.go removes WITHIN a config — the tunnel flaps and neither
+//     end can say why;
+//   - it still holds its outbound connections and keeps probing nodes, so the
+//     log fills with errors attributed to a config that is no longer applied;
+//   - on a 512 MiB router each one costs real memory that is never returned.
+//
+// The generation number is carried in Name so two consecutive status reads can
+// tell "the same stuck generation" from "another one just leaked", and the
+// elapsed time is in the message because a shutdown at 8s and one at 40 minutes
+// are different problems.
+func engineTeardownWarnings(stuck []engine.StuckClose) []Warning {
+	if len(stuck) == 0 {
+		return nil
+	}
+	out := make([]Warning, 0, len(stuck))
+	for _, s := range stuck {
+		config := "unknown config"
+		if len(s.Hash) >= 12 {
+			config = "config " + s.Hash[:12]
+		}
+		out = append(out, Warning{
+			Severity: SeverityCritical,
+			Section:  "engine",
+			Name:     fmt.Sprintf("generation %d", s.Generation),
+			Message: fmt.Sprintf(
+				"a superseded engine instance (%s) has been shutting down for %s and is STILL RUNNING: "+
+					"it keeps its outbound connections and its WireGuard devices, so it can evict the "+
+					"live tunnel at the peer and it keeps writing to the log. The current configuration "+
+					"is applied and running; restart shaterd if this does not clear.",
+				config, s.Elapsed.Round(time.Second)),
+		})
+	}
+	return out
+}
+
 func severityRank(s string) int {
 	switch s {
 	case SeverityCritical:
@@ -240,3 +240,45 @@ func blockGlobals() model.Globals {
 	g.Untunnelable = netplane.UntunnelableBlock
 	return g
 }
+
+// TestCollectWarningsGradesUnreachableRules (B1): a condition-less rule retired by
+// a later condition-less one is graded by CONSEQUENCE, not by the mere fact that a
+// setting is dead.
+//
+//   - the surviving default is `direct` while the retired one wanted a tunnel:
+//     the operator's default policy is not in effect and everything unmatched
+//     leaves on the plain WAN — critical, the panel must show it loudly;
+//   - the surviving default is the tunnel and the dead one was `direct`: a dead
+//     knob, nothing is leaking — warning.
+//
+// Both must be attributed to section "rule" + the rule's name, so the panel can
+// deep-link to the offending row instead of printing prose.
+func TestCollectWarningsGradesUnreachableRules(t *testing.T) {
+	const leak = `rule "default": it has no conditions, so it sets the default for ALL traffic — ` +
+		`but rule "fallback" (order 100) has none either and comes after it, so "direct" wins and ` +
+		`this rule's target "group:auto" is never applied. Everything no other rule matches leaves ` +
+		`over the plain WAN with your real IP address. Delete one of the two, or give this one a condition`
+	const dead = `rule "default": it has no conditions, so it sets the default for ALL traffic — ` +
+		`but rule "default" (order 100) has none either and comes after it, so the default the ` +
+		`router uses is "group:auto" and this rule's target "direct" is never applied. Delete one ` +
+		`of the two, or give this one a condition so it can match something`
+
+	check := func(text, wantSeverity string) {
+		t.Helper()
+		found := false
+		for _, w := range collectWarnings(blockGlobals(), []string{text}, nil, nil) {
+			if w.Section != "rule" || w.Name != "default" {
+				continue
+			}
+			found = true
+			if w.Severity != wantSeverity {
+				t.Errorf("severity = %q, want %q for: %s", w.Severity, wantSeverity, text)
+			}
+		}
+		if !found {
+			t.Fatalf("never-applied warning not attributed to rule %q: %s", "default", text)
+		}
+	}
+	check(leak, SeverityCritical)
+	check(dead, SeverityWarning)
+}
@@ -0,0 +1,206 @@
+// Package buildtags is the contract between what shater DECLARES it supports
+// and the build tags the shipped router binary is actually compiled with.
+//
+// # Why this package exists
+//
+// The router binary is built with a deliberately trimmed tag set (D9/D23,
+// scripts/router-tags.sh) — upstream's full set registers a zoo shater/generate
+// can never emit, and a router pays for every tag in flash and in RAM. Trimming
+// is right; trimming BLIND is not. On 2026-07-25 a production router answered a
+// configured WireGuard node with
+//
+//	create instance: initialize endpoint[0]: create WireGuard device:
+//	gVisor is not included in this build, rebuild with -tags with_gvisor
+//
+// because `with_gvisor` had been trimmed as "unreachable code" (true for the tun
+// inbound we never emit — false for the WireGuard endpoint we ship and declare
+// [MVP]) while `with_wireguard` stayed. Nothing caught it: the test suite builds
+// with the FULL upstream tag set, so the SHIPPED tag combination was, at that
+// point, the one configuration nothing in the repo ever exercised.
+//
+// # What holds it together now
+//
+//  1. Features below names each declared feature and the build tags it needs to
+//     RUN (not merely to compile). shater/buildtags's own test parses
+//     scripts/router-tags.sh and fails if the shipped set does not cover them —
+//     it needs no tags, no Linux and no network, so it runs in every plain
+//     `go test ./...`.
+//  2. shater/generate's TestShippedTagSetConstructsDeclaredProtocols drives one
+//     node of every declared protocol through box.New under whatever tags the
+//     test binary was built with, skipping only what is genuinely not compiled
+//     in. scripts/check-router-tags.sh runs it with the SHIPPED set, so the
+//     combination we ship is proven to construct, not merely to link.
+//
+// (1) catches a trimmed dependency the moment it is trimmed; (2) catches the
+// class of failure (1) cannot model — a tag that is present but insufficient.
+//
+// Adding a protocol to shater/parse + shater/generate means adding a row here.
+package buildtags
+
+import "sort"
+
+// Feature is one capability the product declares, together with the build tags
+// the binary must carry for it to work at runtime.
+type Feature struct {
+	// Name is the feature as a user would name it.
+	Name string
+	// Declared points at where we promise it (docs-shater/FEATURES.md section,
+	// or the generator/registry that emits it).
+	Declared string
+	// Tags are ALL build tags required for the feature to work — including
+	// transitive ones (with_awg alone is useless without with_wireguard, which
+	// is useless without with_gvisor). Listing them transitively is deliberate:
+	// the check must not depend on a dependency graph nobody maintains.
+	Tags []string
+	// Why explains what breaks without those tags, with the code anchor. It is
+	// printed by the failing test, so a future trimmer reads the reason instead
+	// of rediscovering it on a router.
+	Why string
+}
+
+// Features is the authoritative list. Only tag-GATED capabilities belong here:
+// tproxy, routing rules, rule-sets, the DNS filter, nft/policy routing and the
+// panel are compiled unconditionally and cannot be lost to a tag trim.
+var Features = []Feature{
+	{
+		Name:     "WireGuard nodes (wg:// / wireguard:// links, wg-quick .conf import)",
+		Declared: "FEATURES.md §Proxy engine — “VLESS, VMess, Trojan, Shadowsocks, WireGuard” [MVP]",
+		Tags:     []string{"with_wireguard", "with_gvisor"},
+		Why: "with_wireguard registers the endpoint (shater/registry/registry_wireguard.go); " +
+			"with_gvisor supplies the userspace netstack EVERY WireGuard device needs — without it " +
+			"transport/wireguard/device_stack_stub.go returns tun.ErrGVisorNotIncluded from BOTH " +
+			"newStackDevice and newSystemStackDevice, so box.New fails with " +
+			"\"create WireGuard device: gVisor is not included in this build\" and the node is dead. " +
+			"system_interface=true is not an escape hatch: it hits the same stub.",
+	},
+	{
+		Name:     "AmneziaWG obfuscation (awg:// links; jc/jmin/jmax, s1-s4, h1-h4, i1-i5)",
+		Declared: "FEATURES.md §Proxy engine — “AmneziaWG 2.0 … a driving requirement” [MVP]",
+		Tags:     []string{"with_awg", "with_wireguard", "with_gvisor"},
+		Why: "with_awg makes the AWG params reach the device (transport/wireguard/device_awg.go); " +
+			"without it they parse and are silently ignored (option/wireguard.go). It rides on the " +
+			"WireGuard endpoint, so it needs that feature's tags too.",
+	},
+	{
+		Name:     "Hysteria2 nodes (hysteria2:// / hy2://)",
+		Declared: "FEATURES.md §Proxy engine [T1]; shater/registry registerQUICOutbounds",
+		Tags:     []string{"with_quic"},
+		Why:      "hysteria2.RegisterOutbound is compiled only under with_quic (shater/registry/registry_quic.go); without it box.New rejects the outbound as an unknown type.",
+	},
+	{
+		Name:     "TUIC nodes (tuic://)",
+		Declared: "FEATURES.md §Proxy engine [T1]; shater/registry registerQUICOutbounds",
+		Tags:     []string{"with_quic"},
+		Why:      "tuic.RegisterOutbound is compiled only under with_quic (shater/registry/registry_quic.go).",
+	},
+	{
+		Name:     "VLESS/VMess over the QUIC v2ray transport (type=quic)",
+		Declared: "FEATURES.md §Proxy engine — “Transports: TCP/WS/gRPC/HTTPUpgrade/H2/QUIC” [MVP]",
+		Tags:     []string{"with_quic"},
+		Why:      "transport/v2rayquic registers its constructor from an init() blank-imported only under with_quic; without it NewQUICClient returns os.ErrInvalid at dial time.",
+	},
+	{
+		Name:     "QUIC / HTTP3 DNS transports (quic://, h3://)",
+		Declared: "shater/registry registerQUICTransports",
+		Tags:     []string{"with_quic"},
+		Why:      "dns/transport/quic is registered only under with_quic (shater/registry/registry_quic.go).",
+	},
+	{
+		Name:     "REALITY (vless security=reality, pbk/sid)",
+		Declared: "FEATURES.md §Proxy engine — “Reality/XTLS” [MVP]; shater/parse security=reality",
+		Tags:     []string{"with_utls"},
+		Why:      "the REALITY client lives in common/tls/reality_client.go, which is itself `//go:build with_utls`; without the tag a reality config is rejected by the TLS layer.",
+	},
+	{
+		Name:     "uTLS ClientHello fingerprints (fp=chrome/firefox/safari/…)",
+		Declared: "shater/generate/outbound.go TLS mapping (UTLS options)",
+		Tags:     []string{"with_utls"},
+		Why:      "common/tls/utls_client.go is `//go:build with_utls`; the stub (utls_stub.go) refuses a config that sets a fingerprint.",
+	},
+	{
+		Name:     "XHTTP / SplitHTTP transport (type=xhttp, type=splithttp)",
+		Declared: "FEATURES.md §Proxy engine [T1]; shater/parse/sharelink.go case \"xhttp\"",
+		Tags:     []string{"with_xhttp"},
+		Why:      "transport/v2rayxhttp registers the \"xhttp\" transport from an init() blank-imported only under with_xhttp (shater/registry/registry_xhttp.go); without it the transport type is unknown at box.New.",
+	},
+	{
+		Name:     "badtls fast path (zero-copy TLS read-wait / ktls, used by every TLS outbound)",
+		Declared: "common/badtls — linked unconditionally by the TLS client",
+		Tags:     []string{"badlinkname", "tfogo_checklinkname0"},
+		Why: "common/badtls/*.go are `go1.25 && badlinkname`; without the tag the package degrades to read_wait_stub.go. " +
+			"These two tags additionally REQUIRE -checklinkname=0 in the linker flags — the build fails at link time otherwise " +
+			"(\"invalid reference to crypto/tls.(*Conn).handlePostHandshakeMessage\"), which is why " +
+			"scripts/router-tags.sh carries SHATER_ROUTER_LDFLAGS next to the tag set.",
+	},
+}
+
+// RequiredTags is the union of every declared feature's tags, sorted.
+func RequiredTags() []string {
+	seen := map[string]bool{}
+	for _, f := range Features {
+		for _, t := range f.Tags {
+			seen[t] = true
+		}
+	}
+	return sortedKeys(seen)
+}
+
+// Compiled reports the shater-relevant build tags THIS binary was compiled with,
+// sorted. It is populated by the one-line tag_*.go twins in this package; a tag
+// with no file here is simply not tracked (and must not appear in Features).
+func Compiled() []string { return sortedKeys(compiled) }
+
+// Has reports whether this binary was compiled with tag.
+func Has(tag string) bool { return compiled[tag] }
+
+// MissingTags returns the tags f needs that this binary lacks, sorted. Empty
+// means the feature is fully compiled in.
+func MissingTags(f Feature) []string {
+	missing := map[string]bool{}
+	for _, t := range f.Tags {
+		if !compiled[t] {
+			missing[t] = true
+		}
+	}
+	return sortedKeys(missing)
+}
+
+// Tracked reports whether tag has a detector file (tag_*.go) in this package.
+// Features must only reference tracked tags — an untracked tag would silently
+// read as "not compiled" and turn a real check into a skip. TestFeatureTagsAreTracked
+// enforces that, and scripts/check-router-tags.sh additionally proves the
+// detectors match the tag set the compiler was actually handed.
+func Tracked(tag string) bool { return tracked[tag] }
+
+// TrackedTags is every tag this package can observe, i.e. exactly the tags with
+// a tag_*.go detector. Keep the two in sync — the check script fails loudly if
+// they drift.
+func TrackedTags() []string { return sortedKeys(tracked) }
+
+var tracked = map[string]bool{
+	"with_gvisor":          true,
+	"with_quic":            true,
+	"with_wireguard":       true,
+	"with_awg":             true,
+	"with_utls":            true,
+	"with_xhttp":           true,
+	"with_lx_command":      true,
+	"badlinkname":          true,
+	"tfogo_checklinkname0": true,
+}
+
+// compiled is filled by the tag_*.go detectors' init(). A tag with no detector
+// file compiled in is absent from the map, which reads as "not compiled".
+var compiled = map[string]bool{}
+
+// mark records that tag is compiled into this binary.
+func mark(tag string) { compiled[tag] = true }
+
+func sortedKeys(m map[string]bool) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
@@ -0,0 +1,175 @@
+package buildtags
+
+import (
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// repoFile reads a file relative to the repo root (this package sits at
+// <repo>/shater/buildtags).
+func repoFile(t *testing.T, rel string) string {
+	t.Helper()
+	b, err := os.ReadFile(filepath.Join("..", "..", filepath.FromSlash(rel)))
+	if err != nil {
+		t.Fatalf("read %s: %v", rel, err)
+	}
+	return string(b)
+}
+
+// shVar pulls VAR="…" out of a POSIX sh fragment.
+func shVar(t *testing.T, script, name string) string {
+	t.Helper()
+	re := regexp.MustCompile(`(?m)^` + regexp.QuoteMeta(name) + `="([^"]*)"`)
+	m := re.FindStringSubmatch(script)
+	if m == nil {
+		t.Fatalf("scripts/router-tags.sh: %s=\"…\" not found (single line, double quotes)", name)
+	}
+	return m[1]
+}
+
+// routerTagSet returns the shipped tag set as a set, read from the ONE file that
+// defines it.
+func routerTagSet(t *testing.T) map[string]bool {
+	t.Helper()
+	set := map[string]bool{}
+	for _, tag := range strings.Split(shVar(t, repoFile(t, "scripts/router-tags.sh"), "SHATER_ROUTER_TAGS"), ",") {
+		if tag = strings.TrimSpace(tag); tag != "" {
+			set[tag] = true
+		}
+	}
+	if len(set) == 0 {
+		t.Fatal("SHATER_ROUTER_TAGS is empty")
+	}
+	return set
+}
+
+// TestRouterTagSetCoversDeclaredFeatures is THE guard the 2026-07-25 WireGuard
+// outage was missing (D23): it reads the tag set the router binary is actually
+// built with and fails if a feature we DECLARE supported has lost the build tag
+// it needs to run.
+//
+// It deliberately needs no build tags, no Linux, no privileges and no network,
+// so it runs in every plain `go test ./...` — including on the Windows dev host,
+// where nothing else can exercise the shipped configuration. The behavioural
+// half (does the shipped combination actually CONSTRUCT?) is
+// shater/generate.TestShippedTagSetConstructsDeclaredProtocols, run with this
+// same set by scripts/check-router-tags.sh.
+func TestRouterTagSetCoversDeclaredFeatures(t *testing.T) {
+	shipped := routerTagSet(t)
+
+	for _, f := range Features {
+		var missing []string
+		for _, tag := range f.Tags {
+			if !shipped[tag] {
+				missing = append(missing, tag)
+			}
+		}
+		if len(missing) > 0 {
+			t.Errorf("the shipped router binary would NOT support a feature we declare.\n"+
+				"  feature : %s\n"+
+				"  declared: %s\n"+
+				"  missing : %s  (not in SHATER_ROUTER_TAGS, scripts/router-tags.sh)\n"+
+				"  why     : %s\n"+
+				"Either add the tag back, or stop declaring the feature — those are the only two honest options.",
+				f.Name, f.Declared, strings.Join(missing, ", "), f.Why)
+		}
+	}
+}
+
+// TestFeatureTagsAreTracked keeps Features honest: every tag it names must have
+// a tag_*.go detector, or Compiled()/MissingTags() would report it absent even
+// when it is compiled in — and the behavioural test would silently SKIP the
+// feature instead of checking it. A false green is worse than a red.
+func TestFeatureTagsAreTracked(t *testing.T) {
+	for _, f := range Features {
+		for _, tag := range f.Tags {
+			if !Tracked(tag) {
+				t.Errorf("feature %q requires tag %q, which has no detector: add shater/buildtags/tag_%s.go and the entry in the tracked map", f.Name, tag, tag)
+			}
+		}
+	}
+}
+
+// TestTrackedTagsHaveDetectorFiles pairs the tracked map with the files on disk,
+// so a renamed/deleted detector cannot quietly make a tag read as absent.
+func TestTrackedTagsHaveDetectorFiles(t *testing.T) {
+	for _, tag := range TrackedTags() {
+		name := "tag_" + tag + ".go"
+		body, err := os.ReadFile(name)
+		if err != nil {
+			t.Errorf("tracked tag %q has no detector file %s: %v", tag, name, err)
+			continue
+		}
+		if !strings.Contains(string(body), "//go:build "+tag) || !strings.Contains(string(body), `mark("`+tag+`")`) {
+			t.Errorf("%s must be `//go:build %s` and call mark(%q)", name, tag, tag)
+		}
+	}
+	files, err := filepath.Glob("tag_*.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, f := range files {
+		tag := strings.TrimSuffix(strings.TrimPrefix(f, "tag_"), ".go")
+		if !Tracked(tag) {
+			t.Errorf("detector %s exists but %q is not in the tracked map", f, tag)
+		}
+	}
+}
+
+// TestBuildScriptUsesTheSharedTagSet stops the split that caused the outage from
+// coming back: the ship build must SOURCE scripts/router-tags.sh, not carry its
+// own copy of the tag list. A second copy is a second truth, and the second one
+// is the one nobody checks.
+func TestBuildScriptUsesTheSharedTagSet(t *testing.T) {
+	build := repoFile(t, "scripts/build-shaterd.sh")
+	if !strings.Contains(build, "router-tags.sh") {
+		t.Fatal("scripts/build-shaterd.sh must source scripts/router-tags.sh")
+	}
+	if regexp.MustCompile(`(?m)^\s*ROUTER_TAGS="with_`).MatchString(build) {
+		t.Fatal("scripts/build-shaterd.sh re-inlines a literal tag list; the set must come from scripts/router-tags.sh only")
+	}
+}
+
+// TestRouterLdflagsSatisfyTagRequirements: `badlinkname` is not self-contained —
+// the LINK step fails without -checklinkname=0. The flag therefore belongs to
+// the tag set, and lives beside it; assert the pair never separates.
+func TestRouterLdflagsSatisfyTagRequirements(t *testing.T) {
+	script := repoFile(t, "scripts/router-tags.sh")
+	ldflags := shVar(t, script, "SHATER_ROUTER_LDFLAGS")
+	if routerTagSet(t)["badlinkname"] && !strings.Contains(ldflags, "-checklinkname=0") {
+		t.Fatalf("SHATER_ROUTER_TAGS carries badlinkname but SHATER_ROUTER_LDFLAGS (%q) lacks -checklinkname=0: the build will fail at link time", ldflags)
+	}
+	if !strings.Contains(repoFile(t, "scripts/build-shaterd.sh"), "SHATER_ROUTER_LDFLAGS") {
+		t.Fatal("scripts/build-shaterd.sh must use $SHATER_ROUTER_LDFLAGS, not a hand-copied -checklinkname=0")
+	}
+}
+
+// TestCompiledTagsMatchTheShippedSet proves the DETECTORS are telling the truth:
+// when the test binary is compiled with exactly the shipped tag set, Compiled()
+// must equal that set (restricted to tracked tags). Without this, a typo'd or
+// deleted detector would make the behavioural test skip a protocol and pass.
+//
+// It only runs under scripts/check-router-tags.sh (which compiles with that very
+// set and exports SHATER_ROUTER_TAG_CHECK=1); a plain `go test ./...` compiles
+// with no tags at all, where the comparison is meaningless.
+func TestCompiledTagsMatchTheShippedSet(t *testing.T) {
+	if os.Getenv("SHATER_ROUTER_TAG_CHECK") != "1" {
+		t.Skip("not a router-tag-set run; use scripts/check-router-tags.sh")
+	}
+	var want []string
+	for tag := range routerTagSet(t) {
+		if Tracked(tag) {
+			want = append(want, tag)
+		}
+	}
+	sort.Strings(want)
+	got := Compiled()
+	if strings.Join(got, ",") != strings.Join(want, ",") {
+		t.Fatalf("compiled tags do not match the shipped set\n  compiled: %v\n  shipped : %v\n"+
+			"Either the build ran with the wrong -tags, or a tag_*.go detector is broken.", got, want)
+	}
+}
@@ -0,0 +1,7 @@
+//go:build badlinkname
+
+package buildtags
+
+// Detector for the badlinkname build tag — see buildtags.go. There is no !badlinkname twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("badlinkname") }
@@ -0,0 +1,7 @@
+//go:build tfogo_checklinkname0
+
+package buildtags
+
+// Detector for the tfogo_checklinkname0 build tag — see buildtags.go. There is no !tfogo_checklinkname0 twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("tfogo_checklinkname0") }
@@ -0,0 +1,7 @@
+//go:build with_awg
+
+package buildtags
+
+// Detector for the with_awg build tag — see buildtags.go. There is no !with_awg twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_awg") }
@@ -0,0 +1,7 @@
+//go:build with_gvisor
+
+package buildtags
+
+// Detector for the with_gvisor build tag — see buildtags.go. There is no !with_gvisor twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_gvisor") }
@@ -0,0 +1,7 @@
+//go:build with_lx_command
+
+package buildtags
+
+// Detector for the with_lx_command build tag — see buildtags.go. There is no !with_lx_command twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_lx_command") }
@@ -0,0 +1,7 @@
+//go:build with_quic
+
+package buildtags
+
+// Detector for the with_quic build tag — see buildtags.go. There is no !with_quic twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_quic") }
@@ -0,0 +1,7 @@
+//go:build with_utls
+
+package buildtags
+
+// Detector for the with_utls build tag — see buildtags.go. There is no !with_utls twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_utls") }
@@ -0,0 +1,7 @@
+//go:build with_wireguard
+
+package buildtags
+
+// Detector for the with_wireguard build tag — see buildtags.go. There is no !with_wireguard twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_wireguard") }
@@ -0,0 +1,7 @@
+//go:build with_xhttp
+
+package buildtags
+
+// Detector for the with_xhttp build tag — see buildtags.go. There is no !with_xhttp twin:
+// an absent detector means "not compiled in", which is exactly the truth.
+func init() { mark("with_xhttp") }
@@ -0,0 +1,27 @@
+package main
+
+// Control-plane log formatting (the IO-free half, so it is unit-testable on any
+// dev host — same split as profilewatch.go).
+//
+// The daemon's own logger used to be built with a bare log.Formatter, whose
+// DisableColors zero value is false: every control-plane line went to procd's
+// stderr — i.e. straight into syslog — carrying aurora escape sequences. See
+// shater/logsink/color.go for why that is a defect and not a cosmetic.
+
+import (
+	"os"
+	"time"
+
+	"github.com/sagernet/sing-box/log"
+	"github.com/sagernet/sing-box/shater/logsink"
+)
+
+// controlLogFormatter builds the control-plane logger's formatter. out is the
+// process's real stderr (the sink's syslog half): colours are emitted only when
+// that is a terminal, never when it is procd/syslog or a log file.
+func controlLogFormatter(baseTime time.Time, out *os.File) log.Formatter {
+	return log.Formatter{
+		BaseTime:      baseTime,
+		DisableColors: !logsink.IsTTY(out),
+	}
+}
@@ -0,0 +1,37 @@
+package main
+
+import (
+	"context"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/sagernet/sing-box/log"
+)
+
+// TestControlLogFormatterNoANSIOffTTY pins the control-plane half of the syslog
+// colour leak: when the daemon's stderr is not a terminal — which is ALWAYS the
+// case under procd, where stderr is the pipe procd relays to syslog — no line
+// the daemon formats may contain an ESC (0x1b).
+func TestControlLogFormatterNoANSIOffTTY(t *testing.T) {
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatalf("pipe: %v", err)
+	}
+	t.Cleanup(func() { _ = r.Close(); _ = w.Close() })
+
+	f := controlLogFormatter(time.Now(), w)
+	if !f.DisableColors {
+		t.Fatalf("colours enabled for a non-terminal stderr")
+	}
+	// A context ID exercises the second colouring branch of log/format.go (the
+	// 256-colour connection id), which is what produced ESC[38;5;193m on the router.
+	ctx := log.ContextWithNewID(context.Background())
+	for _, level := range []log.Level{log.LevelError, log.LevelWarn, log.LevelInfo, log.LevelDebug, log.LevelTrace} {
+		line := f.Format(ctx, level, "dns", "exchange failed for example.com. IN AAAA: unexpected EOF", time.Now())
+		if i := strings.IndexByte(line, 0x1b); i >= 0 {
+			t.Errorf("level %v: formatted line carries an ANSI escape at byte %d: %q", level, i, line)
+		}
+	}
+}
@@ -79,7 +79,7 @@ func dispatch(args []string) int {
 	case "mint-token":
 		return cmdMintToken()
 	case "nodes":
-		return cmdReadStub("nodes", "[]")
+		return cmdNodes()
 	case "stats":
 		return cmdReadStub("stats", "{}")
 	case "blocklist":
@@ -124,7 +124,7 @@ usage: shaterd <verb>
  rollback               roll back to the last-good config
  status                 print daemon/data-plane status as JSON
  mint-token             mint a single-use panel handoff token (JSON) via the daemon
-  nodes                  print nodes as JSON
+  nodes                  print the configured nodes as JSON (manual + subscription)
  stats                  print stats as JSON
  blocklist update       force a DNS-filter blocklist refresh (reconcile; no-op if down)
  schedule due           re-evaluate time-scheduled rules now (reconcile; no-op if down)
@@ -158,8 +158,11 @@ func cmdRun() int {
 	// The level starts at trace (nothing read yet) and is corrected from
 	// Globals.LogLevel as soon as UCI is read — the control plane now RESPECTS
 	// the configured level instead of the old unconditional trace.
+	// The formatter colours only when stderr is a real terminal
+	// (controlLogFormatter -> logsink.IsTTY): under procd stderr IS syslog, and
+	// ANSI escapes there break `logread | grep ERROR` and every log collector.
 	logFactory := log.NewDefaultFactory(context.Background(),
-		log.Formatter{BaseTime: time.Now()}, sink, "", nil, false)
+		controlLogFormatter(time.Now(), os.Stderr), sink, "", nil, false)
 	log.SetStdLogger(logFactory.Logger())
 	logger := log.StdLogger()

@@ -172,10 +175,23 @@ func cmdRun() int {
 		logFactory.SetLevel(controlLogLevel(g.LogLevel))
 	}

-	// Refuse to start a second daemon (race-safe single-owner guard). procd's
-	// term_timeout ensures the previous instance exits before reload restarts us.
-	if pid, ok := daemonAlive(); ok && pid != os.Getpid() {
-		logger.Error("another shaterd is already running (pid ", pid, ") — refusing to start")
+	// Single-owner guard: never run two daemons at once. On a `restart` procd's
+	// `service delete` is ASYNCHRONOUS — the ubus call returns immediately while
+	// the outgoing instance is still running its honest teardown — and the
+	// following `service add` starts us straight away, so a predecessor being
+	// alive here is the NORMAL restart case, not an error.
+	//
+	// This used to exit(1) on the spot and lean on procd's `respawn ... 5 ...` to
+	// try again five seconds later. That is a blind retry, not synchronisation:
+	// it neither knows nor waits for the predecessor's teardown to finish, and it
+	// turns every restart into at least one logged crash plus a five-second hole
+	// in which the LAN has no plane at all. Waiting for the predecessor to exit
+	// makes `restart` behave exactly like `stop` + pause + `start`: our apply is
+	// then strictly ordered AFTER the previous teardown, which is the whole point
+	// of the guard.
+	if pid, waiting := waitForPredecessor(predecessorBudget, predecessorPoll, logger); waiting {
+		logger.Error("another shaterd is still running (pid ", pid, ") after waiting ",
+			predecessorBudget, " for it to exit — refusing to start")
 		return 1
 	}
 	if err := writePidfile(); err != nil {
@@ -832,6 +848,42 @@ func cmdStatus() int {
 	return 0
 }

+// cmdNodes prints the node inventory as a JSON array (see nodes.go for the
+// shape and for why this is a view rather than the raw model.Node).
+//
+// Shape of the call mirrors cmdStatus: ask the RUNNING daemon first — it is the
+// process that owns the engine, so its answer is the inventory the live box was
+// built from — and fall back to reading the same on-disk desired state directly
+// when there is no daemon (or it did not answer). Both sides call the SAME
+// nodesJSON(), so the fallback cannot report something the daemon would not.
+//
+// The one thing it will never do is print `[]` because it could not find out:
+// a read failure goes to stderr and exits non-zero, so "empty" on stdout with
+// exit 0 means "no nodes are configured" and nothing else.
+func cmdNodes() int {
+	if _, ok := daemonAlive(); ok {
+		resp, err := ctlRequest("nodes")
+		switch {
+		case err != nil:
+			fmt.Fprintf(os.Stderr, "shaterd nodes: %v — reading the on-disk config instead\n", err)
+		case strings.HasPrefix(strings.TrimSpace(resp), "["):
+			fmt.Println(strings.TrimSpace(resp))
+			return 0
+		default:
+			// The daemon answered, but with an error object rather than a list.
+			fmt.Fprintf(os.Stderr, "shaterd nodes: daemon: %s\n", strings.TrimSpace(resp))
+		}
+	}
+	b, err := nodesJSON()
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "shaterd nodes: %v\n", err)
+		fmt.Println("[]") // stdout stays parseable JSON; the exit code carries the failure
+		return 1
+	}
+	fmt.Println(string(b))
+	return 0
+}
+
 // cmdMintToken asks the RUNNING daemon (over the control socket) for a single-use
 // panel handoff token and prints the daemon's JSON reply verbatim — {"token":"..."}
 // on success, {"error":"..."} otherwise. It is the CLI shim the LuCI/rpcd layer
@@ -869,6 +921,13 @@ func printJSONError(msg string) {
 	fmt.Println(string(b))
 }

+// cmdReadStub relays a read-only verb to the daemon and prints def when there is
+// no daemon to ask. It is ONLY valid where def is the truth in the daemon-down
+// case: `stats` counts what the LIVE engine saw, so with no engine there really
+// is nothing counted and `{}` says exactly that. It is NOT a way to make a verb
+// look implemented — `nodes` used to be routed through here with def="[]" while
+// hundreds of nodes sat on disk (see nodes.go). Anything whose data outlives the
+// daemon belongs in its own command that reads that data.
 func cmdReadStub(cmd, def string) int {
 	if _, ok := daemonAlive(); ok {
 		if resp, err := ctlRequest(cmd); err == nil {
@@ -940,7 +999,16 @@ func handleCtl(conn net.Conn, a *apply.Applier, ps *panel.Server, sa stats.Stats
 	case "rollback":
 		writeResult(conn, false, a.Rollback())
 	case "nodes":
-		writeLine(conn, "[]")
+		// Same assembly as the CLI fallback and as GET /api/config: model.ReadUCI
+		// (UCI + the per-subscription caches) projected onto the printable view.
+		b, err := nodesJSON()
+		if err != nil {
+			l.Warn("control socket: nodes: ", err)
+			e, _ := json.Marshal(map[string]string{"error": err.Error()})
+			writeLine(conn, string(e))
+			return
+		}
+		writeLine(conn, string(b))
 	case "stats":
 		if sa == nil {
 			writeLine(conn, "{}")
@@ -0,0 +1,139 @@
+package main
+
+// `shaterd nodes` — the node inventory, as JSON.
+//
+// This verb used to answer `[]` unconditionally (a `cmdReadStub("nodes", "[]")`
+// on the CLI side and a hard-coded `writeLine(conn, "[]")` in the daemon's
+// control-socket handler), while `shaterd --help` advertised "print nodes as
+// JSON". The data was never missing: /etc/shater/subs/*.json holds every
+// subscription-fetched node and `GET /api/config` reports the full merged set —
+// on a live router that is hundreds of nodes answered as "none". An empty list
+// is indistinguishable from a truthful "no nodes are configured", so the verb
+// did not fail loudly, it lied quietly. Same class as the honesty fixes in
+// 9dc954029 / aec82d444; the cure is the same: report what is actually there.
+//
+// # Data path (deliberately the SAME one /api/config uses)
+//
+// model.ReadUCI() = `uci export shater` (manual nodes + everything else) +
+// model.MergeSubCaches (the per-subscription JSON caches). That single call is
+// what the panel's handleConfigGet serves, what generate builds the engine from
+// and what `sub update` writes back — so `shaterd nodes` cannot drift from the
+// panel or from the running engine, because there is no second assembly here to
+// drift. This file only PROJECTS that model onto a small, printable view.
+//
+// # Why a view and not the raw model.Node
+//
+// model.Node carries the share-link URI, and a share link is a credential
+// (uuid/password in the query string). /api/config may return it — that path is
+// session-authenticated and the panel needs the URI to edit a node — but a CLI
+// verb whose output gets piped into support tickets, `logger`, and cron mail
+// must not spray credentials. The view therefore reports the *derived* facts a
+// share link answers (protocol, server, port) and drops the secret-bearing URI.
+// Nodes whose URI does not parse are still listed, with the parse error in
+// `parse_error`: the engine skips exactly those nodes (generate/outbound.go),
+// and a node that is configured-but-unusable is precisely what an operator
+// needs to see — hiding it would be the same lie in a smaller coat.
+
+import (
+	"encoding/json"
+	"strings"
+
+	"github.com/sagernet/sing-box/shater/model"
+	"github.com/sagernet/sing-box/shater/parse"
+)
+
+// nodeView is ONE node as the `nodes` verb reports it. Field set: the model's
+// own facts (name/enabled/sub/egress/stale/fingerprint) plus what the share link
+// decodes to (protocol/server/port). Everything is omitempty-free where a reader
+// would have to distinguish "absent" from "false"/"zero" — `enabled` and `stale`
+// are always present so a consumer never has to guess.
+type nodeView struct {
+	Name    string `json:"name"`
+	Enabled bool   `json:"enabled"`
+	// Sub is the subscription this node came from; "" means a manual node
+	// (model.Node.FromSub semantics, unchanged).
+	Sub string `json:"sub"`
+	// Protocol is the parsed protocol (vless|vmess|trojan|shadowsocks|wireguard…).
+	// When the URI does not parse it falls back to the bare URI scheme so the
+	// operator still sees what kind of thing failed; "" only when the URI is empty.
+	Protocol string `json:"protocol"`
+	Server   string `json:"server,omitempty"`
+	Port     uint16 `json:"port,omitempty"`
+	// Egress is the `config egress` this node's own upstream is bound to
+	// (multi-WAN); "" = default route.
+	Egress string `json:"egress,omitempty"`
+	// Stale marks a cached subscription node whose subscription failed to refresh.
+	Stale       bool   `json:"stale"`
+	Fingerprint string `json:"fingerprint,omitempty"`
+	// ParseError is the reason the engine will SKIP this node, verbatim from
+	// parse.ParseShareLink. Empty on every usable node.
+	ParseError string `json:"parse_error,omitempty"`
+	// ParseWarnings lists what the share link asked for that the engine cannot
+	// do (hysteria2 port hopping, tuic congestion control, …). The node IS
+	// usable — that is the difference from parse_error — but it does not behave
+	// exactly as its link describes, and that gap belongs on screen rather than
+	// in a code comment.
+	ParseWarnings []string `json:"parse_warnings,omitempty"`
+}
+
+// readModel is the model source. A package var so the tests can exercise the
+// assembly without a router's `uci` binary; production always reads the real
+// merged desired state.
+var readModel = model.ReadUCI
+
+// nodesJSON reads the desired state and renders the node inventory as a JSON
+// array. The array is never `null`: an empty configuration marshals to `[]`,
+// which is the ONE case where `[]` is the truth.
+func nodesJSON() ([]byte, error) {
+	m, err := readModel()
+	if err != nil {
+		return nil, err
+	}
+	return json.Marshal(nodeViews(m))
+}
+
+// nodeViews projects the merged model onto the printable view. Pure: no IO, no
+// globals — the whole verb's logic is testable on a dev host.
+func nodeViews(m *model.Model) []nodeView {
+	out := make([]nodeView, 0, len(m.Nodes)) // never nil => never `null`
+	for i := range m.Nodes {
+		n := m.Nodes[i]
+		v := nodeView{
+			Name:        n.Name,
+			Enabled:     n.Enabled,
+			Sub:         n.FromSub,
+			Egress:      n.Egress,
+			Stale:       n.Stale,
+			Fingerprint: n.Fingerprint,
+			Protocol:    uriScheme(n.URI),
+		}
+		// The engine reads a node through exactly this call (generate/outbound.go,
+		// chain.go, group.go); using it here is what makes "protocol" agree with
+		// what will actually be dialled, and the error agree with what will be
+		// skipped.
+		if p, err := parse.ParseShareLink(n.URI); err == nil {
+			if p.Protocol != "" {
+				v.Protocol = p.Protocol
+			}
+			v.Server = p.Server
+			v.Port = p.Port
+			v.ParseWarnings = p.Warnings
+		} else if n.URI != "" {
+			v.ParseError = err.Error()
+		}
+		out = append(out, v)
+	}
+	return out
+}
+
+// uriScheme returns the lower-cased scheme of a share link ("vless://…" ->
+// "vless"), or "" when there is none. It is the fallback protocol for a URI that
+// ParseShareLink refuses, so an unusable node is still described rather than
+// reported as a typeless blank.
+func uriScheme(uri string) string {
+	i := strings.Index(uri, "://")
+	if i <= 0 {
+		return ""
+	}
+	return strings.ToLower(uri[:i])
+}
@@ -0,0 +1,146 @@
+package main
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/sagernet/sing-box/shater/model"
+)
+
+// vlessURI is a syntactically complete share link (host, port, uuid, fragment)
+// so ParseShareLink really succeeds and the derived fields are exercised.
+const vlessURI = "vless://11111111-2222-3333-4444-555555555555@example.net:443?" +
+	"encryption=none&security=tls&sni=example.net&type=ws&path=%2Fws#NL-vless-1"
+
+// writeSubCache points the subscription cache at a temp dir and persists one
+// subscription's nodes there, the way `sub update` does on the router
+// (/etc/shater/subs/<sub>.json). Returns nothing: the point is the side effect
+// that model.MergeSubCaches will pick up.
+func writeSubCache(t *testing.T, sub string, nodes []model.Node) {
+	t.Helper()
+	t.Setenv("SHATER_SUBS_DIR", t.TempDir())
+	if err := model.SaveSubCache(sub, nodes); err != nil {
+		t.Fatalf("SaveSubCache(%q): %v", sub, err)
+	}
+}
+
+// TestNodesJSONReportsCachedSubscriptionNodes is the regression for the bug this
+// verb had: `shaterd nodes` answered `[]` while the subscription cache held
+// hundreds of nodes. A NON-EMPTY cache must produce a NON-EMPTY list.
+func TestNodesJSONReportsCachedSubscriptionNodes(t *testing.T) {
+	writeSubCache(t, "all-qomar", []model.Node{
+		{Name: "NL-vless-1", Enabled: true, URI: vlessURI, FromSub: "all-qomar", Fingerprint: "66b33"},
+		{Name: "NL-vless-2", Enabled: false, URI: vlessURI, FromSub: "all-qomar"},
+	})
+
+	// The model source stands in for `uci export shater` (absent on a dev host);
+	// the subscription half is the REAL model.MergeSubCaches path.
+	restore := readModel
+	readModel = func() (*model.Model, error) {
+		m := &model.Model{Globals: model.DefaultGlobals()}
+		model.MergeSubCaches(m)
+		return m, nil
+	}
+	t.Cleanup(func() { readModel = restore })
+
+	b, err := nodesJSON()
+	if err != nil {
+		t.Fatalf("nodesJSON: %v", err)
+	}
+	var got []nodeView
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatalf("nodesJSON produced invalid JSON %q: %v", b, err)
+	}
+	if len(got) == 0 {
+		t.Fatalf("nodes reported an EMPTY list while the subscription cache holds 2 nodes: %s", b)
+	}
+	if len(got) != 2 {
+		t.Fatalf("got %d nodes, want 2: %s", len(got), b)
+	}
+	byName := map[string]nodeView{}
+	for _, v := range got {
+		byName[v.Name] = v
+	}
+	first, ok := byName["NL-vless-1"]
+	if !ok {
+		t.Fatalf("cached node NL-vless-1 missing from %s", b)
+	}
+	if !first.Enabled {
+		t.Errorf("NL-vless-1.enabled = false, want true")
+	}
+	if first.Sub != "all-qomar" {
+		t.Errorf("NL-vless-1.sub = %q, want %q", first.Sub, "all-qomar")
+	}
+	if first.Protocol != "vless" {
+		t.Errorf("NL-vless-1.protocol = %q, want %q", first.Protocol, "vless")
+	}
+	if first.Server != "example.net" || first.Port != 443 {
+		t.Errorf("NL-vless-1 server:port = %s:%d, want example.net:443", first.Server, first.Port)
+	}
+	if first.ParseError != "" {
+		t.Errorf("NL-vless-1.parse_error = %q, want empty", first.ParseError)
+	}
+	if byName["NL-vless-2"].Enabled {
+		t.Errorf("NL-vless-2.enabled = true, want false (the model says disabled)")
+	}
+	// The credential-bearing share link must not be printed.
+	if strings.Contains(string(b), "11111111-2222-3333-4444-555555555555") {
+		t.Errorf("nodes output leaks the share-link uuid: %s", b)
+	}
+}
+
+// TestNodeViewsShape pins the projection: manual vs subscription, an unparsable
+// URI still being LISTED (with the reason), and an empty model marshalling to
+// `[]` rather than `null`.
+func TestNodeViewsShape(t *testing.T) {
+	m := &model.Model{Nodes: []model.Node{
+		{Name: "manual-1", Enabled: true, URI: vlessURI, Egress: "wan2"},
+		{Name: "broken", Enabled: true, URI: "nosuch://whatever", FromSub: "qomar", Stale: true},
+		{Name: "no-uri", Enabled: false},
+	}}
+	got := nodeViews(m)
+	if len(got) != 3 {
+		t.Fatalf("nodeViews returned %d views, want 3", len(got))
+	}
+	if got[0].Sub != "" {
+		t.Errorf("manual node sub = %q, want empty", got[0].Sub)
+	}
+	if got[0].Egress != "wan2" {
+		t.Errorf("manual node egress = %q, want wan2", got[0].Egress)
+	}
+	if got[1].ParseError == "" {
+		t.Errorf("unparsable node must carry the reason the engine will skip it")
+	}
+	if got[1].Protocol != "nosuch" {
+		t.Errorf("unparsable node protocol = %q, want the bare scheme %q", got[1].Protocol, "nosuch")
+	}
+	if !got[1].Stale {
+		t.Errorf("stale flag lost")
+	}
+	if got[2].Protocol != "" || got[2].ParseError != "" {
+		t.Errorf("URI-less node: got protocol=%q parse_error=%q, want both empty", got[2].Protocol, got[2].ParseError)
+	}
+
+	b, err := json.Marshal(nodeViews(&model.Model{}))
+	if err != nil {
+		t.Fatalf("marshal empty: %v", err)
+	}
+	if string(b) != "[]" {
+		t.Errorf("empty model marshalled to %q, want []", b)
+	}
+}
+
+func TestURIScheme(t *testing.T) {
+	for _, tc := range []struct{ in, want string }{
+		{"vless://x@h:443", "vless"},
+		{"VMESS://payload", "vmess"},
+		{"", ""},
+		{"not-a-uri", ""},
+		{"://leading", ""},
+	} {
+		if got := uriScheme(tc.in); got != tc.want {
+			t.Errorf("uriScheme(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
@@ -0,0 +1,63 @@
+package main
+
+// The startup handshake that stops a restarting daemon from standing a new data
+// plane up on top of the previous one's teardown.
+//
+// Deliberately free of build tags: the logic is pure timing and is exercised by
+// the tests on any host, while the probe it polls (pidfile + kill(pid, 0)) is
+// linux-only and is installed by predecessor_linux.go.
+
+import (
+	"os"
+	"time"
+
+	"github.com/sagernet/sing-box/log"
+)
+
+// predecessorBudget / predecessorPoll bound the wait for an outgoing daemon.
+//
+// The budget must comfortably exceed the slowest honest teardown (engine.Close
+// of a box with a few hundred outbounds plus the cache-file flush to flash, then
+// the nft/ip/sysctl removal) AND the init script's procd term_timeout, which is
+// the hard cap on how long a predecessor can live after its SIGTERM. Overshooting
+// costs nothing on a healthy box — the wait ends the instant the predecessor is
+// gone — while undershooting reintroduces the very overlap this exists to remove.
+//
+// Variables, not constants, so the tests can drive the loop without sleeping.
+var (
+	predecessorBudget = 60 * time.Second
+	predecessorPoll   = 100 * time.Millisecond
+)
+
+// aliveProbe is the seam waitForPredecessor polls. The portable default reports
+// "no predecessor", so a non-linux build (where there is no daemon at all) never
+// waits; predecessor_linux.go replaces it with the real pidfile probe.
+var aliveProbe = func() (int, bool) { return 0, false }
+
+// waitForPredecessor blocks until no OTHER shaterd owns the pidfile, or until the
+// budget runs out.
+//
+// Returns (pid, true) only when a predecessor is STILL alive once the budget
+// expires — the genuine "two daemons" error the caller refuses on. A predecessor
+// that exits within the budget (the restart case) returns (0, false) and startup
+// continues, now guaranteed to be sequenced after its teardown, exactly as it is
+// after a manual `stop` + pause + `start`.
+func waitForPredecessor(budget, poll time.Duration, logger log.ContextLogger) (int, bool) {
+	pid, ok := aliveProbe()
+	if !ok || pid == os.Getpid() {
+		return 0, false
+	}
+	if logger != nil {
+		logger.Info("a previous shaterd (pid ", pid, ") is still shutting down — ",
+			"waiting for its teardown to finish before applying a new data plane")
+	}
+	deadline := time.Now().Add(budget)
+	for time.Now().Before(deadline) {
+		time.Sleep(poll)
+		pid, ok = aliveProbe()
+		if !ok || pid == os.Getpid() {
+			return 0, false
+		}
+	}
+	return pid, true
+}
@@ -0,0 +1,9 @@
+//go:build linux
+
+package main
+
+// Bind the portable predecessor wait to the real single-owner probe. Split out
+// of main.go so waitForPredecessor itself stays build-tag-free and testable on
+// any developer host.
+
+func init() { aliveProbe = daemonAlive }
@@ -0,0 +1,106 @@
+package main
+
+// B3 regression, daemon half.
+//
+// `/etc/init.d/shater restart` is `stop; start`, and procd's `stop` is
+// asynchronous: the ubus `service delete` returns as soon as SIGTERM has been
+// SENT, so `start` re-adds the instance while the outgoing `shaterd run` is
+// still executing its honest teardown. The startup guard used to exit(1) the
+// moment it saw a live predecessor and rely on procd's `respawn ... 5 ...` to
+// try again later — a blind retry that neither knows nor waits for the teardown
+// to finish, and that turns every restart into a logged crash plus a five-second
+// hole with no data plane.
+//
+// The contract these pin: startup BLOCKS until the predecessor is gone (so our
+// apply is strictly ordered after its teardown, exactly as it is after a manual
+// `stop` + pause + `start`), and only refuses when the predecessor outlives the
+// whole budget.
+
+import (
+	"os"
+	"testing"
+	"time"
+)
+
+// scriptAlive installs an aliveProbe that reports a live predecessor for the
+// first n calls and "gone" afterwards, and restores the real probe.
+func scriptAlive(t *testing.T, pid, n int) *int {
+	t.Helper()
+	orig := aliveProbe
+	t.Cleanup(func() { aliveProbe = orig })
+	calls := 0
+	aliveProbe = func() (int, bool) {
+		calls++
+		if calls <= n {
+			return pid, true
+		}
+		return 0, false
+	}
+	return &calls
+}
+
+// TestWaitForPredecessorWaitsForTeardown: a predecessor that is still tearing
+// the plane down must be WAITED for, not refused. Before the fix this returned
+// "still alive" on the first probe and the daemon exited 1.
+func TestWaitForPredecessorWaitsForTeardown(t *testing.T) {
+	calls := scriptAlive(t, 4242, 3)
+
+	pid, stillAlive := waitForPredecessor(2*time.Second, time.Millisecond, nil)
+	if stillAlive {
+		t.Fatalf("a predecessor that exits within the budget must not be refused (pid %d)", pid)
+	}
+	if *calls < 4 {
+		t.Fatalf("expected the guard to keep probing until the predecessor was gone, got %d probes", *calls)
+	}
+}
+
+// TestWaitForPredecessorRefusesAfterBudget: the single-owner invariant is kept —
+// a predecessor that never dies still ends in a refusal, it is just no longer
+// the FIRST answer.
+func TestWaitForPredecessorRefusesAfterBudget(t *testing.T) {
+	orig := aliveProbe
+	defer func() { aliveProbe = orig }()
+	aliveProbe = func() (int, bool) { return 4242, true }
+
+	start := time.Now()
+	pid, stillAlive := waitForPredecessor(30*time.Millisecond, time.Millisecond, nil)
+	if !stillAlive || pid != 4242 {
+		t.Fatalf("an immortal predecessor must still be refused, got pid=%d alive=%v", pid, stillAlive)
+	}
+	if time.Since(start) < 30*time.Millisecond {
+		t.Fatalf("the guard must exhaust its budget before refusing")
+	}
+}
+
+// TestWaitForPredecessorNoPredecessorIsFree: the boot path must pay nothing —
+// one probe, no sleep. A wait that cost a tick on every clean start would show up
+// as a slower boot for no reason.
+func TestWaitForPredecessorNoPredecessorIsFree(t *testing.T) {
+	orig := aliveProbe
+	defer func() { aliveProbe = orig }()
+	calls := 0
+	aliveProbe = func() (int, bool) { calls++; return 0, false }
+
+	start := time.Now()
+	if _, stillAlive := waitForPredecessor(time.Minute, time.Second, nil); stillAlive {
+		t.Fatalf("no predecessor must not be reported as alive")
+	}
+	if calls != 1 {
+		t.Fatalf("expected exactly one probe when nothing is running, got %d", calls)
+	}
+	if time.Since(start) > 500*time.Millisecond {
+		t.Fatalf("the no-predecessor path must not sleep")
+	}
+}
+
+// TestWaitForPredecessorIgnoresOwnPid: a pidfile naming THIS process (a crashed
+// predecessor whose pid we were handed, or a re-exec) is not a predecessor.
+func TestWaitForPredecessorIgnoresOwnPid(t *testing.T) {
+	orig := aliveProbe
+	defer func() { aliveProbe = orig }()
+	aliveProbe = func() (int, bool) { return os.Getpid(), true }
+
+	if _, stillAlive := waitForPredecessor(time.Minute, time.Second, nil); stillAlive {
+		t.Fatalf("our own pid must never count as a predecessor")
+	}
+}
@@ -61,6 +61,25 @@ type Engine struct {
 	current  option.Options // options the running Box was built from
 	hash     string         // stable hash of current (hex sha256 of canonical JSON)

+	// instanceCancel cancels the context THIS box was built on — its own
+	// cancellable child of e.ctx, one per box (see newBox). box.New does not
+	// derive a cancellable context of its own, so without this every goroutine
+	// inside a retired box that waits on ctx.Done() waits forever. Upstream's
+	// runner does exactly this and calls cancel before Close
+	// (cmd/sing-box/cmd_run.go). nil when nothing is running.
+	instanceCancel context.CancelFunc
+	// instanceGen is the 1-based sequence number of the running box, bumped on
+	// every adopted swap. It is what names a generation in the log and in
+	// PendingCloses when a shutdown overruns its budget.
+	instanceGen uint64
+
+	// pending holds the retirements in flight (see teardown.go). Its OWN leaf
+	// lock, deliberately not mu: PendingCloses is read by the status path, and
+	// the moment that read matters most is while an apply is holding mu waiting
+	// out a shutdown that will not finish.
+	pendingMu sync.Mutex
+	pending   []*pendingClose
+
 	// defaultLogWriter, when non-nil, is handed to EVERY box.New this engine
 	// performs (box.Options.DefaultLogWriter): the daemon points it at its
 	// long-lived logsink once, and each Apply-swapped box then logs into that
@@ -184,12 +203,18 @@ func (e *Engine) applyLocked(opts option.Options) (bool, error) {
 		return false, nil
 	}

-	// (2) build + validate. box.New constructs and validates every adapter.
-	nb, err := box.New(box.Options{
-		Context:          e.ctx,
-		Options:          opts,
-		DefaultLogWriter: e.defaultLogWriter,
-	})
+	// (1b) Do not stack generations. A superseded box whose shutdown overran its
+	// budget is still holding its outbound connections and its WireGuard devices;
+	// building another one on top of it is how one process ends up carrying four
+	// live generations. Give any abandoned teardown one more budget to finish
+	// BEFORE we create anything (see teardown.go awaitAbandonedLocked). Placed
+	// after the hash gate on purpose: a no-op reconcile — cron, every minute —
+	// must stay free.
+	stuck := e.awaitAbandonedLocked()
+
+	// (2) build + validate on its OWN cancellable context. box.New constructs and
+	// validates every adapter.
+	nb, nbCancel, err := e.newBox(opts)
 	if err != nil {
 		// Validation failed: keep the running instance, do not swap.
 		return false, E.Cause(err, "create instance")
@@ -211,8 +236,15 @@ func (e *Engine) applyLocked(opts option.Options) (bool, error) {
 	//     close-old-then-start-new DIRECTLY — skipping the stall. (box.New above
 	//     already validated opts, so we never tear the old box down for a config
 	//     that would fail to build.)
-	if e.instance != nil && sharesCacheFileLock(e.current, opts) {
-		return e.closeOldThenStart(nb, opts, newHash)
+	//
+	// A generation that is STILL abandoned after the wait above forces the same
+	// path for a different reason: start-new-first would put a second LIVE box
+	// alongside a third that refuses to die, all three contending for the same
+	// tproxy port, the same cache file and — the expensive one — the same
+	// WireGuard private keys. Closing the current box first keeps the process to
+	// at most one live instance plus the abandoned one.
+	if e.instance != nil && (stuck > 0 || sharesCacheFileLock(e.current, opts)) {
+		return e.closeOldThenStart(nb, nbCancel, opts, newHash)
 	}

 	// (3b) start-new-first (zero-downtime when there is no resource conflict).
@@ -223,25 +255,59 @@ func (e *Engine) applyLocked(opts option.Options) (bool, error) {
 		// close-old-then-start-new then. Any OTHER Start failure keeps today's
 		// behavior: discard the new box, keep the old one running.
 		if e.instance == nil || !isSwapConflict(err) {
+			nbCancel()
 			_ = nb.Close()
 			return false, E.Cause(err, "start instance")
 		}
-		return e.closeOldThenStart(nb, opts, newHash)
+		return e.closeOldThenStart(nb, nbCancel, opts, newHash)
 	}

 	// Swap succeeded. The previously running config becomes last-good.
-	old := e.instance
-	if old != nil {
+	if e.instance != nil {
 		e.lastGood = e.current
 		e.hasLastGood = true
-		_ = old.Close()
+		// Retire the old generation under the hard budget. Its error — including
+		// ErrCloseTimeout — is deliberately NOT returned: the new box is started
+		// and carrying traffic, so this apply SUCCEEDED, and failing it here would
+		// abort the caller's netplane stage and leave a stale ruleset loaded over a
+		// perfectly healthy engine. An abandoned generation is surfaced through
+		// PendingCloses() (critical warning in `shaterd status` and the panel) and
+		// an ERROR line naming it — visible, but not mistaken for a failed apply.
+		_ = e.retireLocked(e.instance, e.instanceCancel, e.instanceGen, e.hash)
 	}
-	e.instance = nb
-	e.current = opts
-	e.hash = newHash
+	e.adoptLocked(nb, nbCancel, opts, newHash)
 	return true, nil
 }

+// newBox builds a box on its OWN cancellable child of the engine context and
+// returns the cancel alongside it. Every goroutine the box starts inherits that
+// context, so cancelling it is what unwinds the ones Close does not reach; see
+// teardown.go for why the shared, never-cancelled context was the defect.
+func (e *Engine) newBox(opts option.Options) (*box.Box, context.CancelFunc, error) {
+	ctx, cancel := context.WithCancel(e.ctx)
+	b, err := box.New(box.Options{
+		Context:          ctx,
+		Options:          opts,
+		DefaultLogWriter: e.defaultLogWriter,
+	})
+	if err != nil {
+		cancel()
+		return nil, nil, err
+	}
+	return b, cancel, nil
+}
+
+// adoptLocked installs a started box as THE running instance and gives it the
+// next generation number. Caller holds e.mu and has already retired whatever was
+// running before.
+func (e *Engine) adoptLocked(b *box.Box, cancel context.CancelFunc, opts option.Options, hash string) {
+	e.instanceGen++
+	e.instance = b
+	e.instanceCancel = cancel
+	e.current = opts
+	e.hash = hash
+}
+
 // closeOldThenStart is the close-old-then-start-new swap. It is taken both
 // proactively (the incoming config shares the running box's cache_file lock, so
 // start-new-first cannot work) and reactively (start-new-first hit a swap
@@ -256,27 +322,36 @@ func (e *Engine) applyLocked(opts option.Options) (bool, error) {
 // already validated by the caller's box.New, so the rebuild below is expected to
 // succeed; the restore path guards the rare case it does not. The caller holds
 // e.mu.
-func (e *Engine) closeOldThenStart(discard *box.Box, opts option.Options, newHash string) (bool, error) {
+func (e *Engine) closeOldThenStart(discard *box.Box, discardCancel context.CancelFunc, opts option.Options, newHash string) (bool, error) {
 	// (a) Drop the pre-built box (a Start-failed box cannot be restarted, and the
 	// proactively-built one must not hold the cache_file lock while we rebuild).
+	// It was never adopted, so it is not a generation — close it inline, but
+	// cancel its context first exactly like a retired one.
+	if discardCancel != nil {
+		discardCancel()
+	}
 	if discard != nil {
 		_ = discard.Close()
 	}

-	// (b) Free the port + cache_file lock by closing the old instance. Remember
-	// its config so we can restore it if the fresh box cannot come up.
+	// (b) Free the port + cache_file lock by closing the old instance, under the
+	// hard budget. Remember its config so we can restore it if the fresh box
+	// cannot come up. An overrun here is reported by retireLocked and tracked in
+	// PendingCloses; we still proceed, because the resources it was supposed to
+	// free are exactly what the operator is waiting on.
 	prevOpts := e.current
 	prevHash := e.hash
 	prevLastGood := e.lastGood
 	prevHasLastGood := e.hasLastGood
-	_ = e.instance.Close()
-	e.instance = nil
+	_ = e.retireLocked(e.instance, e.instanceCancel, e.instanceGen, prevHash)
+	e.instance, e.instanceCancel = nil, nil

 	// (c) Build a FRESH box for opts (the discarded one cannot be reused).
-	nb2, err := box.New(box.Options{Context: e.ctx, Options: opts, DefaultLogWriter: e.defaultLogWriter})
+	nb2, cancel2, err := e.newBox(opts)
 	if err == nil {
 		err = nb2.Start()
 		if err != nil {
+			cancel2()
 			_ = nb2.Close()
 		}
 	}
@@ -284,28 +359,25 @@ func (e *Engine) closeOldThenStart(discard *box.Box, opts option.Options, newHas
 		// (d) Success: the old config we just closed becomes last-good.
 		e.lastGood = prevOpts
 		e.hasLastGood = true
-		e.instance = nb2
-		e.current = opts
-		e.hash = newHash
+		e.adoptLocked(nb2, cancel2, opts, newHash)
 		return true, nil
 	}

 	// (e) The fresh box could not come up and the old one is already closed —
 	// interception is currently down. Try to RESTORE the previous config so we
 	// do not leave the tunnel dead.
-	rb, rerr := box.New(box.Options{Context: e.ctx, Options: prevOpts, DefaultLogWriter: e.defaultLogWriter})
+	rb, rcancel, rerr := e.newBox(prevOpts)
 	if rerr == nil {
 		rerr = rb.Start()
 		if rerr != nil {
+			rcancel()
 			_ = rb.Close()
 		}
 	}
 	if rerr == nil {
 		// Old config restored: keep current/hash/last-good exactly as they were
 		// (do NOT advance them). Report that opts was not applied.
-		e.instance = rb
-		e.current = prevOpts
-		e.hash = prevHash
+		e.adoptLocked(rb, rcancel, prevOpts, prevHash)
 		e.lastGood = prevLastGood
 		e.hasLastGood = prevHasLastGood
 		return false, E.Cause(err, "start instance (config not applied; previous config restored)")
@@ -314,7 +386,7 @@ func (e *Engine) closeOldThenStart(discard *box.Box, opts option.Options, newHas
 	// Restore ALSO failed: the engine is now STOPPED. e.instance stays nil (never
 	// pointing at a closed box). The fail-closed nft kill-switch keeps the LAN
 	// safe (no unproxied leak) even though interception is down.
-	e.instance = nil
+	e.instance, e.instanceCancel = nil, nil
 	e.hash = ""
 	return false, E.Cause(E.Errors(err, rerr), "start instance failed and could not restore previous config; engine stopped")
 }
@@ -410,14 +482,26 @@ func (e *Engine) HasLastGood() bool {
 }

 // Close stops the running instance, if any. It is idempotent.
+//
+// Unlike the swap path, this one DOES return ErrCloseTimeout: here the shutdown
+// is the whole operation, so "it did not stop" is the result, not a footnote. The
+// caller (apply.Teardown) records it as the teardown's error while still
+// completing the netplane teardown — the data plane must come down even when a
+// box will not.
 func (e *Engine) Close() error {
 	e.mu.Lock()
 	defer e.mu.Unlock()
 	if e.instance == nil {
+		// Nothing running, but a previously abandoned generation may still be:
+		// give it a last budget so a teardown followed by a restart does not carry
+		// the leak across.
+		if e.awaitAbandonedLocked() > 0 {
+			return ErrCloseTimeout
+		}
 		return nil
 	}
-	err := e.instance.Close()
-	e.instance = nil
+	err := e.retireLocked(e.instance, e.instanceCancel, e.instanceGen, e.hash)
+	e.instance, e.instanceCancel = nil, nil
 	e.hash = ""
 	return err
 }
@@ -0,0 +1,344 @@
+package engine
+
+// Bounded, deterministic teardown of a SUPERSEDED sing-box instance.
+//
+// # The defect this file exists for
+//
+// sing-box's Box has no live reload, so every applied config change builds a
+// fresh Box and retires the old one (see engine.go). Retiring it used to be one
+// line — `_ = old.Close()` — and that line had three separate problems, all of
+// which had to be true at once for the observed failure:
+//
+//  1. NO CANCELLATION. Upstream's own runner gives every Box its OWN cancellable
+//     context and calls cancel() BEFORE Close (cmd/sing-box/cmd_run.go:136 and
+//     :188-190). This engine handed EVERY box.New the one shared, never-cancelled
+//     context built in New (engine.go), and box.New does not derive a cancellable
+//     child of what it is given (box.go: `ctx := options.Context` and nothing
+//     else). So every goroutine inside a retired box that would have stopped on
+//     context cancellation simply never stopped, and only what each adapter's
+//     Close() explicitly tears down actually went away.
+//
+//  2. NO TIME BUDGET. Box.Close walks its subsystems SEQUENTIALLY and waits for
+//     each one forever; the only thing watching the clock is a taskmonitor that
+//     prints "close endpoint/wireguard[...] take too much time to finish!" after
+//     C.StopTimeout and then keeps waiting anyway (common/taskmonitor/monitor.go).
+//     A wireguard endpoint that will not come down therefore stalls the rest of
+//     the close list, and every subsystem AFTER it in the walk is never reached.
+//
+//  3. NO EVIDENCE. The error was discarded (`_ =`) and the pointer to the old box
+//     was overwritten in the same breath, so after the swap nothing in the process
+//     could tell — or even ask — whether the previous generation had actually
+//     died. On the router this accumulated: up to four generations logging side by
+//     side inside one shaterd process, each with its own outbound connections and,
+//     worse, its own WireGuard devices. Two devices sharing a private key evict
+//     each other at the peer, which is exactly the fault generate/wgdedup.go
+//     removes WITHIN a config — reintroduced here BETWEEN generations.
+//
+// # The contract now
+//
+//   - Every box gets its own cancellable context, cancelled before Close.
+//   - Close runs against a HARD budget (closeBudget). When it expires the apply
+//     moves on: the new box is already started and serving, and blocking the
+//     control plane on a shutdown that is not going to finish would only add an
+//     unresponsive panel to the problem.
+//   - An overrun is LOUD, not swallowed: an ERROR line naming the generation, and
+//     a critical entry in the operator-facing warning set for as long as the
+//     abandoned generation is still running (see PendingCloses).
+//   - Generations do not stack: before an apply builds another instance it gives
+//     any abandoned teardown one more budget to finish, and if one is still alive
+//     the swap is forced onto the close-old-then-start-new path so the process
+//     never holds two LIVE boxes on top of an abandoned one.
+
+import (
+	"errors"
+	"io"
+	"sync"
+	"time"
+
+	box "github.com/sagernet/sing-box"
+)
+
+// ErrCloseTimeout reports that a superseded instance did not finish shutting
+// down inside the close budget and has been ABANDONED (it may still be running).
+//
+// It is deliberately NOT propagated out of Apply. The swap it accompanies
+// succeeded — the new box is built, started and carrying traffic — and returning
+// a failure there would make the caller abort the netplane stage and, with no
+// engine fault to point at, leave a stale ruleset loaded over a healthy engine.
+// The fact is surfaced through PendingCloses() instead, which reaches
+// `shaterd status` and the panel and clears itself the moment the shutdown
+// finally completes. Close()/Teardown DO return it: there the failure to stop is
+// the whole point of the operation.
+var ErrCloseTimeout = errors.New("engine instance did not shut down within the close budget")
+
+// defaultCloseBudget is the HARD wall-clock budget for retiring one superseded
+// box.
+//
+// Five seconds, for three reasons that all point at the same number:
+//
+//   - it is exactly sing-box's own C.StopTimeout — the threshold at which the
+//     engine itself declares a single lifecycle stop to be taking too long. A
+//     close that blows past the budget upstream considers excessive is by
+//     definition not a slow close, it is a stuck one;
+//   - it stays under C.FatalStopTimeout (10s), which is when upstream's CLI gives
+//     up and calls the process unclosable. We want to have already reacted by
+//     then;
+//   - it is paid AFTER the replacement box is started and serving, and only on an
+//     apply that really changed something (the hash gate makes a no-op reconcile
+//     free), so the worst case is five seconds added to one apply — not to the
+//     every-minute cron reconcile, and never to a status poll.
+const defaultCloseBudget = 5 * time.Second
+
+// closeBudget/closeBoxFn are package-wide seams. Production never touches them;
+// tests in this package and in shater/apply install a short budget and a
+// deliberately hung closer to exercise the abandoned-generation path without
+// standing up a box that really refuses to die.
+var (
+	seamMu      sync.RWMutex
+	closeBudget = defaultCloseBudget
+	closeBoxFn  = func(c io.Closer) error { return c.Close() }
+)
+
+// SetCloseBudget overrides the close budget and returns a function restoring the
+// previous value. TEST SEAM — production uses defaultCloseBudget.
+func SetCloseBudget(d time.Duration) (restore func()) {
+	seamMu.Lock()
+	prev := closeBudget
+	closeBudget = d
+	seamMu.Unlock()
+	return func() {
+		seamMu.Lock()
+		closeBudget = prev
+		seamMu.Unlock()
+	}
+}
+
+// SetBoxCloser overrides HOW a superseded instance is closed and returns a
+// function restoring the previous closer. TEST SEAM — production calls
+// Box.Close. A closer that never returns is how the abandoned-generation path is
+// tested.
+func SetBoxCloser(fn func(io.Closer) error) (restore func()) {
+	seamMu.Lock()
+	prev := closeBoxFn
+	closeBoxFn = fn
+	seamMu.Unlock()
+	return func() {
+		seamMu.Lock()
+		closeBoxFn = prev
+		seamMu.Unlock()
+	}
+}
+
+func currentCloseBudget() time.Duration {
+	seamMu.RLock()
+	defer seamMu.RUnlock()
+	return closeBudget
+}
+
+func currentBoxCloser() func(io.Closer) error {
+	seamMu.RLock()
+	defer seamMu.RUnlock()
+	return closeBoxFn
+}
+
+// pendingClose tracks ONE retirement in flight. It lives in Engine.pending from
+// the moment the retirement starts until Close returns, and `abandoned` records
+// whether the budget expired while it was still running — i.e. whether this
+// generation is a leak the operator must be told about, or merely a close that
+// is in progress under a lock the caller is holding anyway.
+type pendingClose struct {
+	gen       uint64
+	hash      string        // config hash the retired box was built from
+	since     time.Time
+	done      chan struct{} // closed when the underlying Close finally returns
+	abandoned bool          // guarded by Engine.pendingMu
+	finished  bool          // guarded by Engine.pendingMu
+}
+
+// StuckClose is the read-side view of a superseded generation that overran the
+// close budget and is STILL running inside this process.
+type StuckClose struct {
+	// Generation is the 1-based sequence number of the box that will not die.
+	// It matches nothing in the sing-box log by itself, but it lets two status
+	// reads tell "the same old generation" from "another one just leaked".
+	Generation uint64
+	// Hash is the config hash the leaked box was built from — the same value
+	// `shaterd status` reported as `hash` while that config was the running one.
+	// It is what connects "an engine is stuck" to WHICH configuration is stuck.
+	Hash string
+	// Since is when its shutdown was started.
+	Since time.Time
+	// Elapsed is how long it has been shutting down, as of the read.
+	Elapsed time.Duration
+}
+
+// retireLocked tears down a superseded box under the close budget. The caller
+// holds e.mu and must clear e.instance itself.
+//
+// Order matters and mirrors upstream: cancel the box's context FIRST so every
+// goroutine keyed on it unwinds, and only then call Close, which is what actually
+// releases the listeners, the WireGuard devices and the cache-file lock.
+//
+// Returns nil on a clean shutdown, ErrCloseTimeout when the budget expired (the
+// close keeps running in the background and is reaped when it finishes), or the
+// close's own error.
+func (e *Engine) retireLocked(b *box.Box, cancel func(), gen uint64, hash string) error {
+	if cancel != nil {
+		cancel()
+	}
+	if b == nil {
+		return nil
+	}
+
+	pc := &pendingClose{gen: gen, hash: hash, since: time.Now(), done: make(chan struct{})}
+	e.pendingMu.Lock()
+	e.pending = append(e.pending, pc)
+	e.pendingMu.Unlock()
+
+	closeFn := currentBoxCloser()
+	errc := make(chan error, 1)
+	go func() {
+		err := closeFn(b)
+		e.pendingMu.Lock()
+		pc.finished = true
+		wasAbandoned := pc.abandoned
+		e.removePendingLocked(pc)
+		e.pendingMu.Unlock()
+		if wasAbandoned && e.log != nil {
+			// The counterpart of the ERROR below: an operator who saw the leak
+			// reported must be able to see it resolve without restarting anything.
+			e.log.Warn("engine generation ", gen, " finally finished shutting down after ",
+				time.Since(pc.since).Round(time.Millisecond), "; it is no longer running")
+		}
+		close(pc.done)
+		errc <- err
+	}()
+
+	budget := currentCloseBudget()
+	timer := time.NewTimer(budget)
+	defer timer.Stop()
+	select {
+	case err := <-errc:
+		return err
+	case <-timer.C:
+	}
+
+	// The budget expired. Claim the generation as abandoned — unless the close
+	// happened to land in the same instant, in which case there is nothing to
+	// report and we take its real result.
+	e.pendingMu.Lock()
+	abandoned := !pc.finished
+	if abandoned {
+		pc.abandoned = true
+	}
+	e.pendingMu.Unlock()
+	if !abandoned {
+		return <-errc
+	}
+
+	if e.log != nil {
+		e.log.Error("engine generation ", gen, " (config ", shortHash(hash), ") did NOT shut down within ", budget,
+			" and has been ABANDONED: it may still hold its outbound connections and its ",
+			"WireGuard devices (two devices with the same private key evict each other at ",
+			"the peer), and it keeps writing to this log. The new configuration is running; ",
+			"restart shaterd if this generation never clears.")
+	}
+	return ErrCloseTimeout
+}
+
+// shortHash renders a config hash the way the log and the panel both need it:
+// enough to identify the configuration, short enough to read in a syslog line.
+func shortHash(h string) string {
+	if len(h) < 12 {
+		return "unknown"
+	}
+	return h[:12]
+}
+
+// removePendingLocked drops pc from the pending list. Caller holds pendingMu.
+func (e *Engine) removePendingLocked(pc *pendingClose) {
+	for i, p := range e.pending {
+		if p == pc {
+			e.pending = append(e.pending[:i], e.pending[i+1:]...)
+			return
+		}
+	}
+}
+
+// pendingSnapshot copies the in-flight retirements. Leaf lock only.
+func (e *Engine) pendingSnapshot() []*pendingClose {
+	e.pendingMu.Lock()
+	defer e.pendingMu.Unlock()
+	return append([]*pendingClose(nil), e.pending...)
+}
+
+// PendingCloses lists the superseded generations that overran their close budget
+// and are still running. Empty is the healthy answer.
+//
+// It takes ONLY the pending leaf lock — never e.mu — so the panel and
+// `shaterd status` can report a leaking teardown even while the apply that
+// produced it is still holding the engine mutex. That is deliberate: the one
+// moment this information matters most is while an apply is stalled behind a
+// shutdown that will not finish.
+func (e *Engine) PendingCloses() []StuckClose {
+	now := time.Now()
+	e.pendingMu.Lock()
+	defer e.pendingMu.Unlock()
+	out := make([]StuckClose, 0, len(e.pending))
+	for _, p := range e.pending {
+		if !p.abandoned {
+			continue
+		}
+		out = append(out, StuckClose{
+			Generation: p.gen,
+			Hash:       p.hash,
+			Since:      p.since,
+			Elapsed:    now.Sub(p.since),
+		})
+	}
+	return out
+}
+
+// Generations reports how many sing-box instances this process is carrying: the
+// running one (0 or 1) plus every superseded generation whose shutdown has not
+// finished. ONE is the healthy answer for a started engine; anything above it is
+// the leak this file exists to make impossible to hide.
+func (e *Engine) Generations() int {
+	e.mu.Lock()
+	live := 0
+	if e.instance != nil {
+		live = 1
+	}
+	e.mu.Unlock()
+	e.pendingMu.Lock()
+	defer e.pendingMu.Unlock()
+	return live + len(e.pending)
+}
+
+// awaitAbandonedLocked gives every already-abandoned generation ONE more close
+// budget to finish, and returns how many are still running afterwards. The caller
+// holds e.mu and is about to build another instance.
+//
+// This is what keeps generations from stacking. Without it, a box that will not
+// come down means the NEXT apply quietly adds a third instance to the process,
+// and the one after that a fourth — which is precisely the accumulation observed
+// on the router. The wait is bounded by one budget for the whole set (not per
+// entry), so a permanently stuck generation costs a single extra budget on an
+// apply that actually changes the config, and nothing at all on the every-minute
+// no-op reconcile, which never gets past the hash gate.
+func (e *Engine) awaitAbandonedLocked() int {
+	pending := e.pendingSnapshot()
+	if len(pending) == 0 {
+		return 0
+	}
+	deadline := time.NewTimer(currentCloseBudget())
+	defer deadline.Stop()
+	for _, p := range pending {
+		select {
+		case <-p.done:
+		case <-deadline.C:
+			return len(e.PendingCloses())
+		}
+	}
+	return len(e.PendingCloses())
+}
@@ -0,0 +1,247 @@
+// lx: pins the invariant that a chain copy of an AmneziaWG node carries the
+// SAME obfuscation parameters as the base node, field for field.
+//
+// Context: on the router a node placed behind an egress hop
+// (egress:ewan -> node:awgout, tag chain-<name>-h1) handshook forever and never
+// passed a transport packet, while the same node standalone was healthy. One
+// candidate explanation was that rebuildNode loses AWG params on the copy — it
+// does not (both the base and the copy go through the same
+// builder.wireguardEndpoint / amneziaOptions, the copy differing only in Tag and
+// DialerOptions.Detour), and this test holds that line. The real cause was the
+// bind swap the detour triggers: a detour makes Endpoint.Start pick ClientBind
+// instead of conn.StdNetBind, and ClientBind unconditionally overwrote bytes 1-3
+// of every datagram, shredding the h4 magic header of transport packets. See
+// transport/wireguard/client_bind.go (hasReserved) and its regression tests.
+//
+// Honest note: this test passes both before and after that fix — it is a pin on
+// a path that was never broken, not the reproducer for the bug.
+//
+// Unlike generate_test.go this file is NOT Linux-gated: it stops at
+// GenerateWithWarnings and never calls engine.Apply / box.New, so it needs no
+// routing_mark validation and runs on every platform.
+package generate
+
+import (
+	"encoding/base64"
+	"fmt"
+	"net/url"
+	"reflect"
+	"testing"
+
+	"github.com/sagernet/sing-box/option"
+	"github.com/sagernet/sing-box/shater/model"
+)
+
+// awgKey returns a valid 32-byte base64 WireGuard key seeded by fill. Local to
+// this file so it does not depend on the Linux-only suite's helpers.
+func awgKey(fill byte) string {
+	b := make([]byte, 32)
+	for i := range b {
+		b[i] = fill + byte(i)
+	}
+	return base64.StdEncoding.EncodeToString(b)
+}
+
+// TestChainCopyPreservesAmneziaWGOptions builds a chain whose terminal hop is an
+// AmneziaWG node and asserts the chain copy's AmneziaWGOptions equals the base
+// endpoint's, comparing every field of the struct (so a field added later
+// without being mapped in amneziaOptions is caught here too).
+func TestChainCopyPreservesAmneziaWGOptions(t *testing.T) {
+	priv := awgKey(1)
+	pub := awgKey(9)
+	// Every knob the option struct carries that the share-link can express:
+	// jc/jmin/jmax, s1-s3, ranged h1-h4 (AWG 2.0), i1-i5. s4 is deliberately left
+	// unset (0) — that is the real-world shape in which the transport magic lands
+	// in bytes 0-3 and the ClientBind bug bit.
+	uri := fmt.Sprintf(
+		"awg://%s@203.0.113.10:51820?publickey=%s&address=10.13.13.2/32&allowedips=0.0.0.0/0"+
+			"&jc=4&jmin=40&jmax=70&s1=86&s2=57&s3=13"+
+			"&h1=1618116899-1618116949&h2=1795397486-1795397536&h3=3333333333&h4=1618116899-1618116949"+
+			"&i1=%s&i2=%s&i3=%s&i4=%s&i5=%s#awgout",
+		url.QueryEscape(priv), url.QueryEscape(pub),
+		url.QueryEscape("<b 0xf0>"), url.QueryEscape("<c>"), url.QueryEscape("<t>"),
+		url.QueryEscape("<r 10>"), url.QueryEscape("<b 0xab>"),
+	)
+
+	node := model.Node{Name: "awgout", Enabled: true, URI: uri}
+
+	// Two SEPARATE configs, not one. wgdedup refuses to materialise a WireGuard
+	// node twice from one private key (two devices sharing a key evict each other's
+	// session), so a single config can hold either the base endpoint or the chain
+	// copy — never both. That is also why on the router this node exists only as
+	// "chain-<name>-h1". The invariant under test is therefore cross-config: the
+	// same node, referenced directly vs referenced through a chain, must yield the
+	// same AmneziaWG parameters.
+	direct := &model.Model{
+		Globals: model.DefaultGlobals(),
+		Nodes:   []model.Node{node},
+		Rules: []model.Rule{
+			{Name: "default", Enabled: true, Order: 100, Target: "node:awgout"},
+		},
+	}
+	chained := &model.Model{
+		Globals: model.DefaultGlobals(),
+		Nodes:   []model.Node{node},
+		Egresses: []model.Egress{
+			{Name: "ewan", Type: "interface", Interface: "wan"},
+		},
+		Chains: []model.Chain{
+			{Name: "viaewan", Hops: []string{"egress:ewan", "node:awgout"}},
+		},
+		Rules: []model.Rule{
+			{Name: "default", Enabled: true, Order: 100, Target: "chain:viaewan"},
+		},
+	}
+
+	directOpts, warns, err := GenerateWithWarnings(direct)
+	if err != nil {
+		t.Fatalf("GenerateWithWarnings(direct): %v (warnings: %v)", err, warns)
+	}
+	chainedOpts, chainWarns, err := GenerateWithWarnings(chained)
+	if err != nil {
+		t.Fatalf("GenerateWithWarnings(chained): %v (warnings: %v)", err, chainWarns)
+	}
+
+	base, ok := awgOptionsForTag(directOpts, "awgout")
+	if !ok {
+		t.Fatalf("base endpoint %q not found; tags = %v (warnings: %v)",
+			"awgout", endpointTags(directOpts), warns)
+	}
+	if !base.IsSet() {
+		t.Fatalf("base endpoint has no AmneziaWG params: %+v", base)
+	}
+
+	// Absolute expectation, not just parity. A "copy == base" assertion alone is
+	// satisfied when BOTH lose a field (they share amneziaOptions), so pin the
+	// literal values the share-link carries. Adding a field to
+	// option.AmneziaWGOptions without mapping it in amneziaOptions fails the
+	// exhaustiveness check below.
+	want := option.AmneziaWGOptions{
+		Jc: 4, Jmin: 40, Jmax: 70,
+		S1: 86, S2: 57, S3: 13, S4: 0,
+		H1: "1618116899-1618116949",
+		H2: "1795397486-1795397536",
+		H3: "3333333333",
+		H4: "1618116899-1618116949",
+		I1: "<b 0xf0>", I2: "<c>", I3: "<t>", I4: "<r 10>", I5: "<b 0xab>",
+	}
+	if base != want {
+		t.Fatalf("base AmneziaWG params lost/garbled in parsing:\n got %+v\nwant %+v", base, want)
+	}
+	// Exhaustiveness: every field of the struct must be exercised above, so a
+	// newly added knob cannot slip through unmapped and unnoticed.
+	assertAWGFieldsCovered(t, want)
+
+	// The chain hop copy: buildHopWrapper tags it chain-<chain>-h<idx>. Find it as
+	// "the wireguard endpoint in the chained config" rather than hardcoding the
+	// index, so a change in hop numbering does not silently no-op this test.
+	var (
+		copyTag     string
+		copyOptions option.AmneziaWGOptions
+		found       bool
+	)
+	for _, endpoint := range chainedOpts.Endpoints {
+		wg, isWG := endpoint.Options.(*option.WireGuardEndpointOptions)
+		if !isWG {
+			continue
+		}
+		copyTag, copyOptions, found = endpoint.Tag, wg.AmneziaWGOptions, true
+		break
+	}
+	if !found {
+		t.Fatalf("no chain copy endpoint emitted; endpoint tags = %v (warnings: %v)",
+			endpointTags(chainedOpts), chainWarns)
+	}
+	if copyTag == "awgout" {
+		t.Fatalf("expected a per-hop chain copy tag, got the base tag %q", copyTag)
+	}
+
+	// Field-by-field, via reflection: any field of AmneziaWGOptions the copy fails
+	// to carry is named explicitly rather than hidden behind one struct diff.
+	baseValue := reflect.ValueOf(base)
+	copyValue := reflect.ValueOf(copyOptions)
+	for i := 0; i < baseValue.NumField(); i++ {
+		field := baseValue.Type().Field(i)
+		want := baseValue.Field(i).Interface()
+		got := copyValue.Field(i).Interface()
+		if !reflect.DeepEqual(want, got) {
+			t.Errorf("chain copy %q lost AmneziaWG field %s: got %#v, want %#v",
+				copyTag, field.Name, got, want)
+		}
+	}
+	if !t.Failed() && base != copyOptions {
+		t.Fatalf("chain copy %q AmneziaWGOptions differ from base: %+v vs %+v",
+			copyTag, copyOptions, base)
+	}
+
+	// The copy must additionally differ from the base in exactly the way the chain
+	// intends: it detours through the egress hop, the base does not.
+	baseDialer, okBase := dialerForTag(directOpts, "awgout")
+	copyDialer, okCopy := dialerForTag(chainedOpts, copyTag)
+	if !okBase || !okCopy {
+		t.Fatalf("dialer options missing: base=%v copy=%v", okBase, okCopy)
+	}
+	if baseDialer.Detour != "" {
+		t.Errorf("base endpoint must dial directly, got Detour=%q", baseDialer.Detour)
+	}
+	if copyDialer.Detour == "" {
+		t.Errorf("chain copy %q must carry the egress hop as Detour, got empty", copyTag)
+	}
+}
+
+// assertAWGFieldsCovered fails if any field of option.AmneziaWGOptions is left
+// at its zero value in the expectation, other than the ones deliberately unset:
+//
+//	S4       — kept 0 on purpose; that is the shape in which the transport magic
+//	           lands in bytes 0-3, i.e. the configuration that broke on the router.
+//	Id/Ip/Ib — WireSock masquerade sugar, mutually exclusive with an explicit I1
+//	           (device_awg.go rejects the combination), and I1 is set here.
+//
+// The point is that adding a knob to the struct without extending this test
+// turns into a failure here rather than silent non-coverage.
+func assertAWGFieldsCovered(t *testing.T, want option.AmneziaWGOptions) {
+	t.Helper()
+	deliberatelyUnset := map[string]bool{"S4": true, "Id": true, "Ip": true, "Ib": true}
+	value := reflect.ValueOf(want)
+	for i := 0; i < value.NumField(); i++ {
+		name := value.Type().Field(i).Name
+		if deliberatelyUnset[name] {
+			continue
+		}
+		if value.Field(i).IsZero() {
+			t.Errorf("AmneziaWGOptions.%s is not exercised by this test "+
+				"(add it to the share-link and to `want`, or to deliberatelyUnset)", name)
+		}
+	}
+}
+
+// awgOptionsForTag returns the AmneziaWGOptions of the wireguard endpoint tagged
+// tag.
+func awgOptionsForTag(opts option.Options, tag string) (option.AmneziaWGOptions, bool) {
+	for _, endpoint := range opts.Endpoints {
+		if endpoint.Tag != tag {
+			continue
+		}
+		wg, ok := endpoint.Options.(*option.WireGuardEndpointOptions)
+		if !ok {
+			return option.AmneziaWGOptions{}, false
+		}
+		return wg.AmneziaWGOptions, true
+	}
+	return option.AmneziaWGOptions{}, false
+}
+
+// dialerForTag returns the DialerOptions of the wireguard endpoint tagged tag.
+func dialerForTag(opts option.Options, tag string) (option.DialerOptions, bool) {
+	for _, endpoint := range opts.Endpoints {
+		if endpoint.Tag != tag {
+			continue
+		}
+		wg, ok := endpoint.Options.(*option.WireGuardEndpointOptions)
+		if !ok {
+			return option.DialerOptions{}, false
+		}
+		return wg.DialerOptions, true
+	}
+	return option.DialerOptions{}, false
+}
@@ -610,44 +610,60 @@ func splitDomainMarker(e string) (marker, value string, ok bool) {
 // reader can tell a deliberate difference from an oversight (R9.3). Verified
 // against the code on both sides, not against upstream docs — this is a fork.
 //
-//	marker      domain lists (this file)      routing rules (ruleMatchers, route.go)
+// There are now only TWO contexts, not three. A routing rule no longer classifies
+// domains at all: `dst_domain` was removed in schema v2 and a rule names its
+// destination through a `config ruleset`, so every domain entry in the system —
+// filter list, device list, inline rule-set — arrives at classifyDomainEntries
+// below. The one caller that adds something on top is inlineRulesetRule
+// (ruleset.go), which peels `regexp:` off first.
+//
+//	marker      DNS filter / device lists     inline rule-sets (inlineRulesetRule)
 //	----------  ----------------------------  --------------------------------------
 //	full:       yes — exact domain            yes — exact domain
 //	suffix:     yes — domain + subdomains     yes — domain + subdomains
 //	.example    SAME AS suffix: (dot stripped) SAME AS suffix: (dot stripped)
 //	keyword:    yes — substring               yes — substring
-//	(bare)      SUFFIX in filter/device/      EXACT domain
-//	            inline-ruleset lists
+//	(bare)      SUFFIX (domain + subdomains)  SUFFIX (domain + subdomains)
 //	regexp:     NO — warns                    yes — DomainRegex (pattern validated)
-//	geosite:    NO — warns                    NO — warns, rule matcher omitted
+//	geosite:    NO — warns                    NO — warns, entry dropped
 //
-// Two corrections this table used to get wrong, both of the "claimed a behaviour
-// that does not exist" kind:
+// A THIRD context exists and has NO row here on purpose: the body of a
+// `source=url` plain-text list. That is a hosts/one-domain-per-line FILE parsed by
+// parseDomainList (ruleset.go), not a list of typed entries, and it understands no
+// marker at all — every line is a domain plus its subdomains. An entry written in
+// the vocabulary above is dropped there (a domain cannot contain ":") and reported
+// per list by warnListEntryVocabulary. The long block above parseDomainList states
+// why unifying the two was rejected; do not read this table as covering it.
 //
-//   - `.example.com` was documented as "subdomains only" on both sides. It is not:
+// The bare-entry row is now the SAME on both sides, and that uniformity is the
+// point of schema v2 — a routing rule used to read a bare entry as an EXACT host
+// while every list read it as a suffix, a difference nothing in the UI showed.
+// model.migrate1to2 is what preserves the old meaning of existing configs: it
+// rewrites a rule's bare `dst_domain` entry as `full:` when it moves it into the
+// generated rule-set.
+//
+// Corrections this table used to get wrong, all of the "claimed a behaviour that
+// does not exist" kind:
+//
+//   - `.example.com` was documented as "subdomains only". It is not:
 //     classifyDomainEntries strips the dot, so it is a synonym of `suffix:` and
 //     matches the apex too. See the leading-dot branch above for why the synonym
 //     is kept rather than the distinction implemented.
 //   - `geosite:` was documented as a working routing matcher. It is not: the
-//     route-rule geosite field was REMOVED from this engine, so route.go warns and
-//     omits the matcher (a rule left with no other matcher is skipped entirely).
-//     Use a `config ruleset` with source=geosite.
+//     route-rule geosite field was REMOVED from this engine. Use a `config
+//     ruleset` with source=geosite and category chips.
 //
-// The two absences in the domain-list column are DELIBERATE, not gaps:
+// The two absences in the FILTER column are DELIBERATE, not gaps:
 //
-//   - regexp: an invalid regular expression is only detected when the rule is
-//     built, where it aborts box.New and takes the whole config down — the exact
-//     fail-open violation this audit spent its time removing. Supporting it here
-//     would require compiling and validating every pattern at generate time.
-//     Warning is the honest answer until that is done.
-//   - geosite: filter lists and rulesets already express geosite properly, via
+//   - regexp: an invalid regular expression aborts box.New and takes the whole
+//     config down, so it may only be accepted where every pattern is compiled and
+//     validated at generate time. Inline rule-sets do exactly that
+//     (peelDomainRegexes, ruleset.go), which is why the right-hand column says
+//     yes; the DNS-filter path has no such validation and warns instead.
+//   - geosite: filter lists and rule-sets already express geosite properly, via
 //     `source=geosite` plus category chips, which fetches the official compiled
 //     .srs. A `geosite:` entry inside an inline list would be a second, weaker
 //     path to the same feature.
-//
-// The BARE-entry difference is also deliberate and long-standing: a blocklist
-// entry is meant to cover subdomains, while a routing rule's bare entry is an
-// exact host. Both are documented at their call sites.

 // toASCIIDomain punycodes a unicode domain entry so it can match the punycoded
 // names that actually arrive in DNS queries. Lenient by design: an entry idna
@@ -185,7 +185,15 @@ func TestDNSFilterRemoteBlocklistHTTPClient(t *testing.T) {
 			{Name: "cf", Type: "doh", Address: "https://1.1.1.1/dns-query", Detour: "direct"},
 		},
 		Blocklists: []model.Blocklist{
-			{Name: "remote-ads", Enabled: true, Source: "url", URL: srv.URL, Response: "nxdomain", UpdateInterval: "24h"},
+			// The ".srs" suffix is LOAD-BEARING, not decoration: ruleSetURLIsEngineNative
+			// (ruleset.go) decides remote-vs-compiled-local by URL EXTENSION alone, and
+			// this test is about the REMOTE path — the engine fetching the compiled set
+			// itself through the direct outbound. httptest.NewServer's bare
+			// "http://127.0.0.1:<port>" has no extension, so it fell into the TEXT-list
+			// path instead: the list was downloaded by generate's own listFetcher, parsed
+			// as a hosts file and compiled into a LOCAL rule-set, which every assertion
+			// below then contradicted. Do not trim the suffix.
+			{Name: "remote-ads", Enabled: true, Source: "url", URL: srv.URL + "/blocklist.srs", Response: "nxdomain", UpdateInterval: "24h"},
 		},
 	}

@@ -289,8 +289,9 @@ func TestFailoverWarnsOnceAcrossChainCopy(t *testing.T) {
 	m := twoNodeGroupModel("failover")
 	m.Nodes = append(m.Nodes, model.Node{Name: "hop", Enabled: true, URI: ss("203.0.113.9")})
 	m.Chains = []model.Chain{{Name: "ch", Hops: []string{"node:hop", "group:g"}}}
+	m.Rulesets = []model.Ruleset{inlineDomainSet("ex", "example.com")}
 	m.Rules = []model.Rule{
-		{Name: "r", Enabled: true, DstDomain: []string{"example.com"}, Target: "chain:ch"},
+		{Name: "r", Enabled: true, DstRuleset: []string{"ex"}, Target: "chain:ch"},
 	}
 	_, warns, err := GenerateWithWarnings(m)
 	if err != nil {
@@ -67,7 +67,7 @@ package generate
 import (
 	"fmt"
 	"net/netip"
-	"sort"
+	"os"
 	"strconv"
 	"strings"
 	"time"
@@ -75,6 +75,7 @@ import (
 	"github.com/sagernet/sing-box/option"
 	"github.com/sagernet/sing/common/json/badoption"

+	"github.com/sagernet/sing-box/shater/logsink"
 	"github.com/sagernet/sing-box/shater/model"
 )

@@ -309,6 +310,19 @@ func GenerateWithWarningsAt(m *model.Model, now time.Time) (option.Options, []st
 		Route:     route,
 		DNS:       dns,
 	}
+
+	// LAST, on the finished config: fold away duplicate WireGuard devices.
+	//
+	// A WG/AWG node may be materialised several times over — the always-emitted
+	// base endpoint (outbound.go), a per-chain hop copy and a chain group hop's
+	// per-member copy (chain.go) — and unlike a TCP proxy copy, each of those is a
+	// real device holding the SAME private key. A WireGuard peer keeps one session
+	// per key, so the copies evict each other and none of them passes traffic:
+	// every chain containing a WG node was dead for exactly this reason. Running
+	// on the assembled opts (rather than at each producer) is what makes the
+	// guarantee hold for all three paths and any future fourth. See wgdedup.go.
+	b.dedupWireGuardEndpoints(&opts)
+
 	return opts, b.warnings, nil
 }

@@ -333,13 +347,31 @@ func Warnings(m *model.Model) []string {
 // it, so a level there could never be honoured and would only invite the reader to
 // think it was. Note that an EMPTY Level on an ENABLED log means LevelTrace, not
 // "default" — so Disabled must never be emitted speculatively.
+//
+// DisableColor is the OTHER half of the engine's log block, and it is not
+// cosmetic. log/log.go turns it into Formatter.DisableColors, which is what
+// stops aurora from painting the level word and the per-connection ID. Left at
+// its zero value it painted them, and the engine's writer is the daemon's shared
+// logsink whose syslog half is procd's stderr — so `logread` filled up with
+//
+//	ESC[31mERRORESC[0m[0026] [ESC[38;5;193m…ESC[0m 70ms] dns: exchange failed …
+//
+// and `grep ERROR` stopped matching. Colour is now emitted only when that
+// destination is a terminal (an interactive `shaterd run`); see
+// shater/logsink/color.go.
 func logOptions(s string) *option.LogOptions {
 	if logSilenced(s) {
 		return &option.LogOptions{Disabled: true}
 	}
-	return &option.LogOptions{Level: logLevel(s)}
+	return &option.LogOptions{Level: logLevel(s), DisableColor: !logColorAllowed()}
 }

+// logColorAllowed reports whether the engine's log lines may carry ANSI colour.
+// A package var so tests pin it instead of depending on how the test binary's
+// stderr happens to be wired; production always asks the real stderr, which is
+// the sink's syslog half.
+var logColorAllowed = func() bool { return logsink.IsTTY(os.Stderr) }
+
 // logSilenced reports whether the operator asked for no log at all.
 //
 // The vocabulary lives in model.SilentLogLevels rather than here because
@@ -487,16 +519,21 @@ func validPortRange(s string) bool {
 // the classifier actually drops.
 //
 // `suffix:` belongs here. The list used to omit it on the theory that suffix:/
-// regexp: are route-rule-only and are peeled off by ruleMatchers before the shared
-// classifier sees them. That is true of route.go — which handles a bare `suffix:`
-// in its own switch and warns there, so this list cannot double-report — but NOT
-// of the classifier itself: classifyDomainEntries has a `case "suffix"`, and every
-// other caller (devices.go, dnsfilter.go) reaches it directly. A lone `suffix:`
-// arriving that way was dropped by add()'s empty-value guard and reported by
-// nobody, which is the one outcome this pair of functions exists to prevent.
+// regexp: were route-rule-only and were peeled off before the shared classifier
+// saw them. classifyDomainEntries has a `case "suffix"`, and every caller
+// (devices.go, dnsfilter.go, inlineRulesetRule) reaches it directly, so a lone
+// `suffix:` was dropped by add()'s empty-value guard and reported by nobody —
+// the one outcome this pair of functions exists to prevent. (The route-rule half
+// of that old reasoning is gone entirely: `dst_domain` was removed in schema v2,
+// so route.go classifies no domains at all and there is no second warner to
+// double-report with.)
 //
-// `regexp:` is correctly absent: the classifier has no case for it, so a bare
-// `regexp:` is an UNRECOGNISED prefix and is reported by unrecognisedDomainPrefix.
+// `regexp:` is correctly absent, for two different reasons depending on the
+// caller. In a filter/device list the classifier has no case for it, so a bare
+// `regexp:` is an UNRECOGNISED prefix reported by unrecognisedDomainPrefix. In an
+// inline rule-set peelDomainRegexes (ruleset.go) strips every `regexp:` entry
+// BEFORE this predicate runs and reports a valueless one itself, so it can never
+// reach here either way.
 var domainMarkers = []string{"keyword:", "full:", "suffix:", "."}

 // isDomainMarkerOnly reports whether an entry is a bare classification marker
@@ -564,15 +601,14 @@ func networkList(tcp, udp bool) option.NetworkList {
 	}
 }

-// sortedByOrder returns rule indices sorted by (Order, original index) so the
+// sortedRuleIndices returns rule indices sorted by (Order, original index) so the
 // engine's first-match order matches the model's declared order, stably.
+//
+// The implementation is model.SortedRuleIndices: the reachability analysis that
+// decides which rules can never fire has to walk the rules in EXACTLY this order
+// to be right, and it lives in model (the leaf both generate and the panel API
+// import). Two copies of "what order do rules run in" is precisely the drift that
+// would make the warning and the panel badge disagree.
 func sortedRuleIndices(rules []model.Rule) []int {
-	idx := make([]int, len(rules))
-	for i := range rules {
-		idx[i] = i
-	}
-	sort.SliceStable(idx, func(a, b int) bool {
-		return rules[idx[a]].Order < rules[idx[b]].Order
-	})
-	return idx
+	return model.SortedRuleIndices(rules)
 }
@@ -227,6 +227,12 @@ func TestAllReachableProtocols(t *testing.T) {
 			{Name: "vmess-ws", Enabled: true, URI: vmessLink},
 			{Name: "trojan1", Enabled: true, URI: "trojan://password@example.com:443?sni=example.com#trojan1"},
 			{Name: "ss1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.5:8388#ss1"},
+			// QUIC protocols travel the same road: share-link -> parse.Proxy ->
+			// outbound -> box.New. Before parse grew these two branches the links
+			// were dropped at the parser and the outbounds below never existed.
+			{Name: "hy2-1", Enabled: true, URI: "hysteria2://secret@example.com:8443/?sni=example.com&alpn=h3#hy2-1"},
+			{Name: "hy2-alias", Enabled: true, URI: "hy2://secret@example.org#hy2-alias"},
+			{Name: "tuic1", Enabled: true, URI: "tuic://22222222-2222-2222-2222-222222222222:secret@example.com:443/?sni=example.com&alpn=h3&congestion_control=cubic&udp_relay_mode=native#tuic1"},
 		},
 	}
 	opts, warns, changed := applyAndClose(t, m)
@@ -236,16 +242,34 @@ func TestAllReachableProtocols(t *testing.T) {
 	if len(warns) != 0 {
 		t.Fatalf("unexpected warnings: %v", warns)
 	}
-	for _, tag := range []string{"vless-ws", "vless-grpc", "vmess-ws", "trojan1", "ss1"} {
+	for _, tag := range []string{"vless-ws", "vless-grpc", "vmess-ws", "trojan1", "ss1", "hy2-1", "hy2-alias", "tuic1"} {
 		if findOutbound(opts, tag) == nil {
 			t.Fatalf("outbound %q not emitted", tag)
 		}
 	}
+	// The hy2 alias link carries no port and no sni: parse must have supplied the
+	// scheme default (443) and the server name, or the node would dial nowhere.
+	if ob := findOutbound(opts, "hy2-alias"); ob != nil {
+		o, ok := ob.Options.(*option.Hysteria2OutboundOptions)
+		if !ok {
+			t.Fatalf("hy2-alias options type = %T", ob.Options)
+		}
+		if o.ServerPort != 443 {
+			t.Fatalf("hy2-alias port = %d, want the scheme default 443", o.ServerPort)
+		}
+		if o.TLS == nil || o.TLS.ServerName != "example.org" {
+			t.Fatalf("hy2-alias tls = %+v", o.TLS)
+		}
+		if o.TLS.UTLS != nil {
+			t.Fatalf("uTLS must never reach a QUIC outbound: %+v", o.TLS.UTLS)
+		}
+	}
 }

 // --- White-box: hysteria2 / tuic / shadowtls option mapping validates. -------
-// These protocols are not yet produced by parse.ParseShareLink, so we drive the
-// mapping directly with synthetic parse.Proxy values and validate via box.New.
+// shadowtls is not produced by parse.ParseShareLink (hysteria2 and tuic now
+// are, see TestAllReachableProtocols), so the mapping is driven directly with
+// synthetic parse.Proxy values and validated via box.New.

 func TestQUICAndShadowTLSMappingValidates(t *testing.T) {
 	b := newBuilder(&model.Model{Globals: model.DefaultGlobals()})
@@ -323,8 +347,9 @@ func TestEgressDPISpoofValidates(t *testing.T) {
 		Globals:  model.DefaultGlobals(),
 		Inbounds: []model.Inbound{{Name: "lan", Enabled: true, Type: "tproxy", TproxyPort: 12366, TCP: true, UDP: true}},
 		Egresses: []model.Egress{{Name: "spf", Type: "direct", DPI: "spoof"}},
+		Rulesets: []model.Ruleset{inlineDomainSet("blocked", "blocked.example")},
 		Rules: []model.Rule{
-			{Name: "spoof-rule", Enabled: true, Order: 10, DstDomain: []string{"blocked.example"}, Target: "egress:spf"},
+			{Name: "spoof-rule", Enabled: true, Order: 10, DstRuleset: []string{"blocked"}, Target: "egress:spf"},
 		},
 	}
 	opts, warns, changed := applyAndClose(t, m)
@@ -346,8 +371,9 @@ func TestByedpiEgressValidates(t *testing.T) {
 		Globals:  model.DefaultGlobals(),
 		Inbounds: []model.Inbound{{Name: "lan", Enabled: true, Type: "tproxy", TproxyPort: 12367, TCP: true, UDP: true}},
 		Egresses: []model.Egress{{Name: "bd", Type: "byedpi", Port: 1080}},
+		Rulesets: []model.Ruleset{inlineDomainSet("blocked", "blocked.example")},
 		Rules: []model.Rule{
-			{Name: "desync", Enabled: true, Order: 10, DstDomain: []string{"blocked.example"}, Target: "egress:bd"},
+			{Name: "desync", Enabled: true, Order: 10, DstRuleset: []string{"blocked"}, Target: "egress:bd"},
 		},
 	}
 	opts, warns, changed := applyAndClose(t, m)
@@ -0,0 +1,73 @@
+package generate
+
+import (
+	"bytes"
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/sagernet/sing-box/log"
+	"github.com/sagernet/sing-box/option"
+	"github.com/sagernet/sing-box/shater/model"
+)
+
+// TestLogOptionsDisableColorOffTTY pins the engine half of the syslog colour
+// leak. The engine's log factory is built by box.New from these options; with
+// DisableColor left false, every engine line reached procd's stderr — i.e.
+// syslog — wrapped in aurora escapes.
+func TestLogOptionsDisableColorOffTTY(t *testing.T) {
+	restore := logColorAllowed
+	logColorAllowed = func() bool { return false } // stderr is procd's pipe
+	t.Cleanup(func() { logColorAllowed = restore })
+
+	for _, level := range []string{"", "info", "error", "warning"} {
+		got := logOptions(level)
+		if got.Disabled {
+			t.Fatalf("logOptions(%q) unexpectedly disabled the log", level)
+		}
+		if !got.DisableColor {
+			t.Errorf("logOptions(%q).DisableColor = false; syslog would get ANSI escapes", level)
+		}
+	}
+
+	// A terminal (an interactive `shaterd run`) keeps its colours.
+	logColorAllowed = func() bool { return true }
+	if logOptions("info").DisableColor {
+		t.Errorf("logOptions on a TTY disabled colour; the gate is supposed to be the destination, not a blanket off")
+	}
+}
+
+// TestGeneratedLogBlockProducesNoANSI is the end-to-end proof: take the log
+// block Generate actually emits, build the engine's factory over it exactly the
+// way box.New does (log.New with DefaultWriter), and check the bytes.
+func TestGeneratedLogBlockProducesNoANSI(t *testing.T) {
+	restore := logColorAllowed
+	logColorAllowed = func() bool { return false }
+	t.Cleanup(func() { logColorAllowed = restore })
+
+	m := &model.Model{Globals: model.DefaultGlobals()}
+	m.Globals.LogLevel = "info"
+	opts, _, err := GenerateWithWarnings(m)
+	if err != nil {
+		t.Fatalf("GenerateWithWarnings: %v", err)
+	}
+	if opts.Log == nil {
+		t.Fatal("generated options carry no log block")
+	}
+
+	var out bytes.Buffer
+	factory, err := log.New(log.Options{Options: option.LogOptions(*opts.Log), DefaultWriter: &out})
+	if err != nil {
+		t.Fatalf("log.New over the generated block: %v", err)
+	}
+	ctx := log.ContextWithNewID(context.Background())
+	factory.Logger().ErrorContext(ctx, "dns: exchange failed for example.com. IN AAAA: unexpected EOF")
+
+	got := out.String()
+	if !strings.Contains(got, "unexpected EOF") {
+		t.Fatalf("engine factory wrote nothing usable: %q", got)
+	}
+	if i := strings.IndexByte(got, 0x1b); i >= 0 {
+		t.Fatalf("engine log line carries an ANSI escape at byte %d: %q", i, got)
+	}
+}
@@ -21,9 +21,10 @@ func TestProfileAppliesCleanly(t *testing.T) {
 		Globals:  g,
 		Inbounds: []model.Inbound{{Name: "lan", Enabled: true, Type: "tproxy", TproxyPort: 12370, TCP: true, UDP: true}},
 		Nodes:    []model.Node{{Name: "ss1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.5:8388#ss1"}},
+		Rulesets: []model.Ruleset{inlineDomainSet("ads", "ads.example")},
 		Rules: []model.Rule{
 			{Name: "lan-proxy", Enabled: true, Order: 10, Src: []string{"192.168.1.0/24"}, Target: "node:ss1"},
-			{Name: "adblock", Enabled: true, Order: 20, DstDomain: []string{"ads.example"}, Target: "block"},
+			{Name: "adblock", Enabled: true, Order: 20, DstRuleset: []string{"ads"}, Target: "block"},
 		},
 		Profiles: []model.Profile{
 			{Name: "home", Enabled: true, Priority: 1, DisableRules: []string{"adblock"}},
@@ -35,7 +36,7 @@ func TestProfileAppliesCleanly(t *testing.T) {
 		t.Fatalf("expected Apply changed==true (warnings: %v)", warns)
 	}
 	// Profile-disabled 'adblock' rule must be absent.
-	if opts.Route == nil || hasDomainRule(opts.Route, "ads.example") {
+	if opts.Route == nil || hasRulesetRule(opts.Route, "ads") {
 		t.Fatalf("profile-disabled rule 'adblock' should not be emitted")
 	}
 	// The lan-proxy rule still routes to ss1.
@@ -49,9 +49,13 @@ func TestNoProfilesUnchanged(t *testing.T) {
 	m := &model.Model{
 		Globals: model.DefaultGlobals(), // kill-switch closed => Final "block"
 		Nodes:   []model.Node{{Name: "ss1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.5:8388#ss1"}},
+		Rulesets: []model.Ruleset{
+			inlineDomainSet("a", "a.example"),
+			inlineDomainSet("b", "b.example"),
+		},
 		Rules: []model.Rule{
-			{Name: "a", Enabled: true, Order: 10, DstDomain: []string{"a.example"}, Target: "direct"},
-			{Name: "b", Enabled: true, Order: 20, DstDomain: []string{"b.example"}, Target: "node:ss1"},
+			{Name: "a", Enabled: true, Order: 10, DstRuleset: []string{"a"}, Target: "direct"},
+			{Name: "b", Enabled: true, Order: 20, DstRuleset: []string{"b"}, Target: "node:ss1"},
 		},
 	}
 	rt, b := buildRouteAt(m, wed12UTC)
@@ -70,7 +74,7 @@ func TestNoProfilesUnchanged(t *testing.T) {
 	if len(rt.Rules) != 4 {
 		t.Fatalf("route rule count = %d, want 4 (sniff+hijack+2 user)", len(rt.Rules))
 	}
-	if !hasDomainRule(rt, "a.example") || !hasDomainRule(rt, "b.example") {
+	if !hasRulesetRule(rt, "a") || !hasRulesetRule(rt, "b") {
 		t.Fatalf("both user rules should survive unchanged")
 	}
 }
@@ -83,9 +87,13 @@ func TestActiveProfileDisablesRule(t *testing.T) {
 	m := &model.Model{
 		Globals: g,
 		Nodes:   []model.Node{{Name: "ss1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.5:8388#ss1"}},
+		Rulesets: []model.Ruleset{
+			inlineDomainSet("blocked", "blocked.example"),
+			inlineDomainSet("keep", "keep.example"),
+		},
 		Rules: []model.Rule{
-			{Name: "blockme", Enabled: true, Order: 10, DstDomain: []string{"blocked.example"}, Target: "block"},
-			{Name: "keep", Enabled: true, Order: 20, DstDomain: []string{"keep.example"}, Target: "direct"},
+			{Name: "blockme", Enabled: true, Order: 10, DstRuleset: []string{"blocked"}, Target: "block"},
+			{Name: "keep", Enabled: true, Order: 20, DstRuleset: []string{"keep"}, Target: "direct"},
 		},
 		Profiles: []model.Profile{
 			// Manual pin: honored regardless of conditions. Disables "blockme".
@@ -94,10 +102,10 @@ func TestActiveProfileDisablesRule(t *testing.T) {
 	}
 	rt, b := buildRouteAt(m, wed12UTC)

-	if hasDomainRule(rt, "blocked.example") {
+	if hasRulesetRule(rt, "blocked") {
 		t.Fatalf("profile disabled rule 'blockme' but its matcher is still emitted")
 	}
-	if !hasDomainRule(rt, "keep.example") {
+	if !hasRulesetRule(rt, "keep") {
 		t.Fatalf("rule 'keep' should be untouched")
 	}
 	if rt.Final != tagBlock {
@@ -115,9 +123,13 @@ func TestActiveProfileDisablesRule(t *testing.T) {
 func TestAutoSelectHighestPriority(t *testing.T) {
 	m := &model.Model{
 		Globals: model.DefaultGlobals(),
+		Rulesets: []model.Ruleset{
+			inlineDomainSet("lo", "lo.example"),
+			inlineDomainSet("hi", "hi.example"),
+		},
 		Rules: []model.Rule{
-			{Name: "rLo", Enabled: true, Order: 10, DstDomain: []string{"lo.example"}, Target: "direct"},
-			{Name: "rHi", Enabled: true, Order: 20, DstDomain: []string{"hi.example"}, Target: "direct"},
+			{Name: "rLo", Enabled: true, Order: 10, DstRuleset: []string{"lo"}, Target: "direct"},
+			{Name: "rHi", Enabled: true, Order: 20, DstRuleset: []string{"hi"}, Target: "direct"},
 		},
 		Profiles: []model.Profile{
 			{Name: "lo", Enabled: true, Priority: 5, DisableRules: []string{"rLo"}},
@@ -125,10 +137,10 @@ func TestAutoSelectHighestPriority(t *testing.T) {
 		},
 	}
 	rt, _ := buildRouteAt(m, wed12UTC)
-	if hasDomainRule(rt, "hi.example") {
+	if hasRulesetRule(rt, "hi") {
 		t.Fatalf("the highest-priority profile must win and disable rHi")
 	}
-	if !hasDomainRule(rt, "lo.example") {
+	if !hasRulesetRule(rt, "lo") {
 		t.Fatalf("only the winning profile applies (rLo should survive)")
 	}
 }
@@ -142,9 +154,13 @@ func TestIfaceProfileSkippedByAutoButHonoredNamed(t *testing.T) {
 		g.ActiveProfile = active
 		return &model.Model{
 			Globals: g,
+			Rulesets: []model.Ruleset{
+				inlineDomainSet("hi", "hi.example"),
+				inlineDomainSet("if", "if.example"),
+			},
 			Rules: []model.Rule{
-				{Name: "rHi", Enabled: true, Order: 10, DstDomain: []string{"hi.example"}, Target: "direct"},
-				{Name: "rIf", Enabled: true, Order: 20, DstDomain: []string{"if.example"}, Target: "direct"},
+				{Name: "rHi", Enabled: true, Order: 10, DstRuleset: []string{"hi"}, Target: "direct"},
+				{Name: "rIf", Enabled: true, Order: 20, DstRuleset: []string{"if"}, Target: "direct"},
 			},
 			Profiles: []model.Profile{
 				{Name: "plain", Enabled: true, Priority: 10, DisableRules: []string{"rHi"}},
@@ -156,19 +172,19 @@ func TestIfaceProfileSkippedByAutoButHonoredNamed(t *testing.T) {

 	// Auto-select (no pin): iface profile skipped (the watcher owns it); 'plain' wins.
 	rt, _ := buildRouteAt(mk(""), wed12UTC)
-	if hasDomainRule(rt, "hi.example") {
+	if hasRulesetRule(rt, "hi") {
 		t.Fatalf("auto-select should apply 'plain' and disable rHi")
 	}
-	if !hasDomainRule(rt, "if.example") {
+	if !hasRulesetRule(rt, "if") {
 		t.Fatalf("iface profile 'wwan' must be skipped by auto-select (rIf should survive)")
 	}

 	// Named explicitly: iface profile honored regardless of its iface condition.
 	rt, _ = buildRouteAt(mk("wwan"), wed12UTC)
-	if hasDomainRule(rt, "if.example") {
+	if hasRulesetRule(rt, "if") {
 		t.Fatalf("explicitly named iface profile must be honored and disable rIf")
 	}
-	if !hasDomainRule(rt, "hi.example") {
+	if !hasRulesetRule(rt, "hi") {
 		t.Fatalf("only the named profile applies; rHi should survive")
 	}
 }
@@ -179,14 +195,15 @@ func TestUnknownActiveProfileFallsBackToAuto(t *testing.T) {
 	g := model.DefaultGlobals()
 	g.ActiveProfile = "ghost"
 	m := &model.Model{
-		Globals: g,
-		Rules:   []model.Rule{{Name: "rX", Enabled: true, Order: 10, DstDomain: []string{"x.example"}, Target: "direct"}},
+		Globals:  g,
+		Rulesets: []model.Ruleset{inlineDomainSet("x", "x.example")},
+		Rules:    []model.Rule{{Name: "rX", Enabled: true, Order: 10, DstRuleset: []string{"x"}, Target: "direct"}},
 		Profiles: []model.Profile{
 			{Name: "auto", Enabled: true, Priority: 1, DisableRules: []string{"rX"}},
 		},
 	}
 	rt, b := buildRouteAt(m, wed12UTC)
-	if hasDomainRule(rt, "x.example") {
+	if hasRulesetRule(rt, "x") {
 		t.Fatalf("fallback auto-select should apply 'auto' and disable rX")
 	}
 	if !hasWarning(b, "falling back to auto-select") {
@@ -208,16 +225,17 @@ func TestUnknownActiveProfileFallsBackToAuto(t *testing.T) {
 // suppress it or to warn about it.
 func TestPlainProfileIsSelectableAfterProbeRemoval(t *testing.T) {
 	m := &model.Model{
-		Globals: model.DefaultGlobals(),
+		Globals:  model.DefaultGlobals(),
+		Rulesets: []model.Ruleset{inlineDomainSet("hi", "hi.example")},
 		Rules: []model.Rule{
-			{Name: "rHi", Enabled: true, Order: 10, DstDomain: []string{"hi.example"}, Target: "direct"},
+			{Name: "rHi", Enabled: true, Order: 10, DstRuleset: []string{"hi"}, Target: "direct"},
 		},
 		Profiles: []model.Profile{
 			{Name: "plain", Enabled: true, Priority: 10, DisableRules: []string{"rHi"}},
 		},
 	}
 	rt, b := buildRouteAt(m, wed12UTC)
-	if hasDomainRule(rt, "hi.example") {
+	if hasRulesetRule(rt, "hi") {
 		t.Fatalf("a plain profile must apply and disable rHi")
 	}
 	// No leftover diagnostic: warning about an option the parser no longer reads
@@ -1,8 +1,6 @@
 package generate

 import (
-	"fmt"
-	"regexp"
 	"sort"
 	"strings"

@@ -30,6 +28,12 @@ func (b *builder) buildRoute() *option.RouteOptions {
 	// never aborts box.New.
 	b.applyProfiles()

+	// Report the rules that cannot fire under this config BEFORE building anything,
+	// so the diagnosis is about the desired state the operator wrote rather than
+	// about whatever survived generation. Diagnosis only — nothing is renamed,
+	// reordered or dropped here.
+	b.warnUnreachableRules()
+
 	// Kill-switch default backstop.
 	final := tagBlock
 	if strings.EqualFold(strings.TrimSpace(b.m.Globals.KillSwitch), "open") {
@@ -240,25 +244,58 @@ func (b *builder) ruleKillFallback(r model.Rule, want string) string {
 // outbound — traffic left over the DEFAULT WAN while the panel (which applies the
 // same "a bare Egress is a target too" rule for display) showed `egress:wan2`.
 // A silent mis-route on exactly the multi-WAN setups the field exists for.
-func effectiveRuleTarget(r model.Rule) string {
-	if t := strings.TrimSpace(r.Target); t != "" {
-		return t
-	}
-	if e := strings.TrimSpace(r.Egress); e != "" {
-		return "egress:" + e
-	}
-	return ""
-}
+//
+// The implementation is model.EffectiveRuleTarget — shared with the reachability
+// analysis, which must resolve a rule's target identically to decide which
+// default the router actually ends up using.
+func effectiveRuleTarget(r model.Rule) string { return model.EffectiveRuleTarget(r) }

 // isCatchAll reports whether a rule carries NO matcher of any kind (src, dst
 // domain/ip/ruleset, port, proto). Such a rule is the default egress.
-func (b *builder) isCatchAll(r model.Rule) bool {
-	return len(r.Src) == 0 &&
-		len(r.DstDomain) == 0 &&
-		len(r.DstIP) == 0 &&
-		len(r.DstRuleset) == 0 &&
-		strings.TrimSpace(r.DstPort) == "" &&
-		strings.TrimSpace(r.Proto) == ""
+//
+// The implementation is model.IsCatchAll — shared with the reachability analysis
+// (and mirrored by the panel), because "is this rule a default?" is the single
+// question both the Final assignment below and the never-fires badge turn on.
+func (b *builder) isCatchAll(r model.Rule) bool { return model.IsCatchAll(r) }
+
+// warnUnreachableRules reports every rule that CANNOT take effect under this
+// config, whatever the traffic.
+//
+// Only one shape is certain enough to report (see model.RuleReachability): two
+// condition-less rules, where the later one wins because the loop above simply
+// overwrites Final. That case is silent today and looks completely healthy — the
+// panel drew both as "default route · final" and the log said nothing — so a
+// config with `default -> direct` at order 20 and `default -> group:auto` at
+// order 100 gave no hint at all that one of the two was doing nothing.
+//
+// Severity is decided by CONSEQUENCE, via the wording (apply/warnings.go
+// classifies generate's text): when the default that actually wins is `direct`
+// while the retired rule asked for a tunnel or a block, the operator's default
+// policy is not in effect and everything unmatched leaves on the plain WAN — a
+// broken protection claim, i.e. critical. Any other combination (a dead `direct`
+// under a live tunnel, one tunnel under another) is a dead setting, not a leak:
+// a warning.
+func (b *builder) warnUnreachableRules() {
+	for _, rr := range model.RuleReachability(b.effectiveRules) {
+		if !rr.Unreachable {
+			continue
+		}
+		dead := effectiveRuleTarget(b.effectiveRules[rr.Index])
+		winner := effectiveRuleTarget(b.effectiveRules[rr.ShadowedByIndex])
+		if winner == tagDirect && dead != tagDirect {
+			b.warnf("rule %q: it has no conditions, so it sets the default for ALL traffic — but "+
+				"rule %q (order %d) has none either and comes after it, so %q wins and this rule's "+
+				"target %q is never applied. Everything no other rule matches leaves over the plain "+
+				"WAN with your real IP address. Delete one of the two, or give this one a condition",
+				rr.Name, rr.ShadowedBy, rr.ShadowedByOrder, winner, dead)
+			continue
+		}
+		b.warnf("rule %q: it has no conditions, so it sets the default for ALL traffic — but "+
+			"rule %q (order %d) has none either and comes after it, so the default the router uses "+
+			"is %q and this rule's target %q is never applied. Delete one of the two, or give this "+
+			"one a condition so it can match something",
+			rr.Name, rr.ShadowedBy, rr.ShadowedByOrder, winner, dead)
+	}
 }

 // ruleMatchers builds the RawDefaultRule matchers the engine can evaluate.
@@ -282,114 +319,14 @@ func (b *builder) ruleMatchers(r model.Rule) (raw option.RawDefaultRule, matched
 		matched = true
 	}

-	// Destination domains. The route-rule-specific prefixes (geosite:/regexp:/
-	// suffix:) are peeled off here; everything else goes through the shared
-	// classifier in dnsfilter.go with bareIsSuffix=FALSE — a bare entry in a
-	// ROUTING rule is an exact domain, unlike the DNS/filter lists where it means
-	// "and all subdomains". classifyDomainEntries is also what drops marker-only
-	// entries ("." / "full:" / "keyword:"), which must never reach the engine:
-	// an empty domain/domain_suffix aborts box.New, and an empty domain_keyword
-	// is strings.Contains(host, "") — a silent match on EVERY host.
-	var explicitSuffix []string
-	var plain []string
-	for _, d := range r.DstDomain {
-		d = strings.TrimSpace(d)
-		if d == "" {
-			continue
-		}
-		switch {
-		case strings.HasPrefix(d, "geosite:"):
-			// A `geosite:` matcher is NOT emittable: the route-rule geosite field was
-			// removed in this engine and route/rule.NewDefaultRule hard-errors on it,
-			// which aborts box.New for the WHOLE config. Mirroring the geoip handling
-			// below, it is treated as inert: warned and omitted, so a legacy/UCI rule
-			// carrying one degrades to "this rule does nothing" instead of taking the
-			// entire tunnel down. Use a `config ruleset` with source=geosite instead.
-			b.warnf("rule %q: geosite matcher %q is removed from this engine — use a ruleset with source=geosite instead, omitted (inert)", r.Name, d)
-		case strings.HasPrefix(d, "regexp:"):
-			// route/rule.NewDomainRegexItem hard-errors on an uncompilable pattern and
-			// takes box.New with it; validate here and drop the bad one with a warning.
-			// A BARE "regexp:" compiles fine but matches every host — same silent
-			// match-all hazard as an empty keyword, so it is dropped too.
-			re := strings.TrimSpace(strings.TrimPrefix(d, "regexp:"))
-			if re == "" {
-				b.warnf("rule %q: %q is a bare matcher marker with no value, omitted (an empty regexp matches EVERY host)", r.Name, d)
-				break
-			}
-			if _, err := regexp.Compile(re); err != nil {
-				b.warnf("rule %q: domain regexp %q is invalid (%v), omitted", r.Name, re, err)
-				break
-			}
-			raw.DomainRegex = append(raw.DomainRegex, re)
-			matched = true
-		case strings.HasPrefix(d, "suffix:"):
-			// Label-aware suffix (apex + subdomains): sing-box domain_suffix stored
-			// in bare form matches both "example.com" and "*.example.com" (see
-			// sing/common/domain matcher). A leading-dot entry, by contrast, matches
-			// subdomains ONLY, so the explicit `suffix:` form is how presets/rules
-			// ask for apex-inclusive suffix matching.
-			if s := strings.TrimSpace(strings.TrimPrefix(d, "suffix:")); s != "" {
-				explicitSuffix = append(explicitSuffix, s)
-			} else {
-				b.warnf("rule %q: %q is a bare matcher marker with no value, omitted (an empty domain token aborts box.New)", r.Name, d)
-			}
-		default:
-			plain = append(plain, d)
-		}
-	}
-	domain, suffix, keyword := classifyDomainEntries(plain, false)
-	for _, d := range plain {
-		if isDomainMarkerOnly(d) {
-			b.warnf("rule %q: %q is a bare matcher marker with no value, omitted (an empty domain token aborts box.New; an empty keyword matches EVERY host)", r.Name, d)
-		}
-	}
-	// An unrecognised `word:` prefix is DROPPED by classifyDomainEntries (a domain
-	// cannot contain ":", so keeping it would load a provably unmatchable literal).
-	// It has to be reported here or the rule silently loses that destination — the
-	// v0.1/xray spelling `domain:example.com` is exactly what someone migrating
-	// writes, and it used to disappear without a trace. geosite:/regexp:/suffix:
-	// were already peeled off above, so `plain` carries only the shared markers.
-	b.warnUnrecognisedPrefixes(fmt.Sprintf("rule %q", r.Name), plain)
-	suffix = append(explicitSuffix, suffix...)
-	if len(domain) > 0 {
-		raw.Domain = badoption.Listable[string](domain)
-		matched = true
-	}
-	if len(suffix) > 0 {
-		raw.DomainSuffix = badoption.Listable[string](suffix)
-		matched = true
-	}
-	if len(keyword) > 0 {
-		raw.DomainKeyword = badoption.Listable[string](keyword)
-		matched = true
-	}
-	// Destination IPs. A `geoip:<code>` entry is NOT a CIDR: route-rule geoip was
-	// removed in this engine (box.New hard-errors on it), so — mirroring how the
-	// ruleset geoip source is handled — it is treated as inert: warned and omitted
-	// (never emitted as an ip_cidr, which would also abort box.New). This keeps a
-	// geoip-driven rule/preset (e.g. the ru-bypass pack) fail-open instead of fatal.
-	var ipcidr []string
-	for _, ip := range r.DstIP {
-		ip = strings.TrimSpace(ip)
-		if ip == "" {
-			continue
-		}
-		if strings.HasPrefix(strings.ToLower(ip), "geoip:") {
-			b.warnf("rule %q: geoip matcher %q is removed from this engine — use a ruleset with source=geoip instead, omitted (inert)", r.Name, ip)
-			continue
-		}
-		ipcidr = append(ipcidr, ip)
-	}
-	// Same fail-open validation as the source list: an unparseable ip_cidr aborts
-	// box.New for the whole config, so drop it with a warning instead.
-	ipcidr, badIP := validPrefixes(ipcidr)
-	for _, s := range badIP {
-		b.warnf("rule %q: destination %q is not a valid IP/CIDR, omitted", r.Name, s)
-	}
-	if len(ipcidr) > 0 {
-		raw.IPCIDR = badoption.Listable[string](ipcidr)
-		matched = true
-	}
+	// Destination: a rule's ONLY destination matcher is DstRuleset (schema v2).
+	// The inline `dst_domain` / `dst_ip` lists that used to be classified here are
+	// gone. A destination list is a `config ruleset` — compiled once into a .srs and
+	// shared by every rule that references it — so the prefix vocabulary
+	// (full:/suffix:/keyword:/regexp:, a leading dot) and the geosite/geoip sources
+	// live in exactly one place (generate/ruleset.go inlineRulesetRule). Existing
+	// configs were rewritten by model.migrate1to2, which preserves each entry's
+	// meaning 1:1.

 	// DstRuleset -> reference the rs-<name> rule-sets materialised by
 	// buildRoutingRuleSets. A dst_ruleset naming an UNDEFINED ruleset is warned +
@@ -8,6 +8,7 @@ import (
 	"strings"
 	"testing"

+	C "github.com/sagernet/sing-box/constant"
 	"github.com/sagernet/sing-box/option"

 	"github.com/sagernet/sing-box/shater/model"
@@ -22,26 +23,45 @@ func killModel(kill string) *model.Model {
 	g := model.DefaultGlobals()
 	g.KillSwitch = "closed"
 	return &model.Model{
-		Globals: g,
-		Nodes:   []model.Node{{Name: "n1", Enabled: false, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
-		Groups:  []model.Group{{Name: "grp", Strategy: "leastping", Nodes: []string{"n1"}}},
+		Globals:  g,
+		Nodes:    []model.Node{{Name: "n1", Enabled: false, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+		Groups:   []model.Group{{Name: "grp", Strategy: "leastping", Nodes: []string{"n1"}}},
+		Rulesets: []model.Ruleset{inlineDomainSet("social", "social.example")},
 		Rules: []model.Rule{
-			{Name: "social", Enabled: true, Order: 10, DstDomain: []string{"social.example"}, Target: "group:grp", Kill: kill},
+			{Name: "social", Enabled: true, Order: 10, DstRuleset: []string{"social"}, Target: "group:grp", Kill: kill},
 			{Name: "catch-tcp", Enabled: true, Order: 20, Proto: "tcp", Target: "direct"},
 		},
 	}
 }

-// domainRuleTarget returns the outbound the rule matching domain routes to.
-func domainRuleTarget(rt *option.RouteOptions, domain string) (string, bool) {
-	for _, r := range generalRules(rt) {
-		for _, d := range r.DefaultOptions.RawDefaultRule.Domain {
-			if d == domain {
-				return r.DefaultOptions.RuleAction.RouteOptions.Outbound, true
-			}
-		}
+// rulesetRuleTarget returns the outbound the rule referencing the rule-set named
+// name routes to. It is the post-schema-v2 replacement for looking a rule up by
+// its dst-domain matcher: a rule's destination is a rule_set reference now, so
+// the matcher that identifies it is the rs-<name> tag.
+func rulesetRuleTarget(rt *option.RouteOptions, name string) (string, bool) {
+	dr := findRouteRuleWithRuleSet(rt, routeRulesetTagPrefix+name)
+	if dr == nil {
+		return "", false
 	}
-	return "", false
+	return dr.RuleAction.RouteOptions.Outbound, true
+}
+
+// soleInlineRule returns the single default headless rule an inline rule-set is
+// expected to carry, failing the ASSERTION (rather than panicking on an index) when
+// the set turns out to be remote/local or to hold a different rule shape.
+func soleInlineRule(t *testing.T, rs option.RuleSet) option.DefaultHeadlessRule {
+	t.Helper()
+	if rs.Type != C.RuleSetTypeInline && rs.Type != "" {
+		t.Fatalf("rule-set %q is type %q, not inline — it has no rules in the config to inspect", rs.Tag, rs.Type)
+	}
+	if len(rs.InlineOptions.Rules) != 1 {
+		t.Fatalf("rule-set %q must carry exactly one headless rule, got %d", rs.Tag, len(rs.InlineOptions.Rules))
+	}
+	hr := rs.InlineOptions.Rules[0]
+	if hr.Type != C.RuleTypeDefault && hr.Type != "" {
+		t.Fatalf("rule-set %q carries a %q headless rule, not a default one", rs.Tag, hr.Type)
+	}
+	return hr.DefaultOptions
 }

 // TestRuleKillDefaultBlocks: kill="" / "default" is fail-closed — the rule is
@@ -54,7 +74,7 @@ func TestRuleKillDefaultBlocks(t *testing.T) {
 		if err != nil {
 			t.Fatalf("%q: Generate: %v", kill, err)
 		}
-		got, ok := domainRuleTarget(opts.Route, "social.example")
+		got, ok := rulesetRuleTarget(opts.Route, "social")
 		if !ok {
 			t.Fatalf("%q: rule must be emitted routing to block, but it was dropped", kill)
 		}
@@ -76,7 +96,7 @@ func TestRuleKillClosedBlocksHere(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Generate: %v", err)
 	}
-	got, ok := domainRuleTarget(opts.Route, "social.example")
+	got, ok := rulesetRuleTarget(opts.Route, "social")
 	if !ok {
 		t.Fatalf("kill=closed must still emit the rule (warnings %v)", warns)
 	}
@@ -103,7 +123,7 @@ func TestRuleKillOpenGoesDirect(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Generate: %v", err)
 	}
-	got, ok := domainRuleTarget(opts.Route, "social.example")
+	got, ok := rulesetRuleTarget(opts.Route, "social")
 	if !ok {
 		t.Fatalf("kill=open must still emit the rule (warnings %v)", warns)
 	}
@@ -122,7 +142,7 @@ func TestRuleKillUnknownBlocks(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Generate: %v", err)
 	}
-	got, ok := domainRuleTarget(opts.Route, "social.example")
+	got, ok := rulesetRuleTarget(opts.Route, "social")
 	if !ok || got != tagBlock {
 		t.Fatalf("unknown kill policy must block, got %q (ok=%v)", got, ok)
 	}
@@ -145,7 +165,7 @@ func TestRuleKillPreservesFailClosedInvariant(t *testing.T) {
 		if opts.Route.Final != tagBlock {
 			t.Fatalf("%q: Final = %q, want block", kill, opts.Route.Final)
 		}
-		if tgt, ok := domainRuleTarget(opts.Route, "social.example"); ok && tgt == tagDirect {
+		if tgt, ok := rulesetRuleTarget(opts.Route, "social"); ok && tgt == tagDirect {
 			t.Fatalf("%q: dead group leaked direct", kill)
 		}
 	}
@@ -158,17 +178,18 @@ func TestRuleKillOnlyAppliesToUnresolvedTargets(t *testing.T) {
 		g := model.DefaultGlobals()
 		g.KillSwitch = "closed"
 		m := &model.Model{
-			Globals: g,
-			Nodes:   []model.Node{{Name: "n1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+			Globals:  g,
+			Nodes:    []model.Node{{Name: "n1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+			Rulesets: []model.Ruleset{inlineDomainSet("social", "social.example")},
 			Rules: []model.Rule{
-				{Name: "ok", Enabled: true, Order: 10, DstDomain: []string{"social.example"}, Target: "node:n1", Kill: kill},
+				{Name: "ok", Enabled: true, Order: 10, DstRuleset: []string{"social"}, Target: "node:n1", Kill: kill},
 			},
 		}
 		opts, _, err := GenerateWithWarnings(m)
 		if err != nil {
 			t.Fatalf("%q: Generate: %v", kill, err)
 		}
-		got, ok := domainRuleTarget(opts.Route, "social.example")
+		got, ok := rulesetRuleTarget(opts.Route, "social")
 		if !ok || got != "n1" {
 			t.Fatalf("%q: healthy target must win, got %q (ok=%v)", kill, got, ok)
 		}
@@ -176,29 +197,54 @@ func TestRuleKillOnlyAppliesToUnresolvedTargets(t *testing.T) {
 }

 // --- R4: marker-only domain entries ------------------------------------------
+//
+// R4 moved with the destination list itself: a rule's domains are an inline
+// `config ruleset` now (schema v2), so the marker-only guard has to hold in
+// inlineRulesetRule rather than in ruleMatchers. The hazard is unchanged.

-// TestRuleMarkerOnlyDomainEntriesDropped: an entry that is nothing but its marker
-// must never reach the engine. An empty domain/domain_suffix makes NewDomainItem
-// return "empty item is not allowed" and aborts box.New (whole LAN down from one
-// stray "."); an empty domain_keyword is SILENT and matches every host.
-func TestRuleMarkerOnlyDomainEntriesDropped(t *testing.T) {
+// TestRulesetMarkerOnlyDomainEntriesDropped: an entry that is nothing but its
+// marker must never reach the engine. An empty domain/domain_suffix makes
+// NewDomainItem return "empty item is not allowed" and aborts box.New (whole LAN
+// down from one stray "."); an empty domain_keyword is SILENT and matches every
+// host. As the sole entry it also leaves the rule-set with nothing usable, so the
+// list is skipped and the rule referencing it is not emitted — never promoted to
+// "matches everything".
+func TestRulesetMarkerOnlyDomainEntriesDropped(t *testing.T) {
 	for _, entry := range []string{".", "full:", "keyword:", "suffix:", "regexp:", "  keyword:  "} {
-		rt, warns := genRules(t, model.Rule{
-			Name: "m", Enabled: true, Order: 10,
-			DstDomain: []string{entry}, Target: "node:n1",
-		})
-		for _, r := range rt.Rules {
-			raw := r.DefaultOptions.RawDefaultRule
-			for _, list := range [][]string{raw.Domain, raw.DomainSuffix, raw.DomainKeyword, raw.DomainRegex} {
-				for _, v := range list {
-					if strings.TrimSpace(v) == "" {
-						t.Fatalf("%q: emitted an EMPTY matcher token", entry)
+		rt, warns := genRulesWithSets(t,
+			[]model.Ruleset{inlineDomainSet("m", entry)},
+			model.Rule{Name: "m", Enabled: true, Order: 10, DstRuleset: []string{"m"}, Target: "node:n1"},
+		)
+		// Only an INLINE rule-set has its rules in the options at all; a remote or
+		// local one (a geosite chip, a compiled url list) carries a URL or a path and
+		// an empty InlineOptions. Indexing Rules[0] unconditionally turned any such
+		// fixture into an index-out-of-range PANIC instead of a failed assertion, so
+		// the shape is checked rather than assumed.
+		for _, rs := range rt.RuleSet {
+			if rs.Type != C.RuleSetTypeInline && rs.Type != "" {
+				continue
+			}
+			for i, hrule := range rs.InlineOptions.Rules {
+				if hrule.Type != C.RuleTypeDefault && hrule.Type != "" {
+					continue // a logical headless rule has no matcher lists of its own
+				}
+				hr := hrule.DefaultOptions
+				for name, list := range map[string][]string{
+					"domain": hr.Domain, "domain_suffix": hr.DomainSuffix,
+					"domain_keyword": hr.DomainKeyword, "domain_regex": hr.DomainRegex,
+				} {
+					for _, v := range list {
+						if strings.TrimSpace(v) == "" {
+							t.Fatalf("%q: rule-set %q rule %d emitted an EMPTY %s token (an empty domain aborts box.New; an empty keyword matches every host)",
+								entry, rs.Tag, i, name)
+						}
 					}
 				}
 			}
 		}
-		// Sole entry => nothing left to match on => the rule must be skipped, never
-		// silently promoted to "matches everything".
+		if _, ok := ruleSetByTag(rt, "rs-m"); ok {
+			t.Fatalf("%q: a marker-only list must not materialise a rule-set", entry)
+		}
 		if n := len(generalRules(rt)); n != 0 {
 			t.Fatalf("%q: marker-only sole entry must skip the rule, got %d rules", entry, n)
 		}
@@ -208,46 +254,55 @@ func TestRuleMarkerOnlyDomainEntriesDropped(t *testing.T) {
 	}
 }

-// TestRuleMarkerOnlyBesideRealEntriesKeepsTheRest: the real entries must survive
-// the marker-only ones.
-func TestRuleMarkerOnlyBesideRealEntriesKeepsTheRest(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "m", Enabled: true, Order: 10,
-		DstDomain: []string{".", "keyword:", "exact.example", "keyword:ads", ".sub.example", "suffix:apex.example"},
-		Target:    "node:n1",
-	})
-	gen := generalRules(rt)
-	if len(gen) != 1 {
-		t.Fatalf("want 1 rule, got %d (warnings %v)", len(gen), warns)
+// TestRulesetMarkerOnlyBesideRealEntriesKeepsTheRest: the real entries must
+// survive the marker-only ones.
+func TestRulesetMarkerOnlyBesideRealEntriesKeepsTheRest(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("m",
+			".", "keyword:", "full:exact.example", "keyword:ads", ".sub.example", "suffix:apex.example")},
+		model.Rule{Name: "m", Enabled: true, Order: 10, DstRuleset: []string{"m"}, Target: "node:n1"},
+	)
+	rs, ok := ruleSetByTag(rt, "rs-m")
+	if !ok {
+		t.Fatalf("the usable entries must keep the list alive (warnings %v)", warns)
 	}
-	raw := gen[0].DefaultOptions.RawDefaultRule
-	if len(raw.Domain) != 1 || raw.Domain[0] != "exact.example" {
-		t.Fatalf("domain = %v, want [exact.example] (bare entry in a ROUTING rule is exact)", raw.Domain)
+	hr := soleInlineRule(t, rs)
+	if len(hr.Domain) != 1 || hr.Domain[0] != "exact.example" {
+		t.Fatalf("domain = %v, want [exact.example]", hr.Domain)
 	}
-	if len(raw.DomainKeyword) != 1 || raw.DomainKeyword[0] != "ads" {
-		t.Fatalf("keyword = %v, want [ads]", raw.DomainKeyword)
+	if len(hr.DomainKeyword) != 1 || hr.DomainKeyword[0] != "ads" {
+		t.Fatalf("keyword = %v, want [ads]", hr.DomainKeyword)
 	}
-	if len(raw.DomainSuffix) != 2 {
-		t.Fatalf("suffix = %v, want both apex.example and sub.example", raw.DomainSuffix)
+	if len(hr.DomainSuffix) != 2 {
+		t.Fatalf("suffix = %v, want both apex.example and sub.example", hr.DomainSuffix)
+	}
+	if findRouteRuleWithRuleSet(rt, "rs-m") == nil {
+		t.Fatalf("the rule must be emitted referencing rs-m; rules=%+v", rt.Rules)
 	}
 }

-// TestRuleBareEntryIsExactDomain pins the routing convention (bare == exact),
-// which deliberately differs from the DNS/filter lists (bare == suffix).
-func TestRuleBareEntryIsExactDomain(t *testing.T) {
-	rt, _ := genRules(t, model.Rule{
-		Name: "b", Enabled: true, Order: 10,
-		DstDomain: []string{"example.com"}, Target: "node:n1",
-	})
-	gen := generalRules(rt)
-	if len(gen) != 1 {
-		t.Fatalf("want 1 rule, got %d", len(gen))
+// TestRulesetBareEntryIsDomainSuffix pins the convention a destination list now
+// follows — and it is the OPPOSITE of the one the old dst_domain used.
+//
+// A bare entry in a routing rule meant one EXACT domain; a bare entry in a
+// rule-set means the domain AND its subdomains (the DNS/filter/device convention,
+// classifyDomainEntries with bareIsSuffix=true). `full:` is how an exact match is
+// written now. model.migrate1to2 rewrites old dst_domain entries accordingly, so
+// this asymmetry is the thing that migration has to get right.
+func TestRulesetBareEntryIsDomainSuffix(t *testing.T) {
+	rt, _ := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("b", "example.com", "full:exact.example")},
+		model.Rule{Name: "b", Enabled: true, Order: 10, DstRuleset: []string{"b"}, Target: "node:n1"},
+	)
+	rs, ok := ruleSetByTag(rt, "rs-b")
+	if !ok {
+		t.Fatalf("rs-b not emitted; route=%+v", rt)
 	}
-	raw := gen[0].DefaultOptions.RawDefaultRule
-	if len(raw.Domain) != 1 || raw.Domain[0] != "example.com" {
-		t.Fatalf("bare entry must be an exact Domain, got domain=%v suffix=%v", raw.Domain, raw.DomainSuffix)
+	hr := soleInlineRule(t, rs)
+	if len(hr.DomainSuffix) != 1 || hr.DomainSuffix[0] != "example.com" {
+		t.Fatalf("bare entry must become a domain_suffix, got suffix=%v domain=%v", hr.DomainSuffix, hr.Domain)
 	}
-	if len(raw.DomainSuffix) != 0 {
-		t.Fatalf("bare entry must NOT become a suffix, got %v", raw.DomainSuffix)
+	if len(hr.Domain) != 1 || hr.Domain[0] != "exact.example" {
+		t.Fatalf("full: must be the exact form, got domain=%v", hr.Domain)
 	}
 }
@@ -18,10 +18,13 @@ import (
 )

 // TestMalformedMatchersStillApply drives ONE model carrying every previously
-// fatal matcher — a geosite: domain, a bad ip_cidr, a bad source cidr, an
+// fatal matcher — a geosite: entry, a bad ip_cidr, a bad source cidr, an
 // uncompilable regexp and a malformed port range — plus a healthy rule, through
 // engine.Apply. It must come up, the healthy rule must survive, and the
-// kill-switch backstop must stay closed.
+// kill-switch backstop must stay closed. The destination lists are inline
+// `config ruleset`s (schema v2), which is where the domain/IP guards live now;
+// a rule-set left with no usable entry is skipped, and so is the rule whose only
+// matcher it was.
 func TestMalformedMatchersStillApply(t *testing.T) {
 	g := model.DefaultGlobals()
 	g.KillSwitch = "closed"
@@ -29,13 +32,19 @@ func TestMalformedMatchersStillApply(t *testing.T) {
 		Globals:  g,
 		Inbounds: []model.Inbound{{Name: "lan", Enabled: true, Type: "tproxy", TproxyPort: 12395, TCP: true, UDP: true}},
 		Nodes:    []model.Node{{Name: "n1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+		Rulesets: []model.Ruleset{
+			inlineDomainSet("geosite", "geosite:youtube"),
+			inlineIPSet("badip", "999.1.1.1/24"),
+			inlineDomainSet("badre", "regexp:*broken("),
+			inlineDomainSet("ok", "ok.example"),
+		},
 		Rules: []model.Rule{
-			{Name: "geosite", Enabled: true, Order: 1, DstDomain: []string{"geosite:youtube"}, Target: "node:n1"},
-			{Name: "badip", Enabled: true, Order: 2, DstIP: []string{"999.1.1.1/24"}, Target: "node:n1"},
+			{Name: "geosite", Enabled: true, Order: 1, DstRuleset: []string{"geosite"}, Target: "node:n1"},
+			{Name: "badip", Enabled: true, Order: 2, DstRuleset: []string{"badip"}, Target: "node:n1"},
 			{Name: "badsrc", Enabled: true, Order: 3, Src: []string{"192.168.0.0/99"}, Target: "node:n1"},
-			{Name: "badre", Enabled: true, Order: 4, DstDomain: []string{"regexp:*broken("}, Target: "node:n1"},
+			{Name: "badre", Enabled: true, Order: 4, DstRuleset: []string{"badre"}, Target: "node:n1"},
 			{Name: "badport", Enabled: true, Order: 5, DstPort: "a-b", Target: "node:n1"},
-			{Name: "healthy", Enabled: true, Order: 6, DstDomain: []string{"ok.example"}, DstPort: "443", Target: "node:n1"},
+			{Name: "healthy", Enabled: true, Order: 6, DstRuleset: []string{"ok"}, DstPort: "443", Target: "node:n1"},
 		},
 	}

@@ -46,14 +55,21 @@ func TestMalformedMatchersStillApply(t *testing.T) {
 	if opts.Route.Final != tagBlock {
 		t.Fatalf("Final = %q, want block", opts.Route.Final)
 	}
-	if !hasDomainRule(opts.Route, "ok.example") {
+	if !hasRulesetRule(opts.Route, "ok") {
 		t.Fatalf("the healthy rule must survive alongside the malformed ones")
 	}
 	if !hasRouteToOutbound(opts, "n1") {
 		t.Fatalf("expected a route rule to n1")
 	}
 	// Every malformed matcher reported itself rather than failing silently.
-	for _, want := range []string{"geosite matcher", "is not a valid IP/CIDR", "domain regexp", "is not a valid port/range"} {
+	for _, want := range []string{
+		"unrecognised prefix",       // geosite: in a domain list
+		"no usable entries",         // ...leaving that list empty
+		"bad ip_cidr entry",         // 999.1.1.1/24
+		"is not a valid IP/CIDR",    // the source cidr
+		"domain regexp",             // regexp:*broken(
+		"is not a valid port/range", // a-b
+	} {
 		if !routeWarnsHave(warns, want) {
 			t.Fatalf("missing diagnostic %q in %v", want, warns)
 		}
@@ -155,10 +171,15 @@ func TestRuleKillPoliciesApply(t *testing.T) {
 		Inbounds: []model.Inbound{{Name: "lan", Enabled: true, Type: "tproxy", TproxyPort: 12398, TCP: true, UDP: true}},
 		Nodes:    []model.Node{{Name: "dead", Enabled: false, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#dead"}},
 		Groups:   []model.Group{{Name: "grp", Strategy: "leastping", Nodes: []string{"dead"}}},
+		Rulesets: []model.Ruleset{
+			inlineDomainSet("k-closed", "closed.example"),
+			inlineDomainSet("k-open", "open.example"),
+			inlineDomainSet("k-default", "default.example"),
+		},
 		Rules: []model.Rule{
-			{Name: "k-closed", Enabled: true, Order: 1, DstDomain: []string{"closed.example"}, Target: "group:grp", Kill: "closed"},
-			{Name: "k-open", Enabled: true, Order: 2, DstDomain: []string{"open.example"}, Target: "group:grp", Kill: "open"},
-			{Name: "k-default", Enabled: true, Order: 3, DstDomain: []string{"default.example"}, Target: "group:grp"},
+			{Name: "k-closed", Enabled: true, Order: 1, DstRuleset: []string{"k-closed"}, Target: "group:grp", Kill: "closed"},
+			{Name: "k-open", Enabled: true, Order: 2, DstRuleset: []string{"k-open"}, Target: "group:grp", Kill: "open"},
+			{Name: "k-default", Enabled: true, Order: 3, DstRuleset: []string{"k-default"}, Target: "group:grp"},
 		},
 	}

@@ -166,13 +187,13 @@ func TestRuleKillPoliciesApply(t *testing.T) {
 	if !changed {
 		t.Fatalf("expected Apply changed==true (warnings: %v)", warns)
 	}
-	if got, ok := domainRuleTarget(opts.Route, "closed.example"); !ok || got != tagBlock {
+	if got, ok := rulesetRuleTarget(opts.Route, "k-closed"); !ok || got != tagBlock {
 		t.Fatalf("kill=closed must emit a rule routed to block, got %q (ok=%v)", got, ok)
 	}
-	if got, ok := domainRuleTarget(opts.Route, "open.example"); !ok || got != tagDirect {
+	if got, ok := rulesetRuleTarget(opts.Route, "k-open"); !ok || got != tagDirect {
 		t.Fatalf("kill=open must emit a rule routed to direct, got %q (ok=%v)", got, ok)
 	}
-	if got, ok := domainRuleTarget(opts.Route, "default.example"); !ok || got != tagBlock {
+	if got, ok := rulesetRuleTarget(opts.Route, "k-default"); !ok || got != tagBlock {
 		t.Fatalf("kill=default must emit a rule routed to block, got %q (ok=%v)", got, ok)
 	}
 	if opts.Route.Final != tagBlock {
@@ -45,6 +45,40 @@ func genRules(t *testing.T, rules ...model.Rule) (*option.RouteOptions, []string
 	return opts.Route, warns
 }

+// ruleModelWithSets is ruleModel plus the `config ruleset` definitions the rules'
+// DstRuleset entries point at. A rule's ONLY destination matcher is a rule-set
+// (schema v2), so every case below that just needs "some destination the engine
+// can match on" declares one here rather than writing an inline domain/IP list.
+func ruleModelWithSets(sets []model.Ruleset, rules ...model.Rule) *model.Model {
+	m := ruleModel(rules...)
+	m.Rulesets = sets
+	return m
+}
+
+// genRulesWithSets is genRules for a model that also carries rule-sets.
+func genRulesWithSets(t *testing.T, sets []model.Ruleset, rules ...model.Rule) (*option.RouteOptions, []string) {
+	t.Helper()
+	opts, warns, err := GenerateWithWarnings(ruleModelWithSets(sets, rules...))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	return opts.Route, warns
+}
+
+// inlineDomainSet builds an inline DOMAIN `config ruleset`.
+//
+// Mind the convention the move to rule-sets brought with it: a BARE entry here is
+// a DomainSuffix (the apex AND its subdomains), whereas the routing rule's old
+// dst_domain read a bare entry as an EXACT domain. `full:` is the exact form.
+func inlineDomainSet(name string, entries ...string) model.Ruleset {
+	return model.Ruleset{Name: name, Type: "domain", Source: "inline", Entries: entries}
+}
+
+// inlineIPSet builds an inline IP-RANGE `config ruleset`.
+func inlineIPSet(name string, entries ...string) model.Ruleset {
+	return model.Ruleset{Name: name, Type: "ipcidr", Source: "inline", Entries: entries}
+}
+
 func routeWarnsHave(warns []string, substr string) bool {
 	for _, w := range warns {
 		if strings.Contains(w, substr) {
@@ -68,69 +102,117 @@ func generalRules(rt *option.RouteOptions) []option.Rule {

 // --- geosite: the landmine ---------------------------------------------------

-// TestGeositeMatcherIsInertNotFatal: route-rule `geosite` was REMOVED from this
-// engine — route/rule.NewDefaultRule returns "geosite database is deprecated ...
-// removed in sing-box 1.12.0" for a non-empty Geosite list, and that error aborts
-// box.New for the whole config. A `geosite:` entry must therefore be warned and
-// omitted (exactly like the geoip matcher below), never emitted.
-func TestGeositeMatcherIsInertNotFatal(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "geo", Enabled: true, Order: 10,
-		DstDomain: []string{"geosite:youtube"}, Target: "node:n1",
-	})
+// TestGeositeEntryInRulesetIsInertNotFatal: route-rule `geosite` was REMOVED from
+// this engine — route/rule.NewDefaultRule returns "geosite database is deprecated
+// ... removed in sing-box 1.12.0" for a non-empty Geosite list, and that error
+// aborts box.New for the whole config. Destinations now live in a `config
+// ruleset`, so a `geosite:` entry lands in an inline domain list, where it is an
+// unrecognised `word:` prefix: dropped by the classifier, reported, and — being
+// the list's only entry — leaving the rule-set with nothing to match, so it is
+// skipped and the rule that referenced it is not emitted either. Nothing about
+// that path may ever put a value in RawDefaultRule.Geosite. (`source=geosite` on
+// the ruleset itself is the working way to route a category; see
+// TestRoutingRuleSetGeositeCategory.)
+func TestGeositeEntryInRulesetIsInertNotFatal(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("geo", "geosite:youtube")},
+		model.Rule{Name: "geo", Enabled: true, Order: 10, DstRuleset: []string{"geo"}, Target: "node:n1"},
+	)
 	for _, r := range rt.Rules {
 		if len(r.DefaultOptions.RawDefaultRule.Geosite) > 0 {
 			t.Fatalf("geosite must never be emitted (aborts box.New), got %v", r.DefaultOptions.RawDefaultRule.Geosite)
 		}
 	}
-	if !routeWarnsHave(warns, "geosite matcher") {
-		t.Fatalf("expected an inert-geosite warning, got %v", warns)
+	if _, ok := ruleSetByTag(rt, "rs-geo"); ok {
+		t.Fatalf("a rule-set with no usable entry must not be emitted (an empty list would match everything)")
+	}
+	if findRouteRuleWithRuleSet(rt, "rs-geo") != nil {
+		t.Fatalf("no rule may reference the skipped rule-set")
+	}
+	if !routeWarnsHave(warns, "unrecognised prefix") {
+		t.Fatalf("expected an unrecognised-prefix warning for geosite:, got %v", warns)
+	}
+	if !routeWarnsHave(warns, "no usable entries") {
+		t.Fatalf("expected a no-usable-entries warning, got %v", warns)
 	}
 }

-// TestGeositeMixedWithRealDomainKeepsTheRest: a rule carrying BOTH a geosite entry
-// and a real domain keeps the real matcher and still routes — only the geosite
-// part is dropped.
+// TestGeositeMixedWithRealDomainKeepsTheRest: a rule-set carrying BOTH a geosite
+// entry and a real domain keeps the real matcher, and the rule referencing it
+// still routes — only the geosite entry is dropped.
 func TestGeositeMixedWithRealDomainKeepsTheRest(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "mixed", Enabled: true, Order: 10,
-		DstDomain: []string{"geosite:youtube", "example.com"}, Target: "node:n1",
-	})
-	gen := generalRules(rt)
-	if len(gen) != 1 {
-		t.Fatalf("want 1 general rule, got %d (warnings %v)", len(gen), warns)
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("mixed", "geosite:youtube", "full:example.com")},
+		model.Rule{Name: "mixed", Enabled: true, Order: 10, DstRuleset: []string{"mixed"}, Target: "node:n1"},
+	)
+	rs, ok := ruleSetByTag(rt, "rs-mixed")
+	if !ok {
+		t.Fatalf("rs-mixed must survive the geosite entry (warnings %v)", warns)
 	}
-	raw := gen[0].DefaultOptions.RawDefaultRule
-	if len(raw.Geosite) != 0 {
-		t.Fatalf("geosite leaked: %v", raw.Geosite)
+	hr := rs.InlineOptions.Rules[0].DefaultOptions
+	if len(hr.Domain) != 1 || hr.Domain[0] != "example.com" {
+		t.Fatalf("real domain matcher lost: %+v", hr)
 	}
-	if len(raw.Domain) != 1 || raw.Domain[0] != "example.com" {
-		t.Fatalf("real domain matcher lost: %+v", raw.Domain)
+	if len(hr.DomainSuffix)+len(hr.DomainKeyword)+len(hr.DomainRegex) != 0 {
+		t.Fatalf("geosite: must be dropped, not reinterpreted: %+v", hr)
 	}
-	if got := gen[0].DefaultOptions.RuleAction.RouteOptions.Outbound; got != "n1" {
+	dr := findRouteRuleWithRuleSet(rt, "rs-mixed")
+	if dr == nil {
+		t.Fatalf("the rule must be emitted referencing rs-mixed; rules=%+v", rt.Rules)
+	}
+	if got := dr.RuleAction.RouteOptions.Outbound; got != "n1" {
 		t.Fatalf("target = %q, want n1", got)
 	}
 }

+// TestGeoipEntryInRulesetIsInertNotFatal is the IP-side twin: model.migrate1to2
+// moves an old `dst_ip geoip:ru` verbatim into an inline type=ipcidr rule-set
+// (deliberately — see TestMigrate1to2KeepsGeoMarkersInert: it must not silently
+// become a working geoip source, because the operator never asked to download
+// anything). Here it is an unparseable prefix, so it must be dropped LOUDLY
+// rather than reach NewIPCIDRItem, which errors and aborts box.New.
+func TestGeoipEntryInRulesetIsInertNotFatal(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineIPSet("geo", "geoip:ru")},
+		model.Rule{Name: "geo", Enabled: true, Order: 10, DstRuleset: []string{"geo"}, Target: "node:n1"},
+	)
+	for _, r := range rt.Rules {
+		if len(r.DefaultOptions.RawDefaultRule.GeoIP) > 0 {
+			t.Fatalf("geoip must never be emitted, got %v", r.DefaultOptions.RawDefaultRule.GeoIP)
+		}
+	}
+	if _, ok := ruleSetByTag(rt, "rs-geo"); ok {
+		t.Fatalf("a list whose only entry is unparseable must not materialise")
+	}
+	if !routeWarnsHave(warns, `bad ip_cidr entry "geoip:ru"`) {
+		t.Fatalf("expected a bad-ip_cidr warning naming the entry, got %v", warns)
+	}
+}
+
 // --- malformed matchers that used to abort box.New ---------------------------

-// TestBadDstCIDRWarnsAndSkips: an unparseable ip_cidr makes
-// route/rule.NewIPCIDRItem error, which aborts box.New. It must be dropped.
-func TestBadDstCIDRWarnsAndSkips(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "bad", Enabled: true, Order: 10,
-		DstIP: []string{"999.1.1.1/24", "198.51.100.0/24"}, Target: "node:n1",
-	})
-	gen := generalRules(rt)
-	if len(gen) != 1 {
-		t.Fatalf("want 1 general rule, got %d", len(gen))
+// TestRulesetBadIPCIDREntryWarnsAndSkips: an unparseable ip_cidr makes
+// route/rule.NewIPCIDRItem error, which aborts box.New. Destination addresses are
+// an inline `type=ipcidr` rule-set now, so the guard lives in inlineRulesetRule:
+// the typo is dropped, the valid entry survives and the rule still routes.
+func TestRulesetBadIPCIDREntryWarnsAndSkips(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineIPSet("bad", "999.1.1.1/24", "198.51.100.0/24")},
+		model.Rule{Name: "bad", Enabled: true, Order: 10, DstRuleset: []string{"bad"}, Target: "node:n1"},
+	)
+	rs, ok := ruleSetByTag(rt, "rs-bad")
+	if !ok {
+		t.Fatalf("one bad entry must not take the whole list down (warnings %v)", warns)
 	}
-	raw := gen[0].DefaultOptions.RawDefaultRule
-	if len(raw.IPCIDR) != 1 || raw.IPCIDR[0] != "198.51.100.0/24" {
-		t.Fatalf("ip_cidr = %v, want only the valid entry", raw.IPCIDR)
+	got := rs.InlineOptions.Rules[0].DefaultOptions.IPCIDR
+	if len(got) != 1 || got[0] != "198.51.100.0/24" {
+		t.Fatalf("ip_cidr = %v, want only the valid entry", got)
 	}
-	if !routeWarnsHave(warns, `destination "999.1.1.1/24" is not a valid IP/CIDR`) {
-		t.Fatalf("expected a bad-destination warning, got %v", warns)
+	if !routeWarnsHave(warns, `bad ip_cidr entry "999.1.1.1/24"`) {
+		t.Fatalf("expected a bad-ip_cidr warning, got %v", warns)
+	}
+	if findRouteRuleWithRuleSet(rt, "rs-bad") == nil {
+		t.Fatalf("the rule must still be emitted referencing rs-bad; rules=%+v", rt.Rules)
 	}
 }

@@ -152,24 +234,30 @@ func TestBadSrcCIDRWarnsAndSkips(t *testing.T) {
 	}
 }

-// TestBadDomainRegexWarnsAndSkips: an uncompilable `regexp:` pattern makes
-// route/rule.NewDomainRegexItem error and abort box.New.
-func TestBadDomainRegexWarnsAndSkips(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "bad", Enabled: true, Order: 10,
-		DstDomain: []string{"regexp:*broken(", `regexp:^ok\.example$`}, Target: "node:n1",
-	})
-	gen := generalRules(rt)
-	if len(gen) != 1 {
-		t.Fatalf("want 1 general rule, got %d", len(gen))
+// TestRulesetBadDomainRegexWarnsAndSkips: an uncompilable `regexp:` pattern makes
+// route/rule.NewDomainRegexItem error and abort box.New. The pattern vocabulary
+// moved into the inline rule-set with the rest of the destination list, so the
+// validation moved with it (peelDomainRegexes): the broken pattern is dropped and
+// the compilable one survives.
+func TestRulesetBadDomainRegexWarnsAndSkips(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("bad", "regexp:*broken(", `regexp:^ok\.example$`)},
+		model.Rule{Name: "bad", Enabled: true, Order: 10, DstRuleset: []string{"bad"}, Target: "node:n1"},
+	)
+	rs, ok := ruleSetByTag(rt, "rs-bad")
+	if !ok {
+		t.Fatalf("one broken pattern must not take the whole list down (warnings %v)", warns)
 	}
-	got := gen[0].DefaultOptions.RawDefaultRule.DomainRegex
+	got := rs.InlineOptions.Rules[0].DefaultOptions.DomainRegex
 	if len(got) != 1 || got[0] != `^ok\.example$` {
 		t.Fatalf("domain_regex = %v, want only the compilable one", got)
 	}
 	if !routeWarnsHave(warns, "domain regexp") {
 		t.Fatalf("expected a bad-regexp warning, got %v", warns)
 	}
+	if findRouteRuleWithRuleSet(rt, "rs-bad") == nil {
+		t.Fatalf("the rule must still be emitted referencing rs-bad; rules=%+v", rt.Rules)
+	}
 }

 // TestBadPortRangeWarnsAndSkips: a malformed range reaches
@@ -911,10 +999,12 @@ func TestRuleProtoKnownValuesAreSilent(t *testing.T) {
 // that would break existing configs either open or closed), but the widening is
 // now reported with its consequence.
 func TestIfaceOnlySourceRuleIsNotSilentlyNetworkWide(t *testing.T) {
-	rt, warns := genRules(t, model.Rule{
-		Name: "guest", Enabled: true, Order: 10,
-		Src: []string{"iface:guest"}, DstDomain: []string{"youtube.com"}, Target: "block",
-	})
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("yt", "youtube.com")},
+		model.Rule{
+			Name: "guest", Enabled: true, Order: 10,
+			Src: []string{"iface:guest"}, DstRuleset: []string{"yt"}, Target: "block",
+		})
 	if !routeWarnsHave(warns, "applies to EVERY client") {
 		t.Fatalf("expected a rule-widening warning, got %v", warns)
 	}
@@ -930,11 +1020,13 @@ func TestIfaceOnlySourceRuleIsNotSilentlyNetworkWide(t *testing.T) {
 // TestMixedSourceRuleReportsTheDroppedHalf: with one usable IP source alongside
 // an unmatchable one, the rule narrows to the IP source only.
 func TestMixedSourceRuleReportsTheDroppedHalf(t *testing.T) {
-	_, warns := genRules(t, model.Rule{
-		Name: "mixed", Enabled: true, Order: 10,
-		Src:       []string{"iface:guest", "aa:bb:cc:dd:ee:ff", "192.168.5.0/24"},
-		DstDomain: []string{"youtube.com"}, Target: "block",
-	})
+	_, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("yt", "youtube.com")},
+		model.Rule{
+			Name: "mixed", Enabled: true, Order: 10,
+			Src:        []string{"iface:guest", "aa:bb:cc:dd:ee:ff", "192.168.5.0/24"},
+			DstRuleset: []string{"yt"}, Target: "block",
+		})
 	if !routeWarnsHave(warns, "could not be used") {
 		t.Fatalf("expected a dropped-source warning, got %v", warns)
 	}
@@ -1041,41 +1133,73 @@ func TestRuleTargetKindsResolve(t *testing.T) {
 	}
 }

-// TestRuleDomainUnrecognisedPrefixWarns: an unknown `word:` prefix is dropped by
-// the shared classifier (a domain cannot contain ":"), so the rule silently lost
-// that destination. `domain:example.com` is the v0.1/xray spelling a migrating
-// user writes, and it must not disappear without a trace.
-func TestRuleDomainUnrecognisedPrefixWarns(t *testing.T) {
+// TestRulesetDomainUnrecognisedPrefixWarns: an unknown `word:` prefix is dropped
+// by the shared classifier (a domain cannot contain ":"), so the list silently
+// lost that destination. `domain:example.com` is the v0.1/xray spelling a
+// migrating user writes, and it must not disappear without a trace — the more so
+// now that a destination list is ALWAYS a rule-set, i.e. the one place a typo can
+// hide.
+func TestRulesetDomainUnrecognisedPrefixWarns(t *testing.T) {
 	for _, entry := range []string{"domain:example.com", "regex:example.com", "ext:foo.dat:cn"} {
-		rt, warns := genRules(t, model.Rule{
-			Name: "mig", Enabled: true, Order: 10,
-			DstDomain: []string{entry, "keep.example"}, Target: "block",
-		})
+		rt, warns := genRulesWithSets(t,
+			[]model.Ruleset{inlineDomainSet("mig", entry, "keep.example")},
+			model.Rule{Name: "mig", Enabled: true, Order: 10, DstRuleset: []string{"mig"}, Target: "block"},
+		)
 		if !routeWarnsHave(warns, "unrecognised prefix") {
 			t.Fatalf("%q: expected an unrecognised-prefix warning, got %v", entry, warns)
 		}
-		gen := generalRules(rt)
-		if len(gen) != 1 {
-			t.Fatalf("%q: want 1 rule, got %d", entry, len(gen))
+		rs, ok := ruleSetByTag(rt, "rs-mig")
+		if !ok {
+			t.Fatalf("%q: the usable entry must keep the rule-set alive; warnings %v", entry, warns)
 		}
-		for _, d := range gen[0].DefaultOptions.RawDefaultRule.Domain {
-			if strings.Contains(d, ":") {
-				t.Fatalf("%q: a prefixed literal reached the matcher: %q", entry, d)
+		hr := rs.InlineOptions.Rules[0].DefaultOptions
+		for _, list := range [][]string{hr.Domain, hr.DomainSuffix, hr.DomainKeyword, hr.DomainRegex} {
+			for _, d := range list {
+				if strings.Contains(d, ":") {
+					t.Fatalf("%q: a prefixed literal reached the matcher: %q", entry, d)
+				}
 			}
 		}
+		if findRouteRuleWithRuleSet(rt, "rs-mig") == nil {
+			t.Fatalf("%q: the rule must still be emitted; rules=%+v", entry, rt.Rules)
+		}
 	}
 }

-// TestRuleDomainKnownPrefixesAreSilent guards the warning against false
-// positives on the vocabulary routing rules really support.
-func TestRuleDomainKnownPrefixesAreSilent(t *testing.T) {
-	_, warns := genRules(t, model.Rule{
-		Name: "ok", Enabled: true, Order: 10, Target: "block",
-		DstDomain: []string{"full:a.example", "suffix:b.example", "keyword:c", `regexp:^d\.`, ".e.example", "f.example"},
-	})
+// TestRulesetDomainKnownPrefixesAreSilent guards the warning against false
+// positives on the vocabulary an inline domain rule-set really supports.
+//
+// `regexp:` is the load-bearing case: the shared classifier has no branch for it,
+// so it WOULD be reported as an unknown prefix — inlineRulesetRule peels the
+// regexes off first (peelDomainRegexes) precisely so it is not. A regression there
+// would both warn about a working matcher and drop it.
+func TestRulesetDomainKnownPrefixesAreSilent(t *testing.T) {
+	rt, warns := genRulesWithSets(t,
+		[]model.Ruleset{inlineDomainSet("ok",
+			"full:a.example", "suffix:b.example", "keyword:c", `regexp:^d\.`, ".e.example", "f.example")},
+		model.Rule{Name: "ok", Enabled: true, Order: 10, DstRuleset: []string{"ok"}, Target: "block"},
+	)
 	if routeWarnsHave(warns, "unrecognised prefix") {
 		t.Fatalf("the supported prefixes must not warn: %v", warns)
 	}
+	rs, ok := ruleSetByTag(rt, "rs-ok")
+	if !ok {
+		t.Fatalf("rs-ok not emitted; warnings %v", warns)
+	}
+	hr := rs.InlineOptions.Rules[0].DefaultOptions
+	if len(hr.Domain) != 1 || hr.Domain[0] != "a.example" {
+		t.Fatalf("full: must be an exact Domain, got %v", hr.Domain)
+	}
+	if len(hr.DomainRegex) != 1 || hr.DomainRegex[0] != `^d\.` {
+		t.Fatalf("regexp: must survive as a domain_regex matcher, got %v", hr.DomainRegex)
+	}
+	// suffix:, the leading dot and the BARE entry all collapse to domain_suffix.
+	if len(hr.DomainSuffix) != 3 {
+		t.Fatalf("domain_suffix = %v, want b/e/f.example (bare entry is a suffix in a rule-set)", hr.DomainSuffix)
+	}
+	if len(hr.DomainKeyword) != 1 || hr.DomainKeyword[0] != "c" {
+		t.Fatalf("domain_keyword = %v, want [c]", hr.DomainKeyword)
+	}
 }

 // TestDeviceDomainUnrecognisedPrefixWarns is the same guarantee in the place it
@@ -0,0 +1,173 @@
+// B1: a routing rule that can never fire, reported instead of applied silently.
+//
+// The field config that prompted this had TWO rules named `default`, both with
+// zero conditions — order 20 -> direct and order 100 -> group:auto. buildRoute
+// points route Final at a condition-less rule and moves on, so the LAST one wins
+// and the other is a dead setting. Nothing anywhere said so: the log was clean and
+// the panel drew both rows identically.
+//
+// These tests pin the diagnosis AND the behaviour it describes, because a warning
+// that disagrees with what the generator actually does is worse than none.
+package generate
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/sagernet/sing-box/shater/model"
+)
+
+// warnAbout returns the warnings mentioning `rule "name"`.
+func warnAbout(warns []string, name string) []string {
+	var out []string
+	for _, w := range warns {
+		if strings.Contains(w, `rule "`+name+`"`) {
+			out = append(out, w)
+		}
+	}
+	return out
+}
+
+func warnsMatching(warns []string, substr string) []string {
+	var out []string
+	for _, w := range warns {
+		if strings.Contains(w, substr) {
+			out = append(out, w)
+		}
+	}
+	return out
+}
+
+const neverApplied = "is never applied"
+
+// twoDefaultsModel reproduces the field config: two condition-less rules sharing
+// the name `default`, differing only in Order and target.
+func twoDefaultsModel(lowTarget, highTarget string) *model.Model {
+	g := model.DefaultGlobals()
+	g.KillSwitch = "closed"
+	return &model.Model{
+		Globals: g,
+		Nodes:   []model.Node{{Name: "n1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+		Groups:  []model.Group{{Name: "auto", Strategy: "leastping", Nodes: []string{"n1"}}},
+		Rules: []model.Rule{
+			{Name: "default", Enabled: true, Order: 20, Target: lowTarget},
+			{Name: "default", Enabled: true, Order: 100, Target: highTarget},
+		},
+	}
+}
+
+// TestTwoCatchAllRulesWarnAndLastWins is the B1 regression. The order-100 rule is
+// the default the engine uses (route Final), and the order-20 one is reported as
+// never applied — naming the rule that supersedes it.
+func TestTwoCatchAllRulesWarnAndLastWins(t *testing.T) {
+	opts, warns, err := GenerateWithWarnings(twoDefaultsModel("direct", "group:auto"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := opts.Route.Final; got != "auto" {
+		t.Fatalf("route Final = %q, want the LAST catch-all's target %q", got, "auto")
+	}
+	got := warnsMatching(warns, neverApplied)
+	if len(got) != 1 {
+		t.Fatalf("want exactly one never-applied warning, got %d: %q", len(got), warns)
+	}
+	// It must name the superseding rule AND its order — with both rules called
+	// `default`, the order is the only thing that tells the two apart.
+	// Targets are quoted as the OPERATOR wrote them (`group:auto`), not as the
+	// engine tag they resolve to (`auto`) — the warning has to be readable next to
+	// the config, not next to the generated JSON.
+	for _, want := range []string{`rule "default"`, "order 100", `"group:auto"`, `"direct"`} {
+		if !strings.Contains(got[0], want) {
+			t.Fatalf("warning must mention %s, got: %s", want, got[0])
+		}
+	}
+	// Diagnosis only: neither rule is renamed, reordered or dropped.
+	if len(opts.Route.Rules) == 0 {
+		t.Fatal("route rules disappeared")
+	}
+}
+
+// TestTwoCatchAllsDirectWinsIsCritical: when the surviving default is `direct` and
+// the retired one asked for a tunnel, everything unmatched leaves on the plain
+// WAN. The warning must say so in the words apply/warnings.go grades critical —
+// this is the case where a healthy-looking panel is a lie.
+func TestTwoCatchAllsDirectWinsIsCritical(t *testing.T) {
+	_, warns, err := GenerateWithWarnings(twoDefaultsModel("group:auto", "direct"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	got := warnsMatching(warns, neverApplied)
+	if len(got) != 1 {
+		t.Fatalf("want exactly one never-applied warning, got %d: %q", len(got), warns)
+	}
+	const marker = "leaves over the plain WAN with your real IP address"
+	if !strings.Contains(got[0], marker) {
+		t.Fatalf("a retired tunnel default under a live direct default must carry %q, got: %s", marker, got[0])
+	}
+}
+
+// TestTwoCatchAllsTunnelWinsIsNotCritical: the field config's actual shape — the
+// dead rule is `direct` and the live default is the tunnel. That is a dead
+// setting, not a leak, so it must NOT carry the critical marker.
+func TestTwoCatchAllsTunnelWinsIsNotCritical(t *testing.T) {
+	_, warns, err := GenerateWithWarnings(twoDefaultsModel("direct", "group:auto"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	for _, w := range warnsMatching(warns, neverApplied) {
+		if strings.Contains(w, "leaves over the plain WAN with your real IP address") {
+			t.Fatalf("a dead direct default under a live tunnel default is not a leak: %s", w)
+		}
+	}
+}
+
+// TestSingleCatchAllDoesNotWarn: the ordinary config — specific rules plus ONE
+// default — must stay silent. A badge on a working rule teaches the operator to
+// ignore the badge.
+func TestSingleCatchAllDoesNotWarn(t *testing.T) {
+	g := model.DefaultGlobals()
+	g.KillSwitch = "closed"
+	m := &model.Model{
+		Globals: g,
+		Nodes:   []model.Node{{Name: "n1", Enabled: true, URI: "ss://aes-256-gcm:secret@203.0.113.1:8388#n1"}},
+		Groups:  []model.Group{{Name: "auto", Strategy: "leastping", Nodes: []string{"n1"}}},
+		Rules: []model.Rule{
+			{Name: "ads", Enabled: true, Order: 10, DstPort: "443", Target: "block"},
+			// A specific rule ordered BELOW the default: still emitted ahead of Final,
+			// so it is not retired either.
+			{Name: "default", Enabled: true, Order: 20, Target: "group:auto"},
+			{Name: "late", Enabled: true, Order: 900, DstPort: "8080", Target: "direct"},
+		},
+	}
+	_, warns, err := GenerateWithWarnings(m)
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := warnsMatching(warns, neverApplied); len(got) != 0 {
+		t.Fatalf("a config with one default must not report anything never-applied, got: %q", got)
+	}
+	if got := warnAbout(warns, "late"); len(got) != 0 {
+		t.Fatalf("a specific rule below the default is not shadowed by it, got: %q", got)
+	}
+}
+
+// TestProfileDisabledCatchAllDoesNotShadow: the active profile switches the later
+// default off, so the earlier one is the live default and must not be badged.
+// Judging the raw config would blame the wrong rule on every profile router.
+func TestProfileDisabledCatchAllDoesNotShadow(t *testing.T) {
+	m := twoDefaultsModel("group:auto", "direct")
+	m.Rules[1].Name = "fallback" // profiles address rules by name
+	m.Globals.ActiveProfile = "home"
+	m.Profiles = []model.Profile{{Name: "home", Enabled: true, DisableRules: []string{"fallback"}}}
+
+	opts, warns, err := GenerateWithWarnings(m)
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := warnsMatching(warns, neverApplied); len(got) != 0 {
+		t.Fatalf("a profile-disabled default shadows nothing, got: %q", got)
+	}
+	if got := opts.Route.Final; got != "auto" {
+		t.Fatalf("route Final = %q, want the surviving default's target %q", got, "auto")
+	}
+}
@@ -25,6 +25,7 @@ import (
 	"net/netip"
 	"os"
 	"path/filepath"
+	"regexp"
 	"runtime/debug"
 	"strings"
 	"sync"
@@ -500,6 +501,47 @@ func ruleSetURLIsEngineNative(rawURL string) (format string, native bool) {
 	}
 }

+// TWO WAYS TO WRITE A DESTINATION LIST, AND WHY THEY DO NOT SHARE A VOCABULARY.
+//
+// D21 promises "one destination mechanism, one vocabulary". The vocabulary half of
+// that promise is about the ENTRIES AN OPERATOR TYPES, and those live in exactly
+// one place: an inline `config ruleset` (inlineRulesetRule -> peelDomainRegexes +
+// classifyDomainEntries), where `full:` / `suffix:` / `keyword:` / `regexp:` / a
+// leading dot / a bare name all mean what docs-shater/DECISIONS.md D21 says.
+//
+// A `source=url` list that is not engine-native is NOT another spelling of that.
+// It is a FILE FORMAT — the hosts / plain-domain / AdBlock-ish text third parties
+// publish — and parseDomainList is a parser for that format, not for our entry
+// vocabulary. `source=file` is a third thing again: a compiled .srs or a rule-set
+// .json handed straight to the engine, which never sees shater's entry syntax at
+// all. (An earlier review read this as "the same list written two ways behaves
+// differently"; it is closer to "a typed list and a downloaded file are different
+// artifacts". The diagnostics below exist so an operator never has to guess which
+// one they are looking at.)
+//
+// Unifying them was considered and REJECTED, on three grounds:
+//
+//   - The formats collide. A real AdGuard/OISD list is full of colon-bearing lines
+//     that are not our markers at all (`example.com##.banner:has(...)`, `$domain=`
+//     options, absolute URL rules). Feeding those through the marker classifier
+//     would either mis-import them or, if we reported every `word:`-shaped token as
+//     an unknown prefix, drown the operator in hundreds of warnings per list — a
+//     louder dishonesty than the quiet one it replaces.
+//   - The shapes collide. A hosts line carries SEVERAL names ("127.0.0.1 a.com
+//     b.com"), so this parser works per TOKEN; the entry vocabulary works per LINE
+//     and allows a space after the marker ("keyword: ads"). There is no split rule
+//     that serves both.
+//   - `regexp:` from a URL is regex supplied by a third party, compiled into the
+//     router's matcher and evaluated per query on a 512 MB box. The inline path can
+//     accept it because the operator typed it; a downloaded list is not that.
+//
+// So the difference STAYS, and is paid for in diagnostics instead: a text list that
+// carries our marker vocabulary is reported per list (see listEntryMarker and
+// warnListEntryVocabulary), naming the entries and where they DO work. The check is
+// narrow on purpose — only the four markers D21 defines, never the general `word:`
+// shape — so it fires on an operator's mistake and stays silent on ordinary filter
+// syntax.
+//
 // parseDomainList extracts domains from the formats public blocklists ship in:
 //
 //   - HOSTS         "0.0.0.0 ads.example.com", "127.0.0.1 a.com b.com"
@@ -517,8 +559,12 @@ func ruleSetURLIsEngineNative(rawURL string) (format string, native bool) {
 // de-duplication map is kept — domain.NewMatcher already de-duplicates internally
 // while building the succinct set, so a second map would just double the largest
 // allocation in the pipeline. See listMaxDomains for the measured budget.
-func parseDomainList(content []byte) []string {
+//
+// The second return reports the inline-vocabulary entries seen on the way past, so
+// the caller can say so instead of dropping them without a word.
+func parseDomainList(content []byte) ([]string, listMarkerNote) {
 	out := make([]string, 0, 4096)
+	var note listMarkerNote
 	scanner := bufio.NewScanner(bytes.NewReader(content))
 	// Public lists are one domain per line; 64 KiB is far beyond any real line, and
 	// an over-long line is skipped rather than aborting the parse.
@@ -544,17 +590,123 @@ func parseDomainList(content []byte) []string {
 			// AdBlock-ish "||domain^" -> domain.
 			f = strings.TrimPrefix(f, "||")
 			f = strings.TrimSuffix(f, "^")
+			if _, isMarker := listEntryMarker(f); isMarker {
+				// normaliseListDomain would drop this silently (a domain cannot contain
+				// ":"). Record it so the caller can name it; it is the one class of junk
+				// in a text list that is provably an operator mistake rather than filter
+				// syntax we simply do not import.
+				note.record(f)
+				continue
+			}
 			d, ok := normaliseListDomain(f)
 			if !ok {
 				continue
 			}
 			out = append(out, d)
 			if len(out) >= listMaxDomains {
-				return out
+				return out, note
 			}
 		}
 	}
-	return out
+	return out, note
+}
+
+// listEntryVocabulary is EXACTLY the marker set an INLINE rule-set entry may use
+// (D21). It is deliberately NOT the general `word:` shape unrecognisedDomainPrefix
+// tests for: a published filter list legitimately contains hundreds of colon-
+// bearing tokens, and reporting those would make the diagnostic useless. These
+// four, by contrast, appear in a downloaded text list only when a human wrote them
+// there expecting shater to honour them.
+var listEntryVocabulary = []string{"full:", "suffix:", "keyword:", "regexp:"}
+
+// listEntryMarker reports whether a text-list token is written in the inline entry
+// vocabulary, and which marker it used.
+func listEntryMarker(token string) (string, bool) {
+	lower := strings.ToLower(strings.TrimSpace(token))
+	for _, m := range listEntryVocabulary {
+		if strings.HasPrefix(lower, m) {
+			return m, true
+		}
+	}
+	return "", false
+}
+
+// listMarkerSamples bounds how many offending entries a warning quotes. A list is
+// remote content: it must not be able to write an unbounded amount into our log.
+const listMarkerSamples = 5
+
+// listMarkerNote records the inline-vocabulary entries one plain-text list carried.
+type listMarkerNote struct {
+	Count   int
+	Samples []string
+}
+
+func (n *listMarkerNote) record(entry string) {
+	n.Count++
+	if len(n.Samples) < listMarkerSamples {
+		n.Samples = append(n.Samples, strings.TrimSpace(entry))
+	}
+}
+
+func (n listMarkerNote) empty() bool { return n.Count == 0 }
+
+// listMarkerMemo remembers, per list URL, what the last COMPILATION of that list
+// found. Without it the diagnostic would exist for exactly one reconcile — the one
+// that happened to refresh the artifact — and then vanish for a whole
+// update_interval, which is precisely the "reported to nobody" failure it is meant
+// to fix. Same shape (and same reasoning) as ruleSetProbeCache above; a refresh
+// that finds nothing clears the entry, so fixing the list silences it.
+var (
+	listMarkerMu   sync.Mutex
+	listMarkerMemo = map[string]listMarkerNote{}
+)
+
+func rememberListMarkers(url string, note listMarkerNote) {
+	listMarkerMu.Lock()
+	if note.empty() {
+		delete(listMarkerMemo, url)
+	} else {
+		listMarkerMemo[url] = note
+	}
+	listMarkerMu.Unlock()
+}
+
+func recallListMarkers(url string) listMarkerNote {
+	listMarkerMu.Lock()
+	defer listMarkerMu.Unlock()
+	return listMarkerMemo[url]
+}
+
+// resetListMarkerMemo clears the memo (tests).
+func resetListMarkerMemo() {
+	listMarkerMu.Lock()
+	listMarkerMemo = map[string]listMarkerNote{}
+	listMarkerMu.Unlock()
+}
+
+// warnListEntryVocabulary tells the operator that entries written in the INLINE
+// entry vocabulary were found in a downloaded TEXT list, where they mean nothing.
+// See the parseDomainList block above for why the two vocabularies are separate
+// and why saying so is the whole of the fix.
+func (b *builder) warnListEntryVocabulary(diag, url string) {
+	note := recallListMarkers(url)
+	if note.empty() {
+		return
+	}
+	b.warnf("%s: %q is a plain-text list (a hosts file or one domain per line), but %d of its entries are written in the "+
+		"inline rule-set vocabulary (e.g. %s) — a plain-text list has NO markers, so every line is read as a domain plus its "+
+		"subdomains and anything containing \":\" is dropped, because a domain name cannot contain one. Those entries match NOTHING. "+
+		"Put them in a rule-set with source=inline, which is the one place full:/suffix:/keyword:/regexp: are honoured.",
+		diag, url, note.Count, quoteList(note.Samples))
+}
+
+// quoteList renders sample entries for a diagnostic.
+func quoteList(in []string) string {
+	out := make([]string, 0, len(in))
+	for _, s := range in {
+		out = append(out, fmt.Sprintf("%q", s))
+	}
+	return strings.Join(out, ", ")
 }

 // hostsBoilerplate are the names every hosts file carries for its own bookkeeping.
@@ -675,6 +827,11 @@ func (b *builder) compiledListRuleSet(tag, url, updateInterval, diag string) (op
 		}
 	}

+	// Reported on EVERY generate, not only on the one that refreshed the artifact:
+	// an entry that matches nothing is exactly as wrong the day after it was
+	// compiled as the moment it was.
+	b.warnListEntryVocabulary(diag, url)
+
 	return option.RuleSet{
 		Type:         C.RuleSetTypeLocal,
 		Tag:          tag,
@@ -716,8 +873,11 @@ func (b *builder) refreshCompiledList(path, url, diag string) error {
 	if err != nil {
 		return err
 	}
-	domains := parseDomainList(body)
+	domains, markers := parseDomainList(body)
 	body = nil // release the source text before the matcher allocates
+	// Remember (or clear) what this compilation saw, so the diagnostic survives the
+	// reconciles that do no I/O at all. See warnListEntryVocabulary.
+	rememberListMarkers(url, markers)

 	if len(domains) == 0 {
 		return fmt.Errorf("no usable domains found at %s (fetched %s, but nothing in it parsed as a domain)", url, "the file")
@@ -1163,7 +1323,8 @@ func (b *builder) buildRoutingRuleSetRaw(rs model.Ruleset) ([]option.RuleSet, []
 // its Entries, keyed by Type: an ipcidr ruleset fills ip_cidr; a domain ruleset
 // (the default) is classified with inlineDomainRule — bare entry => DomainSuffix
 // (so subdomains match), full: => Domain, keyword: => DomainKeyword, . => suffix
-// — the same classification the DNS filter uses. ok=false when nothing usable.
+// — the same classification the DNS filter uses, plus `regexp:` (see
+// peelDomainRegexes). ok=false when nothing usable.
 func (b *builder) inlineRulesetRule(rs model.Ruleset) (option.DefaultHeadlessRule, bool) {
 	b.warnUnknownRuleSetType(fmt.Sprintf("ruleset %q", rs.Name), rs.Type)
 	if ruleSetTypeIsIPCIDR(rs.Type) {
@@ -1190,10 +1351,73 @@ func (b *builder) inlineRulesetRule(rs model.Ruleset) (option.DefaultHeadlessRul
 		return option.DefaultHeadlessRule{IPCIDR: badoption.Listable[string](cidrs)}, true
 	}
 	// "domain" (and empty, and anything unrecognised => domain, warned above).
+	// `regexp:` is peeled off first: it is a routing-rule matcher the DNS-filter
+	// classifier does not know, and it must not be reported as an unknown prefix.
+	diag := fmt.Sprintf("ruleset %q", rs.Name)
+	rest, regexes := b.peelDomainRegexes(diag, rs.Entries)
+	// R4, on the path every destination list now takes. classifyDomainEntries
+	// DROPS an entry that is nothing but its marker (".", "full:", "keyword:"),
+	// silently — and the silence is the dangerous half: an empty domain token
+	// aborts box.New for the whole config, and an empty keyword is
+	// strings.Contains(host, "") i.e. EVERY host. The drop is right; not saying so
+	// is not. (devices.go reports the same class for a device's own lists; the bare
+	// `regexp:` form is reported by peelDomainRegexes above, which is why it is
+	// peeled off before this loop and cannot be double-reported.)
+	for _, e := range rest {
+		if isDomainMarkerOnly(e) {
+			b.warnf("%s: entry %q is a bare matcher marker with no value, omitted (an empty domain token aborts box.New; an empty keyword would match EVERY host)", diag, strings.TrimSpace(e))
+		}
+	}
 	// Only the DOMAIN branch reports unknown `word:` prefixes — the ipcidr branch
 	// above is full of legitimate colons (IPv6) and must never be checked (R9.2).
-	b.warnUnrecognisedPrefixes(fmt.Sprintf("ruleset %q", rs.Name), rs.Entries)
-	return inlineDomainRule(rs.Entries)
+	b.warnUnrecognisedPrefixes(diag, rest)
+	hr, ok := inlineDomainRule(rest)
+	if len(regexes) > 0 {
+		hr.DomainRegex = badoption.Listable[string](regexes)
+		ok = true
+	}
+	return hr, ok
+}
+
+// peelDomainRegexes splits `regexp:<pattern>` entries out of a domain rule-set's
+// entry list, returning the remaining entries and the validated patterns.
+//
+// It exists because a destination list is now ALWAYS a rule-set (schema v2), so
+// every matcher a `dst_domain` used to express has to be expressible here —
+// including the regex form, which the shared DNS-filter classifier
+// (classifyDomainEntries) deliberately does not know about. Validation mirrors
+// what the routing rule did before the move, and for the same reason:
+// route/rule.NewDomainRegexItem returns an error for an uncompilable pattern and
+// that aborts box.New for the WHOLE config, so a bad pattern must degrade to a
+// warning. A BARE `regexp:` compiles fine but matches every host — the same
+// silent match-all hazard as an empty keyword — so it is dropped too.
+//
+// SCOPE: this is the INLINE entry path only, and deliberately so. A `source=url`
+// text list is a hosts/plain-domain FILE, parsed by parseDomainList, which has no
+// marker vocabulary at all — see the block above parseDomainList for why the two
+// are not unified and how an entry written in the wrong one is reported.
+func (b *builder) peelDomainRegexes(diag string, entries []string) (rest, regexes []string) {
+	for _, e := range entries {
+		e = strings.TrimSpace(e)
+		if e == "" {
+			continue
+		}
+		if !strings.HasPrefix(strings.ToLower(e), "regexp:") {
+			rest = append(rest, e)
+			continue
+		}
+		re := strings.TrimSpace(e[len("regexp:"):])
+		if re == "" {
+			b.warnf("%s: %q is a bare matcher marker with no value, omitted (an empty regexp matches EVERY host)", diag, e)
+			continue
+		}
+		if _, err := regexp.Compile(re); err != nil {
+			b.warnf("%s: domain regexp %q is invalid (%v), omitted", diag, re, err)
+			continue
+		}
+		regexes = append(regexes, re)
+	}
+	return rest, regexes
 }

 // Ruleset.Type — the two shapes a rule-set can match, and the accepted spellings.
@@ -33,6 +33,14 @@ func findRouteRuleWithRuleSet(rt *option.RouteOptions, tag string) *option.Defau
 	return nil
 }

+// hasRulesetRule reports whether any emitted route rule references the rule-set
+// named name (tag rs-<name>). Since schema v2 a rule's destination is ALWAYS a
+// rule-set reference, so this is how a test says "that rule was emitted" — the
+// former "does any rule carry this dst domain" question has no answer any more.
+func hasRulesetRule(rt *option.RouteOptions, name string) bool {
+	return findRouteRuleWithRuleSet(rt, routeRulesetTagPrefix+name) != nil
+}
+
 func ruleSetByTag(rt *option.RouteOptions, tag string) (option.RuleSet, bool) {
 	if rt == nil {
 		return option.RuleSet{}, false
@@ -117,6 +125,50 @@ func TestRoutingRuleSetInlineDomain(t *testing.T) {
 	}
 }

+// TestRoutingRuleSetInlineDomainRegex: `regexp:` is a matcher the shared domain
+// classifier does NOT know — it belongs to the routing plane, and it used to be
+// peeled off inside ruleMatchers, which no longer sees any domains at all. It
+// therefore had to move into the inline rule-set with the rest of the destination
+// vocabulary (peelDomainRegexes), or every migrated `regexp:` entry would have
+// been reported as an unknown prefix and silently dropped: a routing rule that
+// looks configured and matches nothing.
+func TestRoutingRuleSetInlineDomainRegex(t *testing.T) {
+	m := &model.Model{
+		Globals: model.DefaultGlobals(),
+		Rulesets: []model.Ruleset{
+			{Name: "ads", Type: "domain", Source: "inline", Entries: []string{`regexp:^ads\.`}},
+		},
+		Rules: []model.Rule{
+			{Name: "block-ads", Enabled: true, Order: 10, DstRuleset: []string{"ads"}, Target: "block"},
+		},
+	}
+	opts, warns, err := GenerateWithWarnings(m)
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if len(warns) != 0 {
+		t.Fatalf("a valid regexp: entry must not warn: %v", warns)
+	}
+	rs, ok := ruleSetByTag(opts.Route, "rs-ads")
+	if !ok {
+		t.Fatalf("a regexp-only rule-set must still materialise; route=%+v", opts.Route)
+	}
+	hr := rs.InlineOptions.Rules[0].DefaultOptions
+	if len(hr.DomainRegex) != 1 || hr.DomainRegex[0] != `^ads\.` {
+		t.Fatalf("domain_regex = %+v, want [^ads\\.]", hr.DomainRegex)
+	}
+	if len(hr.Domain)+len(hr.DomainSuffix)+len(hr.DomainKeyword)+len(hr.IPCIDR) != 0 {
+		t.Fatalf("the regexp entry must not leak into another matcher: %+v", hr)
+	}
+	dr := findRouteRuleWithRuleSet(opts.Route, "rs-ads")
+	if dr == nil {
+		t.Fatalf("no route rule references rs-ads; rules=%+v", opts.Route.Rules)
+	}
+	if dr.RuleAction.RouteOptions.Outbound != tagBlock {
+		t.Fatalf("route rule must route to %q, got %+v", tagBlock, dr.RuleAction)
+	}
+}
+
 // TestRoutingRuleSetInlineIPCIDR: an ipcidr ruleset fills ip_cidr (not domain*),
 // and the route rule references it.
 func TestRoutingRuleSetInlineIPCIDR(t *testing.T) {
@@ -1501,7 +1553,7 @@ func TestURLBlocklistNotRefetchedWhileFresh(t *testing.T) {
 // wildcards, IPs and bare labels can never match a domain query, so importing them
 // would be a silent dud (the R9 lesson applied to fetched content).
 func TestParseDomainListRejectsJunk(t *testing.T) {
-	got := parseDomainList([]byte(strings.Join([]string{
+	got, _ := parseDomainList([]byte(strings.Join([]string{
 		"good.example.com",
 		"*.wildcard.example",          // wildcard syntax
 		"/regex/",                     // regex rule
@@ -1532,12 +1584,12 @@ func TestParseDomainListRejectsJunk(t *testing.T) {

 // TestParseDomainListHostsEdgeCases covers the messy real-world shapes.
 func TestParseDomainListHostsEdgeCases(t *testing.T) {
-	got := parseDomainList([]byte(stevenBlackSample))
+	got, _ := parseDomainList([]byte(stevenBlackSample))
 	if len(got) != 5 {
 		t.Fatalf("expected 5 domains from the sample, got %d (%v)", len(got), got)
 	}
 	// Unicode is punycoded, matching what actually arrives in a DNS query.
-	uni := parseDomainList([]byte("0.0.0.0 реклама.рф\n"))
+	uni, _ := parseDomainList([]byte("0.0.0.0 реклама.рф\n"))
 	if len(uni) != 1 || !strings.HasPrefix(uni[0], "xn--") {
 		t.Fatalf("a unicode entry must be punycoded, got %v", uni)
 	}
--- a/Show More
+++ b/Show More