docs: refresh the TrueNAS compatibility matrix

Run CI as a single job so concurrent jobs cannot race
Two CI failures shared one cause: four jobs starting together on the self-hosted runner. act caches each action as one shared clone under /root/.cache/act/<hash> and re-pulls it per job, so concurrent jobs fight over that directory and the loser dies with "lstat .../<file>: no such file or directory" before any test runs — a different victim each push. And the runner force-pulls its base image per job, so four jobs meant four anonymous Docker Hub pulls per push; that hit 429 Too Many Requests and every job started failing before it began, including the shell job nothing had touched. Installing uv without an action only shrank the surface, since every job still used actions/checkout. Concurrency is the ingredient, so this removes it: one job cannot race itself whatever actions it uses, and one job is one pull. The version sweep moves inside the job and still runs every version after one fails, preserving what fail-fast: false bought.
2026-09-08 06:25:24 +00:00 · 2026-08-26 05:34:11 +00:00 · 2026-08-26 05:28:26 +00:00 · 2026-08-26 05:20:23 +00:00 · 2026-08-26 04:57:33 +00:00 · 2026-08-26 04:37:50 +00:00
36 changed files with 5709 additions and 420 deletions
@@ -0,0 +1,2 @@
+github: sudolulo
+ko_fi: sudolulo
@@ -9,9 +9,31 @@ on:
 permissions:
  contents: read

+# ONE job, deliberately. This was four (shell + a 3-way python matrix) and they
+# started within the same second on the self-hosted Gitea runner, which is what
+# made CI unreliable in two separate ways:
+#
+#   1. The act action-cache race. `act` caches each ACTION as a single shared git
+#      clone under /root/.cache/act/<hash> and re-pulls it per job, so concurrent
+#      jobs using the same action fight over that directory and the loser dies
+#      with `lstat /root/.cache/act/<hash>/<file>: no such file or directory` --
+#      a red `main` with zero suite output, and a different victim each push
+#      (3.12 on one, 3.11 on the next). Dropping one action only shrank the
+#      surface: every job still used actions/checkout. Concurrency is the actual
+#      ingredient, so removing it removes the whole class -- a single job cannot
+#      race itself, no matter which actions it uses.
+#
+#   2. Docker Hub 429s. The runner force-pulls its base image per job, so four
+#      jobs meant four anonymous pulls per push. A few pushes and re-runs in an
+#      afternoon exhausted the anonymous limit and every job failed before it
+#      started -- including the shell job, which nothing had touched. One job is
+#      one pull.
+#
+# The cost is wall-clock parallelism, and this repo does not need it: the suite
+# is ~1.5s, so container start and interpreter downloads dominate either way.
 jobs:
-  shell:
-    name: shell (shellcheck + syntax)
+  ci:
+    name: ci (shell + python 3.11-3.13)
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
@@ -31,32 +53,57 @@ jobs:
        env:
          SHELLCHECK_OPTS: -S warning -e SC1091

-  python:
-    name: python ${{ matrix.python }}
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        # TrueNAS SCALE middleware runs 3.11+; keep the patch importable across
-        # the versions it may be injected into.
-        python: ["3.11", "3.12", "3.13"]
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: install dev deps
-        run: python -m pip install --upgrade pip pytest ruff
+      # uv-managed interpreters instead of actions/setup-python: the prebuilt-CPython
+      # download path setup-python relies on does not work on the self-hosted Gitea
+      # runner (all three matrix jobs failed at setup there while passing on GitHub);
+      # uv works identically on both.
+      #
+      # Installed by a plain `run:` step rather than astral-sh/setup-uv: one fewer
+      # action is one fewer thing to go wrong, and the action was only ever
+      # fetching a binary -- the interpreter is chosen per command by `uvx
+      # --python`, never by the action.
+      #
+      # Pinned for the same reason ruff is pinned below: an unpinned uv means any
+      # upstream release can turn main red with no code change here.
+      - name: install uv
+        env:
+          UV_VERSION: "0.11.21"
+        run: |
+          curl -LsSf "https://astral.sh/uv/${UV_VERSION}/install.sh" | sh
+          echo "$HOME/.local/bin" >> "$GITHUB_PATH"

      - name: ruff
-        run: ruff check patch tests tools
+        # Pinned: an unpinned ruff means any upstream release can turn main red
+        # with no code change.
+        run: uvx ruff@0.16.1 check patch tests tools

+      # TrueNAS SCALE middleware runs 3.11+; keep the patch importable across the
+      # versions it may be injected into. Every version runs even after one
+      # fails -- that is what `fail-fast: false` bought when this was a matrix,
+      # and losing it would mean a 3.11 break hides whether 3.12 and 3.13 are
+      # fine, which is exactly the information you want at that moment.
      - name: pytest
-        run: pytest tests -v
+        env:
+          PYTHONS: "3.11 3.12 3.13"
+        run: |
+          fail=0
+          for v in $PYTHONS; do
+            echo "::group::pytest on python $v"
+            uvx --python "$v" pytest tests -v \
+              || { echo "::error::suite failed on python $v"; fail=1; }
+            echo "::endgroup::"
+          done
+          exit $fail

      - name: verify injected middleware blocks compile
        # Belt-and-braces: the *_BLOCK strings are appended into live middlewared
        # modules. A syntax error there would break the box at boot.
-        run: pytest tests/test_apply_blocks.py -v
+        env:
+          PYTHONS: "3.11 3.12 3.13"
+        run: |
+          fail=0
+          for v in $PYTHONS; do
+            uvx --python "$v" pytest tests/test_apply_blocks.py -v \
+              || { echo "::error::injected blocks failed to compile on python $v"; fail=1; }
+          done
+          exit $fail
@@ -24,10 +24,18 @@ on:
    paths:
      # The manifest itself changed -- re-check immediately rather than waiting a day.
      - "tools/compat.py"
+      # ...and so did the thing that PUBLISHES the finding. This was missing, and it
+      # showed: the commit teaching the bot to refresh a stale report body touched only
+      # compat_publish.py, so no run fired, and the report stayed stale until the next
+      # scheduled one. A fix nobody runs is a fix nobody has.
+      - "tools/compat_publish.py"
      - ".github/workflows/compat.yml"

 permissions:
-  contents: read
+  # write, because the matrix refresh pushes a branch and opens a PR. It does NOT get
+  # to move `main` -- that is the whole reason it is a PR. See the refresh step below.
+  contents: write
+  pull-requests: write
  issues: write

 jobs:
@@ -130,17 +138,27 @@ jobs:
      - name: matrix
        run: cat /tmp/matrix.md

-      # Keep the README's table true. A support matrix that quietly goes stale is not
-      # a stale doc -- it is a false promise to somebody deciding whether to trust
-      # this with their backups.
+      # Keep the README's table true — as a PULL REQUEST, on the CANONICAL forge.
      #
-      # Only ever touches the block between the COMPAT MATRIX markers, and only on
-      # the canonical host (Gitea) so the two forges cannot race each other. The
-      # `paths:` trigger above does not include README.md, so this cannot re-trigger
-      # itself; and a README change is documentation-only, which by design raises no
-      # update alert on anyone's box.
-      - name: refresh the README matrix
+      # Two things this gets right that the obvious version gets wrong:
+      #
+      #  1. It is a PR, not a push to main. This used to `git push origin HEAD:main`
+      #     from CI. An unattended write to main is exactly what the release barrier
+      #     exists to prevent — a bot that can move main can move it somewhere nobody
+      #     looked. Nothing lands by itself.
+      #
+      #  2. It runs on GITEA, not GitHub. GitHub is a one-way MIRROR: a PR merged there
+      #     would be silently clobbered by the next `fleet-repos mirror` push from Gitea.
+      #     A bot opening PRs against a mirror is a bot doing nothing, slowly.
+      #
+      # A stale support matrix is not a stale doc — it is a false promise to somebody
+      # deciding whether to trust this with their backups. So it is refreshed daily; it
+      # just asks first.
+      - name: refresh the README matrix (PR on the canonical forge)
        if: ${{ github.event_name == 'schedule' && !contains(github.server_url, 'github.com') }}
+        env:
+          TOKEN: ${{ secrets.GITEA_TOKEN || github.token }}
+          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
        run: |
          python3 - <<'PY'
          import json, sys
@@ -151,66 +169,73 @@ jobs:
          print("changed" if compat.update_readme(rows) else "unchanged")
          PY

-          if ! git diff --quiet -- README.md; then
-            git config user.name  "truecloud-patch bot"
-            git config user.email "bot@onetick.ninja"
-            git add README.md
-            git commit -m "docs: refresh the TrueNAS compatibility matrix"
-            git push origin HEAD:main
-          fi
+          git diff --quiet -- README.md && { echo "matrix unchanged — nothing to propose"; exit 0; }

-      # A broken SHIPPED release is an outage: users are on it right now.
+          git config user.name  "truecloud-patch bot"
+          git config user.email "bot@onetick.ninja"
+
+          BRANCH=bot/compat-matrix
+          git checkout -B "$BRANCH"
+          git add README.md
+          git commit -m "docs: refresh the TrueNAS compatibility matrix"
+          git push -f origin "$BRANCH"
+
+          # ONE long-lived PR, force-pushed in place — not a new one every morning.
+          # (A daily PR is the same mistake as a daily comment, wearing a hat.)
+          BRANCH="$BRANCH" python3 - <<'PY'
+          import json, os, urllib.error, urllib.request
+
+          api, token, branch = os.environ["API"], os.environ["TOKEN"], os.environ["BRANCH"]
+          h = {"Authorization": f"token {token}", "Content-Type": "application/json"}
+
+          def call(url, method="GET", data=None):
+              r = urllib.request.Request(
+                  url, method=method, headers=h,
+                  data=json.dumps(data).encode() if data else None)
+              with urllib.request.urlopen(r) as resp:  # noqa: S310
+                  return json.load(resp) if resp.length != 0 else {}
+
+          existing = [
+              p for p in call(f"{api}/pulls?state=open")
+              if p["head"]["ref"] == branch
+          ]
+          if existing:
+              print(f"PR #{existing[0]['number']} already open; the force-push updated it")
+          else:
+              pr = call(f"{api}/pulls", "POST", {
+                  "head": branch, "base": "main",
+                  "title": "docs: refresh the TrueNAS compatibility matrix",
+                  "body": (
+                      "The daily compatibility check found that the support matrix in "
+                      "the README no longer matches iXsystems' actual middleware.\n\n"
+                      "This only touches the block between the `COMPAT MATRIX` markers. "
+                      "It is regenerated by `tools/compat.py --matrix --update-readme` "
+                      "and force-pushed, so it always reflects the latest run."
+                  ),
+              })
+              print(f"opened PR #{pr['number']}")
+          PY
+
+      # ONE bug report, kept in sync. It is edited in place when the findings change and
+      # says NOTHING when they do not.
+      #
+      # The first version commented on every run and left 11 identical 3,000-character
+      # comments on one issue in a single day. A bot that repeats itself daily gets
+      # muted, and then the next real finding is scrolled past — which defeats the whole
+      # reason for building it.
+      #
+      # Runs on whichever forge it lands on; compat_publish.py handles both, so the two
+      # cannot drift.
+      - name: file / update / close the bug report
+        env:
+          TOKEN: ${{ secrets.GITEA_TOKEN || github.token }}
+          API: ${{ contains(github.server_url, 'github.com') && 'https://api.github.com' || format('{0}/api/v1', github.server_url) }}/repos/${{ github.repository }}
+        run: python3 tools/compat_publish.py --api "$API" --token "$TOKEN" --matrix /tmp/matrix.json
+
+      # A broken SHIPPED release is an outage: users are on it right now. Fails LAST, so
+      # the report is filed before the job goes red.
      - name: fail if a shipped release is broken
        if: ${{ steps.check.outputs.shipped_broken != '0' }}
        run: |
          echo "::error::The patch is broken on a SHIPPED TrueNAS release."
          exit 1
-
-      - name: file a bug report (GitHub)
-        if: ${{ steps.report.outputs.broken == '1' && contains(github.server_url, 'github.com') }}
-        env:
-          GH_TOKEN: ${{ github.token }}
-          TITLE: "Incompatible with upcoming TrueNAS: ${{ steps.report.outputs.refs }}"
-        run: |
-          # One issue per set of broken refs, reopened/updated rather than duplicated
-          # daily -- a bot that files the same issue every morning gets muted, and
-          # then it is not a warning system any more.
-          existing="$(gh issue list --state all --search "$TITLE" \
-                        --json number,title \
-                        --jq '.[] | select(.title == env.TITLE) | .number' | head -1)"
-
-          if [ -n "$existing" ]; then
-            gh issue comment "$existing" --body-file /tmp/issue.md
-            gh issue reopen "$existing" 2>/dev/null || true
-          else
-            gh issue create --title "$TITLE" --body-file /tmp/issue.md
-          fi
-
-      - name: file a bug report (Gitea)
-        if: ${{ steps.report.outputs.broken == '1' && !contains(github.server_url, 'github.com') }}
-        env:
-          TOKEN: ${{ secrets.GITEA_TOKEN || github.token }}
-          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
-          TITLE: "Incompatible with upcoming TrueNAS: ${{ steps.report.outputs.refs }}"
-        run: |
-          body="$(jq -Rs . < /tmp/issue.md)"
-          title="$(printf '%s' "$TITLE" | jq -Rs .)"
-
-          # Same title => same issue. Comment on it instead of filing a new one.
-          number="$(curl -sf -H "Authorization: token $TOKEN" \
-            "$API/issues?state=all&type=issues" \
-            | jq -r --arg t "$TITLE" '.[] | select(.title == $t) | .number' | head -1)"
-
-          if [ -n "$number" ]; then
-            curl -sS -X POST "$API/issues/$number/comments" \
-              -H "Authorization: token $TOKEN" -H "Content-Type: application/json" \
-              -d "$(printf '{"body":%s}' "$body")" -o /dev/null -w 'comment -> %{http_code}\n'
-            curl -sS -X PATCH "$API/issues/$number" \
-              -H "Authorization: token $TOKEN" -H "Content-Type: application/json" \
-              -d '{"state":"open"}' -o /dev/null -w 'reopen -> %{http_code}\n'
-          else
-            curl -sS -X POST "$API/issues" \
-              -H "Authorization: token $TOKEN" -H "Content-Type: application/json" \
-              -d "$(printf '{"title":%s,"body":%s}' "$title" "$body")" \
-              -o /dev/null -w 'create -> %{http_code}\n'
-          fi
@@ -124,7 +124,7 @@ jobs:
          echo "--- release body ---"
          cat /tmp/notes.md

-      # This repo is canonically hosted on Gitea (git.onetick.ninja) and mirrored to
+      # This repo is canonically hosted on Gitea (git.arch.fyi) and mirrored to
      # GitHub, and BOTH run this workflow -- Gitea reads .github/workflows too. So
      # the publish step has to work on whichever forge it lands on. Everything
      # above is forge-agnostic; only the "create a release" API differs.
@@ -164,27 +164,45 @@ jobs:
            *-rc*|*-beta*|*-alpha*) prerelease=true ;;
          esac

-          # jq -Rs so the notes are JSON-encoded properly: the changelog is full of
-          # quotes, backticks and newlines, and hand-built JSON would mangle them.
-          body="$(jq -Rs . < /tmp/notes.md)"
-          payload="$(printf '{"tag_name":%s,"name":%s,"body":%s,"prerelease":%s}' \
-            "$(printf '%s' "$TAG" | jq -Rs .)" \
-            "$(printf '%s' "$TAG" | jq -Rs .)" \
-            "$body" "$prerelease")"
+          # python3, not jq. The changelog is full of quotes, backticks and newlines,
+          # so the body must be properly JSON-encoded -- but `jq` is not guaranteed on
+          # a self-hosted Gitea runner, and a publish step that dies on a missing tool
+          # leaves a tag with no release behind it. python3 is guaranteed: setup-python
+          # ran above.
+          python3 - "$TAG" "$API" "$TOKEN" "$prerelease" <<'PY'
+          import json, sys, urllib.error, urllib.request

-          existing="$(curl -sf -H "Authorization: token $TOKEN" \
-            "$API/releases/tags/$TAG" 2>/dev/null || true)"
+          tag, api, token, prerelease = sys.argv[1:5]
+          with open("/tmp/notes.md", encoding="utf-8") as fh:
+              body = fh.read()

-          if [ -n "$existing" ]; then
-            id="$(printf '%s' "$existing" | jq -r .id)"
-            echo "Release $TAG exists (id=$id) — updating notes."
-            curl -sS -X PATCH "$API/releases/$id" \
-              -H "Authorization: token $TOKEN" \
-              -H "Content-Type: application/json" \
-              -d "$payload" -o /dev/null -w 'PATCH -> %{http_code}\n'
-          else
-            curl -sS -X POST "$API/releases" \
-              -H "Authorization: token $TOKEN" \
-              -H "Content-Type: application/json" \
-              -d "$payload" -o /dev/null -w 'POST -> %{http_code}\n'
-          fi
+          payload = {
+              "tag_name": tag, "name": tag, "body": body,
+              "prerelease": prerelease == "true",
+          }
+          headers = {
+              "Authorization": f"token {token}",
+              "Content-Type": "application/json",
+          }
+
+          def call(url, method, data=None):
+              req = urllib.request.Request(
+                  url, method=method, headers=headers,
+                  data=json.dumps(data).encode() if data else None)
+              with urllib.request.urlopen(req) as r:  # noqa: S310
+                  return r.status, json.load(r) if r.length != 0 else {}
+
+          try:
+              _, existing = call(f"{api}/releases/tags/{tag}", "GET")
+          except urllib.error.HTTPError as e:
+              if e.code != 404:
+                  raise
+              existing = None
+
+          if existing:
+              status, _ = call(f"{api}/releases/{existing['id']}", "PATCH", payload)
+              print(f"updated release {tag} -> {status}")
+          else:
+              status, _ = call(f"{api}/releases", "POST", payload)
+              print(f"created release {tag} (prerelease={payload['prerelease']}) -> {status}")
+          PY
@@ -3,6 +3,10 @@
 /apply.log.1
 /apply.log.2
 /hook_status.json
+# The resolved middlewared directory, recorded by apply.sh so wait_restart.sh can
+# check whether the patched modules are still on the live path without
+# re-deriving site-packages.
+/.mw_dir
 /disabled
 /nested_snapshots_enabled
 # Written by apply.sh when a module's assumptions no longer fit the installed
@@ -6,6 +6,349 @@ is deliberate: see [Releasing](docs/releasing.md). Twelve releases were cut on
 live, every one of those interrupts every user. An alert people learn to ignore is
 worse than no alert, because one day it carries a security fix.

+## v0.8.0 — 2026-08-26
+### Changed
+
+- **CI's python matrix is green on the self-hosted Gitea runner again.** The real
+  failure was that the Gitea runner image executes jobs as root, and the two
+  unreadable-sidecar tests build their scenario with `chmod(0)` — which cannot make
+  a file unreadable for root (`CAP_DAC_OVERRIDE`). Those two tests now skip as root
+  with that reason; GitHub's non-root runner still exercises them. The matrix also
+  moved to uv-managed interpreters (one toolchain across both runners) and ruff is
+  pinned to 0.16.1 so an upstream ruff release can't turn `main` red without a code
+  change.
+- **README badges point at the public GitHub mirror** (workflow status and
+  releases) instead of the private forge. The release badge had also been reading
+  the stale Gitea v0.6.1 release instead of the current v0.7.0 on GitHub.
+
+- **`master` is now labelled `27-dev`, because it is not the next release.** iX
+  branches each major onto its own `release/` line and master rolls straight on to the
+  one after — on 2026-07-14 every recent commit on master targeted `27.0.0-BETA.1`
+  while 26 was still in beta. So a **BROKEN** master row, rendered as
+  "master _(unreleased)_", read as *"the version you are about to install is broken"*
+  when the breakage was a major release away on a line nobody can download. In a table
+  whose entire job is helping somebody decide whether to trust this with their backups,
+  that is a false alarm in the worst possible place. The label is derived from the
+  newest major in the matrix plus one, so it rolls over to `28-dev` by itself once 27
+  branches.
+
+  For the record, the breakage is `NAS-141498` (2026-06-24), "Convert cloud_backup
+  plugin to the typesafe pattern": it re-signatures `restic_backup` and
+  `get_restic_config`, splitting `entry`/`credentials` out of the `cloud_backup` dict.
+  It is deliberately not being chased while the 27 line is still churning.
+
+### Fixed
+
+- **CI turned `main` red on two of three pushes without running a single test,
+  then stopped running at all.** Two failures, one cause: four concurrent jobs on
+  one self-hosted runner.
+
+  `act`, the engine behind the Gitea runner, caches each *action* as one shared
+  git clone under `/root/.cache/act/<hash>` and re-pulls it per job. Jobs
+  starting in the same second fight over that directory, and the loser dies with
+  `lstat /root/.cache/act/<hash>/.npmrc: no such file or directory` — before any
+  suite output exists, with a different victim each push (3.12 on one, 3.11 on
+  the next). Separately, the runner force-pulls its base image per job, so four
+  jobs meant four anonymous Docker Hub pulls per push; a few pushes and re-runs
+  in one afternoon hit `429 Too Many Requests` and *every* job began failing
+  before it started — including the shell job, which nothing had touched.
+
+  CI is now a single job. Dropping one action (uv is installed by a `run:` step
+  rather than `astral-sh/setup-uv`, which was only ever fetching a binary) only
+  shrank the surface, because every job still used `actions/checkout`.
+  Concurrency is the actual ingredient, so removing it removes the whole class:
+  one job cannot race itself whatever actions it uses, and one job is one pull.
+  The Python sweep moved inside that job and still runs every version after one
+  fails — that is what `fail-fast: false` bought, and losing it would mean a 3.11
+  break hides whether 3.12 and 3.13 are fine. The cost is wall-clock
+  parallelism, which this repo does not need: the suite is ~1.5s, so container
+  start and interpreter downloads dominate either way.
+
+  A red gate that is usually noise is worse than no gate, because the one time
+  it means something, nobody looks.
+
+- **The patch survived being applied and then silently stopped existing, because
+  something else remounted `/usr` four seconds later.** On a box running
+  TrueNAS 25.10.6 the boot of 2026-08-19 went: 16:41:56 `apply.sh` mounts its
+  overlay on `/usr/lib/python3/dist-packages`, patches `b2.py`/`restic.py`, logs
+  every step `OK`; **16:42:00** a second PREINIT hook runs `systemd-sysext
+  refresh` over `/usr` — `Unmerged '/usr'.` / `Merged extensions into '/usr'.` —
+  and our overlay, which lives *inside* that hierarchy, is torn off with it;
+  16:47:24 our own deferred restart fires exactly as designed and middlewared
+  imports the **stock** modules. Every B2 TrueCloud Backup task then failed with
+  `NotImplementedError` from stock `rclone/base.py` for nineteen hours, across
+  four scheduled runs, while `apply.log` and `hook_status.json` both said the
+  patch was active.
+
+  Nothing in the patch was wrong, which is the point: applying at PREINIT and
+  restarting later is only sound if the patched files are still on the live path
+  when middlewared re-imports them, and **that is not something PREINIT can
+  guarantee**. Init scripts run sequentially in id order, so any hook registered
+  after ours always wins. Worse, hook ordering cannot fix it either —
+  middlewared's own `docker.configure_nvidia` merges a sysext over `/usr` at
+  *runtime*, long after every PREINIT hook is finished.
+
+  So the deferred restart no longer trusts the PREINIT pass. `wait_restart.sh`
+  now re-applies immediately before it restarts middlewared — after boot has
+  settled, which is also after every sysext merge and docker nvidia
+  configuration — and verifies the marker is genuinely on the live path before
+  restarting. It is no longer `exec systemctl try-restart middlewared`, because
+  something has to run afterwards.
+
+  What runs afterwards deliberately does **not** restart again. `try-restart`
+  returns as soon as middlewared is READY, and middlewared then brings docker up
+  — `docker.configure_nvidia` merges the stock nvidia sysext over `/usr` at that
+  point, detaching the overlay *after* the patched modules have already been
+  imported. A disk check there reports "missing" on a perfectly healthy system,
+  and restarting on that signal would restart a correctly-patched middlewared
+  straight back into the same race. So the overlay is re-mounted for the benefit
+  of the next restart, and the question of whether *this* middlewared actually
+  holds the patch is left to the one thing that can answer it exactly — the
+  in-process alert below. That re-mount preserves `hook_status.json`'s
+  `patched_at`: `create_task.py verify` decides "loaded" by comparing
+  middlewared's start time against that stamp, so a re-apply running *after* the
+  restart would have made the stamp newer than the process which correctly
+  imported the patch, and `verify` would have reported FAIL forever on every
+  boot where the sysext merge detaches the overlay. Caught on hardware while
+  validating the candidate — a new lying status introduced by the fix for a
+  lying status.
+
+  Two supporting fixes fell out of the same failure. `_ensure_writable` treated
+  "one of our overlays is listed on this directory" as "already done" — but it
+  only ever reaches that check when the directory is **not** writable, and a live
+  overlay of ours always is. A shadowed overlay was therefore indistinguishable
+  from a healthy one; it is now detached and re-mounted, reusing the same
+  upperdir so everything patched earlier in the boot reappears intact, with a
+  fresh workdir because overlayfs refuses one left behind by a detached mount.
+
+- **middlewared now says so when it is running stock.** The gap that let this
+  cost nineteen hours was not the remount, it was that nothing could tell the
+  difference between "patched on disk" and "patched in the running process".
+  `apply.log` can only ever report the first. A new CRITICAL alert asks the
+  second question from inside middlewared, hourly, where it is exact: the patch
+  stamps the objects it replaces, so a missing stamp means this interpreter
+  imported stock code. It checks both halves — `restic.py`'s `_truecloud_patched`
+  marker and whether `B2RcloneRemote.get_restic_config` is still the base class's
+  — since either can go missing alone. It stays quiet when the kill switch is
+  set or the providers module has been retired as native, and it is deliberately
+  **not** silenced by `update_alerts_disabled`: that mutes release notifications,
+  not a broken backup path.
+
+  Boot-time diagnosis also no longer depends on the journal. `wait_restart.sh`
+  logged only to the journal, and journald retention on a busy box is easily
+  shorter than the interval between reboots — the 2026-08-19 boot had already
+  rotated away by the time it was investigated. It now writes to `apply.log`
+  alongside everything else.
+
+- **The next maintenance release was never checked, and it is the one that reaches
+  users.** Shipped versions were discovered from `TS-*` tags and unreleased ones from
+  `release/*` branches carrying `-BETA`/`-RC`. A branched-but-untagged *maintenance*
+  release is neither: `release/25.10.5` has no tag, and its line has already shipped,
+  so the "a prerelease of a shipped line is history" filter discarded it. It was
+  invisible — and it is precisely what a 25.10.4 box gets on its next update. A break
+  there would have reached real users before the daily check ever looked at it, on the
+  only line anybody is actually running.
+
+  A plain `release/X.Y.Z` branch is now checked when its line **has** shipped and it
+  sorts **newer** than that line's newest tag. Both things that must stay out fall out
+  of the same rule: `release/24.10-RC.2` sorts older than `TS-24.10.2.4` (history, not
+  a warning), and iX's typo branch `release/25.20.2.2` is on a line that has no tag at
+  all, so it is not a release line. This immediately surfaced two refs that had never
+  been checked — `release/25.10.5` and `release/24.10.2.5` — both of which pass.
+
+  `is_unreleased()` now keys off where a ref came from (branch = not yet shipped)
+  rather than looking for `-BETA`/`-RC` in its name. Otherwise `release/25.10.5` would
+  count as shipped and a break in it would fail the build as a live outage — on a
+  version nobody is running yet.
+
+- **An unchanged fingerprint froze the bug report's body, not just its comments.** Two
+  questions were sharing one answer. *Have the findings changed?* gates **comments** —
+  they notify, and a daily "still broken, same as yesterday" is what teaches everyone
+  to ignore the one that finally matters. *Is the body still true?* gates the **body** —
+  and editing an issue body notifies nobody on either forge, so keeping it honest is
+  free. Conflated, the report could never be corrected while the findings held steady,
+  and the fingerprint deliberately ignores everything that moves on its own — healthy
+  rows, the hardware-verified column, point releases, and how a row is labelled. The
+  `master` → `27-dev` relabel above would have reached the README and never the issue
+  anybody actually opens. The body is now rewritten whenever it is out of date (after
+  normalising line endings, so a forge round-tripping `\r\n` does not cause a rewrite
+  every run) and comments remain strictly a changelog of real changes.
+
+- **A change to the publisher did not re-run the check.** `compat.yml`'s `push:` paths
+  listed `tools/compat.py` but not `tools/compat_publish.py` — so the very commit that
+  taught the bot to refresh a stale report body triggered no run, and the report stayed
+  stale until the next scheduled one. A fix nobody runs is a fix nobody has.
+
+- **The compatibility bot filed a new duplicate bug report on every Gitea run.**
+  `find_issue()` skipped pull requests by testing for the *presence* of the
+  `pull_request` key. GitHub omits that key on a plain issue; Gitea sends it as
+  `null`. So on Gitea every issue was discarded as a PR, the lookup always came back
+  empty, and the bot took the "nothing filed yet" branch and opened a fresh report
+  each run — **nine copies on the canonical forge**, four of them filed *after* the
+  commit that was meant to stop precisely this. The mirror was fine, which is why it
+  went unnoticed: GitHub's payload shape is the one the filter was written against.
+
+  It is the same failure the anti-spam fix was written to prevent, moved from
+  comments to issues, and it survived because `find_issue` was the only function in
+  `compat_publish.py` with no test. It now has one, per forge, and the daily cron —
+  which had not yet run once — no longer accumulates a report a day.
+
+  The issue list is also requested with **both** paging parameters (`per_page` for
+  GitHub, `limit` for Gitea). Each forge ignores the other's, and Gitea's default page
+  is 30, so the lookup would have started missing the report again once the pile it
+  was creating grew past one page.
+
+## v0.7.0 — 2026-07-14
+### Added
+
+- **TrueNAS 26 support, verified on a real TrueNAS 26 install.** 26 deletes
+  `plugins/zfs_/` outright, taking the private `zfs.dataset.query`,
+  `zfs.snapshot.query` and `zfs.snapshot.delete` with it. Every one of those was on
+  the nested module's critical path, so nested snapshots were **BROKEN** on 26 and
+  `apply.sh` correctly refused to apply the module there.
+
+  Snapshot **deletion** now resolves its namespace at runtime — `pool.snapshot` on
+  25.10 and 26, `zfs.snapshot` on 24.10 and 25.04, because no single namespace spans
+  every supported release. `tools/compat.py` checks the same list the runtime uses,
+  so what CI verifies and what runs cannot drift apart.
+
+  Hardware-verified on TrueNAS 26.0.0-BETA.1: a 274-snapshot recursive backup of a
+  292-dataset pool, then a **byte-identical restore of a four-level-deep child
+  dataset**.
+
+### Fixed
+
+- **Enumeration no longer trusts middleware's dataset and snapshot queries — they
+  are filtered.** This is the important one, and it is the bug that a test VM caught
+  and no amount of source analysis ever could have.
+
+  The obvious port of the deleted private `zfs.dataset.query` was the public
+  `pool.dataset.query`. It exists, it is documented, it is covered by iX's
+  deprecation policy — and it is **not a like-for-like replacement**. It applies a
+  *visibility policy*: it hides the datasets TrueNAS considers its own — `ix-apps/*`,
+  `.system/*`, `.ix-virt/*`. On a real pool that is **84 of 270 datasets**, and
+  `ix-apps` holds **live application data**.
+
+  Staging from that view would have silently omitted every one of them. Worse,
+  `plan_staging()` would never have seen them, so they would not have appeared in its
+  `skipped` list either — no warning, no failure, just a green backup quietly missing
+  data. That is precisely the failure this module exists to prevent. The snapshot
+  query lies the same way (205 of 274), so the sweep would have orphaned one snapshot
+  per hidden dataset, on every run, forever.
+
+  The module now **reads the truth from ZFS and makes changes through middleware**:
+  enumeration is `zfs list`, which no policy can filter and which behaves identically
+  on every release; mutation stays a middleware call, so TrueNAS's own bookkeeping
+  stays consistent. A failing `zfs list` raises rather than returning an empty list —
+  "no datasets" and "the command broke" must never look the same.
+
+  **No shipped release is affected.** v0.6.1 and earlier call the *private*
+  `zfs.dataset.query`, which returns all 270 datasets. The bug existed only in the
+  unreleased TrueNAS 26 port.
+
+- **The patch now owns the snapshot sweep even when it does not stage anything.**
+  Stock decides whether to take a *recursive* snapshot by its own rule, and on
+  TrueNAS 26 that rule stopped being ours.
+
+  Up to 25.10, stock's `create_snapshot` called `get_dataset_recursive()` — the same
+  function this module vendors — so "stock went recursive" and "we have something to
+  stage" were the *same question*, and stock's non-recursive delete was correct for
+  everything the patch declined to stage. TrueNAS 26 uses `filesystem.statfs`:
+  `recursive = (path == the dataset's mountpoint)`. The two rules now disagree for a
+  dataset whose only descendants are **ZVOLs** or **legacy/none-mountpoint** datasets
+  — stock snapshots it recursively, while the patch sees nothing to stage.
+
+  The patch then handed the snapshot back to stock, which destroys the parent only.
+  With no staging tree there was no sidecar, and the garbage collector only ever ran
+  from the staging path — so nothing on the box would ever have found the children.
+  Reproduced on the test VM: one orphaned snapshot per zvol, on every run, forever,
+  with the backup reporting success. Ownership of the sweep is no longer conditional
+  on staging.
+
+- **The runtime resolved a *namespace*; the checker verified a *method*.** Those are
+  different questions, and the gap is a false "ok". `get_service()` only proves a
+  namespace is registered — it says nothing about whether `delete` still exists on it.
+  So if iX guts the method while keeping the service (they have already done exactly
+  that to `pool.snapshot.do_update` on master), `tools/compat.py` would fall through
+  to `zfs.snapshot`, report the box healthy, and let the patch apply — while the
+  runtime picked `pool.snapshot` and failed *every* delete, orphaning the whole tree.
+  Both sides now ask the same question, and a test binds the two lists together.
+
+- `query_filesystems()` **dropped malformed `zfs list` rows silently** — the last
+  remaining silent-omission path, and a direct contradiction of this module's cardinal
+  rule. It raises now. A missing `zfs` binary raised `FileNotFoundError` rather than
+  `ZfsError`; also fixed.
+
+- The snapshot retry loop **discarded the delete error** and reported every survivor
+  as "(still busy?)" — naming the one cause that is benign and self-healing, and
+  hiding the ones that are permanent. It keeps and reports the real error.
+
+- The staging-failure handler could **lose the original exception** if its own cleanup
+  sweep raised. An error handler must not be able to lose the error.
+
+- **A snapshot delete that returns cleanly is not proof that anything was deleted.**
+  The recursive sweep's fast path took the call's word for it and returned "no
+  survivors" — so `cleanup_task` read that as a clean sweep and removed the sidecar,
+  the only record the tree ever existed. Roughly 250 snapshots would have been orphaned
+  on every run, with nothing left able to find them, and the backup reporting success.
+
+  This is not a hypothetical about a well-behaved API: iX has already gutted
+  `pool.snapshot.do_update` on master into a no-op whose body is commented out and
+  which returns `None`. A source check still sees the `def`; a runtime check still sees
+  a callable method. Only asking ZFS can tell. The sweep now confirms against ZFS, and
+  where it *cannot* confirm it keeps owning the tree rather than claiming success — a
+  false survivor self-heals on the next run, a lost record never does.
+
+- `_write_sidecar` **swallowed `OSError`**. The sidecar is the only thing that survives
+  a middlewared restart; failing to write it is not fatal, but it must never be
+  invisible. `_read_sidecar` had the mirror bug — it conflated "there is no sidecar"
+  with "I could not read the sidecar", and `cleanup_task` then took the empty branch
+  and **unlinked the only record** of a tree it had failed to read.
+
+- **A dataset from another tree, mounted inside the backup path, was omitted
+  silently.** The staging plan scopes by dataset *name*, which is correct — a dataset
+  with no mountpoint cannot be scoped by path at all. But ZFS lets any dataset mount
+  anywhere, so one from an unrelated tree can sit inside the path:
+
+      Tank/photos   mountpoint=/mnt/Tap/apps/photos
+
+  It holds data inside the backed-up path, and `zfs snapshot -r Tap@…` does **not**
+  cover it: recursion follows the dataset tree, not the directory tree. So there is no
+  snapshot of it to stage, and no way to capture it consistently with the rest. It fell
+  out of the name filter and vanished — not staged, not in `skipped`, no error, backup
+  green. Stock has the same blind spot, but stock also refuses the nested config
+  outright; this patch is what relaxes that guard, so the hole is this patch's to close.
+  It now refuses, and names the offending datasets.
+
+## v0.6.1 — 2026-07-13
+### Fixed
+
+- **A reboot mid-backup orphaned the entire snapshot tree, permanently.** The sidecar
+  is the record of which snapshots a run pinned — and it lives in `/run`, which is
+  **tmpfs**. A reboot (or a crash) between taking the recursive snapshot and cleaning
+  it up destroyed that record, leaving one snapshot per descendant dataset — **250+ on
+  a real pool** — with nothing left pointing at them. Nothing would ever have found
+  them again.
+
+  `gc_stale_snapshots()` is the backstop: it identifies leftovers **by name**, so it
+  works when the record is gone. It runs at the start of every backup, after the
+  sidecar reclaim — the recorded path stays authoritative, and the collector only ever
+  mops up what the record lost.
+
+  Because it deletes data on a *name match* — a weaker claim than a recorded fact — the
+  selection is a **pure function** with the harshest tests in the suite. A snapshot is
+  collected only if **all** of these hold:
+
+  | | |
+  | --- | --- |
+  | name is exactly `<dataset>@<task>-<YYYYMMDDHHMMSS>` | so `cloud_backup-5` never matches `cloud_backup-50`, an `auto-*` periodic snapshot, or anything a human made |
+  | it is not the current run's | parent *and* children are excluded |
+  | **nothing is mounted from it** | an in-flight run pins its own snapshots — this, not the age guard, is what protects a concurrent backup |
+  | it is **over an hour old** | covers the seconds-long window where a live run has snapshotted but not yet mounted |
+
+  Verified against the real pool: of **4,728** snapshots — including **2,341** periodic
+  ones — it selects exactly the orphans of the task being run, and nothing else.
+
 ## v0.6.0 — 2026-07-13
 ### Added

@@ -66,6 +409,50 @@ worse than no alert, because one day it carries a security fix.

 ### Fixed

+- **A few snapshots leaked on every nested run, forever.** Found on real hardware, in
+  the one place it could be: a 256-snapshot backup of `/mnt/Tap` swept 253 cleanly and
+  left **3 behind** with `dataset is busy`.
+
+  The cause is ZFS's own automount. Reading anything under
+  `<dataset>/.zfs/snapshot/<snap>/` makes ZFS **automount that snapshot**, and it stays
+  mounted for `zfs_expire_snapshot` seconds (**300** by default) after the last access.
+  `teardown()` unmounts *our* bind mounts — but not the automount underneath — so
+  `zfs destroy` refuses for exactly the datasets restic read most recently. Then
+  `cleanup_task()` removed the sidecar anyway, destroying the only record that those
+  snapshots existed. Nothing would ever have reclaimed them.
+
+  Three changes, and the third is the one that makes it safe rather than merely
+  unlikely:
+  - `release_snapdirs()` unmounts ZFS's own `.zfs/snapshot` automounts (deepest first)
+    before deleting, so the snapshots are not busy in the first place.
+  - `delete_snapshot_tree()` **retries** the transient busy, and **returns the
+    snapshots it could not delete** instead of swallowing them.
+  - **The sidecar is now removed only on a confirmed-clean sweep** — including on the
+    staging-failure path, which used to remove it *before* the caller swept. The
+    asymmetry is deliberate: a sidecar left behind when the tree is already gone costs
+    one no-op delete on the next run, while a sidecar removed while the tree still
+    exists is unrecoverable. Survivors are reclaimed by the next run.
+
+  **Expect the occasional straggler, and expect it to clean itself up.** On a
+  256-snapshot tree this reliably sweeps ~255 immediately and may leave **one**: it is
+  whatever restic read last, so its 300-second window has barely opened. That one is
+  logged, its sidecar is kept, and the next run reclaims it before doing anything else.
+  The leak is bounded at a single cycle rather than growing without limit — which is
+  the property that actually matters. Blocking a backup job for five minutes to chase
+  the last snapshot would be a worse trade, so it is not made.
+
+- **Installing the patch permanently blocked updating it.** `install.sh` does
+  `chmod +x update.sh`, and git recorded `update.sh` as `100644` — so the chmod was a
+  *tracked modification*, and `update.sh` refuses to run over a dirty tree. Install
+  once and you could never update again; the error even told you to run
+  `git checkout -- .`, which just undoes the exec bit so the next install can re-dirty
+  it. A real box sat on an old version for exactly this reason.
+
+  Fixed on both sides: the scripts `install.sh` chmods are now executable in git (so
+  the chmod is a no-op), and `update.sh`'s dirty check now looks at **content**, not
+  file mode — `git diff --numstat` reports `0 0` for a mode-only change. A test
+  asserts every script in `install.sh`'s chmod loop is already `100755` in git.
+
 - **Nested snapshots were broken on TrueNAS 24.10 and 25.04, and had been all
  along.** `SYNC_BLOCK`'s wrapper spelled out the stock signature and forwarded five
  arguments — but those releases declare `restic_backup(middleware, job,
@@ -0,0 +1,4 @@
+truenas-truecloud-patch
+
+Parts of this project were written with AI assistance (Claude); all of it is
+reviewed and tested before release.
@@ -1,5 +1,7 @@
 # truenas-truecloud-patch

+[![CI](https://git.arch.fyi/flan/truenas-truecloud-patch/actions/workflows/ci.yml/badge.svg)](https://git.arch.fyi/flan/truenas-truecloud-patch/actions)
+
 Extends TrueNAS SCALE's **TrueCloud Backup** to:

 - back up to **Backblaze B2 and any S3-compatible provider**, not just Storj;
@@ -22,7 +24,7 @@ Clone it onto a **pool** (not the boot device — that is wiped on TrueNAS upgra
 then run `install.sh` as root:

 ```bash
-git clone https://git.onetick.ninja/flan/truenas-truecloud-patch.git \
+git clone https://github.com/sudolulo/truenas-truecloud-patch.git \
    /mnt/tank/truenas-truecloud-patch          # replace `tank` with your pool
 cd /mnt/tank/truenas-truecloud-patch
 sudo bash install.sh
@@ -59,9 +61,10 @@ If something is wrong, the reason is in `apply.log` — start at
 | --- | --- | --- | --- |
 | 24.10.2.4 | ok | ok | — |
 | 25.04.2.6 | ok | ok | — |
-| 25.10.4 | ok | ok | nested + providers; 252-snapshot recursive backup of /mnt/Tap, 18m |
-| 26.0.0-BETA.3 _(unreleased)_ | ok | **BROKEN** | — |
-| master _(unreleased)_ | **BROKEN** | **BROKEN** | — |
+| 25.10.7 | ok | ok | — |
+| 24.10.2.5 _(unreleased)_ | ok | ok | — |
+| 26.0.0-RC.1 _(unreleased)_ | ok | ok | — |
+| master _(27-dev)_ | **BROKEN** | **BROKEN** | — |

 | verdict | meaning |
 | --- | --- |
@@ -72,17 +75,44 @@ If something is wrong, the reason is in `apply.log` — start at
 "ok" means *the patch's assumptions hold*, checked automatically against iX's
 source. It does not mean a human ran a backup on it — that is the
 **Hardware-verified** column, which is filled in by hand and only by doing it.
+
+**`master` is not the next release.** iX branches each major off to its own
+`release/` line and master rolls straight on to the one after — so master is
+`27-dev` while 26 is still in beta. A **BROKEN** master means iX has changed
+something that will reach users *a major release from now*, not in the version you
+are about to install. Read the numbered rows for that.
+
+A row like `25.10.5 _(unreleased)_` is the next maintenance release: branched by iX,
+not tagged yet, and the very next thing a 25.10.4 box gets. It is checked precisely
+because it is the one unshipped ref that reaches real users without warning.
 <!-- END COMPAT MATRIX -->

 The table is **regenerated daily by CI** against iXsystems' actual middleware source
 — it is not a claim somebody typed once and forgot.

-**TrueNAS 26: nested snapshots are not supported yet, and upgrading will not break
-you.** 26 rewrites `cloud_backup` and deletes the ZFS methods this module calls. On
-26 `apply.sh` finds that the assumptions no longer hold and **does not apply the
-module**: TrueNAS is left stock, B2/S3 keeps working, nested datasets are simply not
-covered, and the reason is named in `apply.log`. A broken backup is worse than a
-missing feature. Details: [How it works](docs/how-it-works.md#truenas-26).
+It is also **static analysis**: it proves the patch's assumptions still hold, which is
+a weaker claim than "a backup ran and a restore came back". For what has actually been
+run — which tasks, on which hardware, and the md5 of the file that came back — see
+[docs/verification.md](docs/verification.md).
+
+**TrueNAS 26 is supported** as of v0.7.0, and was verified on a real
+**26.0.0-BETA.1** install: a 274-snapshot recursive backup of a 292-dataset pool, and a
+byte-identical restore of a four-level-deep child dataset. The *Hardware-verified*
+column tracks the newest beta iX has tagged (currently BETA.3), so it does not carry
+that mark — a build nobody has actually run a backup on does not get credit for one.
+
+26 rewrites `cloud_backup` from async to synchronous and deletes the private ZFS
+methods this module used to call, so getting there took real work: the patch now
+injects the wrapper flavour that matches the installed middleware, reads dataset and
+snapshot lists **from ZFS rather than middleware** (whose queries hide TrueNAS's own
+datasets — 84 of 270 on a real pool, including live app data), and owns the snapshot
+sweep even when it stages nothing (26 decides `recursive` by a rule this patch does not
+share, and would otherwise orphan one snapshot per zvol on every run).
+
+**And if a future TrueNAS breaks it, you get a missing feature, not a broken backup.**
+`apply.sh` re-checks the patch's assumptions at every boot and **refuses to apply a
+module whose assumptions no longer hold** — TrueNAS is left stock, and the reason is
+named in `apply.log`. Details: [How it works](docs/how-it-works.md#truenas-26).

 ---

@@ -125,6 +155,17 @@ alert, which both take the newest plain `vX.Y.Z` tag. That is what lets debuggin
 happen in `-rc` tags instead of in your notification bell — see
 [Releasing](docs/releasing.md).

+**A second alert reports the patch not being loaded**, and this one you cannot
+turn off with `--no-update-alerts` — it is CRITICAL, hourly, and it means B2/S3
+backup tasks are about to fail. Being patched *on disk* and being patched *in the
+running middlewared* are different facts, and only middlewared can answer the
+second one: the patch stamps the objects it replaces, so a missing stamp means
+the process imported stock code. It fires if something detaches the patch overlay
+(a `systemd-sysext` merge over `/usr`, for instance) and the self-healing re-apply
+in the deferred restart could not put it back. `bash install.sh` clears it. It
+stays quiet when the kill switch is set, or when the providers module has been
+retired because TrueNAS went native.
+
 The changelog is read from whichever forge `origin` points at, derived from the
 remote rather than hard-coded. That is not cosmetic: when the changelog cannot be
 read, the alert deliberately fires **anyway** rather than risk hiding a security
@@ -207,5 +248,7 @@ and restores the original UI bundle from backup.
 - Filing a TrueNAS bug? **Remove the patch first** and reproduce on a stock system.
 - Provided as-is, no warranty. See LICENSE.

-Parts of this project were written with AI assistance (Claude); all of it is reviewed
-and tested before release. Bugs are mine.
+## Support
+
+If truenas-truecloud-patch is useful to you, consider supporting development via
+[GitHub Sponsors](https://github.com/sponsors/sudolulo) or [Ko-fi](https://ko-fi.com/sudolulo).
@@ -74,22 +74,46 @@ Two different things must survive two different events:
   and creates a transient systemd unit (`truecloud-mw-restart`, via
   `systemd-run --no-block`) running `patch/wait_restart.sh` — detached so it
   cannot disrupt the remainder of the boot sequence.
-5. **Once boot has settled, middlewared restarts once** and imports the
-   patched modules from the overlay. `wait_restart.sh` holds the restart until
-   the systemd boot job queue has drained (so in-flight `ix-*` units like
-   `ix-reporting` finish first) *and* middlewared's docker/apps startup has
-   reached a terminal state — plain unit ordering cannot see either, and
-   restarting middlewared while they run kills apps and dashboard reporting
+5. **Once boot has settled, the patch is re-applied and middlewared restarts
+   once**, importing the patched modules from the overlay. `wait_restart.sh`
+   holds the restart until the systemd boot job queue has drained (so in-flight
+   `ix-*` units like `ix-reporting` finish first) *and* middlewared's docker/apps
+   startup has reached a terminal state — plain unit ordering cannot see either,
+   and restarting middlewared while they run kills apps and dashboard reporting
   for the whole boot. S3/B2 backup support is then active until the next
   reboot, when the cycle repeats.

+   The **re-apply** in that sentence is load-bearing, not a safety blanket. The
+   overlay from step 3 sits *inside* `/usr`, so anything that remounts that
+   hierarchy detaches it, and two ordinary things do exactly that after our hook
+   has finished: another PREINIT script running `systemd-sysext merge`/`refresh`
+   over `/usr` (an out-of-tree nvidia driver, say), and middlewared's own
+   `docker.configure_nvidia` when it brings docker up. Init scripts run
+   sequentially in id order, so a hook registered after ours always wins — and
+   ordering them differently would still not help, because `docker.configure_nvidia`
+   fires at runtime. `wait_restart.sh` therefore re-runs `apply.sh` at the point
+   where boot has settled and every such remount is behind it, re-mounting the
+   overlay if it was torn off (same upper layer, so files patched in step 3
+   reappear intact), then verifies the patch is really on the live path, restarts,
+   and verifies again — retrying once if it was lost in between.
+
+   This is the failure that made it necessary: on 2026-08-19 the overlay was
+   mounted at 16:41:56 and a sysext refresh unmerged and remerged `/usr` four
+   seconds later. The restart at 16:47:24 loaded stock modules, and every B2
+   backup failed for nineteen hours while `apply.log` said `OK` — because
+   `apply.log` can only report what was written to disk, never what the restart
+   imported. That second question is now asked from inside middlewared by an
+   hourly CRITICAL alert (see [Update alerts](../README.md#update-alerts)).
+
 What you will observe: one middlewared restart shortly after every boot (a
 brief web UI/API blip; running services are unaffected). Between steps 3
 and 5 there is a short window — typically well under a minute — where the UI
 already shows S3/B2 (the JS bundle is read from disk per request) but the
 backend is still stock. A backup job that fires inside that window fails once
 with `NotImplementedError` and succeeds on its next run; see
-[Troubleshooting](recovery.md) if it persists beyond boot.
+[Troubleshooting](recovery.md) if it persists beyond boot. If the backend is
+still stock an hour after boot, middlewared raises the "installed but NOT
+loaded" alert rather than leaving you to notice via a failed backup.

 Manual runs of `bash patch/apply.sh` never trigger the restart — that only
 happens in boot context. `install.sh` and `recover.sh` perform their own
@@ -158,7 +182,7 @@ python3 /mnt/tank/truenas-truecloud-patch/patch/create_task.py verify
 | What you see | What it means |
 |---|---|
 | `[OK] providers`, `[OK]`/`[SKIP] nested_snapshots` | Fine. Nothing to do. |
-| `WARNING: … pattern not found` (UI) | The Angular bundle changed. The UI dropdown reverts to Storj-only, but **backups keep working** — create tasks with `create_task.py` meanwhile, and [open an issue](https://git.onetick.ninja/flan/truenas-truecloud-patch/issues) with your TrueNAS version. |
+| `WARNING: … pattern not found` (UI) | The Angular bundle changed. The UI dropdown reverts to Storj-only, but **backups keep working** — create tasks with `create_task.py` meanwhile, and [open an issue](https://github.com/sudolulo/truenas-truecloud-patch/issues) with your TrueNAS version. |
 | `WARNING: truecloud-patch is NOT COMPATIBLE with this TrueNAS version` | This TrueNAS changed middleware underneath the patch, and the named module was **deliberately not applied** — see `incompatible.json` for exactly which assumption broke. TrueNAS is left stock, so nothing is half-patched. Check [TrueNAS compatibility](../README.md#truenas-compatibility), then `bash update.sh` once a release supports your version; it re-applies itself on the next boot. This is **not** the kill switch and needs no manual reset. |
 | `[FAIL] providers` | **Your B2/S3 backups will not run.** middlewared is fine, but the credential/URL handling is gone. Open an issue with your version. |
 | `[FAIL] nested_snapshots` | The stock guard is back, so tasks with `snapshot = true` on a nested dataset will fail validation. Turn the option off on those tasks until it's fixed. |
@@ -187,23 +211,71 @@ backup-breaking**, and none of them is visible from the `cloud_backup` files:
 | `get_dataset_recursive()` **deleted** from `plugins/cloud/snapshot.py` | `NameError` — the injected block called it out of the host module's namespace |
 | `plugins/zfs_/dataset.py` and `zfs_/snapshot.py` **deleted** | `zfs.dataset.query`, `zfs.snapshot.query` and `zfs.snapshot.delete` all vanish. 26 uses `filesystem.statfs` and `zfs.resource.*` |

-The first two are fixed: the patch reads which flavour of `cloud_backup` your box
-declares and injects the wrapper that matches (one implementation of the real logic,
-two thin wrappers), and it carries its own copy of the deleted helper.
+All three are fixed as of **v0.7.0**, and 26 is supported.

-The third is **not** fixed, and is why 26 reports BROKEN. Porting it means rewriting
-the module's ZFS calls onto 26's new API, and no single API spans 24.10 through 26 —
-so it needs a real 26 box to verify against, not a plausible-looking diff. Shipping a
-port nobody has run is exactly the failure this project exists to avoid.
+The first two were straightforward: the patch reads which flavour of `cloud_backup`
+your box declares and injects the wrapper that matches (one implementation of the real
+logic, two thin wrappers), and it carries its own copy of the deleted helper.

-It is also the row that would have hurt most. `zfs.snapshot.delete` is what sweeps the
-recursive snapshot; without it, **every run would orphan one snapshot per descendant
-dataset — 250 on a real pool — forever.** The compatibility check caught it only
-because it now asserts the middleware *methods the patch calls*, not just the symbols
-it wraps.
+The third was not, and it is the one that would have hurt most — `zfs.snapshot.delete`
+is what sweeps the recursive snapshot, and without it **every run would orphan one
+snapshot per descendant dataset (250 on a real pool), forever, while reporting
+success.**

-`master` (development after 26) reports BROKEN too: iXsystems are still reshaping
+The obvious port is to the public `pool.dataset.query` / `pool.snapshot.query`. Those
+methods exist, are documented, and are covered by iX's deprecation policy — and they
+are **not like-for-like replacements**. They apply a *visibility policy*: they hide the
+datasets TrueNAS considers its own (`ix-apps/*`, `.system/*`, `.ix-virt/*`). On a real
+pool that is **84 of 270 datasets, including live application data.** Staging from that
+view would have omitted every one of them from the backup — and the planner would never
+have seen them, so they would not have appeared in its "skipped" list either. A green
+backup, quietly missing data. The snapshot query lies the same way, so the sweep would
+have orphaned one snapshot per hidden dataset.
+
+No source analysis could have caught that. The methods are all present and correctly
+shaped. Only running it could, which is why it took a real 26 box.
+
+So the module now follows one rule:
+
+> **Read the truth from ZFS. Make changes through middleware.**
+
+Enumeration is `zfs list` — no policy can filter it, and it behaves identically on every
+release, which also means one code path instead of a version conditional. Mutation stays
+a middleware call, so TrueNAS's own bookkeeping stays consistent; an exact-name delete
+works fine even on a dataset the query hides. It is only enumeration that lies.
+
+The snapshot *delete* still needs a namespace, and no single one spans every release —
+24.10 and 25.04 have `zfs.snapshot`, 26 has only `pool.snapshot`, 25.10 has both. So it
+is resolved at runtime, by asking whether the namespace can actually delete. `tools/compat.py`
+asks the identical question against iX's source, and a test binds the two lists together,
+so what CI verifies and what runs cannot drift apart.
+
+### The one 26 changed that nothing warned about
+
+Stock decides whether to take a **recursive** snapshot by its own rule, and on 26 that
+rule stopped being ours:
+
+| | decides `recursive` by |
+| --- | --- |
+| stock ≤ 25.10 | `get_dataset_recursive()` — the same function this patch vendors |
+| **stock 26** | `filesystem.statfs`: `recursive = (path == the dataset's mountpoint)` |
+| this patch | `get_dataset_recursive()` — is a mounted *filesystem* child under the path? |
+
+Up to 25.10 those were the *same question*, so a snapshot the patch declined to stage
+provably had no children and stock's non-recursive delete was correct. On 26 they
+disagree: a dataset whose only descendants are **zvols** or **legacy-mountpoint**
+datasets gets a recursive snapshot, while the patch sees nothing to stage. Stock then
+destroys the parent only — and with no staging tree there was no sidecar, and the
+garbage collector only ever ran from the staging path. Nothing on the box would ever
+have found the children.
+
+It was reproduced on a 26 VM (one orphan per zvol, every run, backup green) and closed:
+**ownership of the sweep is no longer conditional on staging.**
+
+`master` (development after 26) **does** report BROKEN: iXsystems are still reshaping
 these functions there, renaming `middleware` → `context` and `cloud_backup` → `entry`
 and adding a required `credentials` parameter. That is a moving target and is
 deliberately not chased; the check keeps reporting it until it settles into a beta,
-which is when it becomes worth fixing.
+which is when it becomes worth fixing. Until then, a box running master would simply
+not get the modules — `apply.sh` refuses to apply a module whose assumptions no longer
+hold, and says why in `apply.log`.
@@ -95,6 +95,40 @@ re-scan each time.

 ### Snapshot lifecycle

+> **Two mechanisms clean up, and the second exists because the first can be destroyed.**
+>
+> 1. **The sidecar** records exactly which snapshots a run pinned, and is removed only
+>    on a confirmed-clean sweep. Precise, and it survives a middlewared restart.
+> 2. **The garbage collector** finds leftovers by *name*, so it still works when the
+>    sidecar is gone — and it can be: **the sidecar lives in `/run`, which is tmpfs.** A
+>    reboot mid-backup takes it, and with it the only record of a 250-snapshot tree.
+>
+> The collector runs at the start of every backup, after the sidecar reclaim. It will
+> only touch a snapshot named `<dataset>@<task>-<timestamp>` that is not the current
+> run's, has **nothing mounted from it** (which is what protects a concurrently-running
+> backup), and is **over an hour old**. Periodic `auto-*` snapshots, other tasks'
+> snapshots, and anything you made by hand are structurally out of reach.
+
+
+> **A snapshot may survive a run, and that is expected.** ZFS **automounts**
+> `<dataset>/.zfs/snapshot/<snap>` the moment it is read, and holds it for
+> `zfs_expire_snapshot` seconds (**300** by default) after the last access. So
+> whatever restic read *last* is still pinned when we try to destroy it, and
+> `zfs destroy` refuses with `dataset is busy`.
+>
+> The patch unmounts those automounts itself and retries, which clears ~255 of 256 on
+> a real pool. The one that remains is **logged, its sidecar is kept, and the next run
+> reclaims it before doing anything else** — so the leak is bounded at a single cycle
+> instead of growing forever. Seeing one `could not delete snapshot … it will be
+> reclaimed on the next run` in the log is normal. Seeing the count *grow* run over run
+> is not, and would be a bug.
+>
+> This is why the sidecar is removed **only on a confirmed-clean sweep**: it is the
+> only record those snapshots exist, and a run that dropped it while they were still
+> around would orphan them permanently. That is precisely what happened before this was
+> fixed.
+
+
 `zfs.snapshot.delete` defaults to **`recursive=False`**, and stock
 `restic_backup()` calls it with no options. Stock is safe only because its
 validation means a *recursive* snapshot never actually happens in the field.
@@ -109,7 +109,7 @@ If a module shows `[FAIL]`:
   risk.
 4. **If the detail says the module doesn't exist**, a TrueNAS update renamed
   or restructured the internal API.
-   [Open an issue](https://git.onetick.ninja/flan/truenas-truecloud-patch/issues)
+   [Open an issue](https://github.com/sudolulo/truenas-truecloud-patch/issues)
   with your TrueNAS version number and the full verify output.

 ---
@@ -120,8 +120,9 @@ If a module shows `[FAIL]`:

 The traceback ends in `rclone/base.py` → `raise NotImplementedError` and
 contains no `_tc_` frames: the running middlewared is executing stock code.
-Either the deferred restart never fired, or the patch never landed on disk
-this boot. Diagnose in this order:
+Either the deferred restart never fired, the patch never landed on disk this
+boot, or it landed and was then torn off before the restart. Diagnose in this
+order:

 ```bash
 # Did apply.sh run this boot, at which version, and did it schedule the restart?
@@ -130,11 +131,21 @@ tail -40 /mnt/tank/truenas-truecloud-patch/apply.log
 # Full check — compares the running process against the patch timestamp
 python3 /mnt/tank/truenas-truecloud-patch/patch/create_task.py verify

-# Did the deferred restart unit run, fail, or never get created?
-systemctl status truecloud-mw-restart.service
+# What the deferred restart did -- re-apply, restart, and what it verified.
+# apply.log is the durable record; journald retention on a busy box is often
+# shorter than the gap between reboots, so the journal may have nothing left.
+grep wait_restart /mnt/tank/truenas-truecloud-patch/apply.log | tail -20
 journalctl -u truecloud-mw-restart.service --no-pager | tail -20
+
+# Did something remount /usr and detach the patch overlay?
+systemd-sysext status
+findmnt -o TARGET,SOURCE /usr/lib/python3/dist-packages
 ```

+`systemctl status truecloud-mw-restart.service` reporting *"could not be
+found"* is **normal** — the unit is transient and is collected once it exits.
+It is not evidence that the restart was skipped.
+
 - `verify` reports the process started **before** the patch → the restart
  didn't happen. `systemctl restart middlewared` fixes it immediately; the
  journal output above tells you why it was missed.
@@ -145,6 +156,22 @@ journalctl -u truecloud-mw-restart.service --no-pager | tail -20
 - `apply.log` header shows `[v0.0.3]` or older → update:
  `git pull && bash install.sh` (v0.0.4 fixed patches not loading after
  reboot).
+- `apply.log` says the patch applied, but `findmnt` shows no `truecloud-mw`
+  overlay on the dist-packages path → something remounted `/usr` after our
+  PREINIT hook and detached it. `systemd-sysext status` names the culprit if it
+  is a sysext (the `SINCE` column will sit a few seconds *after* the `apply.log`
+  timestamp). Releases from 2026-08-26 on re-apply and verify immediately before
+  the restart, so this should self-heal; if you are seeing it, update first.
+
+**TrueNAS raises "truecloud-patch is installed but NOT loaded"**
+
+The definitive symptom, and it does not depend on a backup failing first: the
+running middlewared has stock cloud_backup modules even though the patch is
+installed and its providers module is meant to be active. `bash install.sh`
+re-applies and restarts. The alert clears within the hour. It is silent when the
+kill switch is set or the providers module has been retired as native, and it is
+deliberately not muted by `update_alerts_disabled` — that silences release
+notifications, not a broken backup path.

 **Apply log** (check after each reboot or install):
 ```bash
@@ -25,7 +25,7 @@ are worth calling out, because nothing else would catch what they catch:
  newest CHANGELOG entry. `VERSION=` had silently drifted to three different
  values across the scripts before anything checked.

-The project is hosted on **Gitea** (`git.onetick.ninja/flan/truenas-truecloud-patch`)
+The project is hosted on **Gitea** (`git.arch.fyi/flan/truenas-truecloud-patch`)
 and mirrored to GitHub. Both run the same workflows — Gitea reads
 `.github/workflows/` too — so a change is checked twice, on two independent runners.

@@ -0,0 +1,77 @@
+# What has actually been run
+
+The support matrix in the README is **static analysis**: it proves the patch's
+assumptions about middlewared still hold. That is a strictly weaker claim than "a
+backup ran and a restore came back". This file is the stronger claim, and it is
+maintained by hand, because the only way to fill it in is to do it.
+
+If you are deciding whether to trust this with your backups, read this file, not the
+matrix.
+
+---
+
+## v0.7.0 — TrueNAS 25.10.4 (production hardware)
+
+Six live TrueCloud tasks, all `snapshot = true`, backing up to Backblaze B2. Three were
+exercised end to end, chosen to cover the three shapes the code handles differently:
+
+| Task | Path | Shape | Result |
+| --- | --- | --- | --- |
+| 5 | `/mnt/Tap` | 191 nested datasets staged, 282-snapshot recursive tree | SUCCESS |
+| 7 | `/mnt/Tank/backups` | 215 filesystems **+ 2 zvols** | SUCCESS |
+| 9 | `/mnt/Tank/flan` | **no** nested filesystem children | SUCCESS |
+
+After every run: **0 orphaned snapshots, 0 leaked bind mounts, 0 stale sidecars.**
+
+**The restore.** `apps/vaultwarden/data/config.json` — a file inside a *child* dataset,
+which is exactly what stock TrueNAS cannot capture — was restored from B2 and compared
+against the live file:
+
+    live      f809df6ba231986b1ba824044228a03a   1808 bytes
+    restored  f809df6ba231986b1ba824044228a03a   1808 bytes
+    => byte-identical
+
+**The collector earned its keep on real data.** The pool was already carrying an orphan:
+`Tap/apps/prometheus@cloud_backup-5-20260713202355`, left behind by an earlier run when
+ZFS's automount held the snapshot busy past all four retries. The first v0.7.0 run found
+it by name, reclaimed it, and the pool's snapshot count went 2148 → 2147. That is the
+garbage collector doing the job it was written for, against a leak that was already
+there and that nothing else would ever have found.
+
+**Boot path.** `apply.sh` is registered as a PREINIT `initshutdownscript`; it was
+re-run against the live middleware and left exactly one `TRUECLOUD_PATCH` marker in
+each patched module (a second copy stacked into a live middlewared module would break
+the box at boot). It correctly detected the box as **async** (`cloud_backup is async
+(TrueNAS <= 25.10)`) and injected the matching wrappers.
+
+**Upgrade path.** `update.sh` was used to move the box from the release candidate to
+the stable tag, in detached HEAD at `v0.7.0`, which is how a user's box actually
+upgrades.
+
+## v0.7.0 — TrueNAS 26.0.0-BETA.1 (VM)
+
+A throwaway VM whose pool reproduces the production pool's *shape* — 292 datasets, 26
+`legacy` mountpoints, nesting five deep — because every bug found on the real box came
+from the shape of the pool, not the bytes in it. MinIO was not used; `rclone serve s3`
+(already on the box) provided the S3 target, so no real B2 credential ever entered the
+VM.
+
+* 274-snapshot recursive backup of the 292-dataset pool. 0 orphans, 0 leaked mounts.
+* Restored `ix-apps/app_mounts/vaultwarden/pgData` — **four levels deep, and a dataset
+  that middleware's own `pool.dataset.query` hides from itself** — byte-identical.
+* The zvol-orphan case was **reproduced with the fix disabled** (one orphan per zvol,
+  every run, backup green), then **closed with it enabled**. See the CHANGELOG entry
+  for why TrueNAS 26 decides `recursive` by a different rule than this patch decides
+  `nested`.
+
+## What is NOT covered
+
+* **24.10 and 25.04** are `ok` in the matrix — the assumptions hold, checked against
+  iX's source — but nobody has run a backup on them. The matrix says so.
+* **master** is BROKEN, and correctly reports so: iX renamed the leading parameters of
+  `get_restic_config` and `restic_backup`. It is not a shipped release; the daily
+  compatibility bot files it, and `apply.sh` would refuse to apply the modules on a box
+  running it.
+* A **reboot** of the production box has not been done on v0.7.0. `apply.sh` was
+  re-executed by hand against the live middleware, which exercises the same code path,
+  but the PREINIT ordering itself has only been proven on earlier versions.
@@ -4,7 +4,7 @@
 # Prerequisites: run as root on TrueNAS SCALE with middlewared running.
 # Clone this repository to a persistent ZFS pool first:
 #
-#   git clone https://git.onetick.ninja/flan/truenas-truecloud-patch \
+#   git clone https://github.com/sudolulo/truenas-truecloud-patch \
 #       /mnt/<pool>/truenas-truecloud-patch
 #   cd /mnt/<pool>/truenas-truecloud-patch && bash install.sh
 #
@@ -18,7 +18,7 @@

 set -euo pipefail

-VERSION="0.6.0"
+VERSION="0.8.0"

 # The directory containing install.sh is the permanent install location.
 PATCH_DIR="$(cd "$(dirname "$0")" && pwd)"
@@ -72,7 +72,7 @@ done
 if [ ! -f "$PATCH_DIR/patch/apply.sh" ]; then
    echo "ERROR: patch files not found at $PATCH_DIR/patch/" >&2
    echo "Run install.sh from a clone of the repository on a persistent pool:" >&2
-    echo "  git clone https://git.onetick.ninja/flan/truenas-truecloud-patch \\" >&2
+    echo "  git clone https://github.com/sudolulo/truenas-truecloud-patch \\" >&2
    echo "      /mnt/<pool>/truenas-truecloud-patch" >&2
    echo "  cd /mnt/<pool>/truenas-truecloud-patch && bash install.sh" >&2
    exit 1
@@ -21,6 +21,7 @@ the way a `git fetch` from middlewared (running as root) would.

 import datetime
 import importlib.util
+import json
 import logging
 import os
 import re
@@ -47,8 +48,8 @@ _VERSION_RE = re.compile(r'^VERSION="([^"]+)"', re.M)
 #: owner/repo out of any of:
 #:   git@github.com:sudolulo/repo.git
 #:   https://github.com/sudolulo/repo.git
-#:   ssh://git@git.onetick.ninja:55214/flan/repo.git
-#:   https://git.onetick.ninja/flan/repo.git
+#:   ssh://git@git.arch.fyi:55214/flan/repo.git
+#:   https://git.arch.fyi/flan/repo.git
 #: The SSH port is deliberately not captured: it is not the web port.
 _REMOTE_RE = re.compile(
    r"^(?:\w+://)?(?:[^@/]+@)?([^:/]+)(?::\d+)?[:/]([^/]+)/([^/]+?)(?:\.git)?/?$"
@@ -77,6 +78,107 @@ class TrueCloudPatchSecurityUpdateAlertClass(AlertClass):
    )


+class TrueCloudPatchNotLoadedAlertClass(AlertClass):
+    category = AlertCategory.SYSTEM
+    level = AlertLevel.CRITICAL
+    title = "truecloud-patch is installed but NOT loaded"
+    text = (
+        "truecloud-patch patched middlewared on disk, but this middlewared is "
+        "running the STOCK cloud_backup modules -- B2 and S3 TrueCloud Backup "
+        "tasks will fail with NotImplementedError. Something remounted /usr "
+        "after the patch was applied (a systemd-sysext merge, or "
+        "docker.configure_nvidia), detaching the patch overlay. Re-apply with:  "
+        "bash %(dir)s/install.sh"
+    )
+
+
+class TrueCloudPatchNotLoadedAlertSource(ThreadedAlertSource):
+    """Does the middlewared running this check actually have the patch in it?
+
+    This is the one question apply.log cannot answer. apply.sh reports what it
+    wrote to disk; whether the restart that followed imported those files is a
+    separate fact, and on 2026-08-19 the two disagreed silently for nineteen
+    hours while every B2 backup task failed. Asking from inside the process is
+    exact -- the patch stamps the objects it replaces, so a missing stamp means
+    this interpreter imported stock code.
+
+    Deliberately NOT silenced by the update-alert marker: that mutes release
+    notifications, not a broken backup path. Only the patch's own kill switch
+    (the `disabled` file, meaning the operator turned the patch off) stops it.
+    """
+
+    schedule = IntervalSchedule(datetime.timedelta(hours=1))
+    run_on_backup_node = False
+
+    def check_sync(self):
+        try:
+            return self._check()
+        except Exception:
+            # An alert source must never take middlewared down with it.
+            logger.debug("truecloud-patch loaded check failed", exc_info=True)
+            return None
+
+    # -- internals ------------------------------------------------------------
+
+    def _check(self):
+        if os.path.exists(os.path.join(PATCH_DIR, "disabled")):
+            return None
+
+        # Only the providers module puts B2/S3 on the restic path. If it was
+        # never applied here, or TrueNAS went native and it was retired, then
+        # "not loaded" is the correct state and not a fault.
+        status = self._hook_status()
+        if not status:
+            return None
+        providers = status.get("patches", {}).get("providers", {})
+        if not providers.get("active"):
+            return None
+
+        if self._providers_loaded():
+            return None
+
+        return Alert(
+            TrueCloudPatchNotLoadedAlertClass,
+            {"dir": PATCH_DIR},
+            key=None,
+        )
+
+    def _hook_status(self):
+        try:
+            with open(os.path.join(PATCH_DIR, "hook_status.json")) as f:
+                return json.load(f)
+        except (OSError, ValueError):
+            return None
+
+    def _providers_loaded(self):
+        """True when THIS interpreter holds the patched provider objects.
+
+        Two independent stamps, because the two halves are written separately
+        and either can be missing on its own:
+
+        * restic.py -- apply.sh sets `_truecloud_patched` on the wrapper it
+          installs over `get_restic_config`.
+        * b2.py -- apply.sh binds a B2-specific `get_restic_config` onto
+          `B2RcloneRemote`. Comparing it against the base implementation is
+          exact and survives renames of the patch's own helper.
+        """
+        try:
+            from middlewared.plugins.cloud_backup.restic import get_restic_config
+        except Exception:
+            return False
+        if not getattr(get_restic_config, "_truecloud_patched", False):
+            return False
+
+        try:
+            from middlewared.rclone.base import BaseRcloneRemote
+            from middlewared.rclone.remote.b2 import B2RcloneRemote
+        except Exception:
+            return False
+        base = getattr(BaseRcloneRemote, "get_restic_config", None)
+        b2 = getattr(B2RcloneRemote, "get_restic_config", None)
+        return b2 is not None and b2 is not base
+
+
 class TrueCloudPatchUpdateAlertSource(ThreadedAlertSource):
    schedule = IntervalSchedule(datetime.timedelta(hours=24))
    run_on_backup_node = False
@@ -32,7 +32,7 @@
 # Derive PATCH_DIR from this script's location (parent of the patch/ directory).
 PATCH_DIR="$(cd "$(dirname "$0")/.." && pwd)"
 LOG="$PATCH_DIR/apply.log"
-VERSION="0.6.0"
+VERSION="0.8.0"

 # Rotate log at 512 KB to avoid unbounded growth on a system volume.
 # Keep two prior generations (.1 and .2) so the last three boots are always available.
@@ -64,17 +64,45 @@ _ensure_writable() {
        rm -f "$dir/.truecloud-probe"
        return 0
    fi
-    # Already our overlay on this exact directory from an earlier run this boot?
+    # Not writable, so any overlay of ours listed on this directory is a
+    # SHADOWED leftover rather than a working mount: something remounted the
+    # hierarchy above it -- a systemd-sysext merge/refresh over /usr, or
+    # middlewared's own docker.configure_nvidia -- and buried it. A live
+    # overlay of ours is always writable, so this must never be treated as
+    # "already done"; doing so is what let a buried overlay pass for a healthy
+    # one and left the backend patch on disk but never loaded.
    if mount | grep -qF "truecloud-${tag} on ${dir} "; then
-        return 0
+        echo "NOTICE: a previous truecloud-${tag} overlay on $dir is shadowed --"
+        echo "NOTICE: the hierarchy above it was remounted. Detaching and re-mounting."
+        umount -l "$dir" 2>/dev/null
    fi
+    # Keep the SAME upperdir across re-mounts: it holds everything patched
+    # earlier this boot, so re-mounting restores those files intact instead of
+    # re-deriving them. The workdir is scratch and must be empty, so it is
+    # recreated -- a stale one left behind by a detached mount fails the mount.
    local upper="/run/truecloud-${tag}-upper" work="/run/truecloud-${tag}-work"
-    mkdir -p "$upper" "$work"
+    mkdir -p "$upper"
+    rm -rf "$work" 2>/dev/null
+    mkdir -p "$work"
    if mount -t overlay "truecloud-${tag}" \
           -o "lowerdir=$dir,upperdir=$upper,workdir=$work" "$dir" 2>/dev/null; then
        echo "OK: Mounted writable overlay on $dir"
        return 0
    fi
+    # A lazily-detached overlay releases its workdir only once its last user is
+    # gone, and overlayfs refuses a workdir that is still in use. That would turn
+    # the re-mount this function exists to perform into a hard failure, so retry
+    # once on a private workdir. It is scratch in /run (tmpfs) and goes away at
+    # the next boot; the upperdir, which holds the patched files, is unchanged.
+    work="/run/truecloud-${tag}-work.$$"
+    rm -rf "$work" 2>/dev/null
+    mkdir -p "$work"
+    if mount -t overlay "truecloud-${tag}" \
+           -o "lowerdir=$dir,upperdir=$upper,workdir=$work" "$dir" 2>/dev/null; then
+        echo "OK: Mounted writable overlay on $dir (fresh workdir)"
+        return 0
+    fi
+    rmdir "$work" 2>/dev/null
    echo "WARNING: overlay mount failed on $dir — backend patch will be skipped."
    return 1
 }
@@ -196,6 +224,13 @@ _tc_native_nested=$(printf '%s' "$_tc_info" | sed -n '2p')
 SITE_PKG=$(printf '%s' "$_tc_info"          | sed -n '3p')
 _MW_DIR=$(printf '%s' "$_tc_info"           | sed -n '4p')

+# Record the resolved middlewared directory so patch/wait_restart.sh can check,
+# without re-deriving any of this, whether the patched modules are still on the
+# live filesystem path at the moment it restarts middlewared.
+if [ -n "$_MW_DIR" ]; then
+    printf '%s\n' "$_MW_DIR" > "$PATCH_DIR/.mw_dir" 2>/dev/null
+fi
+
 # Nested support is opt-in; if it was never enabled, it cannot be the reason to
 # keep the patch alive.
 if [ -f "$PATCH_DIR/nested_snapshots_enabled" ]; then
@@ -535,8 +570,8 @@ if _tc_nested is not None:
        # snapshot=true, not just ours. Two consequences, and the second is worse:
        #
        #   * everything below is a NEW failure mode for tasks that worked before we
-        #     were installed. A `zfs.dataset.query` that errors would break a
-        #     CloudSync job we have no business touching.
+        #     were installed. A `zfs list` that errors would break a CloudSync job
+        #     we have no business touching.
        #   * if a CloudSync task ever were staged, nothing would ever tear it down:
        #     the teardown is wired into cloud_backup's restic_backup finally, and
        #     CRUD_BLOCK deliberately leaves CloudSync's nesting guard intact. The
@@ -556,18 +591,28 @@ if _tc_nested is not None:
            # our staging plan would not -- silently omitting it from the backup.
            # Read afterwards, an unsnapshotted dataset instead trips the isdir()
            # check in plan_staging and fails the run loudly. Loud beats silent.
-            datasets = middleware.call_sync(
-                "zfs.dataset.query", [["type", "=", "FILESYSTEM"]]
-            )
+            # query_filesystems() reads ZFS directly. It deliberately does NOT use
+            # pool.dataset.query: that applies a visibility policy and hides
+            # TrueNAS-internal datasets (ix-apps/*, .system/*, .ix-virt/*) -- 84 of
+            # 270 on a real pool, including live app data. Staging from the filtered
+            # view omits them silently, which is the one thing this must never do.
+            datasets = _tc_nested.query_filesystems(middleware)
            # OUR copy of get_dataset_recursive, not the host module's: TrueNAS 26
            # deleted that helper (create_snapshot uses filesystem.statfs now), so
            # calling it out of the module namespace is a NameError there.
            dataset, nested = _tc_nested.get_dataset_recursive(datasets, path)

            if not nested:
-                # No children: stock behaviour, untouched. Stock's `finally` owns
-                # the snapshot from here (its non-recursive delete is correct,
-                # because a non-nested snapshot has no children).
+                # Nothing to STAGE -- but we still own the SWEEP, and that is not a
+                # formality. Stock decides `recursive` by its own rule, and on 26 that
+                # rule is no longer ours: it snapshots recursively whenever the backup
+                # path IS the dataset's mountpoint (filesystem.statfs), while
+                # get_dataset_recursive() sees nothing to stage when the only
+                # descendants are ZVOLs or legacy/none-mountpoint datasets. Stock then
+                # deletes the PARENT ONLY. Without this, one snapshot per descendant is
+                # orphaned on every run, forever, with no sidecar and no GC to find it --
+                # and the backup still reports success.
+                _tc_nested.own_snapshot(middleware, name, snapshot, logger=_logger)
                return snapshot, snap_path

            staging_root = _tc_nested.stage_nested(
@@ -581,7 +626,19 @@ if _tc_nested is not None:
            # stays None and its `finally` deletes NOTHING. Sweep the tree ourselves
            # or leak the parent plus one snapshot per descendant dataset (160+ here)
            # on every failed run.
-            _tc_nested.delete_snapshot_tree(middleware, snapshot, logger=_logger)
+            #
+            # The sweep is itself wrapped: a cleanup that raises would REPLACE the
+            # original exception with its own, hiding why the backup actually failed.
+            # An error handler must not be able to lose the error.
+            try:
+                _tc_nested.delete_snapshot_tree(middleware, snapshot, logger=_logger)
+            except Exception as _tc_sweep_err:
+                if _logger:
+                    _logger.error(
+                        "truecloud-patch: could not sweep %s after a staging failure "
+                        "(%r) -- it is orphaned and must be deleted by hand",
+                        snapshot, _tc_sweep_err,
+                    )
            raise

        return snapshot, staging_root
@@ -977,8 +1034,13 @@ fi
 # runs (install.sh, recovery) never trigger a restart.
 #
 # The unit runs wait_restart.sh, which blocks until boot has actually
-# settled (systemd job queue drained, docker/apps state terminal) before
-# restarting. systemd ordering alone (After=multi-user.target, ≤ v0.0.4)
+# settled (systemd job queue drained, docker/apps state terminal), then
+# RE-APPLIES this script before restarting. The re-apply is not belt-and-
+# braces: our overlay lives inside /usr, and a systemd-sysext merge or
+# middlewared's docker.configure_nvidia remounts /usr *after* PREINIT and
+# detaches it, so what we patch here can be gone by restart time (seen
+# 2026-08-19). wait_restart.sh re-mounts and re-verifies at the moment it
+# matters. systemd ordering alone (After=multi-user.target, ≤ v0.0.4)
 # fired while ix-reporting and the docker/apps startup were still in flight
 # and killed both — apps and dashboard stats stayed down until the next
 # boot. No Type=oneshot: a oneshot's start job would hold the boot queue
@@ -993,7 +1055,12 @@ echo "--- deferred restart ---"
 #
 # "No module active at all" cannot reach here: that is the kill-switch branch
 # above, which exits.
-if ! grep -aq middlewared "/proc/$PPID/cmdline" 2>/dev/null; then
+if [ "${TRUECLOUD_REAPPLY:-0}" = "1" ]; then
+    # Invoked by patch/wait_restart.sh as its pre-restart re-apply pass. That
+    # unit already exists to do the restart and verifies the result, so
+    # scheduling another one here would be a loop.
+    echo "Re-apply pass from wait_restart.sh — that unit owns the restart."
+elif ! grep -aq middlewared "/proc/$PPID/cmdline" 2>/dev/null; then
    echo "Manual run (parent is not middlewared) — no restart scheduled."
 elif [ "$_backend_ok" != "1" ]; then
    echo "Nothing landed on disk — no restart scheduled (nothing new to load)."
@@ -52,7 +52,7 @@ import subprocess
 import sys
 import time

-__version__ = "0.6.0"
+__version__ = "0.8.0"

 _PATCH_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 _STATUS_FILE = os.path.join(_PATCH_DIR, "hook_status.json")
@@ -111,7 +111,7 @@ def main():
        print(
            "[truecloud-patch] WARNING: filterByProviders pattern not found in any JS bundle.\n"
            "[truecloud-patch] The TrueNAS webui may have been restructured in this version.\n"
-            "[truecloud-patch] File an issue at https://git.onetick.ninja/flan/truenas-truecloud-patch\n"
+            "[truecloud-patch] File an issue at https://github.com/sudolulo/truenas-truecloud-patch\n"
            f"[truecloud-patch] TrueNAS version info: {_tnversion()}"
        )
        return
@@ -136,7 +136,7 @@ def main():
            f"[truecloud-patch] WARNING: {count} replacement(s) in {path}; "
            f"expected exactly 1 — skipping write to avoid corrupting the bundle.\n"
            f"[truecloud-patch] File an issue at "
-            f"https://git.onetick.ninja/flan/truenas-truecloud-patch"
+            f"https://github.com/sudolulo/truenas-truecloud-patch"
        )
        return

@@ -154,7 +154,7 @@ def main():
            "[truecloud-patch] The UI is UNCHANGED and still works. This means the "
            "pattern no longer fits this TrueNAS build.\n"
            "[truecloud-patch] File an issue at "
-            "https://git.onetick.ninja/flan/truenas-truecloud-patch"
+            "https://github.com/sudolulo/truenas-truecloud-patch"
        )
        return

@@ -27,6 +27,50 @@
 # --wait` below waits for that same queue to drain — the unit would deadlock
 # on itself until the timeout. apply.sh schedules this with the default
 # service type, whose start job completes at fork.
+#
+# THE RE-APPLY PASS (added 2026-08-26). Applying the patch at PREINIT and
+# restarting later is only sound if the patched files are still on the live
+# path at the moment middlewared re-imports them. They may not be: our patch
+# lives in an overlay mounted *inside* /usr, and anything that remounts the
+# hierarchy above it detaches or buries that overlay. Two things on a normal
+# TrueNAS box do exactly that, both AFTER our PREINIT hook has run:
+#
+#   - `systemd-sysext merge/refresh` over /usr (an nvidia sysext, for
+#     instance) — `Unmerged '/usr'` then `Merged extensions into '/usr'`;
+#   - middlewared's own `docker.configure_nvidia`, which merges the stock
+#     nvidia sysext over /usr when it brings docker up.
+#
+# PREINIT scripts run sequentially in id order, so a hook registered after
+# ours always wins the race, silently. Observed 2026-08-19: our overlay was
+# mounted at 16:41:56 and a sysext refresh tore /usr down four seconds later;
+# the restart at 16:47:24 then loaded stock modules and every B2 cloud_backup
+# job failed for the next nineteen hours while apply.log said "OK".
+#
+# Ordering the hooks cannot fix this — docker.configure_nvidia re-merges at
+# runtime, long after every PREINIT hook is done. So instead of trusting the
+# PREINIT pass, re-apply immediately before the restart (apply.sh is
+# idempotent and re-mounts a lost overlay, keeping the same upperdir so
+# already-patched files survive), verify the marker is really on the live
+# path, and verify again afterwards.
+
+PATCH_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+LOG="$PATCH_DIR/apply.log"
+
+_log() { echo "[wait_restart] $*" >> "$LOG" 2>/dev/null; }
+
+# Is the providers patch visible on the live filesystem path -- i.e. would a
+# middlewared starting right now import it? Reads the marker apply.sh leaves
+# in restic.py. Returns 0 when patched, 1 when stock, 2 when we cannot tell
+# (no recorded middlewared dir yet, or the file is gone).
+_patch_visible() {
+    local mw_dir restic_py
+    mw_dir=$(cat "$PATCH_DIR/.mw_dir" 2>/dev/null)
+    [ -n "$mw_dir" ] || return 2
+    restic_py="$mw_dir/plugins/cloud_backup/restic.py"
+    [ -f "$restic_py" ] || return 2
+    grep -q "TRUECLOUD_PATCH" "$restic_py" 2>/dev/null && return 0
+    return 1
+}

 # 1. systemd layer: wait for the boot job queue to drain. This covers every
 #    ix-* oneshot still activating, including ix-reporting's in-flight midclt
@@ -39,6 +83,9 @@ timeout 900 systemctl is-system-running --wait > /dev/null 2>&1
 #    transitional states (PENDING/INITIALIZING/STOPPING/MIGRATING — see
 #    middlewared/plugins/docker/state_utils.py). An empty answer means
 #    midclt could not respond at all; keep waiting. Cap at 10 minutes.
+#    This also covers docker.configure_nvidia, the runtime /usr re-merge:
+#    waiting for docker to reach a terminal state means the merge that would
+#    bury our overlay has already happened by the time we re-apply below.
 for _ in $(seq 1 120); do
    _status=$(midclt call docker.status 2>/dev/null \
                  | grep -oE '"status": "[A-Z_]+"' | cut -d'"' -f4)
@@ -52,4 +99,70 @@ done
 #    queryable state (smb.configure and friends). Bounded insurance.
 sleep 30

-exec systemctl try-restart middlewared
+# 4. Re-apply pass. Boot has settled, so every sysext merge and docker nvidia
+#    configuration that could bury our overlay is behind us. Re-running
+#    apply.sh is cheap and idempotent: it re-mounts the overlay if it was
+#    detached (same upperdir, so files patched at PREINIT reappear intact)
+#    and re-patches anything that reverted to stock.
+_patch_visible
+case $? in
+    0) _log "providers patch still visible on the live path before restart" ;;
+    1) _log "PATCH LOST since PREINIT (something remounted /usr) — re-applying" ;;
+    *) _log "cannot confirm patch state before restart — re-applying anyway" ;;
+esac
+
+TRUECLOUD_REAPPLY=1 /bin/bash "$PATCH_DIR/patch/apply.sh"
+
+if ! _patch_visible; then
+    _log "WARNING: patch is STILL not on the live path after the re-apply pass;"
+    _log "WARNING: restarting anyway, but middlewared will load stock modules."
+fi
+
+# 5. The restart itself.
+systemctl try-restart middlewared
+
+# 6. Record what the restart landed on -- but do NOT restart again on a miss.
+#
+# `try-restart` returns as soon as middlewared is READY; it then brings docker
+# up asynchronously, and `docker.configure_nvidia` merges the stock nvidia
+# sysext over /usr at that point. That detaches our overlay AFTER the new
+# middlewared has already imported the patched modules -- so a disk check here
+# can report "missing" on a perfectly healthy system. Restarting on that signal
+# would restart a correctly-patched middlewared and then hit the same race
+# again, so the disk is deliberately not treated as a verdict after the restart.
+#
+# The authoritative answer is whether the running process holds the patch, and
+# only middlewared can answer that. The alert source installed by apply.sh
+# checks exactly that, in-process and hourly, and is what reports a genuine
+# miss. What is still worth doing here is putting the overlay back, so the next
+# middlewared restart -- whenever and whyever it happens -- finds patched files.
+if _patch_visible; then
+    _log "OK: providers patch present on the live path across the restart"
+else
+    _log "overlay detached again after the restart (expected when docker's"
+    _log "nvidia sysext merge follows it) -- re-mounting for the next restart."
+    _log "Whether THIS middlewared loaded the patch is answered in-process by"
+    _log "the 'installed but NOT loaded' alert, not by this check."
+
+    # Preserve hook_status.json's patched_at across this re-mount.
+    #
+    # create_task.py verify decides "loaded" by comparing middlewared's start
+    # time against patched_at. This re-apply restores the SAME patch the boot
+    # pass already applied, but it runs *after* the restart -- so letting it
+    # re-stamp would make patched_at newer than the process that correctly
+    # imported the patch, and verify would report FAIL forever, on every boot
+    # where docker's sysext merge detaches the overlay. That is precisely the
+    # lying-status failure this release exists to remove, so do not introduce a
+    # new one. The snapshot lives in /run, never in the repo: a leftover file
+    # there would leave the tree dirty and update.sh refuses to run over that.
+    _saved_status=/run/truecloud-hook_status.pre
+    cp -p "$PATCH_DIR/hook_status.json" "$_saved_status" 2>/dev/null
+
+    TRUECLOUD_REAPPLY=1 /bin/bash "$PATCH_DIR/patch/apply.sh"
+
+    if [ -f "$_saved_status" ]; then
+        mv -f "$_saved_status" "$PATCH_DIR/hook_status.json" 2>/dev/null
+    fi
+fi
+
+_log "=== deferred restart complete ==="
@@ -17,7 +17,7 @@
 #   bash /mnt/tank/truenas-truecloud-patch/patch/apply.sh
 #   systemctl restart middlewared

-VERSION="0.6.0"
+VERSION="0.8.0"

 PATCH_DIR="$(cd "$(dirname "$0")" && pwd)"

@@ -140,19 +140,60 @@ class TestSnapshotLeak:
    every path that creates one must also sweep the whole tree.
    """

-    def test_staging_failure_deletes_the_snapshot_tree(self):
-        # On a staging failure, sync.py's `snapshot, local_path = await
-        # create_snapshot(...)` never completes, so its local `snapshot` stays
-        # None and its finally deletes nothing. We must sweep it ourselves.
-        block = extract_blocks()["SNAPSHOT_ASYNC"]
-        assert "except Exception:" in block
-        assert "delete_snapshot_tree" in block
-        assert "raise" in block
+    # The behaviour these once asserted as substrings -- the sweep, the re-raise, the
+    # teardown in the finally -- is now asserted STRUCTURALLY, against the parsed
+    # block: see TestTheStagingFailurePathReallyReRaises and
+    # TestTheSyncBlockAlwaysTearsDown. As substring checks they were satisfied by
+    # COMMENTS ("a cleanup that raises...", "cleanup_task gets logger=None"), so
+    # deleting the actual `raise` and the actual cleanup call both left the suite
+    # green -- reinstating a silently-empty backup and ~250 orphans per run.

-    def test_sync_block_cleans_up_on_every_path(self):
-        block = extract_blocks()["SYNC_ASYNC"]
-        assert "finally:" in block
-        assert "cleanup_task" in block
+    def test_the_snapshot_block_still_owns_the_snapshot_when_not_staging(self):
+        # The TrueNAS 26 zvol/legacy orphan: stock decides `recursive` by its own rule
+        # (path == mountpoint) and deletes only the parent, so we must record the
+        # snapshot even on the path where we stage nothing.
+        for name in ("SNAPSHOT_ASYNC", "SNAPSHOT_SYNC"):
+            stage = functions(tree_of(name), "_tc_stage")[0]
+            assert calls_to(stage, "_tc_nested.own_snapshot"), (
+                f"{name} hands an unstaged snapshot back to stock, whose delete is "
+                f"non-recursive -- every zvol/legacy child is orphaned, every run"
+            )
+
+    def test_the_staging_plan_is_enumerated_from_ZFS(self):
+        for name in ("SNAPSHOT_ASYNC", "SNAPSHOT_SYNC"):
+            stage = functions(tree_of(name), "_tc_stage")[0]
+            assert calls_to(stage, "_tc_nested.query_filesystems"), (
+                "the staging plan must come from query_filesystems() (which reads ZFS "
+                "unfiltered); middleware's query hides ix-apps/*, .system/*, .ix-virt/*"
+            )
+            assert not calls_to(stage, "middleware.call_sync"), (
+                "the block calls middleware directly again -- its dataset/snapshot "
+                "queries are FILTERED and silently omit 84 of 270 datasets"
+            )
+
+    def test_the_vendored_helper_is_used_not_the_host_module(self):
+        # TrueNAS 26 DELETED get_dataset_recursive from plugins/cloud/snapshot.py, so
+        # calling it out of the host module's namespace is a NameError there.
+        for name in ("SNAPSHOT_ASYNC", "SNAPSHOT_SYNC"):
+            stage = functions(tree_of(name), "_tc_stage")[0]
+            assert calls_to(stage, "_tc_nested.get_dataset_recursive"), (
+                "must call OUR vendored copy: TrueNAS 26 deleted the host's"
+            )
+
+    def test_datasets_are_enumerated_AFTER_the_snapshot(self):
+        # A dataset created between the listing and the snapshot would be captured by
+        # the recursive snapshot but missing from the staging plan -- silently omitted.
+        # Read afterwards, it instead trips plan_staging's probe and fails loudly.
+        for name in ("SNAPSHOT_ASYNC", "SNAPSHOT_SYNC"):
+            src = extract_blocks()[name]
+            code = "\n".join(
+                ln for ln in src.splitlines() if not ln.lstrip().startswith("#")
+            )
+            # _tc_stage receives `snapshot` as a parameter -- i.e. it is taken by the
+            # caller, before any of this runs. If the enumeration ever moves ahead of
+            # create_snapshot it can only do so by leaving _tc_stage.
+            assert "def _tc_stage(middleware, path, name, snapshot, snap_path)" in code
+            assert "query_filesystems" in code


 def test_crud_block_is_scoped_to_cloud_backup():
@@ -438,8 +479,275 @@ class TestOnlyOurOwnTasksAreTouched:
        # The point is to add NO new failure mode to a CloudSync task. If any
        # middleware call happened before the bail-out, we would already have broken
        # the thing we are trying not to touch.
+        #
+        # Checked against whichever interactions the block ACTUALLY contains, not a
+        # fixed list: the dataset query moved behind `_tc_nested.query_filesystems()`
+        # when it switched to the public pool.* API, and a hardcoded
+        # `middleware.call_sync(` simply stopped being found -- a test that silently
+        # stops testing is worse than no test.
        block = extract_blocks()[name]
        gate = block.index('if not name.startswith("cloud_backup"):')
-        for call in ("middleware.call_sync(", "_tc_nested.stage_nested(",
-                     "_tc_nested.delete_snapshot_tree("):
+
+        interactions = [
+            "middleware.call_sync(",
+            "_tc_nested.query_filesystems(",
+            "_tc_nested.stage_nested(",
+            "_tc_nested.delete_snapshot_tree(",
+        ]
+        present = [c for c in interactions if c in block]
+        assert present, "found no middleware interaction at all -- the test is vacuous"
+        for call in present:
            assert gate < block.index(call), f"{call} runs before the cloud_backup gate"
+
+
+# ── structural assertions ────────────────────────────────────────────────────
+#
+# `assert "raise" in block` was TRUE because a COMMENT in the block says "a cleanup
+# that raises would replace the original exception". `assert "cleanup_task" in block`
+# was TRUE because a comment says "cleanup_task gets logger=None". Deleting the actual
+# `raise`, and deleting the actual cleanup call from the `finally`, both left the suite
+# green -- while reinstating, respectively, a silently-empty backup and ~250 orphaned
+# snapshots per run.
+#
+# A test that a comment can satisfy is not a test. These parse the block and assert on
+# the CODE.
+
+def tree_of(name):
+    return ast.parse(textwrap.dedent(extract_blocks()[name]))
+
+
+def functions(tree, name):
+    return [
+        n for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef | ast.AsyncFunctionDef) and n.name == name
+    ]
+
+
+def calls_to(node, dotted):
+    """Every Call in `node` whose callee renders as `dotted` (e.g. a.b.c)."""
+    out = []
+    for n in ast.walk(node):
+        if isinstance(n, ast.Call):
+            try:
+                if ast.unparse(n.func) == dotted:
+                    out.append(n)
+            except Exception:                                    # noqa: BLE001
+                pass
+    return out
+
+
+class TestTheStagingFailurePathReallyReRaises:
+    """If staging fails and we swallow it, restic backs up the UN-STAGED path.
+
+    That is the silently-empty backup this entire module exists to prevent: stock
+    points the tool at the parent's `.zfs/snapshot/`, where child datasets are
+    invisible. The exception MUST propagate.
+    """
+
+    @pytest.mark.parametrize("name", ["SNAPSHOT_ASYNC", "SNAPSHOT_SYNC"])
+    def test_the_handler_sweeps_the_snapshot_and_re_raises(self, name):
+        stage = functions(tree_of(name), "_tc_stage")
+        assert stage, "_tc_stage is gone"
+
+        handlers = [
+            h for t in ast.walk(stage[0]) if isinstance(t, ast.Try)
+            for h in t.handlers
+        ]
+        assert handlers, "the staging failure handler is gone"
+
+        sweeps = any(calls_to(h, "_tc_nested.delete_snapshot_tree") for h in handlers)
+        assert sweeps, (
+            "a staging failure no longer sweeps the snapshot. sync.py's `snapshot` "
+            "local stays None, so ITS finally deletes nothing -- the whole tree leaks "
+            "on every failed run."
+        )
+
+        # A bare `raise` directly in the handler body -- not one nested inside the
+        # defensive try/except that wraps the sweep.
+        reraises = any(
+            any(isinstance(s, ast.Raise) and s.exc is None for s in h.body)
+            for h in handlers
+        )
+        assert reraises, (
+            "the staging failure is SWALLOWED. restic then runs against the un-staged "
+            "path and uploads a near-empty tree, reporting SUCCESS."
+        )
+
+
+class TestTheSyncBlockAlwaysTearsDown:
+    """The teardown is what unmounts the staging tree and sweeps the snapshot.
+
+    It must run on EVERY exit from restic_backup -- success, failure, or exception --
+    or the bind mounts pin the snapshot and the tree is orphaned.
+    """
+
+    @pytest.mark.parametrize("name", ["SYNC_ASYNC", "SYNC_SYNC"])
+    def test_cleanup_runs_in_a_finally(self, name):
+        fns = functions(tree_of(name), "restic_backup")
+        assert fns, "the restic_backup wrapper is gone"
+
+        tries = [t for t in ast.walk(fns[0]) if isinstance(t, ast.Try) and t.finalbody]
+        assert tries, "restic_backup no longer has a try/finally"
+
+        cleans = any(
+            "cleanup_task" in ast.unparse(stmt)
+            for t in tries for stmt in t.finalbody
+        )
+        assert cleans, (
+            "cleanup_task is not called in the finally. The staging tree is never torn "
+            "down, its bind mounts pin the snapshot, and ~250 snapshots leak per run."
+        )
+
+
+class TestTheBlockingWorkNeverRunsOnTheEventLoop:
+    """`zfs list` and `call_sync` are BLOCKING. On <=25.10 these blocks are async.
+
+    Running them directly on middlewared's event loop stalls the whole daemon.
+    """
+
+    @pytest.mark.parametrize("name,fn", [
+        ("SNAPSHOT_ASYNC", "create_snapshot"),
+        ("SYNC_ASYNC", "restic_backup"),
+    ])
+    def test_the_async_flavour_hops_to_a_thread(self, name, fn):
+        fns = functions(tree_of(name), fn)
+        assert fns and isinstance(fns[0], ast.AsyncFunctionDef)
+        assert calls_to(fns[0], "middleware.run_in_thread"), (
+            f"{name}.{fn} does the blocking work on the asyncio event loop"
+        )
+
+    @pytest.mark.parametrize("name,fn", [
+        ("SNAPSHOT_SYNC", "create_snapshot"),
+        ("SYNC_SYNC", "restic_backup"),
+    ])
+    def test_the_sync_flavour_does_not(self, name, fn):
+        # On 26 stock already runs this in the thread pool; hopping again would be
+        # wrong (and there is no event loop to protect).
+        fns = functions(tree_of(name), fn)
+        assert fns and isinstance(fns[0], ast.FunctionDef)
+        assert not calls_to(fns[0], "middleware.run_in_thread")
+
+
+def test_the_flavour_mapping_is_not_inverted():
+    # `_snapshot_block = SNAPSHOT_ASYNC if _flavour else SNAPSHOT_SYNC` -- inverting it
+    # injects an async wrapper on 26 (a coroutine gets unpacked as a tuple) or a sync
+    # one on 25.10 (the event loop blocks). Every nested backup breaks, both ways.
+    with open(APPLY_SH, encoding="utf-8") as fh:
+        code = " ".join(
+            ln for ln in fh.read().splitlines() if not ln.lstrip().startswith("#")
+        )
+    code = re.sub(r"\s+", " ", code)          # the assignments are space-aligned
+    for block in ("SNAPSHOT", "CRUD", "SYNC"):
+        assert f"{block}_ASYNC if _flavour else {block}_SYNC" in code, (
+            f"the {block} flavour mapping is missing or inverted: _flavour is True for "
+            f"an ASYNC middleware, so it must select {block}_ASYNC"
+        )
+
+
+# ── the compat preflight ─────────────────────────────────────────────────────
+#
+# This is the guard that stands between a broken middleware and a live NAS: at every
+# boot, apply.sh checks the patch's assumptions against the middlewared actually
+# installed, and REFUSES to apply a module whose assumptions no longer hold.
+#
+# It had no test. An audit turned it into a no-op eight different ways -- `verdict()`
+# always returning 'ok', the broken branch never firing, the kill switch never honoured
+# -- and the suite stayed green every time. The most consequential safety net in the
+# project was unguarded.
+
+def preflight_heredoc():
+    """The preflight's Python, lifted out of apply.sh and made runnable.
+
+    Extracted, not reimplemented: a reimplementation would happily pass while the
+    SHIPPED preflight stayed broken, which is exactly the failure being guarded.
+    """
+    with open(APPLY_SH, encoding="utf-8") as fh:
+        sh = fh.read()
+    # Line-based: the compat heredoc opens with `<<'PYEOF'` on the _tc_compat line and
+    # closes at the next bare PYEOF. (A regex that matched `<< 'PYEOF'` silently found
+    # the OTHER heredoc and ran a different script entirely.)
+    lines = sh.splitlines()
+    start = next(
+        i for i, ln in enumerate(lines)
+        if ln.startswith("_tc_compat=$(") and "<<'PYEOF'" in ln
+    )
+    end = next(i for i in range(start + 1, len(lines)) if lines[i].strip() == "PYEOF")
+    m = "\n".join(lines[start + 1:end])
+    assert m, "could not find the compat preflight heredoc in apply.sh"
+    return m
+
+
+def run_preflight(result, tmp_path):
+    """Run the SHIPPED preflight against a fake compat.check_tree result.
+
+    The heredoc does `import sys`, so a fake `sys` in the namespace is immediately
+    rebound to the real module -- drive the real one instead.
+    """
+    import contextlib
+    import io
+    import sys
+    import types
+
+    src = preflight_heredoc()
+    fake = types.ModuleType("compat")
+    fake.check_tree = lambda _mw: result
+
+    saved_mod = sys.modules.get("compat")
+    saved_argv = sys.argv
+    sys.modules["compat"] = fake
+    sys.argv = ["x", "/patch", "/mw", str(tmp_path / "compat.json")]
+
+    buf = io.StringIO()
+    try:
+        with contextlib.redirect_stdout(buf):
+            exec(compile(src, "apply.sh:preflight", "exec"), {"__name__": "__main__"})  # noqa: S102
+    except SystemExit:
+        pass
+    finally:
+        sys.argv = saved_argv
+        if saved_mod is not None:
+            sys.modules["compat"] = saved_mod
+        else:
+            sys.modules.pop("compat", None)
+    return buf.getvalue().splitlines()
+
+
+def _mod(ok=True, native=False, unknown=False, problems=()):
+    return {"ok": ok, "native": native, "unknown": unknown, "problems": list(problems)}
+
+
+class TestTheBootPreflightRefusesABrokenMiddleware:
+    def test_a_healthy_tree_is_ok(self, tmp_path):
+        out = run_preflight({"providers": _mod(), "nested": _mod()}, tmp_path)
+        assert out[:2] == ["ok", "ok"]
+
+    def test_a_broken_module_is_reported_broken(self, tmp_path):
+        out = run_preflight({
+            "providers": _mod(),
+            "nested": _mod(ok=False, problems=[
+                {"id": "x", "detail": "gone", "why": "orphans every run"},
+            ]),
+        }, tmp_path)
+        assert "broken" in out, (
+            "the preflight did not report a module whose assumptions FAILED. It would "
+            "be injected into a middleware it does not fit -- broken backups, "
+            "discovered at restore time."
+        )
+
+    def test_a_module_that_went_NATIVE_is_also_not_applied(self, tmp_path):
+        # 'native' answers "do we still need it?", 'ok' answers "is it safe to inject?".
+        # Applying a module TrueNAS now implements itself is not safe either.
+        out = run_preflight({
+            "providers": _mod(),
+            "nested": _mod(ok=False, native=True),
+        }, tmp_path)
+        assert "broken" in out
+
+    def test_an_UNKNOWN_verdict_is_not_reported_as_broken(self, tmp_path):
+        # A network error or an unreadable file is not iX deleting our symbols. Calling
+        # it broken would switch a working module off on a healthy box.
+        out = run_preflight({
+            "providers": _mod(unknown=True),
+            "nested": _mod(unknown=True),
+        }, tmp_path)
+        assert "broken" not in out
@@ -15,6 +15,7 @@ dangerous thing this file can say -- it means "TrueNAS does this now, retire the
 module" -- and it rests on nothing more than a substring match.
 """

+import json
 import os
 import sys

@@ -23,6 +24,7 @@ import pytest
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "tools"))

 import compat  # noqa: E402
+import compat_publish  # noqa: E402
 from compat import (  # noqa: E402
    NESTED,
    PROVIDERS,
@@ -51,24 +53,44 @@ GOOD = {
        "async def restic_backup(middleware, job, cloud_backup, dry_run=False, "
        "rate_limit=None):\n    pass\n"
    ),
-    # The middlewared METHODS the injected code calls. TrueNAS 26 deleted both of
-    # these files, taking zfs.dataset.query / zfs.snapshot.query / zfs.snapshot.delete
-    # with them -- see TestMiddlewareMethodsWeCall.
-    "plugins/zfs_/dataset.py": (
-        "class ZFSDataset(CRUDService):\n"
+    # The middlewared METHODS the injected code calls. These now go through the
+    # PUBLIC pool.* API: TrueNAS 26 deleted plugins/zfs_/ outright, taking the whole
+    # private zfs.* service with it -- see TestMiddlewareMethodsWeCall.
+    #
+    # This default tree is a MODERN box (25.10/26): it has pool.snapshot and no
+    # zfs.snapshot. The older shape is built explicitly where it is tested.
+    # Not a plugin: a method on the middleware OBJECT. `snapshot_service()` resolves
+    # the snapshot namespace through it, so if it vanishes the module cannot sweep the
+    # snapshot it just took.
+    "utils/plugins.py": (
+        "class LoadPluginsMixin:\n"
+        "    def get_service(self, name):\n        pass\n"
+    ),
+    "plugins/pool_/dataset.py": (
+        "class PoolDatasetService(CRUDService):\n"
        "    class Config:\n"
-        "        namespace = 'zfs.dataset'\n"
+        "        namespace = 'pool.dataset'\n"
        "    def query(self, filters, options):\n        pass\n"
    ),
-    "plugins/zfs_/snapshot.py": (
-        "class ZFSSnapshot(CRUDService):\n"
+    "plugins/pool_/snapshot.py": (
+        "class PoolSnapshotService(CRUDService):\n"
        "    class Config:\n"
-        "        namespace = 'zfs.snapshot'\n"
+        "        namespace = 'pool.snapshot'\n"
        "    def query(self, filters, options):\n        pass\n"
        "    def delete(self, id_, options={}):\n        pass\n"
    ),
 }

+#: A 24.10/25.04 box: `pool.snapshot` does not exist yet and the snapshot CRUD
+#: service still answers to the (then-public) `zfs.snapshot`.
+ZFS_ERA_SNAPSHOT = (
+    "class ZFSSnapshot(CRUDService):\n"
+    "    class Config:\n"
+    "        namespace = 'zfs.snapshot'\n"
+    "    def query(self, filters, options):\n        pass\n"
+    "    def do_delete(self, id_, options={}):\n        pass\n"
+)
+

 def loader(files):
    def load(path):
@@ -284,9 +306,23 @@ class TestAsyncFlavour:
        broken["plugins/cloud_backup/sync.py"] = Unreadable("HTTP 429")
        assert compat.async_flavour(loader(broken)) is None

-    def test_the_real_truenas_versions(self):
-        # Pinning the actual fact this whole port exists for.
-        assert compat.async_flavour(loader(GOOD)) is True
+    def test_it_reads_STOCK_source_not_our_own_injected_block(self):
+        # This was a byte-identical copy of test_async_middleware_is_detected under a
+        # name that promised more. The fact worth pinning: apply.sh re-runs on an
+        # ALREADY-PATCHED overlay, so the probe must cut our block off first -- our own
+        # SNAPSHOT_SYNC wrapper is a plain `def create_snapshot`, and reading it would
+        # report a 25.10 box as synchronous and inject the wrong flavour.
+        patched = dict(GOOD)
+        patched["plugins/cloud/snapshot.py"] = (
+            GOOD["plugins/cloud/snapshot.py"]
+            + "\n# TRUECLOUD_PATCH\n"
+            + 'def create_snapshot(middleware, path, name="x"):\n    return "s", "p"\n'
+        )
+        assert compat.async_flavour(loader(patched)) is True, (
+            "the flavour probe read our own injected block and concluded the box is "
+            "synchronous -- it would then inject a sync wrapper into an async "
+            "middleware, and every nested backup would break"
+        )


 class TestMiddlewareMethodsWeCall:
@@ -303,52 +339,86 @@ class TestMiddlewareMethodsWeCall:
    per descendant dataset (250 on a real pool) on every run, forever.
    """

-    ZFS_SNAPSHOT = (
-        "class ZFSSnapshot(CRUDService):\n"
-        "    class Config:\n"
-        "        namespace = 'zfs.snapshot'\n"
-        "    def query(self, filters, options):\n        pass\n"
-        "    def delete(self, id_, options={}):\n        pass\n"
-    )
-    ZFS_DATASET = (
-        "class ZFSDataset(CRUDService):\n"
-        "    class Config:\n"
-        "        namespace = 'zfs.dataset'\n"
-        "    def query(self, filters, options):\n        pass\n"
-    )
+    POOL_SNAPSHOT = GOOD["plugins/pool_/snapshot.py"]

    def _tree(self, **over):
        files = dict(GOOD)
-        files["plugins/zfs_/snapshot.py"] = self.ZFS_SNAPSHOT
-        files["plugins/zfs_/dataset.py"] = self.ZFS_DATASET
        files.update(over)
        return files

+    #: A 26 box: pool.snapshot only.
+    def _modern(self, **over):
+        return self._tree(**over)
+
+    #: A 24.10/25.04 box: zfs.snapshot only -- plugins/pool_/snapshot.py does not
+    #: exist yet.
+    def _zfs_era(self, **over):
+        return self._tree(**{
+            "plugins/pool_/snapshot.py": None,
+            "plugins/zfs_/snapshot.py": ZFS_ERA_SNAPSHOT,
+            **over,
+        })
+
    def test_present_methods_are_ok(self):
-        r = check_files(self._tree())
+        r = check_files(self._modern())
        assert r[NESTED]["ok"], r[NESTED]["problems"]

-    def test_a_deleted_plugin_file_is_broken(self):
-        # Literally TrueNAS 26: plugins/zfs_/snapshot.py does not exist.
-        r = check_files(self._tree(**{"plugins/zfs_/snapshot.py": None}))
+    def test_the_OLD_zfs_era_snapshot_service_also_satisfies_the_call(self):
+        # 24.10 and 25.04 have no `pool.snapshot` at all -- the CRUD service is the
+        # then-public `zfs.snapshot`. Pinning only the modern spelling marked both of
+        # those releases BROKEN and would have switched nested snapshots OFF on boxes
+        # where they work perfectly. The runtime picks the same way; see
+        # pick_snapshot_service().
+        r = check_files(self._zfs_era())
+        assert r[NESTED]["ok"], r[NESTED]["problems"]
+
+    def test_it_is_broken_only_when_NEITHER_namespace_exists(self):
+        # The real failure: middleware drops the last spelling we know how to call.
+        r = check_files(self._tree(**{
+            "plugins/pool_/snapshot.py": None,
+            "plugins/zfs_/snapshot.py": None,
+        }))
        assert is_broken(r[NESTED])
        details = " ".join(p["detail"] for p in r[NESTED]["problems"])
-        assert "zfs.snapshot.delete" in details
+        assert "pool.snapshot.delete" in details
+        assert "zfs.snapshot.delete" in details, (
+            "the report must say BOTH spellings were tried, or whoever reads it will "
+            "think we simply never looked for the one their box has"
+        )
+
+    def test_we_do_NOT_depend_on_a_middleware_dataset_query_at_all(self):
+        # iX could delete plugins/pool_/dataset.py tomorrow and the patch would not
+        # care, because the staging plan is enumerated from ZFS, not from middleware.
+        #
+        # That is deliberate, and it was expensive to learn. `pool.dataset.query`
+        # exists and is correctly shaped -- and it LIES: it applies a visibility
+        # policy that hides ix-apps/*, .system/* and .ix-virt/* (84 of 270 datasets
+        # on the real pool, including live app data). No source check could ever
+        # have caught that; only running it could. So there is no assumption here
+        # left to break.
+        r = check_files(self._modern(**{"plugins/pool_/dataset.py": None}))
+        assert r[NESTED]["ok"], r[NESTED]["problems"]
+
+        ids = {c.id for c in compat.MIDDLEWARE_CALLS}
+        assert not any("dataset" in i or "query" in i for i in ids), (
+            "a dataset/snapshot QUERY assumption crept back into the manifest -- "
+            "middleware's queries are filtered; enumerate from ZFS"
+        )

    def test_a_renamed_namespace_is_broken(self):
        r = check_files(self._tree(**{
-            "plugins/zfs_/snapshot.py": self.ZFS_SNAPSHOT.replace(
-                "'zfs.snapshot'", "'zfs.resource.snapshot'"),
+            "plugins/pool_/snapshot.py": self.POOL_SNAPSHOT.replace(
+                "'pool.snapshot'", "'zfs.resource.snapshot'"),
+            "plugins/zfs_/snapshot.py": None,
        }))
        assert is_broken(r[NESTED])

    def test_the_CRUDService_do_prefix_is_accepted(self):
-        # 24.10 and 25.04 declare `do_delete`; 25.10 renamed it to `delete`. BOTH
-        # answer to zfs.snapshot.delete. Accepting only the literal name reported the
-        # two older releases as broken -- a false BROKEN that would have switched off
-        # nested snapshots on boxes where they work perfectly.
-        r = check_files(self._tree(**{
-            "plugins/zfs_/snapshot.py": self.ZFS_SNAPSHOT.replace(
+        # A CRUDService exposes `delete` from a method NAMED `do_delete`. Both
+        # spellings are live across the matrix. Accepting only the literal name
+        # reported working releases as broken.
+        r = check_files(self._modern(**{
+            "plugins/pool_/snapshot.py": self.POOL_SNAPSHOT.replace(
                "def delete(", "def do_delete("),
        }))
        assert r[NESTED]["ok"], r[NESTED]["problems"]
@@ -356,6 +426,383 @@ class TestMiddlewareMethodsWeCall:
    def test_the_snapshot_delete_reason_names_the_orphan_risk(self):
        # If this ever regresses, whoever reads the bug report must understand that
        # it is not a cosmetic failure.
-        r = check_files(self._tree(**{"plugins/zfs_/snapshot.py": None}))
+        r = check_files(self._tree(**{
+            "plugins/pool_/snapshot.py": None,
+            "plugins/zfs_/snapshot.py": None,
+        }))
        whys = " ".join(p["why"] for p in r[NESTED]["problems"])
        assert "orphan" in whys
+
+
+class TestTheMethodCheckIsNotJustANamespaceCheck:
+    """compat must verify the METHOD, not merely that the namespace still exists.
+
+    Deleting the method check entirely used to leave all 304 tests green -- so the
+    "namespace AND method" claim was unenforced and silently revertible. It is the
+    half of the predicate that catches iX gutting a method while keeping its service,
+    which they have already done to `pool.snapshot.do_update` on master.
+    """
+
+    def test_a_namespace_that_no_longer_defines_delete_is_broken(self):
+        gutted = (
+            "class PoolSnapshotService(CRUDService):\n"
+            "    class Config:\n"
+            "        namespace = 'pool.snapshot'\n"
+            "    def query(self, filters, options):\n        pass\n"
+            # do_delete is GONE -- the service is still registered and still a
+            # CRUDService, so it still INHERITS a callable `delete`.
+        )
+        r = check_files(with_(**{
+            "plugins/pool_/snapshot.py": gutted,
+            "plugins/zfs_/snapshot.py": None,      # no fallback either
+        }))
+        assert is_broken(r[NESTED]), (
+            "a namespace with no delete must be BROKEN. Checking only that the "
+            "namespace exists would apply the patch to a box that cannot sweep its "
+            "own snapshots."
+        )
+
+    def test_the_alternative_still_saves_it_when_only_the_primary_is_gutted(self):
+        gutted = (
+            "class PoolSnapshotService(CRUDService):\n"
+            "    class Config:\n"
+            "        namespace = 'pool.snapshot'\n"
+            "    def query(self, filters, options):\n        pass\n"
+        )
+        r = check_files(with_(**{
+            "plugins/pool_/snapshot.py": gutted,
+            "plugins/zfs_/snapshot.py": ZFS_ERA_SNAPSHOT,
+        }))
+        assert r[NESTED]["ok"], r[NESTED]["problems"]
+
+
+class TestUnreadableIsNeverOkAndNeverBroken:
+    """A rate limit is not a regression, and it is not a clean bill of health either.
+
+    compat runs ~30 unauthenticated GitHub requests per matrix; 429 is a real outcome.
+    It also runs at BOOT against the installed tree, where a read can fail with EACCES.
+
+      * treating unreadable as BROKEN repaints the README, files a bug report, and
+        makes apply.sh refuse the module on a box where it works.
+      * treating it as OK injects a module whose delete may be gone.
+
+    Both mutations used to pass the whole suite.
+    """
+
+    def test_both_spellings_unreadable_is_unknown_not_broken(self):
+        r = check_files(with_(**{
+            "plugins/pool_/snapshot.py": Unreadable("HTTP 429"),
+            "plugins/zfs_/snapshot.py": Unreadable("HTTP 429"),
+        }))
+        assert not is_broken(r[NESTED]), "a 429 is not iX deleting the snapshot service"
+        assert r[NESTED]["unknown"]
+
+    def test_an_unreadable_primary_with_a_healthy_alternative_is_ok(self):
+        r = check_files(with_(**{
+            "plugins/pool_/snapshot.py": Unreadable("HTTP 429"),
+            "plugins/zfs_/snapshot.py": ZFS_ERA_SNAPSHOT,
+        }))
+        assert r[NESTED]["ok"], r[NESTED]["problems"]
+        assert not r[NESTED]["unknown"], (
+            "one spelling answered the question; the other's 429 is irrelevant"
+        )
+
+    def test_a_missing_primary_with_an_unreadable_alternative_is_unknown(self):
+        # We cannot tell whether the box is broken. Saying either would be a guess.
+        r = check_files(with_(**{
+            "plugins/pool_/snapshot.py": None,
+            "plugins/zfs_/snapshot.py": Unreadable("HTTP 429"),
+        }))
+        assert not is_broken(r[NESTED])
+        assert r[NESTED]["unknown"]
+
+
+class TestGetServiceIsChecked:
+    """The runtime resolves the snapshot namespace through `middleware.get_service`.
+
+    It is not a plugin method, so the manifest had no way to express it and never
+    checked it. If it vanishes, `_can_delete` reports BOTH namespaces unusable and
+    every nested backup fails -- on a box the preflight had declared healthy.
+    """
+
+    def test_a_middleware_without_get_service_is_broken(self):
+        r = check_files(with_(**{"utils/plugins.py": None}))
+        assert is_broken(r[NESTED])
+        details = " ".join(p["detail"] for p in r[NESTED]["problems"])
+        assert "get_service" in details
+
+
+class TestATransientNetworkBlipDoesNotWakeAnybody:
+    """The fingerprint must digest what iX BROKE, not what GitHub failed to serve.
+
+    `unknown` problems (a 429 on one of ~30 unauthenticated fetches, an EACCES at boot)
+    used to be folded into an already-broken module's problem list, so one blip flipped
+    the fingerprint, `compat_publish` rewrote the issue body, and the next clean run
+    rewrote it back. Daily churn is what teaches people to ignore the bot -- which is
+    the whole thing this fingerprint exists to prevent.
+    """
+
+    def _rows(self, files):
+        return [{"ref": "master", "modules": check_files(files)}]
+
+    def test_an_unreadable_file_does_not_change_the_fingerprint_of_a_broken_ref(self):
+        # The blip must land in the SAME module that is broken. Put it in `providers`
+        # (which is healthy) and `fingerprint()` skips the whole module via
+        # `is_broken(m)` -- so the `state` filter under test never runs and the test
+        # passes no matter what the code does. `nested` is the broken one here, so the
+        # unreadable file goes in `nested` too.
+        broken = with_(**{
+            "plugins/cloud/snapshot.py":
+                "async def create_snapshot(name, path, middleware):\n    return 1, 2\n",
+        })
+        clean = compat.fingerprint(self._rows(broken))
+
+        blipped = dict(broken)
+        blipped["plugins/cloud_backup/sync.py"] = Unreadable("HTTP 429")   # nested
+        assert compat.fingerprint(self._rows(blipped)) == clean, (
+            "a rate-limited fetch changed the fingerprint, so the bot rewrites the "
+            "issue body and then rewrites it back tomorrow"
+        )
+
+    def test_a_REAL_new_finding_still_changes_it(self):
+        # ...and the anti-noise measure must not have made it deaf.
+        broken = with_(**{
+            "plugins/cloud/snapshot.py":
+                "async def create_snapshot(name, path, middleware):\n    return 1, 2\n",
+        })
+        worse = dict(broken)
+        worse["plugins/cloud_backup/restic.py"] = (
+            "class ResticConfig:\n    cmd: list\n\n"
+            "def get_restic_config(entry, credentials):\n    pass\n"
+        )
+        assert compat.fingerprint(self._rows(worse)) != compat.fingerprint(self._rows(broken))
+
+
+class TestTheBotFindsItsOwnIssueOnBOTHForges:
+    """`find_issue` decides "have I already filed this?" -- and it ran on two forges.
+
+    It used to skip pull requests with `"pull_request" not in i`. GitHub omits that key
+    on a plain issue; **Gitea sends it as `null`**. So on Gitea every issue looked like
+    a PR, the match list was always empty, and the bot took the "nothing filed yet"
+    branch on EVERY run: nine duplicate copies of the same report on the canonical
+    forge, four of them filed after the commit that was supposed to stop exactly this.
+
+    It is the same failure the spam fix was written to prevent, moved from comments to
+    issues -- and it survived because `find_issue` was the one function here with no
+    test. So the payload shapes are pinned, per forge, by hand.
+    """
+
+    TITLE = compat_publish.TITLE
+
+    def _find(self, monkeypatch, payload):
+        monkeypatch.setattr(compat_publish, "_call", lambda *a, **k: payload)
+        return compat_publish.find_issue("https://forge/api", "tok", self.TITLE)
+
+    def test_gitea_sends_pull_request_as_null_and_the_issue_is_still_found(self, monkeypatch):
+        found = self._find(monkeypatch, [
+            {"number": 7, "title": self.TITLE, "state": "open", "pull_request": None},
+            {"number": 1, "title": self.TITLE, "state": "open", "pull_request": None},
+        ])
+        assert found is not None, (
+            "find_issue missed a Gitea issue, so the bot files a NEW duplicate report "
+            "every run -- which is how nine of them piled up"
+        )
+        assert found["number"] == 1, "lowest-numbered wins"
+
+    def test_github_omits_the_key_entirely_and_the_issue_is_still_found(self, monkeypatch):
+        found = self._find(monkeypatch, [
+            {"number": 2, "title": self.TITLE, "state": "open"},
+        ])
+        assert found is not None and found["number"] == 2
+
+    def test_a_real_PR_with_the_same_title_is_still_skipped_on_both(self, monkeypatch):
+        # The reason the filter exists at all: both forges list PRs on /issues, and
+        # commenting on a PR instead of the bug report would be worse than useless.
+        assert self._find(monkeypatch, [
+            {"number": 3, "title": self.TITLE, "state": "open",           # Gitea PR
+             "pull_request": {"merged": False}},
+            {"number": 4, "title": self.TITLE, "state": "open",           # GitHub PR
+             "pull_request": {"url": "https://api.github.com/..."}},
+        ]) is None
+
+    def test_an_unrelated_issue_is_not_mistaken_for_the_report(self, monkeypatch):
+        assert self._find(monkeypatch, [
+            {"number": 1, "title": "TypeError when create B2 backup on Electric Eel",
+             "state": "closed", "pull_request": None},
+        ]) is None
+
+
+class TestTheNextMaintenanceReleaseIsChecked:
+    """`release/25.10.5` fell through every sieve, and it is the one that reaches users.
+
+    Shipped versions come from `TS-*` TAGS; unreleased ones come from `release/*`
+    BRANCHES that carry `-BETA`/`-RC`. A branched-but-untagged MAINTENANCE release is
+    neither: no tag, and its line (25.10) has already shipped, so the "prereleases of
+    a shipped line are history" filter threw it out. It was invisible.
+
+    That is backwards. `release/24.10-RC.2` is history -- nobody can install it. But
+    `release/25.10.5` is the FUTURE of a shipped line: it is what a 25.10.4 box gets
+    on its next update. A break there ships to real users before the daily check has
+    ever looked at it.
+    """
+
+    TAGS = ["TS-24.10.2.4", "TS-25.04.2.6", "TS-25.10.4"]
+    HEADS = [
+        "release/25.10.4.1",
+        "release/25.10.5",          # branched, untagged -- the next maintenance release
+        "release/24.10-RC.2",       # history: its line shipped long ago
+        "release/25.20.2.2",        # iX's typo branch: 25.20 is not a TrueNAS version
+        "release/26.0.0-BETA.3",
+        "master",
+    ]
+
+    def _refs(self, monkeypatch):
+        monkeypatch.setattr(
+            compat, "_ls_remote",
+            lambda remote, what: self.TAGS if what == "--tags" else self.HEADS)
+        return compat.discover_refs("origin")
+
+    def test_the_next_maintenance_release_is_checked(self, monkeypatch):
+        assert "release/25.10.5" in self._refs(monkeypatch), (
+            "the next thing a 25.10.4 box updates to is not checked, so a break in it "
+            "reaches users before the bot ever sees it"
+        )
+
+    def test_a_superseded_maintenance_branch_is_not(self, monkeypatch):
+        # 25.10.4.1 sorts OLDER than the newest tag TS-25.10.4? No -- it is NEWER, and
+        # both are on the 25.10 line, so only the newest branch on the line is taken.
+        refs = self._refs(monkeypatch)
+        assert "release/25.10.4.1" not in refs, "only the newest branch per line"
+
+    def test_the_typo_branch_stays_out(self, monkeypatch):
+        # 25.20 has no TS tag, so it is not a release line at all. A typo branch in the
+        # matrix reads as a real supported release we are silently broken on.
+        assert "release/25.20.2.2" not in self._refs(monkeypatch)
+
+    def test_a_prerelease_of_an_already_shipped_line_stays_out(self, monkeypatch):
+        assert "release/24.10-RC.2" not in self._refs(monkeypatch)
+
+    def test_an_untagged_branch_counts_as_UNRELEASED(self, monkeypatch):
+        # The exit code keys off this. Calling 25.10.5 "shipped" would fail the build
+        # as a live outage on a version nobody is running yet.
+        assert compat.is_unreleased("release/25.10.5")
+        assert compat.is_unreleased("master")
+        assert not compat.is_unreleased("TS-25.10.4")
+
+
+class TestMasterIsNotTheNextRelease:
+    """A red `master` row used to read as "the version you are about to install".
+
+    On 2026-07-14 master was 27-dev -- every recent commit targeted 27.0.0-BETA.1 --
+    while 26 was still in beta on its own branches. So `master BROKEN` meant "iX will
+    break us a major release from now", but the matrix said "master _(unreleased)_",
+    which any reader takes as the next thing out the door. For a table whose whole job
+    is helping somebody decide whether to trust this with their backups, that is a
+    false alarm in the worst possible place.
+    """
+
+    def _rows(self, refs):
+        return [{"ref": r, "unreleased": compat.is_unreleased(r), "modules": {}}
+                for r in refs]
+
+    def test_master_is_labelled_with_the_major_AFTER_the_newest_known_one(self):
+        rows = self._rows(["TS-25.10.4", "release/26.0.0-BETA.3", "master"])
+        assert compat.dev_label(rows) == "27-dev"
+
+    def test_it_rolls_over_on_its_own_when_the_next_beta_branches(self):
+        # Derived, not hardcoded: when release/27.0.0-BETA.1 appears, master is 28-dev.
+        rows = self._rows(["TS-26.0.0", "release/27.0.0-BETA.1", "master"])
+        assert compat.dev_label(rows) == "28-dev"
+
+    def test_the_rendered_matrix_says_dev_not_unreleased(self):
+        healthy = check_files(with_())
+        rows = [
+            {"ref": r, "unreleased": compat.is_unreleased(r), "modules": healthy}
+            for r in ("TS-25.10.4", "release/26.0.0-BETA.3", "master")
+        ]
+        md = compat.render_markdown(rows)
+        assert "master _(27-dev)_" in md
+        assert "master _(unreleased)_" not in md
+        # ...and the ordinary rows are untouched.
+        assert "| 25.10.4 |" in md
+        assert "| 26.0.0-BETA.3 _(unreleased)_ |" in md
+
+
+class TestTheBodyIsTruthAndCommentsAreTheChangelog:
+    """An unchanged fingerprint used to freeze the BODY, not just silence the comments.
+
+    Two different questions were sharing one answer. "Have the findings changed?" gates
+    COMMENTS -- they notify, and a daily "still broken, same as yesterday" is what
+    teaches people to ignore the one that finally matters. But "is the body still
+    true?" gates the BODY, and editing a body notifies nobody, so keeping it honest
+    costs nothing.
+
+    Conflated, an unchanged fingerprint meant the report could never be corrected --
+    and the fingerprint deliberately ignores everything that moves on its own, which
+    includes how a row is LABELLED. Relabelling master `27-dev` would have reached the
+    README and never the issue anybody opens.
+    """
+
+    ROWS = [{"ref": "master", "unreleased": True,
+             "modules": check_files(with_(**{
+                 "plugins/cloud_backup/restic.py":
+                     "class ResticConfig:\n    cmd: list\n\n"
+                     "def get_restic_config(entry, credentials):\n    pass\n",
+             }))}]
+
+    def _run(self, monkeypatch, tmp_path, existing_body):
+        calls = []
+
+        def fake(url, token, method="GET", data=None):
+            calls.append((method, url, data))
+            if url.endswith("/issues?state=all&per_page=100&limit=100"):
+                return [{"number": 1, "title": compat_publish.TITLE,
+                         "state": "open", "body": existing_body,
+                         "pull_request": None}]
+            return {"number": 1}
+
+        monkeypatch.setattr(compat_publish, "_call", fake)
+        matrix = tmp_path / "m.json"
+        matrix.write_text(json.dumps(self.ROWS))
+        compat_publish.main([
+            "prog", "--api", "https://forge/api", "--token", "t",
+            "--matrix", str(matrix)])
+        return calls
+
+    def _writes(self, calls):
+        patched = [c for c in calls if c[0] == "PATCH"]
+        commented = [c for c in calls if c[0] == "POST" and c[1].endswith("/comments")]
+        return patched, commented
+
+    def test_identical_body_and_findings_touches_nothing(self, monkeypatch, tmp_path):
+        body = compat.render_issue(self.ROWS)
+        patched, commented = self._writes(self._run(monkeypatch, tmp_path, body))
+        assert not patched and not commented, "a quiet run must be completely silent"
+
+    def test_a_relabel_refreshes_the_body_but_says_NOTHING(self, monkeypatch, tmp_path):
+        # Same findings (same fingerprint), different rendering -- the exact shape of
+        # the master -> 27-dev relabel.
+        stale = compat.render_issue(self.ROWS).replace("27-dev", "unreleased")
+        assert compat.extract_fingerprint(stale) == compat.fingerprint(self.ROWS)
+
+        patched, commented = self._writes(self._run(monkeypatch, tmp_path, stale))
+        assert patched, "the body was left stale, so the issue keeps telling lies"
+        assert "27-dev" in patched[0][2]["body"]
+        assert not commented, (
+            "a rendering change is not news -- commenting on it is how the bot gets "
+            "muted before the next real finding"
+        )
+
+    def test_a_REAL_findings_change_still_comments(self, monkeypatch, tmp_path):
+        # ...and the fix must not have made it mute.
+        stale = compat.render_issue(self.ROWS).replace(
+            compat.fingerprint(self.ROWS), "0" * 16)
+        patched, commented = self._writes(self._run(monkeypatch, tmp_path, stale))
+        assert patched and commented, "a genuine change must still notify"
+
+    def test_a_body_differing_only_by_CRLF_is_not_rewritten(self, monkeypatch, tmp_path):
+        # Forges round-trip line endings. Without normalising, every run would rewrite
+        # the body -- silent, but it churns updated_at and looks freshly touched daily.
+        body = compat.render_issue(self.ROWS).replace("\n", "\r\n")
+        patched, _ = self._writes(self._run(monkeypatch, tmp_path, body))
+        assert not patched
@@ -94,3 +94,40 @@ class TestTheReadmeStaysAReadme:
        assert "24.10" in text[:text.index("## Install")], (
            "the minimum TrueNAS version must be visible above the install steps"
        )
+
+
+class TestInstallDoesNotDirtyTheCheckout:
+    """install.sh chmod +x's scripts. If git records them as 100644, that chmod is a
+    TRACKED MODIFICATION -- and update.sh refuses to run over a dirty tree.
+
+    So installing once permanently blocked updating, for every user, with a message
+    telling them to `git checkout -- .` (which would just undo the exec bit and let
+    the next install re-dirty it). Found on a real box that had been stuck on an old
+    version for exactly this reason.
+
+    Every script install.sh makes executable must already be executable in git.
+    """
+
+    def test_every_chmodded_script_is_already_executable_in_git(self):
+        import re
+        import subprocess
+
+        with open(os.path.join(ROOT, "install.sh"), encoding="utf-8") as fh:
+            m = re.search(r"^for _exe in (.+?); do", fh.read(), re.M)
+        assert m, "could not find install.sh's chmod loop"
+        scripts = m.group(1).split()
+
+        out = subprocess.run(
+            ["git", "ls-files", "-s", *scripts],
+            cwd=ROOT, capture_output=True, text=True, check=True,
+        ).stdout
+
+        not_exec = [
+            line.split("\t")[-1] for line in out.strip().splitlines()
+            if not line.startswith("100755")
+        ]
+        assert not not_exec, (
+            "install.sh chmod +x's these, but git records them as non-executable — "
+            "so installing dirties the checkout and update.sh then refuses to run:\n  "
+            + "\n  ".join(not_exec)
+        )
@@ -0,0 +1,210 @@
+"""Behavioural tests for the "installed but NOT loaded" alert.
+
+apply.log can only report what was written to disk. Whether the middlewared that
+restarted afterwards actually imported those files is a different fact, and when
+the two disagree nothing else notices: on 2026-08-19 every B2 backup failed for
+nineteen hours while the log said OK. This alert is the only thing that closes
+that gap, so it is tested against real objects rather than by reading source.
+
+The middlewared package does not exist off-box, so the modules the alert source
+imports are stubbed here.
+"""
+
+import importlib.util
+import json
+import os
+import sys
+import types
+
+import pytest
+
+ALERT_SRC = os.path.join(os.path.dirname(__file__), "..", "patch", "alert_source.py")
+
+
+class _StubAlertClass:
+    pass
+
+
+class _StubThreadedAlertSource:
+    pass
+
+
+class _StubAlert:
+    def __init__(self, klass, args=None, key=None):
+        self.klass = klass
+        self.args = args
+        self.key = key
+
+
+def _module(name):
+    mod = types.ModuleType(name)
+    sys.modules[name] = mod
+    return mod
+
+
+@pytest.fixture
+def alert_source(monkeypatch, tmp_path):
+    """Load patch/alert_source.py against stubbed middlewared modules."""
+    for name in list(sys.modules):
+        if name == "middlewared" or name.startswith("middlewared."):
+            monkeypatch.delitem(sys.modules, name, raising=False)
+
+    _module("middlewared")
+    _module("middlewared.alert")
+    base = _module("middlewared.alert.base")
+    base.Alert = _StubAlert
+    base.AlertClass = _StubAlertClass
+    base.ThreadedAlertSource = _StubThreadedAlertSource
+    base.AlertCategory = types.SimpleNamespace(SYSTEM="SYSTEM")
+    base.AlertLevel = types.SimpleNamespace(
+        INFO="INFO", WARNING="WARNING", CRITICAL="CRITICAL"
+    )
+    schedule = _module("middlewared.alert.schedule")
+    schedule.IntervalSchedule = lambda delta: ("interval", delta)
+
+    spec = importlib.util.spec_from_file_location("_tc_alert_source", ALERT_SRC)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    mod.PATCH_DIR = str(tmp_path)
+    return mod
+
+
+def _write_status(tmp_path, providers_active=True):
+    payload = {
+        "patched_at": "2026-08-26T00:00:00Z",
+        "patches": {
+            "providers": {"ok": True, "active": providers_active, "detail": "x"},
+            "nested_snapshots": {"ok": True, "active": True, "detail": "x"},
+        },
+    }
+    (tmp_path / "hook_status.json").write_text(json.dumps(payload))
+
+
+def _install_provider_modules(monkeypatch, *, restic_patched, b2_patched):
+    """Stub the two modules the alert inspects, in the requested state."""
+    plugins = _module("middlewared.plugins")
+    _module("middlewared.plugins.cloud_backup")
+    restic = _module("middlewared.plugins.cloud_backup.restic")
+
+    def get_restic_config(task):
+        return None
+
+    if restic_patched:
+        get_restic_config._truecloud_patched = True
+    restic.get_restic_config = get_restic_config
+
+    rclone_base = _module("middlewared.rclone.base")
+    _module("middlewared.rclone")
+    _module("middlewared.rclone.remote")
+    b2_mod = _module("middlewared.rclone.remote.b2")
+
+    class BaseRcloneRemote:
+        def get_restic_config(self, task):
+            raise NotImplementedError
+
+    class B2RcloneRemote(BaseRcloneRemote):
+        pass
+
+    if b2_patched:
+        B2RcloneRemote.get_restic_config = staticmethod(lambda task: ("url", {}))
+
+    rclone_base.BaseRcloneRemote = BaseRcloneRemote
+    b2_mod.B2RcloneRemote = B2RcloneRemote
+    b2_mod.BaseRcloneRemote = BaseRcloneRemote
+    plugins.__path__ = []
+
+    for name in (
+        "middlewared.plugins",
+        "middlewared.plugins.cloud_backup",
+        "middlewared.plugins.cloud_backup.restic",
+        "middlewared.rclone",
+        "middlewared.rclone.base",
+        "middlewared.rclone.remote",
+        "middlewared.rclone.remote.b2",
+    ):
+        monkeypatch.setitem(sys.modules, name, sys.modules[name])
+
+
+def _source(alert_source):
+    cls = alert_source.TrueCloudPatchNotLoadedAlertSource
+    return cls.__new__(cls)
+
+
+def test_no_alert_when_patch_is_loaded(alert_source, monkeypatch, tmp_path):
+    _write_status(tmp_path)
+    _install_provider_modules(monkeypatch, restic_patched=True, b2_patched=True)
+    assert _source(alert_source)._check() is None
+
+
+def test_alert_when_middlewared_loaded_stock_modules(alert_source, monkeypatch, tmp_path):
+    """The exact 2026-08-19 state: patched on disk, stock in the process."""
+    _write_status(tmp_path)
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=False)
+    alert = _source(alert_source)._check()
+    assert alert is not None
+    assert alert.klass is alert_source.TrueCloudPatchNotLoadedAlertClass
+
+
+def test_alert_when_only_b2_half_is_missing(alert_source, monkeypatch, tmp_path):
+    # b2.py is the half that supplies B2's get_restic_config. restic.py alone
+    # being patched still means every B2 task raises NotImplementedError.
+    _write_status(tmp_path)
+    _install_provider_modules(monkeypatch, restic_patched=True, b2_patched=False)
+    assert _source(alert_source)._check() is not None
+
+
+def test_alert_when_only_restic_half_is_missing(alert_source, monkeypatch, tmp_path):
+    _write_status(tmp_path)
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=True)
+    assert _source(alert_source)._check() is not None
+
+
+def test_silent_when_the_kill_switch_is_set(alert_source, monkeypatch, tmp_path):
+    # The operator turned the patch off on purpose; stock is the intended state.
+    _write_status(tmp_path)
+    (tmp_path / "disabled").write_text("")
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=False)
+    assert _source(alert_source)._check() is None
+
+
+def test_silent_when_providers_module_is_retired(alert_source, monkeypatch, tmp_path):
+    # TrueNAS went native for B2: not loading our providers patch is correct.
+    _write_status(tmp_path, providers_active=False)
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=False)
+    assert _source(alert_source)._check() is None
+
+
+def test_silent_when_the_patch_was_never_applied_here(alert_source, monkeypatch, tmp_path):
+    # No hook_status.json at all -- nothing claims a patch, so nothing is broken.
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=False)
+    assert _source(alert_source)._check() is None
+
+
+def test_update_alert_silencer_does_not_mute_a_broken_backup_path(
+    alert_source, monkeypatch, tmp_path
+):
+    # update_alerts_disabled mutes release notifications. It must not hide the
+    # fact that TrueCloud backups are silently running stock.
+    _write_status(tmp_path)
+    (tmp_path / "update_alerts_disabled").write_text("")
+    _install_provider_modules(monkeypatch, restic_patched=False, b2_patched=False)
+    assert _source(alert_source)._check() is not None
+
+
+def test_check_sync_never_raises(alert_source, monkeypatch, tmp_path):
+    """An alert source that raises is polled forever inside middlewared."""
+    _write_status(tmp_path)
+
+    def boom(self):
+        raise RuntimeError("provider import exploded")
+
+    monkeypatch.setattr(
+        alert_source.TrueCloudPatchNotLoadedAlertSource, "_check", boom, raising=True
+    )
+    assert _source(alert_source).check_sync() is None
+
+
+def test_alert_is_critical_and_names_the_recovery_command(alert_source):
+    klass = alert_source.TrueCloudPatchNotLoadedAlertClass
+    assert klass.level == "CRITICAL"
+    assert "install.sh" in klass.text
@@ -9,6 +9,7 @@ across install.sh / uninstall.sh / recover.sh / apply.sh and nothing noticed.
 """

 import os
+import re
 import sys

 import pytest
@@ -205,3 +206,88 @@ class TestSignificance:
        assert version_tuple("0.4.2") > version_tuple("0.4.1")
        # Pre-release suffixes are dropped, not ranked above the release.
        assert version_tuple("v0.5.0-rc1") == version_tuple("v0.5.0")
+
+
+class TestCandidateNotesResolveToTheBaseVersion:
+    """`notes v0.6.0-rc1` must return v0.6.0's section.
+
+    A candidate ships the same code as the release it is a candidate for, and the
+    CHANGELOG only ever has the one section. Without this, the release workflow cut
+    v0.6.0-rc1, passed every gate, and then died extracting the body -- so the tag
+    existed but nothing was ever published. Caught in an rc, which is the entire
+    point of having them.
+    """
+
+    CHANGELOG = "# C\n\n## v0.6.0 — 2026-07-13\n\n### Added\n- the thing\n\n## v0.5.1 — 2026-07-13\n\n- older\n"
+
+    def test_an_rc_resolves_to_its_base_version(self):
+        body = extract_notes(self.CHANGELOG, "v0.6.0-rc1")
+        assert "the thing" in body
+        assert "older" not in body
+
+    def test_rc10_too(self):
+        assert "the thing" in extract_notes(self.CHANGELOG, "v0.6.0-rc10")
+
+    def test_the_plain_version_still_works(self):
+        assert "the thing" in extract_notes(self.CHANGELOG, "v0.6.0")
+
+    def test_a_genuinely_missing_section_still_raises(self):
+        with pytest.raises(KeyError):
+            extract_notes(self.CHANGELOG, "v9.9.9-rc1")
+
+
+class TestTheChangelogIsStructurallySound:
+    """The release body IS this file, so a mangled section ships to every user.
+
+    It has been mangled once: an edit matched the literal `## Unreleased` inside a
+    backticked phrase in a prose bullet and spliced a whole new section into the middle
+    of it, splitting the sentence in half.
+    """
+
+    def changelog(self):
+        with open(os.path.join(REPO, "CHANGELOG.md"), encoding="utf-8") as fh:
+            return fh.read()
+
+    def test_no_version_section_is_empty(self):
+        text = self.changelog()
+        for v in changelog_versions(text):
+            assert extract_notes(text, v).strip(), f"v{v} has an empty section"
+
+    def test_versions_are_in_descending_order(self):
+        from release_notes import version_tuple
+        versions = changelog_versions(self.changelog())
+        assert versions == sorted(versions, key=version_tuple, reverse=True), (
+            "CHANGELOG versions are out of order — a section was spliced in wrong"
+        )
+
+    def test_headings_are_at_the_start_of_a_line_and_not_inside_prose(self):
+        # A `### Fixed` that ends up indented under a bullet is a section nobody sees.
+        for i, line in enumerate(self.changelog().splitlines(), 1):
+            if line.lstrip().startswith(("## ", "### ")) and line != line.lstrip():
+                raise AssertionError(
+                    f"line {i}: heading is indented, so it is inside a list item "
+                    f"rather than being a section: {line!r}"
+                )
+
+    def test_every_bullet_that_opens_a_bold_phrase_closes_it(self):
+        # The splice cut `- **A stable release ... under \`## Unreleased` in half,
+        # leaving an unterminated ** and a dangling sentence.
+        #
+        # A bullet is the `- ` line plus everything up to the next top-level bullet or
+        # heading -- bold phrases routinely wrap across lines, so a per-line check
+        # would flag every long bullet in the file.
+        text = self.changelog()
+        bullets = re.split(r"^(?=- |#{2,3} )", text, flags=re.M)
+        bad = []
+        for b in bullets:
+            if not b.startswith("- "):
+                continue
+            # Code spans are not markup: `*args, **kwargs` is a literal, not a bold
+            # phrase, and counting its ** would flag a perfectly well-formed bullet.
+            prose = re.sub(r"`[^`]*`", "", b)
+            if prose.count("**") % 2:
+                bad.append(b.splitlines()[0][:70])
+        assert not bad, (
+            "unbalanced ** in a bullet — a section was probably spliced into the "
+            "middle of it:\n  " + "\n  ".join(bad)
+        )
@@ -0,0 +1,199 @@
+"""The deferred restart must re-apply the patch before it restarts middlewared.
+
+Patching at PREINIT and restarting minutes later is only sound while the patched
+files are still on the live path when middlewared re-imports them. They may not
+be: the patch lives in an overlay mounted inside /usr, and anything that
+remounts that hierarchy detaches it. On 2026-08-19 a systemd-sysext refresh over
+/usr ran four seconds after apply.sh mounted its overlay; the deferred restart
+then loaded stock modules and every B2 cloud_backup job failed for nineteen
+hours while apply.log reported "OK".
+
+These tests pin the ordering that makes that non-recoverable failure impossible:
+re-apply, verify, restart, verify again.
+"""
+
+import os
+import re
+import subprocess
+
+import pytest
+
+HERE = os.path.dirname(__file__)
+WAIT_RESTART = os.path.join(HERE, "..", "patch", "wait_restart.sh")
+APPLY_SH = os.path.join(HERE, "..", "patch", "apply.sh")
+
+
+def wait_restart_source():
+    with open(WAIT_RESTART, encoding="utf-8") as fh:
+        return fh.read()
+
+
+def apply_source():
+    with open(APPLY_SH, encoding="utf-8") as fh:
+        return fh.read()
+
+
+def test_wait_restart_is_executable():
+    # apply.sh schedules it as `/bin/bash <script>`, but install.sh ships exec
+    # bits and a mode-only diff once blocked update.sh outright (v0.6.0).
+    assert os.access(WAIT_RESTART, os.X_OK)
+
+
+def test_wait_restart_is_syntactically_valid():
+    subprocess.run(["bash", "-n", WAIT_RESTART], check=True)
+
+
+def test_reapply_runs_before_the_restart():
+    src = wait_restart_source()
+    reapply = src.index("TRUECLOUD_REAPPLY=1")
+    restart = src.index("systemctl try-restart middlewared")
+    assert reapply < restart, "the re-apply pass must precede the restart"
+
+
+def test_restart_is_not_exec_so_verification_can_follow():
+    # Up to v0.7.0 the script ended in `exec systemctl try-restart middlewared`,
+    # which replaces the shell -- nothing could run afterwards. The post-restart
+    # verification only exists if the restart is a plain call.
+    src = wait_restart_source()
+    assert not re.search(r"^\s*exec\s+systemctl", src, re.M)
+
+
+def test_middlewared_is_restarted_exactly_once():
+    """No restart loop.
+
+    `try-restart` returns at READY; middlewared then brings docker up, and
+    docker.configure_nvidia merges the nvidia sysext over /usr right about then
+    -- detaching the overlay AFTER the patched modules are already imported. A
+    disk check after the restart therefore false-negatives on a healthy system,
+    and restarting on that signal would restart a correctly-patched middlewared
+    straight back into the same race.
+    """
+    src = wait_restart_source()
+    assert src.count("systemctl try-restart middlewared") == 1
+
+
+def test_patch_is_verified_after_the_restart():
+    src = wait_restart_source()
+    restart = src.index("systemctl try-restart middlewared")
+    assert "_patch_visible" in src[restart:], (
+        "the script must check what the restart actually loaded"
+    )
+
+
+def test_verification_reads_the_marker_apply_sh_writes():
+    # _patch_visible greps restic.py for TRUECLOUD_PATCH; apply.sh must still be
+    # the thing that puts it there, or the check silently always fails.
+    assert "TRUECLOUD_PATCH" in wait_restart_source()
+    assert "TRUECLOUD_PATCH" in apply_source()
+
+
+def test_verification_uses_the_recorded_middlewared_dir():
+    # wait_restart.sh must not re-derive site-packages; apply.sh records it.
+    assert ".mw_dir" in wait_restart_source()
+    assert ".mw_dir" in apply_source()
+
+
+def test_apply_sh_records_the_middlewared_dir():
+    src = apply_source()
+    assert re.search(r'>\s*"\$PATCH_DIR/\.mw_dir"', src), (
+        "apply.sh must write the resolved middlewared dir for wait_restart.sh"
+    )
+
+
+def test_reapply_pass_does_not_schedule_another_restart():
+    # wait_restart.sh owns the restart. If the re-apply pass scheduled its own
+    # transient unit, each boot would spawn restarts recursively.
+    src = apply_source()
+    guard = src.index('if [ "${TRUECLOUD_REAPPLY:-0}" = "1" ]; then')
+    systemd_run = src.index("systemd-run --no-block")
+    assert guard < systemd_run, (
+        "the TRUECLOUD_REAPPLY branch must short-circuit before systemd-run"
+    )
+
+
+def test_shadowed_overlay_is_remounted_not_accepted():
+    """A buried overlay must never pass for a healthy one.
+
+    _ensure_writable reaches its mount-table check only when the directory is
+    NOT writable -- and a live overlay of ours is always writable. So a
+    truecloud mount listed at that point is shadowed, and returning 0 there is
+    exactly how a detached overlay used to masquerade as applied.
+    """
+    src = apply_source()
+    start = src.index("_ensure_writable()")
+    end = src.index("\n}", start)
+    body = src[start:end]
+
+    check = body.index('mount | grep -qF "truecloud-${tag} on ${dir} "')
+    following = body[check:]
+    # The old code did `return 0` immediately inside this branch.
+    branch_end = following.index("fi")
+    assert "return 0" not in following[:branch_end]
+    assert "umount -l" in following[:branch_end]
+
+
+def test_workdir_is_recreated_before_mounting():
+    # overlayfs refuses a workdir left behind by a detached mount, so a stale
+    # one would turn every re-mount attempt into "overlay mount failed".
+    src = apply_source()
+    start = src.index("_ensure_writable()")
+    end = src.index("\n}", start)
+    body = src[start:end]
+    assert re.search(r'rm -rf "\$work"', body)
+
+
+def test_upperdir_is_preserved_across_remounts():
+    # The upperdir holds everything patched earlier this boot; reusing it is
+    # what lets a re-mount restore those files instead of re-deriving them.
+    src = apply_source()
+    start = src.index("_ensure_writable()")
+    end = src.index("\n}", start)
+    body = src[start:end]
+    assert 'rm -rf "$upper"' not in body
+
+
+@pytest.mark.parametrize("state", ["0", "1", "2"])
+def test_patch_visible_returns_three_distinct_states(state):
+    # patched / stock / cannot-tell must stay distinguishable: "cannot tell"
+    # has to re-apply rather than assume the patch is fine.
+    src = wait_restart_source()
+    assert f"return {state}" in src or f") return {state}" in src
+
+
+def test_mount_retries_on_a_private_workdir():
+    """A lazily-detached overlay can still pin the shared workdir.
+
+    overlayfs refuses a workdir that is in use, so without a retry the re-mount
+    this whole fix depends on would fail exactly when it is most needed.
+    """
+    src = apply_source()
+    start = src.index("_ensure_writable()")
+    end = src.index("\n}", start)
+    body = src[start:end]
+    assert body.count("mount -t overlay") == 2, "expected a retry mount"
+    assert 'work="/run/truecloud-${tag}-work.$$"' in body
+
+
+def test_post_restart_remount_preserves_the_patched_at_stamp():
+    """create_task.py verify compares middlewared's start time to patched_at.
+
+    The post-restart re-mount restores the same patch the boot pass applied, so
+    letting apply.sh re-stamp would make patched_at newer than the process that
+    correctly imported it -- verify would then report FAIL forever on every boot
+    where docker's sysext merge detaches the overlay.
+    """
+    src = wait_restart_source()
+    tail = src[src.index("systemctl try-restart middlewared"):]
+    assert "hook_status.json" in tail
+    save = tail.index("/run/truecloud-hook_status.pre")
+    reapply = tail.index("TRUECLOUD_REAPPLY=1")
+    restore = tail.rindex("hook_status.json")
+    assert save < reapply < restore, "snapshot must bracket the re-apply"
+
+
+def test_status_snapshot_is_not_written_into_the_repo():
+    """A leftover file in the repo dir leaves the tree dirty, and update.sh
+    refuses to run over a dirty tree -- that once made the patch un-updatable."""
+    src = wait_restart_source()
+    assert "/run/truecloud-hook_status.pre" in src
+    assert '"$PATCH_DIR/hook_status.json.pre' not in src
@@ -10,7 +10,8 @@ import re

 import pytest

-WORKFLOWS = os.path.join(os.path.dirname(__file__), "..", ".github", "workflows")
+ROOT = os.path.join(os.path.dirname(__file__), "..")
+WORKFLOWS = os.path.join(ROOT, ".github", "workflows")


 def workflow_files():
@@ -84,11 +85,77 @@ class TestBothForges:
        assert "if: ${{ contains(github.server_url, 'github.com') }}" in src
        assert "if: ${{ !contains(github.server_url, 'github.com') }}" in src

-    def test_compat_files_an_issue_on_each_forge(self):
+    def test_compat_files_its_report_through_ONE_implementation(self):
+        # It used to be two near-identical shell steps, one per forge. Two copies of
+        # "find the issue, decide whether to comment, post it" is two chances to drift,
+        # and the Gitea one duplicated an issue for real.
        with open(os.path.join(WORKFLOWS, "compat.yml"), encoding="utf-8") as fh:
            src = fh.read()
-        assert "file a bug report (GitHub)" in src
-        assert "file a bug report (Gitea)" in src
+        assert "tools/compat_publish.py" in src
+        assert "file a bug report (GitHub)" not in src
+        assert "file a bug report (Gitea)" not in src
+
+
+class TestTheBotDoesNotSpam:
+    """It left 11 identical 3,000-character comments on one issue in a single day.
+
+    A bot that repeats itself daily gets muted — and then the next REAL finding is
+    scrolled past, which defeats the entire reason for building it.
+    """
+
+    def publisher(self):
+        with open(os.path.join(ROOT, "tools", "compat_publish.py"), encoding="utf-8") as fh:
+            return fh.read()
+
+    def test_it_compares_a_fingerprint_before_saying_anything(self):
+        src = self.publisher()
+        assert "extract_fingerprint" in src
+        assert "staying quiet" in src
+
+    def test_the_body_is_edited_in_place_not_appended_to(self):
+        src = self.publisher()
+        assert '"PATCH"' in src, "the issue body must be updated, not commented onto"
+
+    def test_it_closes_the_issue_when_everything_is_fixed(self):
+        src = self.publisher()
+        assert '"state": "closed"' in src
+
+    def test_the_matrix_refresh_opens_a_PR_rather_than_pushing_to_main(self):
+        # An unattended push to main from CI is exactly what the release barrier exists
+        # to prevent: a bot that can move main can move it somewhere nobody looked.
+        #
+        # Checked against CODE, not comments — the step's own commentary explains what
+        # it replaced, and that mention must not read as the thing itself.
+        with open(os.path.join(WORKFLOWS, "compat.yml"), encoding="utf-8") as fh:
+            src = fh.read()
+        code = "\n".join(
+            ln for ln in src.splitlines() if not ln.lstrip().startswith("#")
+        )
+        assert "/pulls" in code, "the matrix refresh must open a PR"
+        assert "HEAD:main" not in code, "CI still pushes straight to main"
+
+    def test_the_matrix_PR_targets_the_CANONICAL_forge_not_the_mirror(self):
+        # GitHub is a one-way mirror: a PR merged there would be silently clobbered by
+        # the next `fleet-repos mirror` push from Gitea. A bot opening PRs against a
+        # mirror is a bot doing nothing, slowly.
+        with open(os.path.join(WORKFLOWS, "compat.yml"), encoding="utf-8") as fh:
+            src = fh.read()
+        i = src.index("refresh the README matrix")
+        step = src[i:i + 400]
+        assert "!contains(github.server_url, 'github.com')" in step, (
+            "the matrix PR must be opened on Gitea (canonical), not GitHub (mirror)"
+        )
+
+    def test_the_workflow_has_the_permissions_its_steps_actually_need(self):
+        # It shipped with `contents: read` while the step pushed a branch and opened a
+        # PR — it would have died with a 403 on the first scheduled run, and I would
+        # have had a bot that silently never worked.
+        with open(os.path.join(WORKFLOWS, "compat.yml"), encoding="utf-8") as fh:
+            src = fh.read()
+        perms = src[src.index("permissions:"):src.index("jobs:")]
+        assert "contents: write" in perms, "pushing a branch needs contents: write"
+        assert "pull-requests: write" in perms, "opening a PR needs pull-requests: write"
+        assert "issues: write" in perms


 class TestCompatCannotSilentlyPass:
@@ -105,3 +172,81 @@ class TestCompatCannotSilentlyPass:
        with open(os.path.join(WORKFLOWS, "compat.yml"), encoding="utf-8") as fh:
            src = fh.read()
        assert "steps.check.outputs.shipped_broken != '0'" in src
+
+
+class TestActionCacheRace:
+    """CI must not run concurrent jobs on the self-hosted runner.
+
+    `act` caches each ACTION as one shared clone under /root/.cache/act/<hash>
+    and re-pulls it per job, so jobs starting together fight over that directory
+    and the loser dies with `lstat .../<file>: no such file or directory` before
+    any test runs -- a red `main` with zero suite output and a different victim
+    each push. The runner also force-pulls its base image per job, so job count
+    is also Docker Hub pull count, and four-per-push exhausted the anonymous
+    limit in an afternoon. Both problems have the same cure: one job.
+    """
+
+    def _ci(self):
+        with open(os.path.join(WORKFLOWS, "ci.yml"), encoding="utf-8") as fh:
+            return fh.read()
+
+    def test_ci_runs_as_exactly_one_job(self):
+        """The fix is the absence of concurrency, not the absence of one action.
+
+        Dropping astral-sh/setup-uv only shrank the surface -- every job still
+        used actions/checkout. A single job cannot race itself whatever actions
+        it uses, which is why this, and not the action count, is the invariant.
+        """
+        ci = self._ci()
+        # Scope to the jobs: block -- `on:` has two-space keys of its own
+        # (push/pull_request/workflow_dispatch) that look identical otherwise.
+        body = ci[ci.index("\njobs:"):]
+        jobs = re.findall(r"^  (\w[\w-]*):$", body, re.M)
+        assert len(jobs) == 1, (
+            f"ci.yml defines {len(jobs)} jobs ({jobs}); concurrent jobs on the "
+            "self-hosted runner race on act's shared action cache and multiply "
+            "Docker Hub pulls. Keep CI to one job."
+        )
+
+    def test_no_matrix_reintroduces_parallel_jobs(self):
+        ci = self._ci()
+        assert "strategy:" not in ci and "matrix:" not in ci, (
+            "a matrix fans out into concurrent jobs again -- sweep versions "
+            "inside one job instead"
+        )
+
+    def test_every_python_version_still_runs_after_one_fails(self):
+        """`fail-fast: false` is what the loop has to preserve.
+
+        A 3.11 break must not hide whether 3.12 and 3.13 are fine; that is
+        precisely the information you want at that moment.
+        """
+        ci = self._ci()
+        assert 'PYTHONS: "3.11 3.12 3.13"' in ci
+        assert ci.count("fail=1") >= 2, "the sweeps must collect failures, not exit early"
+
+    def test_uv_is_installed_without_an_action(self):
+        """Checks `uses:` directives, not prose.
+
+        The comment in ci.yml names the action it deliberately avoids, and that
+        explanation is the most useful thing in the file -- a test that greps the
+        raw text would forbid documenting the very lesson it enforces. Parsed
+        with a regex rather than PyYAML on purpose: CI runs `uvx pytest`, whose
+        environment holds pytest and nothing else, so a third-party import here
+        fails on the runner while passing locally.
+        """
+        ci = self._ci()
+        used = re.findall(r"^\s*-?\s*uses:\s*(\S+)", ci, re.M)
+        assert not [u for u in used if "setup-uv" in u], (
+            "the action was only fetching a binary; a run: step does the same "
+            "with one less moving part"
+        )
+        assert "astral.sh/uv/" in ci
+
+    def test_the_uv_version_is_pinned(self):
+        ci = self._ci()
+        assert re.search(r'UV_VERSION:\s*"\d+\.\d+\.\d+"', ci), (
+            "an unpinned uv lets any upstream release turn main red with no "
+            "code change here -- the same rule ruff is pinned under"
+        )
+        assert "https://astral.sh/uv/${UV_VERSION}/install.sh" in ci
@@ -43,6 +43,7 @@ from __future__ import annotations

 import argparse
 import ast
+import hashlib
 import json
 import os
 import sys
@@ -132,9 +133,43 @@ ASSUMPTIONS = [
        params=["middleware", "job", "cloud_backup"],
        why="SYNC_BLOCK wraps it to tear down bind mounts in a finally",
    ),
+    Assumption(
+        # Not a plugin method -- a method on the middleware OBJECT itself, which the
+        # manifest had no way to express and therefore never checked.
+        #
+        # The nested module calls `middleware.get_service(<ns>)` to decide whether to
+        # sweep snapshots through `pool.snapshot` or `zfs.snapshot` (see
+        # SNAPSHOT_SERVICES). If it ever disappears, `_can_delete()` catches the
+        # AttributeError, reports BOTH namespaces unusable, and every nested backup
+        # fails -- loudly, but only at RUN time, on a box the preflight had already
+        # declared healthy. Checking it costs one file read.
+        "get-service", NESTED, "utils/plugins.py",
+        "LoadPluginsMixin.get_service", kind="method",
+        params=["self", "name"],
+        why="snapshot_service() resolves the snapshot namespace through it; without "
+            "it the module cannot sweep the snapshot it just took",
+    ),
 ]


+def accepted_spellings(name):
+    """The method names that satisfy a call to `<namespace>.<name>`.
+
+    A CRUDService exposes `create`/`update`/`delete` from methods NAMED
+    `do_create`/`do_update`/`do_delete`. Both are live across the matrix: 24.10 and
+    25.04 declare `do_delete`, 25.10 renamed it to `delete`, and all of them answer
+    to `<ns>.delete`. Accepting only the literal name reported working releases as
+    BROKEN and would have switched nested snapshots off on boxes where they work.
+    """
+    return (name, f"do_{name}")
+
+
+#: The spellings that satisfy `<ns>.delete`. A test binds this to the runtime's
+#: `truecloud_nested.DELETE_METHODS`, so the checker and the patch cannot come to
+#: disagree about what "can delete" means on the same box.
+DELETE_NAMES = accepted_spellings("delete")
+
+
 class MiddlewareCall:
    """A middlewared METHOD the injected code calls at runtime.

@@ -161,86 +196,166 @@ class MiddlewareCall:
    the whole design: declining is always the cheaper mistake.
    """

-    def __init__(self, ident, module, method, path, why=""):
+    def __init__(self, ident, module, method, path, why="", also=()):
        self.id = ident
        self.module = module
-        self.method = method              # "zfs.snapshot.delete"
+        self.method = method              # "pool.snapshot.delete"
        self.path = path                  # plugin file that declares it
        self.why = why
+        #: Equally acceptable spellings of the SAME call, as (method, path) pairs.
+        #:
+        #: No single snapshot namespace spans every supported release. 24.10 and
+        #: 25.04 expose the CRUD service as the public `zfs.snapshot`; 25.10
+        #: promoted it to `pool.snapshot` and demoted `zfs.snapshot` to private;
+        #: 26 deleted `plugins/zfs_/` entirely. Pinning either one alone marks
+        #: half the matrix BROKEN and declines to apply on versions that work
+        #: perfectly well.
+        #:
+        #: The call is satisfied if ANY option is present. The runtime picks the
+        #: same way -- see `pick_snapshot_service()` in the nested module -- so
+        #: what this checks and what the patch does cannot drift apart.
+        self.also = tuple(also)

    @property
-    def namespace(self):
-        return self.method.rsplit(".", 1)[0]
+    def options(self):
+        """Every (method, path) that would satisfy this call, best first."""
+        return ((self.method, self.path), *self.also)

-    @property
-    def name(self):
-        return self.method.rsplit(".", 1)[1]
+    @staticmethod
+    def namespace_of(method):
+        return method.rsplit(".", 1)[0]
+
+    @staticmethod
+    def name_of(method):
+        return method.rsplit(".", 1)[1]


 #: Every middlewared method the nested module calls at runtime.
+#: The middleware methods the nested module CALLS.
+#:
+#: These used to be the PRIVATE `zfs.*` service (`zfs.dataset.query`,
+#: `zfs.snapshot.delete`, `zfs.snapshot.query`). TrueNAS 26 deleted
+#: `plugins/zfs_/` outright and every one of them vanished -- silently, because a
+#: private service carries no stability contract and nothing warned us. The patch
+#: would have applied cleanly and then failed on the first backup.
+#:
+#: The replacements are the PUBLIC `pool.*` API, and switching to it is not merely
+#: a TrueNAS 26 fix -- it is the correct call on every version:
+#:
+#:   * It is public, documented, and covered by iX's deprecation policy, so it
+#:     cannot be deleted from under us the way `zfs.*` just was.
+#:   * The same methods, in the same files, exist on 24.10 through 26. One code
+#:     path, no version conditionals.
+#:   * Both spellings take `recursive`, so ONE call sweeps the whole tree instead
+#:     of ~250 individual deletes, any of which could be missed.
 MIDDLEWARE_CALLS = [
    MiddlewareCall(
-        "call-zfs-dataset-query", NESTED, "zfs.dataset.query",
-        "plugins/zfs_/dataset.py",
-        why="SNAPSHOT_BLOCK enumerates FILESYSTEM datasets to build the staging plan",
-    ),
-    MiddlewareCall(
-        "call-zfs-snapshot-delete", NESTED, "zfs.snapshot.delete",
-        "plugins/zfs_/snapshot.py",
+        "call-snapshot-delete", NESTED, "pool.snapshot.delete",
+        "plugins/pool_/snapshot.py",
+        also=[("zfs.snapshot.delete", "plugins/zfs_/snapshot.py")],
        why="delete_snapshot_tree() sweeps the recursive snapshot. Without it every "
            "run orphans one snapshot per descendant dataset (250 on a real pool)",
    ),
-    MiddlewareCall(
-        "call-zfs-snapshot-query", NESTED, "zfs.snapshot.query",
-        "plugins/zfs_/snapshot.py",
-        why="delete_snapshot_tree()'s fallback sweep enumerates the tree by name",
-    ),
 ]

+# There is deliberately NO entry here for a dataset or snapshot QUERY.
+#
+# The patch used to call `zfs.dataset.query` / `zfs.snapshot.query` (private, and
+# deleted in TrueNAS 26). The obvious port was to the public `pool.dataset.query` /
+# `pool.snapshot.query` -- and that port was WRONG in a way no source check could
+# ever have caught, because the methods are all present and correctly shaped.
+#
+# They are simply filtered. On a real box they return 205 of 274 datasets and 205
+# of 274 snapshots, hiding `ix-apps/*`, `.system/*` and `.ix-virt/*` -- 84 of 270
+# on the production pool, including live application data. Staging from that view
+# silently omits them; sweeping from it orphans one snapshot per hidden dataset,
+# forever.
+#
+# So the module enumerates from ZFS itself and there is no middleware assumption
+# left to check. That is the point: the fewer things we assume about middleware,
+# the less there is for iX to break. Only the MUTATION is still a middleware call,
+# and that is the one entry above.
+
+
+def check_call(c: MiddlewareCall, src: str | None,
+               method: str | None = None, path: str | None = None,
+               ) -> tuple[str, str | None]:
+    """Is `method` still registered by middlewared?
+
+    `method`/`path` name WHICH spelling of the call is being tried -- a call may
+    have several equally acceptable ones (see MiddlewareCall.also). They default
+    to the preferred spelling.
+    """
+    method = method or c.method
+    path = path or c.path
+    namespace = MiddlewareCall.namespace_of(method)
+    name = MiddlewareCall.name_of(method)

-def check_call(c: MiddlewareCall, src: str | None) -> tuple[str, str | None]:
-    """Is `c.method` still registered by middlewared?"""
    if src is None:
        return "broken", (
-            f"{c.path} no longer exists, so `{c.method}` is gone"
+            f"{path} no longer exists, so `{method}` is gone"
        )

    try:
        tree = ast.parse(_stock(src))
    except SyntaxError as e:
-        return "unknown", f"{c.path} does not parse: {e}"
+        return "unknown", f"{path} does not parse: {e}"

-    # namespace = 'zfs.snapshot' on some Service class in this file...
-    namespaces = {
-        n.value.value
-        for n in ast.walk(tree)
-        if isinstance(n, ast.Assign)
-        and isinstance(n.value, ast.Constant)
-        and isinstance(n.value.value, str)
-        and any(isinstance(t, ast.Name) and t.id == "namespace" for t in n.targets)
-    }
-    if c.namespace not in namespaces:
-        return "broken", (
-            f"{c.path} no longer declares namespace {c.namespace!r} "
-            f"(found: {sorted(namespaces) or 'none'}), so `{c.method}` is gone"
-        )
-
-    # ...and it defines the method.
+    # Find the CLASS that declares this namespace, and look for the method THERE.
+    #
+    # Not anywhere in the file. `ast.walk` over the whole module made *any* function
+    # called `delete` satisfy the check -- one on an unrelated class, or even a nested
+    # local function inside `do_query`. That is a FALSE OK, and it breaks the one
+    # invariant this checker and the runtime share: `_defines_delete()` looks in
+    # `vars(klass)` for a PLUGIN class on the service's MRO. If iX gutted
+    # `PoolSnapshotService.do_delete` while some other class in the same file still had
+    # a `delete`, compat would say ok, apply.sh would patch, and the runtime would then
+    # correctly refuse `pool.snapshot`, fall through to a `zfs.snapshot` that does not
+    # exist on 26, and fail every nested backup on a box the preflight called healthy.
+    #
+    # Same question on both sides: does the class that OWNS this namespace define the
+    # method?
    #
    # A CRUDService exposes `create`/`update`/`delete` from methods NAMED
-    # `do_create`/`do_update`/`do_delete`. Both spellings are live right now:
-    # 24.10 and 25.04 declare `do_delete`, 25.10 renamed it to `delete`, and all
-    # three answer to `zfs.snapshot.delete`. Accepting only the literal name reported
-    # the two older releases as broken -- a false BROKEN that would have switched off
-    # nested snapshots on boxes where they work.
-    defined = {
-        n.name for n in ast.walk(tree)
-        if isinstance(n, ast.FunctionDef | ast.AsyncFunctionDef)
-    }
-    if c.name not in defined and f"do_{c.name}" not in defined:
-        return "broken", f"{c.path} no longer defines `{c.method}`"
+    # `do_create`/`do_update`/`do_delete`. Both spellings are live: 24.10 and 25.04
+    # declare `do_delete`, 25.10 renamed it to `delete`, and all answer to
+    # `<ns>.delete`. Accepting only the literal name reported working releases as
+    # broken.
+    owners = []
+    all_namespaces = set()
+    for cls in (n for n in ast.walk(tree) if isinstance(n, ast.ClassDef)):
+        declared = {
+            n.value.value
+            for n in ast.walk(cls)
+            if isinstance(n, ast.Assign)
+            and isinstance(n.value, ast.Constant)
+            and isinstance(n.value.value, str)
+            and any(isinstance(t, ast.Name) and t.id == "namespace" for t in n.targets)
+        }
+        all_namespaces |= declared
+        if namespace in declared:
+            owners.append(cls)

-    return "ok", None
+    if not owners:
+        return "broken", (
+            f"{path} no longer declares namespace {namespace!r} "
+            f"(found: {sorted(all_namespaces) or 'none'}), so `{method}` is gone"
+        )
+
+    wanted = accepted_spellings(name)
+    for cls in owners:
+        # Direct members of the class, not its nested scopes: a `def delete` inside
+        # another method is a local function, not a service method.
+        if any(
+            isinstance(n, ast.FunctionDef | ast.AsyncFunctionDef) and n.name in wanted
+            for n in cls.body
+        ):
+            return "ok", None
+
+    return "broken", (
+        f"{path} still declares namespace {namespace!r}, but its class no longer "
+        f"defines `{'` or `'.join(wanted)}` -- so `{method}` is gone"
+    )


 #: Things that mean iX has done the job themselves and the module should RETIRE,
@@ -403,7 +518,10 @@ def check_source(a: Assumption, src: str | None) -> tuple[str, str | None]:
    breaks a box that was working.
    """
    if src is None:
-        return "broken", f"{a.path} does not exist"
+        # Name the SYMBOL, not just the file. Whoever reads the bug report needs to
+        # know what the patch can no longer reach, and "utils/plugins.py does not
+        # exist" does not tell them that `get_service` is gone.
+        return "broken", f"{a.path} does not exist, so `{a.symbol}` is gone"

    try:
        tree = ast.parse(src)
@@ -538,6 +656,7 @@ def check(loader, modules=None) -> dict:
            out[a.module]["unknown"] = True
            out[a.module]["problems"].append({
                "id": a.id, "detail": f"could not read {a.path}: {e}", "why": a.why,
+                "state": "unknown",
            })
            continue

@@ -545,37 +664,57 @@ def check(loader, modules=None) -> dict:
        if status == "broken":
            out[a.module]["ok"] = False
            out[a.module]["problems"].append({
-                "id": a.id, "detail": detail, "why": a.why,
+                "id": a.id, "detail": detail, "why": a.why, "state": "broken",
            })
        elif status == "unknown":
            out[a.module]["unknown"] = True
            out[a.module]["problems"].append({
-                "id": a.id, "detail": detail, "why": a.why,
+                "id": a.id, "detail": detail, "why": a.why, "state": "unknown",
            })

    # The methods the injected code CALLS, not just the symbols it wraps.
+    #
+    # A call may have several equally acceptable spellings, because no single
+    # snapshot namespace spans every supported release (24.10 has `zfs.snapshot`,
+    # 26 has only `pool.snapshot`). It is satisfied if ANY of them is present --
+    # exactly as the runtime resolves it -- and BROKEN only when they all vanish.
    for c in MIDDLEWARE_CALLS:
        if c.module not in out:
            continue
-        try:
-            text = src(c.path)
-        except Unreadable as e:
-            out[c.module]["unknown"] = True
-            out[c.module]["problems"].append({
-                "id": c.id, "detail": f"could not read {c.path}: {e}", "why": c.why,
-            })
+
+        satisfied, unknown, details = False, False, []
+        for method, path in c.options:
+            try:
+                text = src(path)
+            except Unreadable as e:
+                unknown = True
+                details.append(f"could not read {path}: {e}")
+                continue
+
+            status, detail = check_call(c, text, method, path)
+            if status == "ok":
+                satisfied = True
+                break
+            if status == "unknown":
+                unknown = True
+            details.append(detail)
+
+        if satisfied:
            continue

-        status, detail = check_call(c, text)
-        if status == "broken":
-            out[c.module]["ok"] = False
-            out[c.module]["problems"].append({
-                "id": c.id, "detail": detail, "why": c.why,
-            })
-        elif status == "unknown":
+        # Every spelling failed. If we could not READ one of them we do not know
+        # that it is broken -- a rate-limited fetch is not a regression.
+        if unknown:
            out[c.module]["unknown"] = True
            out[c.module]["problems"].append({
-                "id": c.id, "detail": detail, "why": c.why,
+                "id": c.id, "detail": "; ".join(details), "why": c.why,
+                "state": "unknown",
+            })
+        else:
+            out[c.module]["ok"] = False
+            out[c.module]["problems"].append({
+                "id": c.id, "detail": "; ".join(details), "why": c.why,
+                "state": "broken",
            })

    for module, (path, phrase, native_when_present) in NATIVE_PROBES.items():
@@ -732,6 +871,24 @@ def discover_refs(remote: str = REPO) -> list[str]:
      * UNRELEASED comes from the BRANCHES, because that is where a beta appears
        first: `release/26.0.0-BETA.3` had no tag yet while it was the newest beta.
        Catching breakage here, before it ships, is the whole point of this file.
+
+    THE NEXT MAINTENANCE RELEASE IS ALSO A BRANCH, and it used to fall through both
+    sieves. `release/25.10.5` is branched but not yet tagged, and 25.10 has already
+    shipped -- so it is not in `shipped` (no tag) and it was excluded from `upcoming`
+    (its line is in `shipped_lines`). It was invisible. That is the one ref a 25.10.4
+    user is actually about to be upgraded onto, so a break there reaches people
+    BEFORE the daily check ever looks at it -- the exact hole this file exists to
+    close, on the only line anybody is running.
+
+    The rule that separates it from the two things we must NOT report:
+
+      * `release/24.10-RC.2` -- a prerelease of an already-shipped line. It is
+        history, not a warning; it sorts OLDER than TS-24.10.2.4, so it is dropped.
+      * `release/25.20.2.2` -- iX's typo branch. 25.20 never shipped, so it has no
+        TS tag, so it is not a line at all and is dropped.
+
+    ...which is: a plain `release/X.Y.Z` branch counts only if its line HAS shipped
+    and it sorts NEWER than that line's newest tag. Both exclusions fall out of it.
    """
    tags = _ls_remote(remote, "--tags")
    heads = _ls_remote(remote, "--heads")
@@ -739,26 +896,45 @@ def discover_refs(remote: str = REPO) -> list[str]:
    shipped = _newest_per_line([
        t for t in tags if t.startswith("TS-") and "-BETA" not in t and "-RC" not in t
    ])
+    newest_shipped = {_version_of(t)[0][:2]: _version_of(t) for t in shipped}
+
+    release_heads = [h for h in heads if h.startswith("release/")]

    # A prerelease of a line that has ALREADY shipped is history, not a warning:
    # release/24.10-RC.2 still exists, and the nested module does not apply to it,
    # but 24.10 shipped long ago and TS-24.10.2.4 is fine. Reporting it would be a
    # standing red row in the matrix for a version nobody can install.
-    shipped_lines = {_version_of(t)[0][:2] for t in shipped}
    upcoming = [
        h for h in _newest_per_line([
-            h for h in heads
-            if h.startswith("release/") and ("-BETA" in h or "-RC" in h)
+            h for h in release_heads if "-BETA" in h or "-RC" in h
        ])
-        if _version_of(h)[0][:2] not in shipped_lines
+        if _version_of(h)[0][:2] not in newest_shipped
    ]

-    return [*shipped, *upcoming, "master"]
+    # The next maintenance release of a line that HAS shipped: branched, untagged,
+    # and the very next thing those users get. See the docstring.
+    pending = []
+    for h in _newest_per_line([
+        h for h in release_heads if "-BETA" not in h and "-RC" not in h
+    ]):
+        v = _version_of(h)
+        tagged = newest_shipped.get(v[0][:2])
+        if tagged and v > tagged:
+            pending.append(h)
+
+    return [*shipped, *pending, *upcoming, "master"]


 def is_unreleased(ref: str) -> bool:
-    """master and any BETA/RC. Breakage here is early warning, not an outage."""
-    return ref == "master" or "-BETA" in ref or "-RC" in ref
+    """Anything iX has not TAGGED. Breakage here is early warning, not an outage.
+
+    Keyed on where the ref came from, not on its name: `discover_refs` takes shipped
+    releases from `TS-*` TAGS and everything else from BRANCHES, so a `release/*` ref
+    is by construction something iX has not released yet. Testing for `-BETA`/`-RC`
+    instead would call `release/25.10.5` SHIPPED, and a break there would fail the
+    build as a live outage -- on a version nobody is running yet.
+    """
+    return ref == "master" or ref.startswith("release/")


 def matrix(refs=None, remote: str = REPO) -> list[dict]:
@@ -802,7 +978,16 @@ def is_broken(r: dict) -> bool:
 #: is static analysis of iX's source, which proves the patch's assumptions hold --
 #: a strictly weaker claim than "a restore worked". Add a row only after doing it.
 HARDWARE_VERIFIED = {
-    "25.10.4": "nested + providers; 252-snapshot recursive backup of /mnt/Tap, 18m",
+    "25.10.4": (
+        "v0.7.0: 3 live tasks — 191-dataset nested backup of /mnt/Tap, a "
+        "215-filesystem/2-zvol backup of /mnt/Tank/backups, and a non-nested one; "
+        "0 orphans, 0 leaked mounts, byte-identical restore; the collector also "
+        "reclaimed a real orphan the pool had been carrying"
+    ),
+    "26.0.0-BETA.1": (
+        "v0.7.0: 274-snapshot recursive backup of a 292-dataset pool; restored a "
+        "4-deep child dataset byte-identical; zvol-orphan case reproduced then closed"
+    ),
 }

 _LEGEND = """
@@ -815,6 +1000,16 @@ _LEGEND = """
 "ok" means *the patch's assumptions hold*, checked automatically against iX's
 source. It does not mean a human ran a backup on it — that is the
 **Hardware-verified** column, which is filled in by hand and only by doing it.
+
+**`master` is not the next release.** iX branches each major off to its own
+`release/` line and master rolls straight on to the one after — so master is
+`27-dev` while 26 is still in beta. A **BROKEN** master means iX has changed
+something that will reach users *a major release from now*, not in the version you
+are about to install. Read the numbered rows for that.
+
+A row like `25.10.5 _(unreleased)_` is the next maintenance release: branched by iX,
+not tagged yet, and the very next thing a 25.10.4 box gets. It is checked precisely
+because it is the one unshipped ref that reaches real users without warning.
 """


@@ -862,17 +1057,42 @@ def update_readme(rows: list[dict], path: str = README) -> bool:
    return True


+def dev_label(rows: list[dict]) -> str:
+    """What `master` is a development line FOR -- e.g. "27-dev".
+
+    `master` is NOT the next release, and labelling it "master _(unreleased)_" said
+    it was. On 2026-07-14 every recent commit on master targeted 27.0.0-BETA.1 while
+    26 was still in beta on its own `release/26.0.0-BETA.*` branches: master had
+    already rolled over to the major AFTER the one that has not shipped yet. So a red
+    `master` row read as "the version you are about to install is broken" when the
+    breakage was a year out, on a line nobody can even download. That is a false alarm
+    aimed squarely at the person deciding whether to trust this with their backups.
+
+    Derived, not hardcoded: the newest major we can see anywhere, plus one. iX branches
+    `release/N.0.0-BETA.1` off master and master immediately becomes N+1 -- so when 27
+    betas appear, this says 28-dev on its own.
+    """
+    majors = [
+        v[0][0] for v in (_version_of(r["ref"]) for r in rows if r["ref"] != "master")
+        if v
+    ]
+    return f"{max(majors) + 1}-dev" if majors else "unreleased"
+
+
 def render_markdown(rows: list[dict]) -> str:
    """The matrix, for the README."""
    out = [
        "| TrueNAS | B2/S3 providers | Nested snapshots | Hardware-verified |",
        "| --- | --- | --- | --- |",
    ]
+    dev = dev_label(rows)
    for row in rows:
        m = row["modules"]
        ref = row["ref"]
        label = ref.removeprefix("TS-").removeprefix("release/")
-        if row["unreleased"]:
+        if ref == "master":
+            label = f"master _({dev})_"
+        elif row["unreleased"]:
            label = f"{label} _(unreleased)_"

        cells = []
@@ -892,6 +1112,96 @@ def render_markdown(rows: list[dict]) -> str:
    return "\n".join(out) + "\n" + _LEGEND


+FINGERPRINT = "<!-- compat-fingerprint:"
+
+
+def fingerprint(rows: list[dict]) -> str:
+    """A stable digest of WHAT IS BROKEN, and nothing else.
+
+    The bug report must be updated when the findings change and stay silent when they
+    do not. Without this the workflow commented on every run -- it left **11 identical
+    3,000-character comments** on one issue in a single day, which is not a warning
+    system, it is a mute button with extra steps.
+
+    Deliberately excludes anything that moves on its own: the matrix's `ok` rows, the
+    hardware-verified column, and the exact TrueNAS point-release (`TS-25.10.4` ->
+    `TS-25.10.5` is not news). Only the broken (ref, module, problem-id) triples count.
+    """
+    findings = sorted(
+        (r["ref"], mod, p["id"])
+        for r in rows
+        for mod, m in r["modules"].items()
+        if is_broken(m)
+        for p in m["problems"]
+        # `unknown` problems are things we could not READ (a 429, an EACCES), not
+        # things iX changed. On a ref that is broken for some other reason they would
+        # otherwise join the digest, so one transient network blip rewrites the issue
+        # body and the next clean run rewrites it back. That is the daily-noise
+        # failure this fingerprint exists to prevent, wearing a different hat.
+        if p.get("state", "broken") == "broken"
+    )
+    return hashlib.sha256(repr(findings).encode()).hexdigest()[:16]
+
+
+def extract_fingerprint(body: str) -> str | None:
+    """The fingerprint a previous run left in the issue body, if any."""
+    if not body:
+        return None
+    i = body.find(FINGERPRINT)
+    if i == -1:
+        return None
+    return body[i + len(FINGERPRINT):].split("-->", 1)[0].strip() or None
+
+
+def render_issue(rows: list[dict]) -> str:
+    """The bug report body: what is broken, why it matters, and nothing else up front.
+
+    Short by design. The full matrix and the healthy versions go in a fold -- somebody
+    opening this wants to know what broke and whether it can hurt them, not to re-read
+    a table they can see in the README.
+    """
+    broken = [r for r in rows if any(is_broken(m) for m in r["modules"].values())]
+
+    out = [
+        "`tools/compat.py` checks what this patch assumes about middlewared against "
+        "iXsystems' actual source, every day. Those assumptions no longer hold on the "
+        "versions below.",
+        "",
+        "**This does not break anyone today.** `apply.sh` re-checks on every boot and "
+        "**declines to apply** a module whose assumptions fail, so TrueNAS is left "
+        "stock rather than half-patched. The cost is the module's feature, not a "
+        "broken backup.",
+        "",
+    ]
+
+    for r in broken:
+        out.append(f"### `{r['ref']}`")
+        out.append("")
+        for mod, m in sorted(r["modules"].items()):
+            if not is_broken(m):
+                continue
+            out.append(f"**{mod}**")
+            out.append("")
+            for p in m["problems"]:
+                out.append(f"- {p['detail']}")
+                out.append(f"  <br><sub>{p['why']}</sub>")
+            out.append("")
+
+    out += [
+        "<details><summary>Full support matrix</summary>",
+        "",
+        render_markdown(rows),
+        "</details>",
+        "",
+        "_Filed and kept up to date by "
+        "[`compat.yml`](.github/workflows/compat.yml). It edits this body when the "
+        "findings change, and stays quiet when they do not._",
+        "",
+        f"{FINGERPRINT} {fingerprint(rows)} -->",
+    ]
+    return "\n".join(out)
+
+
 def render_matrix(rows: list[dict]) -> str:
    """A support table.

@@ -900,7 +1210,13 @@ def render_matrix(rows: list[dict]) -> str:
    hardware-verified column lives in COMPATIBILITY.md and is maintained by hand,
    because nothing else can honestly fill it in.
    """
-    w = max((len(r["ref"]) for r in rows), default=10)
+    dev = dev_label(rows)
+    # Same relabel as the README: a red `master` is a warning about the major AFTER
+    # next, and "master" alone reads as "the release you are about to install".
+    names = {r["ref"]: (f"master ({dev})" if r["ref"] == "master" else r["ref"])
+             for r in rows}
+
+    w = max((len(n) for n in names.values()), default=10)
    lines = [
        f"{'TrueNAS'.ljust(w)}  {'providers':<10}  {'nested':<10}",
        f"{'-' * w}  {'-' * 10}  {'-' * 10}",
@@ -908,7 +1224,7 @@ def render_matrix(rows: list[dict]) -> str:
    for row in rows:
        m = row["modules"]
        lines.append(
-            f"{row['ref'].ljust(w)}  "
+            f"{names[row['ref']].ljust(w)}  "
            f"{_verdict(m[PROVIDERS]):<10}  {_verdict(m[NESTED]):<10}"
        )
    return "\n".join(lines)
@@ -0,0 +1,185 @@
+#!/usr/bin/env python3
+"""Keep ONE bug report in sync with what compat.py currently finds.
+
+WHY THIS IS NOT JUST "POST A COMMENT"
+-------------------------------------
+The first version commented on every run that found a break. In one day it left
+**11 identical 3,000-character comments** on the same issue. That is not a warning
+system; it is a mute button with extra steps. The next real finding would have been
+scrolled past, which defeats the entire point of building it.
+
+So:
+
+  * **The issue body is the current truth.** It is edited in place, never appended to.
+  * **Comments are a changelog of CHANGES.** A run whose findings are identical to the
+    last one says nothing at all -- no comment, no edit, no notification.
+  * A fingerprint of the findings (broken ref/module/problem triples only) is embedded
+    in the body. It deliberately ignores things that move on their own -- healthy rows,
+    the hardware-verified column, TrueNAS point releases -- so `TS-25.10.4` becoming
+    `TS-25.10.5` is not news, and does not wake anybody up.
+
+  * When everything is fixed, the issue is **closed** with a comment saying so.
+
+Works against GitHub and Gitea, which differ only in the auth header and the issue
+list URL. One implementation, so the two cannot drift.
+
+    python3 tools/compat_publish.py --api <url> --token <tok> --matrix /tmp/matrix.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+import urllib.error
+import urllib.request
+
+sys.path.insert(0, __file__.rsplit("/", 1)[0])
+
+from compat import (  # noqa: E402
+    extract_fingerprint,
+    fingerprint,
+    is_broken,
+    render_issue,
+)
+
+TITLE = "TrueNAS compatibility: the patch's assumptions no longer hold"
+
+
+def _call(url, token, method="GET", data=None):
+    req = urllib.request.Request(
+        url, method=method,
+        headers={
+            # Gitea wants `token <t>`; GitHub accepts `Bearer <t>`. GitHub also
+            # accepts `token <t>`, so one header serves both.
+            "Authorization": f"token {token}",
+            "Content-Type": "application/json",
+            "Accept": "application/vnd.github+json",
+        },
+        data=json.dumps(data).encode() if data else None,
+    )
+    with urllib.request.urlopen(req) as r:  # noqa: S310
+        return json.load(r) if r.length != 0 else {}
+
+
+def _same(a, b):
+    """Is the issue body already what we would write?
+
+    Compared after normalising line endings and trailing space: forges are free to
+    round-trip `\r\n`, and a body that only "differs" by that would be rewritten on
+    every single run -- a silent edit, but a pointless one that churns `updated_at`
+    and makes the issue look freshly touched every morning.
+    """
+    def norm(s):
+        return "\n".join(line.rstrip() for line in (s or "").replace("\r\n", "\n").split("\n")).strip()
+    return norm(a) == norm(b)
+
+
+def find_issue(api, token, title):
+    """The LOWEST-numbered issue with this title, open or closed.
+
+    Lowest, not "whichever the API returns first": two issues with the same title
+    existed once (an earlier version put the ref list in the title, so the identity
+    changed whenever that set changed), and an order-dependent pick would alternate
+    between them -- reopening one while commenting on the other.
+
+    Both forges list PRs alongside issues, but they SAY SO DIFFERENTLY: GitHub omits
+    the `pull_request` key on a plain issue, Gitea sends it as `null`. Testing for the
+    KEY therefore discards every Gitea issue as if it were a PR -- so this returned
+    None on every Gitea run, and the bot filed a brand-new duplicate report each time
+    instead of editing the one it already had. Test the VALUE; it is the only form
+    that is true on both.
+    """
+    issues = _call(f"{api}/issues?state=all&per_page=100&limit=100", token)
+    mine = [
+        i for i in issues
+        if i.get("title") == title and not i.get("pull_request")
+    ]
+    return min(mine, key=lambda i: i["number"]) if mine else None
+
+
+def main(argv):
+    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    ap.add_argument("--api", required=True, help="…/repos/<owner>/<repo>")
+    ap.add_argument("--token", required=True)
+    ap.add_argument("--matrix", required=True, help="compat.py --matrix --json output")
+    args = ap.parse_args(argv[1:])
+
+    with open(args.matrix, encoding="utf-8") as fh:
+        rows = json.load(fh)
+
+    broken = [r for r in rows if any(is_broken(m) for m in r["modules"].values())]
+    issue = find_issue(args.api, args.token, TITLE)
+
+    # ── everything is healthy ────────────────────────────────────────────────
+    if not broken:
+        if issue and issue["state"] == "open":
+            _call(f"{args.api}/issues/{issue['number']}/comments", args.token, "POST",
+                  {"body": "All of the patch's assumptions hold again on every "
+                           "checked TrueNAS version. Closing."})
+            _call(f"{args.api}/issues/{issue['number']}", args.token, "PATCH",
+                  {"state": "closed"})
+            print(f"closed #{issue['number']} — nothing is broken any more")
+        else:
+            print("nothing broken; no open report to close")
+        return 0
+
+    body = render_issue(rows)
+    want = fingerprint(rows)
+
+    # ── nothing to file yet ──────────────────────────────────────────────────
+    if issue is None:
+        made = _call(f"{args.api}/issues", args.token, "POST",
+                     {"title": TITLE, "body": body})
+        print(f"filed #{made['number']}")
+        return 0
+
+    have = extract_fingerprint(issue.get("body") or "")
+    n = issue["number"]
+
+    # ── two different questions, and they were being answered with one answer ────
+    #
+    #   * IS THE BODY STILL TRUE?     -> if not, rewrite it. Editing an issue body
+    #     notifies NOBODY on either forge, so keeping it honest is free.
+    #   * HAVE THE FINDINGS CHANGED?  -> only then comment. Comments DO notify, and a
+    #     daily "still broken, same as yesterday" is what teaches everyone to ignore
+    #     the one that finally matters.
+    #
+    # Conflating them meant an unchanged FINGERPRINT froze the BODY. The fingerprint
+    # deliberately ignores everything that moves on its own -- healthy rows, the
+    # hardware-verified column, point releases, how a row is LABELLED -- so none of
+    # that could ever reach the report. Relabelling master `27-dev` (it is not the
+    # next release; a red row there was reading as "the version you are about to
+    # install is broken") would have shipped to the README and never to the issue
+    # anybody actually opens.
+    body_is_current = _same(issue.get("body"), body)
+
+    if have == want and issue["state"] == "open" and body_is_current:
+        print(f"#{n} is already current ({want}) — staying quiet")
+        return 0
+
+    _call(f"{args.api}/issues/{n}", args.token, "PATCH", {"body": body, "state": "open"})
+
+    if have != want:
+        refs = ", ".join(f"`{r['ref']}`" for r in broken)
+        note = (
+            "The findings changed — the report above has been updated.\n\n"
+            f"Currently broken on: {refs}."
+            if have else
+            "This report is now kept up to date automatically: the body above always "
+            "reflects the current findings, and a comment is only added when they "
+            "change."
+        )
+        _call(f"{args.api}/issues/{n}/comments", args.token, "POST", {"body": note})
+        print(f"updated #{n}: {have} -> {want}")
+    elif issue["state"] != "open":
+        print(f"reopened #{n}")
+    else:
+        # Same findings, new rendering. Silent by design: nothing has changed that
+        # anybody needs waking up for, but the report should not be telling lies.
+        print(f"#{n}: findings unchanged ({want}); body refreshed silently")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv))
@@ -107,10 +107,15 @@ def changelog_versions(text: str) -> list[str]:
 def extract_notes(text: str, version: str) -> str:
    """The body of one version's section, without its heading.

+    A release candidate resolves to its BASE version: v0.6.0-rc2 ships the same code
+    as v0.6.0 and therefore the same notes, and the CHANGELOG only ever has the one
+    section. Without this, the release workflow cut the tag, passed every gate, and
+    then died extracting the body -- so the candidate existed but was never published.
+
    Raises KeyError if the version has no section -- a release with an empty or
    wrong body is worse than a failed release.
    """
-    want = normalise(version)
+    want = base_version(version)
    lines = text.splitlines()

    start = None
@@ -3,7 +3,7 @@

 set -euo pipefail

-VERSION="0.6.0"
+VERSION="0.8.0"

 PATCH_DIR="$(cd "$(dirname "$0")" && pwd)"
 _HOOK_COMMENT='TrueCloud provider patch (S3/B2)'
@@ -124,6 +124,11 @@ if [ -f "$PATCH_DIR/nested_snapshots_enabled" ]; then
    rm -f "$PATCH_DIR/nested_snapshots_enabled"
    echo "  Removed nested-snapshot opt-in marker."
 fi
+
+# Runtime breadcrumb recorded by apply.sh for wait_restart.sh. Harmless, but a
+# stale path left in an uninstalled tree is exactly the sort of thing that reads
+# as state later.
+rm -f "$PATCH_DIR/.mw_dir"
 echo ""

 if [ "$_restore_failed" -eq 1 ]; then
@@ -19,7 +19,7 @@

 set -euo pipefail

-VERSION="0.6.0"
+VERSION="0.8.0"

 PATCH_DIR="$(cd "$(dirname "$0")" && pwd)"
 _PREV_FILE="$PATCH_DIR/.update_previous"
@@ -123,7 +123,7 @@ cd "$PATCH_DIR"

 if ! git rev-parse --git-dir >/dev/null 2>&1; then
    echo "ERROR: $PATCH_DIR is not a git clone — nothing to update." >&2
-    echo "  Re-clone from https://git.onetick.ninja/flan/truenas-truecloud-patch" >&2
+    echo "  Re-clone from https://github.com/sudolulo/truenas-truecloud-patch" >&2
    exit 1
 fi

@@ -136,9 +136,21 @@ fi

 # A dirty tree means someone edited or scp'd files in place; merging over that
 # silently loses their changes, or conflicts halfway through.
-if [ -n "$(git status --porcelain --untracked-files=no)" ]; then
+#
+# CONTENT changes only. A mode-only change (100644 -> 100755) is not somebody's work
+# and must not block an update -- and it is not hypothetical: install.sh chmod +x's
+# these very scripts, so on any version where git recorded one as 100644, INSTALLING
+# dirtied the checkout and update.sh then refused to run. Install once, and updating
+# was blocked forever, with an error telling the user to `git checkout -- .` (which
+# merely undoes the exec bit so the next install can re-dirty it). A real box sat on
+# an old version for exactly this reason.
+#
+# `git diff --numstat` reports "0  0  file" for a mode-only change, so anything with a
+# nonzero insert or delete count is a genuine edit.
+_dirty=$(git diff --numstat HEAD -- . | awk '$1 != 0 || $2 != 0 { print $3 }')
+if [ -n "$_dirty" ]; then
    echo "ERROR: the working tree has uncommitted changes:" >&2
-    git status --short --untracked-files=no >&2
+    printf '  M %s\n' $_dirty >&2
    echo "" >&2
    echo "  Refusing to update over them. Commit, stash, or discard them first:" >&2
    echo "    git -C $PATCH_DIR checkout -- ." >&2
@@ -0,0 +1,3 @@
+version = 1
+revision = 3
+requires-python = ">=3.12"