ersatztv/docs/ci-cd.md

# CI/CD for the ErsatzTV Fork

The fork builds its own Docker image via **Gitea Actions** on the homelab and pushes
to the **Gitea container registry**. Runner + registry were provisioned in
server-management#172; the build pipeline is ersatztv#4; test/prod containers are
server-management#481.

## Hosts (read this before trusting a hostname below)

| Host | IP | Role |
|---|---|---|
| **jazz** | 192.168.1.29 | **Docker host for the media transcoders** — prod `ersatztv` (8409), `ersatztv-test` (8410), Jellyfin. Release scans (`security-scan.sh`) and the prod-copy `migration-smoke.sh` run here. |
| **bumblebee** | 192.168.1.99 | CI runners (`bumblebee-runner`, `small-runner`), plus every other Docker stack. All the memory/lane measurements below were taken here. |
| **ci-runner** | VM 127 (pve4) | The other `ubuntu-latest` CI runner. |

The transcoders moved bumblebee → jazz on **2026-07-20** (server-management#633).

> **Name-reuse trap.** `jazz` was an **earlier name for the .99 host**. Anything written before
> 2026-07-20 that says "jazz" means today's **bumblebee** — resolve hostnames by **IP**, not by name,
> and don't "fix" a historical bumblebee reference into jazz.

## Versioning & releases

The fork inherits upstream ErsatzTV's scheme: **`vYY.<release-seq>.<patch>`** (lightweight, `v`-prefixed git tags).

- **`YY`** — two-digit year.
- **`<release-seq>`** — a sequential release counter **within the year**, reset at each year boundary. It is **not** the calendar month. (Evidence: `v25.2.0` shipped in June 2025, `v25.5.0` in Sep, `v26.3.0` in Feb 2026 — minors don't track months; and `v25.9.0` → `v26.1.0` shows the year-reset.)
- **`<patch>`** — a small follow-up/hotfix on the *same* release line (e.g. `v26.1.0` → `v26.1.1`, days later).

Upstream's final release was **`v26.3.0`** (archived). Our line continues from there:

| Tag | Meaning |
|-----|---------|
| `v26.3.1` | Upstream 26.3.0 **rebuilt on our infra** (Gitea CI/registry, fork ffmpeg base) — **no application changes**. A patch bump, because nothing functional changed. |
| `v26.4.0` | First fork release carrying application changes. Later 2026 releases continue `26.5.0`, `26.6.0`, …; a new year resets to `27.1.0`. |
| `v26.7.0` | Blazor-removal release: ChicoryTV became the only UI. |
| `v26.8.0` | Secured/versioned ChicoryTV SPA + REST API go-live release (#335). |
| `v26.9.0` | Configurable advertised IPTV base URL for M3U/XMLTV (#340) + SPA shell/routing + playouts modularization (#247/#245); on-air/Plex/library-path fixes (#99/#345/#371); coverage + functional-E2E CI (#15/#299). |
| `v26.10.0` | Auto-Tune channel workflow (#69) + weighted content distribution (#70); scheduling refactors, health-check remediation UX (#164), HLS cold-start instrumentation (#350), security hardening (#293/#376/#308). |
| `v26.11.0` | **QSV profiles decode via VA-API** — `QsvPreferNativeDecoder`, default **on**, fixes ~50% channel cold-start failures on Intel (#498); unified logo/on-screen bug via a shared watermark preset (#67). Media-scanner resilience: Jellyfin mixed-content libraries (#489), music-video scan correctness (#488/#494/#497), remote-stream probing before ffmpeg (#473/#480); weighted-distribution SPA (#404). **First release deployed to `jazz`** (server-management#633). |
| `v26.12.0` | **`ErsatzTV.Mcp` MCP server** — read + cautious-write over `/api/v1`, `ERSATZTV_ALLOW_WRITES`-gated (#58). **External channel-logo URLs download + cache at save time** (#525), with the on-screen bug now rendered for external-URL logos (#502). HLS cold-start hardening: burst-read the first segments so start isn't `-readrate`-bound (#350) and floor QSV extra hardware frames so an unthrottled read can't exhaust the pool (#529); remote graphics-engine image fetches bounded — timeout, size cap, decode cap, redirects, pooling (#511). Decision-lifecycle tooling + parallel-orientation startup rewrite (#520/#521); CI `docker build` lane rebalance (#508). |
| `v26.13.0` | **RuleBuilder maturation** — arbitrary-depth group nesting (#436), inline smart-query authoring in Channel Builder (#437), DB-sourced facet typeahead + relative-date operators + validation (#434/#435/#438), and an artist typeahead covering music-video/song credits with `album_artist` no longer 404ing (#578). **Per-channel On Now/Next transient overlay** (#74/#570) and **per-schedule clock-boundary padding** (#392); in-browser channel preview (#60); Auto-Tune per-source weight steppers + exclude/add-untagged (#440). Library-browse pickers now resolve by search instead of a 100-row window, closing several silent at-cap truncations (#644/#650/#651/#634). Correctness: one watermark resolver for all four attachment points, incl. `MiddleCenter` (#503/#510); QSV HDR tonemaps through OpenCL because `vpp_qsv=tonemap` is a silent no-op (#505); `LibraryFolder` unique index + concurrent-insert tolerance (#491); per-library music-video identity with soft trash (#496); Jellyfin Album/Track music-video projection (#177); metadata-collection dedup (#500); accented facet values via a registered Unicode fold on SQLite (#668); `WorkAheadSlots` atomic slot claim, never a negative count (#536/#539); on-demand guide rebuild on thaw (#68). Process/CI: the H10 review-verdict gate became a sha-bound **required** commit status and was hardened through its false-open chain (#622/#629/#632/#648/#649/#672/#698), the decision corpus split to one YAML-frontmatter record per file (#610/#620), and headless Playwright UI-E2E flows landed (#445/#533). Five dual-provider migrations. |
| `v26.14.0` | **Live TV no longer starves on embedded bitmap subtitles** — `-readrate` paces an input off its *furthest-behind* stream, and a PGS/DVD subtitle read through the video's own `-i` is sparse enough to drag the whole process to **0.53x realtime** against the 1.0x a client consumes, draining the buffer until the channel stalls. Fixed with a capability-gated `-readrate_catchup` (ffmpeg 8.0+) on realtime inputs, keeping `-readrate` on the frame-producing path so the `ffmpeg.qsv-extra-hw-frames-floor` bound is untouched; measured 0.533x → 1.067x on QSV and software, with a 240s QSV soak clean of allocation errors (#726). Affects items carrying an embedded bitmap subtitle matching the channel's subtitle mode — 3,182 of 24,646 media versions on prod, and a property of the *item*, not the channel, which is why the stall presented as random. Process/CI: the H10 review-verdict gate's repair sentinel became a fixed point and its write is now fenced on the timeline retarget count, closing a raced-sentinel false-open (#706/#707/#711). **The decisions validator now cross-checks its dependency-free frontmatter parse against PyYAML** and reports both the truncating unquoted `` #`` and the scalar-closing bare apostrophe as errors, so a record whose `rule:` silently halves under PyYAML fails the local gate instead of CI (#674/#688) — the ceiling-calibration claim was also split so the suite pins what the derivation MEANS rather than live-corpus order statistics. Dependencies: CliWrap 3.10.4, JetBrains.ReSharper.GlobalTools 2025.3.5. |
| `v26.15.0` | **`main` became PR-only, which is what makes `review-verdict/h10` load-bearing** — branch protection now carries `enable_push: false` *and* `block_admin_merge_override: true`, so a direct `git push origin HEAD:main` is refused at pre-receive for every account including a site admin, and an admin cannot force-merge past a missing or red required context (#743/#749). Gitea only evaluates `status_check_contexts` on the PR merge path, so before this the whole gate was skippable with no forgery needed. The gate was then hardened through the rest of its false-open chain: CI's registry credential scoped so a head-resolved workflow cannot forge the status (#697), a stray expression delimiter **in a comment** that killed the gate outright and is now banned on the release path itself by a `scan` job `build` depends on (#751/#767), a positive verdict inherited only from an allow-listed reviewer (#742), the PR-timeline `HEAD` alias fenced on the `pull_push` count (#803/#664), and both `/statuses/{sha}` reads paged to a validated terminator (#763). App-facing: **FFmpeg profile numeric fields are bounded with a 422 that names the bound and its consequence** instead of being silently rewritten, and readrate pacing is exposed as a profile field (#735 — one dual-provider migration, `Add_FFmpegProfile_ReadRatePacing`). **On Now / Next gets a boxed background and is on by default** (#732). Field-level progressive disclosure lands as a shared `FieldHelp` trigger + portalled panel, adopted on FFmpeg Profiles and documented as `docs/spa-conventions.md` §15 (#734). Auto-Tune proposal rows align on a grid (#721); `AddItemsDialog` resolves by search instead of windowing a whole media-library type, with stale-response and is-mounted guards (#685/#740). Correctness: a paged handler now counts **the same query it pages**, so a filtered page stops reporting the unfiltered total across playouts, rerun collections and six sibling handlers (#690/#758); nullable primitive collections are guarded at the read site rather than mutated on a tracked entity — `SongMetadata.Artists` in both search indexers (#701/#824) and the scheduling `DaysOfMonth`/`MonthsOfYear`/`DaysOfWeek` columns, where a runtime NULL reads as **unrestricted** (#823); SPA full-replace bodies are built as `Complete<T>` so an omitted optional DTO field is a type error (#807); and every `ErsatzTV.Mcp` tool is pinned to its OpenAPI contract, with `graphicsElementIds` and `padToNearestMinute` declared so an MCP channel update stops silently stripping the overlay (#754/#757). Process/CI: one rule for how a guard derives its population and one for the proof it ships with, both enforced (#774/#775) — populations come from the git index, never a filesystem walk (#806/#819), every `MUTATION` grade is executed each suite rather than asserted in prose (#790), and verification code is itself code under test (#796). Workflows declare their own per-job metadata and guard populations derive from it (#786/#789); all six declare `permissions:` (#748); `ci-image.yml` publishes from `main` only and the persisted checkout credential is gone (#744/#746/#835). Hooks report their own execution and the report is measured (#776), and `docs.no-session-narrative` was generalized to every durable artifact and the sites its sweep IDENTIFIED were remediated as judgements rather than by regex — no closure is claimed over a phrasing space (#784/#812). A ruff config is committed and enforced in CI (#780). Dependencies: Scriban.Signed 7.2.6. **Cut at `736649b3b`, not at `main`'s head**: `completeAnnotations.guard.test.ts` (merged in the very next commit) needs a git index the Dockerfile's `web-build` stage does not have, so every image build after `736649b3b` fails — including a tag build (#887). `736649b3b` is the newest commit whose `Build & push image (amd64)` actually ran green. The scheduling-recurrence fix #880 therefore slips to the next release. |

**Before cutting a release — sweep `docs/decisions.md` + `docs/decisions/`** (ersatztv#521, supersedes
the ersatztv#303 H9 append-only ritual). Supersession/retirement is now a same-PR act (add the new
active record, relocate the predecessor to `docs/decisions/archive/` with reciprocal
`supersedes`/`superseded-by` links), not a release-boundary batch job — most of the old "consolidate"
step is now continuous. The release boundary is instead where you:
1. Run `PYTHONPATH=. python3 scripts/decisions_validate.py` — confirms lifecycle metadata is
   well-formed and every `supersedes`/`superseded-by` link resolves both ways. Since **ersatztv#674**
   it also cross-checks its dependency-free frontmatter parse against **PyYAML when PyYAML is
   importable**, failing on any file PyYAML rejects (a bare apostrophe in a single-quoted value) or
   reads differently (an unquoted ` #`, which YAML truncates as a comment). Where PyYAML is absent —
   the `decisions-guard` job, the Husky hooks — the cross-check is **skipped with a `::notice::`**
   and every other check still runs; the read path stays dependency-free.
2. Confirm every record already classified `superseded`/`retired` actually lives under
   `docs/decisions/archive/` (the validator fails this, but eyeball it at the boundary too).
3. Regenerate the active catalog: `PYTHONPATH=. python3 scripts/build_decisions_catalog.py` and
   commit any drift.
4. Read the corpus size signals. Since **ersatztv#620** these are two separate things:
   - a **per-record prose ceiling** (`decisions_validate.py --record-ceiling <n>`, default **60**)
     — a **non-blocking `::warning::`** naming every record over it. This is the actionable signal:
     it points at a file. The 60 is derived from the distribution, not picked as a round number.
     Its **calibration is guarded in two pieces of different robustness** (ersatztv#688), because
     four earlier single-assertion versions all failed — the first two by being vacuous or
     accepting an absurd ceiling, the last two by ratcheting:
     - **blocking** (`script-tests`) — only the coarse property that the ceiling flags a
       **meaningful minority** of records (`0.02 <= fraction_over <= 0.25`). One record moves a
       fraction by at most 1/N, so no SINGLE ordinary addition can cross it. This is measured
       headroom, not immunity: from today's 18/183 it takes 38 consecutive over-ceiling additions to
       breach the cap, 718 short ones to dilute below the floor, or — the tightest arm —
       consolidating 15 of the 18 offenders away. The floor is
       a fraction rather than "at least one record", which would accept any ceiling up to 229 on the
       live corpus; as a fraction the accepted range is 43..180.
     - **reported, never asserted against the LIVE corpus** — the fine claim that the ceiling sits
       between the **90th and 95th percentile**, i.e. at the tail boundary. `main()` prints a
       `::notice::` when it drifts; the tests assert it only on distributions they own.
       It is an order statistic over a sparse distribution, so a single new record could move p90 by
       21 lines and red the blocking job for whoever wrote it; a ceiling going out of date is
       the passage of corpus growth, not a defect in the commit under test, so it is treated like
       `stale-after`. Re-derive the constant when the notice says so.
   - the **aggregate prose total**, printed every run as an unthresholded `::notice::` **trend**.
     It has no pass/fail. A total over a monotonically growing corpus can only ratchet: the old
     4800→5600 budget went quiet at 5228 after #610 changed the metric and was back over at 5658
     **three and a half hours later the same evening**, with nobody consolidating anything — the
     "permanently red = no signal" failure, not in slow motion at all. It reports record prose and
     non-record scaffolding separately, because they are not the same unit. The generated catalog is no longer counted at all — it
     gains one row per record and cannot be consolidated away.

   **Being listed by the ceiling is an invitation to check for redundancy, not an instruction to
   cut.** A long record that is entirely distinct findings is a legitimate decline — say so in the
   record and move on. (`--budget` is still accepted and ignored, so old invocations keep working.)
5. Report the remaining `legacy-unmigrated` count (the validator prints it as a `::notice::`) so the
   backlog is visible, even though it isn't required to hit zero before a release.
A genuine rationale-prose rewrite still needs a `Decisions-Edit: yes` git trailer on a **non-merge**
commit in the range (see the `decisions.md` header) — routine lifecycle metadata writes above do not.

**Cutting a release:** keep build and promotion as two explicit phases (#335):

1. Confirm `main` CI is green; run the full local gate plus `dotnet list package --vulnerable
   --include-transitive`; then push a `vYY.N.P` tag on that exact `main` commit.

   > **`v*` tags are protected — push as `timothy`.** Since ersatztv#885 the repo carries one
   > tag-protection rule (`name_pattern: v*`, `whitelist_usernames: ["timothy"]`,
   > `release.tag-protection-v-star`), because a `v*` push builds and publishes `:prod`. A push from
   > any other account is expected to be refused — configured, not exercised, since only the
   > `timothy` credential is available here. **The positive half is UNVERIFIED too**: the rule was
   > applied 2026-09-04 and only its negative half was measured (a non-`v*` tag still pushes), because
   > pushing a `v*` tag outside a real cut would publish `:prod`. This cut is its first real
   > exercise. If the tag push is refused, unblock with
   > `DELETE /api/v1/repos/timothy/ersatztv/tag_protections/1`, push, then re-`POST` the rule
   > (prefer a `PATCH` adding the account if the cut has moved to a different operator, so the
   > protection is never left off) — and record the outcome on `release.tag-protection-v-star`,
   > which is waiting for it.
2. Wait for tag CI to build `:prod` + the immutable `:<version>` + `:<sha>` images. Run
   `scripts/security-scan.sh` on jazz against **the immutable `:<version>` image**, not a
   moving tag, and triage every ZAP/semgrep finding.
3. Only after the candidate passes, manually `DeployStack jazz-media` and observe its pre-deploy
   output. Prod's compose deliberately follows floating `:prod` (Timothy's 2026-07-11 decision), so
   no CI push or pin bump is needed.

> **The Komodo stack is `jazz-media`, not `media-servers`** (verified live 2026-07-20 during the
> v26.11.0 cut). The *compose project* is still `media-servers` — which is what the container labels
> show — but the **Komodo stack name** changed with the move to jazz. A stack named `media-servers`
> still exists on bumblebee and is `unhealthy` (the stopped migration leftovers), so
> `DeployStack media-servers` silently targets the **dead** stack. Confirm with
> `/read ListStacks` before deploying.
>
> **A third stack, literally named `ersatztv`, is also live and is NOT prod**: it owns
> `ersatztv-test`, tracks the floating `:latest` tag, and runs with `auto_update: true`. (Stack
> name, compose project, config-file path and image tag re-confirmed from the container labels on
> jazz 2026-08-11; the `auto_update` flags and the bumblebee `unhealthy` status are carried from
> the earlier checks noted above, not re-measured that day.)
> `DeployStack ersatztv` succeeds and looks healthy — it just promotes nothing to prod. Three names,
> one target:
>
> | Stack name | project_name | auto_update | owns |
> |---|---|---|---|
> | `jazz-media` | `media-servers` | `false` | `ersatztv` (**prod**, tracks `:prod`) |
> | `ersatztv` | `ersatztv` | `true` | `ersatztv-test` (tracks `:latest`) |
> | `media-servers` | — | — | dead stack on bumblebee, `unhealthy` |
>
> Don't resolve the prod stack by name at all — resolve it by the **container's**
> `com.docker.compose.project.config_files` label (e.g. via `docker inspect`), which points at the
> actual compose file path and can't be confused by a same-named stack.
>
> **There is no Global Auto Update fallback anymore**: `jazz-media` has `auto_update: false`
> (`poll_for_updates: true` only), so nothing promotes `:prod` on a timer — promotion is manual,
> full stop. The old "don't cut a tag near the 03:00 run" caveat no longer applies.

server-management#585 source-confirmed that Global Auto Update invokes the **same** `DeployStack`
execution as a manual promotion, and extended the #553 pre-deploy hook to detect a floating-tag
**digest** change. Either path now takes the fail-closed prod backup; server-management#589 then
wired `migration-smoke.sh` into that hook, against the exact candidate and the backup it just made.
A backup, fetch, or migration-smoke failure aborts before the live container is recreated. See
`homelab-docs/Docker/ErsatzTV.md` for the operational evidence and rollback procedure.

**Gotcha:** never put a `[skip ci]` token in a commit you intend to tag — Gitea reads skip-ci from the *tagged* commit and will **suppress the release build**. (Also, `workflow_dispatch` on a tag ref isn't supported on this Gitea version, so the tag *push* must do the triggering.) Release commits, and anything you'll tag, must not contain skip-ci.

**Also avoid firing several pushes back-to-back** (e.g. a `[skip ci]` commit, then `main`, then a tag, all within ~1s). Observed once on this Gitea instance: the later events were silently dropped — no `ActionRun` records created at all, even though the runner was online and the workflow `active`. Pushing again, spaced out, created the runs normally. If a push/tag doesn't produce a run, re-push (or push an empty commit) rather than assuming the runner is broken.

## Workflow token scope (`permissions:`, ersatztv#748)

Every workflow in this repo declares `permissions:`. Five declare `code: read` at the top level;
`review-verdict.yml` declares `code: write` + `issues: read` + `pull-requests: read` on its
`set-verdict-status` job and nothing at the top level. That is the whole convention — a new workflow
should declare the narrowest set it needs, and say at the declaration why it needs write if it does.

**Declaring `permissions:` is exhaustive, not additive.** A unit omitted is a unit not granted, and
that holds *at any owner-level default* — it is not conditional on the instance default being
Restricted. This is the property that makes the block load-bearing rather than decorative, and it is
also what makes a wrong one dangerous: the gate job's API calls are `curl -sf` under
`set -euo pipefail`, so a missing unit is a 403 that kills the step, and `review-verdict/h10` is
never posted. That blocks **every** merge to `main`, including the PR that would repair it, and
since ersatztv#743 (`block_admin_merge_override: true`) there is no force-merge escape.

`code: write` and not something narrower because **Gitea has no `statuses` scope**.
`POST /repos/{o}/{r}/statuses/{sha}` is gated by `reqRepoWriter(unit.TypeCode)`, so status-write *is*
Code-write (`ci.actions-credential-scoping`). `contents: write` would also work but additionally
grants `releases: write`, which nothing here uses. The two read units are not spare: the job GETs
`/issues/{n}/timeline` (the ersatztv#706 retarget fence) and, through
`scripts/pr-changed-files.sh`, `/pulls/{n}` and `/pulls/{n}/files` (the ersatztv#698 exemption path).

No workflow needs a `packages:` unit. Since ersatztv#885 the five `container:` blocks in
`docker-build.yml` carry no `credentials:` at all: they run on the `pull_request` route, so they may
hold no stored secret, and the runner pulls the toolchain image ANONYMOUSLY — this registry issues a
pull token for a public package with no credential (measured at the HTTP level 2026-09-04, and at the
daemon level 2026-09-05 — a `docker pull` of the pinned tag with an empty docker config succeeds on
the runner host, which is the same daemon that creates the job containers; see
`ci.pr-route-carries-no-stored-credential`, whose `mechanics:` names both measurements and the two
things neither exercises). So no job pulls that image with the injected
token, and none pulls it with the `REGISTRY_*` PAT either.

### The binding was measured, not assumed (2026-08-27, Gitea 1.27.1)

`ci.actions-credential-scoping` recorded for three weeks that whether `permissions:` is honored on
this instance was **unprobed** — 1.26.0 was the version it was said to start binding at, from
release notes rather than from an experiment. It was settled with the scratch-base probe below, run
as a matched pair that differed **only** in one unit:

| Control | Gate job unit | Result |
|---|---|---|
| Positive | `code: write` | job `success`; `Posted review-verdict/h10-PROBE=success`; `creator: null` |
| Negative | `code: read` | job `failure`; `curl` exit 22 (HTTP ≥400 under `-f`) at the POST; **no status written** |

Both runs reached the classify step's start marker, so neither failure was the ersatztv#751
dropped-step shape. What shows each read unit was exercised is a different observation for each, and
they are not interchangeable:

- **`issues: read`** — both runs logged `Retarget fence: 0 retarget event(s) … (trusted=yes)`, which
  requires the timeline walk to have read `/issues/{n}/timeline`. That log line and the function
  behind it were renamed by ersatztv#803, which added the head axis to the same walk: the line now
  reads `Mutation fence: N retarget event(s) and M push event(s) … (trusted=…)` and the function is
  `count_pr_mutations`. The observation above is left as it was recorded — it is what those runs
  printed — but grep for the new strings when reproducing it.
- **`pull-requests: read`** — the positive control posted `=success`, and a `success` is reachable
  only through `complete=yes`, which requires `scripts/pr-changed-files.sh` to have paged
  `/pulls/{n}` and `/pulls/{n}/files` to a validated end. The fence and the start marker do **not**
  show this; the `success` does.

The negative control is the load-bearing half: it is what shows the declaration *binds* rather than
being a no-op that happens to sit next to a permissive default. Both controls carried the block **at
job level**, on `set-verdict-status` — the same level and the same job as the shipped
`review-verdict.yml`, so the gate's own shape is what was measured.

Both ran while the owner default was still `permissive`, so the binding holds **today**, without any
flip. One scope limit, stated rather than glossed: the other five workflows declare `permissions:` at
**top level**, a form the probe did not exercise. That is inferred, not measured — but the failure
direction is safe (a top-level key Gitea ignored would be a security no-op, never a brick), and
`docker-build.yml` and `pr-checks.yml` are `on: pull_request`, so this change's own CI exercises
theirs.

**What the probe does NOT establish — read this before flipping the owner default.** It says nothing
about what a declared `code: write` receives under a **Restricted** default. GitHub's semantics let
`permissions:` only *narrow*, never widen past the default; if Gitea copies that, Restricted would
cap the gate job at read and `review-verdict/h10` would become unwritable — precisely the failure
this section warns about, with no force-merge escape. Both controls ran under `permissive`, so that
case is untested. server-management#714 stays open for exactly this: re-run the scratch-base probe
**under Restricted** before trusting the flip.

`creator: null` on the positive control matters separately: it is the provenance signal
`review-verdict.yml` relies on to tell a machine-written status from a reviewer's
(`ci.exemption-provenance`, ersatztv#742). Approach A — scoping the built-in token — preserves it.
Moving status writes onto a PAT (approach B) would not, which is why B stays blocked.

**What this does and does not close.** Scoping bounds what a job may *do*; it never bounds whether
attacker YAML *runs* — that needs a trigger filter (ersatztv#744). And `review-verdict.yml` still
holds repo-write. The property gained is not "workflows can no longer assert the verdict"; it is
that *only a workflow visibly opting into repo-write can*, instead of every job on the instance
holding write implicitly. Real, and weaker than closed.

⚠️ **Changing `review-verdict.yml`'s block requires the scratch-base probe**, not CI-green on the
editing PR — that workflow is base-resolved and is never exercised by its own PR, so its own checks
say nothing about the edit. See the probe recipe in the review-verdict gate section below.

The other five split two ways, and only one half is covered by an editing PR's own CI:

- `docker-build.yml` and `pr-checks.yml` run on `pull_request`, so an editing PR exercises their
  blocks directly. (`docker-build.yml`'s `build` job carries
  `if: github.event_name != 'pull_request'`, so its registry-push path is first exercised on the
  post-merge push to `main` — that reddens `main`, not the gate, and is recoverable.)
- `ci-image.yml` (`push` to `main` only since ersatztv#744, plus `schedule`/`workflow_dispatch`),
  `dependency-scan.yml` and `renovate.yml` are **not** exercised by an editing PR at all — their
  first run is after merge or on the next schedule. None is a required check, so a wrong block there reddens a non-blocking run rather than the gate, but do not read an
  editing PR's green as covering them.

**`ci-image.yml` is included, and only just.** It was the one workflow this change could not
originally reach: editing it re-pointed `ci-image-pin`'s `expected` at the editing commit and
reddened a **blocking** job, and its own `paths:` made the edit publish a toolchain image. ersatztv#748 briefly
carried a documented one-name exemption for it. ersatztv#744 (`ci.toolchain-image-publish-is-a-dispatch`)
took that path out of both, so the exemption was **deleted rather than documented** — which is the
form ersatztv#835 asked for, and it is why `test_pr_changed_files.py` asserts over the whole derived
population with nothing carved out.

## Checkouts drop the persisted credential (ersatztv#746)

**Every** `actions/checkout` in this repo sets `persist-credentials: false` — with no exemption,
since ersatztv#744 removed the one that existed (see "The last exemption, and how it closed" below),
and ersatztv#835 now holds the rule mechanically rather than by convention:
`scripts/tests/test_workflow_persist_credentials.py` derives the workflow population from the git
index and asserts the flag on every checkout step it finds. Without it the action leaves an
`Authorization` header in
`.git/config`, and a head-resolved job then hands that credential to every script it runs, not just
to the checkout. That header is write-capable *only while the job's token is*: this instance's Actions default
token permission is `permissive`, where it stands after a Restricted trial was reverted on
2026-08-05 — but since ersatztv#748 that default no longer decides it. Every workflow declares
`permissions:` (see the section above), so every persisted header in this repo is read-only by
declaration. Re-applying the owner-level default is server-management#714. No job that checks out
loses anything: none of them push with git, and anonymous access covers the fetches because the repo
is public — `private: false`, and `GET /info/refs?service=git-upload-pack` answers unauthenticated,
measured 2026-08-26 from the LAN. That measures the repository, not a runner's network path to it,
and the runner path is deliberately not asserted here: it does not need to be, because the fetches
in the jobs below are unmasked: an anonymous fetch that failed from a runner reddens `api-docs`,
`Formatting` and `decisions lifecycle` instead of being swallowed. (The two masked fetches further
down are covered by their own argument, not by this one.) Renovate does push branches, but
`renovate.yml` has no `actions/checkout` step at all and carries its own `RENOVATE_TOKEN`.

The base-ref fetches that feed the changed-file skip logic — in `docs-reminder`, `decisions
lifecycle`, `api-docs` and `Formatting` — therefore run **unmasked**: a failed `git fetch` — and a failed
`git diff` at the three steps that run one inline — exit the step with an `::error::` instead of
yielding an empty changed set. That ordering is the general rule, not a fact about one
change: unmask a dependent fetch *before* you drop a persisted credential anywhere. Under `|| true`
a credential regression looked like "nothing changed" on the three steps that also masked
`git diff`, so those checks skipped themselves silently on a green job; on the other two it handed
a possibly-absent `origin/<base>` to a Python validator. Neither shape is one you want a credential
change to land on top of.

**What this does and does not make fail-closed.** A failed *fetch* now reddens `api-docs`,
`Formatting` and `decisions lifecycle`, which is the credential/network case this change exists for.
It does not make the two validator steps fail-closed in general: they run no inline `git diff`, and
if the base resolves badly despite a successful fetch, `scripts/decisions_validate.py` emits a
`::warning::` and skips its diff-dependent checks while `scripts/check-doc-narrative.py` prints
`SCANNED NOTHING` and exits 0. Both fail open there by their own design and are out of scope here —
do not read a green `decisions lifecycle` as proof its diff-dependent checks ran. Note which side of
that line the `Decisions-Edit: yes` trailer check falls on: `_edit_trailer_armed` is reached only
after the merge base resolves, so it goes quiet with the rest of them rather than surviving.

**Loud is not the same as blocking, and the split is by job.** `api-docs`, `Formatting` and
`decisions lifecycle` gate on this diff, so a failed fetch reddens them. `docs-reminder` does not —
it is a non-blocking nudge, and an advisory red still joins the combined status the merge gate
reads. All three of its non-checkout steps therefore carry `continue-on-error: true` — the parity
nudge, `Set up Python` and the doc-narrative warning — so the `::error::` reaches the log without
turning a warning into a merge block. `Checkout` is deliberately still blocking; with no tree there
is nothing to warn about.

**None of these fetches passes `--depth`.** All five run in jobs that check out `fetch-depth: 0`,
and `git fetch --depth=N` on a *complete* clone grafts it shallow: the history is still there but
cut off, so `git diff origin/<base>...HEAD` fails with `no merge base` once the branch point is
further back than N. Measured 2026-08-26: a 150-commit repo branched 130 back, `--depth=100`,
`.git/shallow` created, three-dot diff `fatal: no merge base`; the same fetch without `--depth`
returns the correct file list. Under the old `|| true` that surfaced as a silent skip; unmasked it
would have reddened `api-docs` and `Formatting` on any branch not rebased recently, sent
`docs-reminder`'s parity nudge advisory-red, and left `decisions lifecycle` green with its
diff-dependent checks silently skipped — so the depth came off with the mask. The same graft,
reached through the detector script rather than a workflow step, is what stamped `:latest` images
`InformationalVersion 0.0.0-<sha>` until ersatztv#836 — see below.

**Two masked fetches that stay masked, deliberately.** `scripts/ci-detect-docs-only.sh` and
`scripts/ci-detect-already-validated.sh` both swallow a failed fetch — and should. In both, a failed
fetch can only leave the base revision unresolvable, and an unresolvable base emits `false`: *run
the full matrix / full validation*. Neither can turn a failed fetch into a skip. Masking there costs
runner time and nothing else, which is the opposite direction from the workflow steps above, where
the same masking bought a silently skipped check. Judge these by which way the failure falls, not by
the `|| true`.

That clears their **masking** and nothing else. Their `--depth` was a separate question with a
separate answer, and it is the one that shipped a defect (ersatztv#836).

**A depth is passed only to a checkout that is already shallow.** `ci-detect-docs-only.sh` routes
both its fetches through `fetch_ref`, which asks `git rev-parse --is-shallow-repository` and drops
the `--depth` on any answer but `true`. The script has **exactly four consumers** — `test`,
`migrations` and `functional-e2e` at `fetch-depth: 2`, and `build` at `fetch-depth: 0`. Enumerate
them with `grep -rn ci-detect-docs-only .gitea/`, not by looking for a step with `id: detect`:
`api-docs` and `format` have one of those too, and it runs their own inline diff.

So `build` is the only complete-clone consumer, and the push arm is the only arm it takes. That is
where the cost landed: its `--depth=2` grafted `build`'s complete clone, the `git describe --tags`
in the immediately following step found no reachable tag, and a `|| echo v0.0.0` fallback that
cannot fail turned that into a version. Every `:latest` image published from `main` carried
`InformationalVersion 0.0.0-<sha>` instead of `26.x.y-<sha>` — read off the live test container as
`0.0.0-5a7a4ed7` on 2026-08-26, six weeks after the depth landed with #416 on 2026-07-17, because
nothing went red.

The PR arm's `--depth=200` had **no complete-clone consumer at all** (on `pull_request` only the
three shallow jobs run this script), so it was inert rather than latently firing. It goes through
the same helper anyway, because the clause is shared and the next `fetch-depth: 0` consumer of that
arm would inherit the graft silently. Keeping the depth for the shallow three is likewise
conservative rather than required: a depth-less fetch into a `--depth=2` clone leaves `.git/shallow`
in place, still writes `FETCH_HEAD`, and still reports the parents the push arm reads.

Both halves are now held by something that executes. `scripts/tests/test_docs_only_detector_clone_depth.py`
drives the real script over real `file://` clones and asserts the clone is still complete and
`git describe` still resolves afterwards, with a negative control that requires the fixture to be
able to graft at all — a plain path clone silently ignores `--depth`, which would make the whole
file vacuous. And `Compute version and tags` no longer defaults: a failed `git describe` on the
non-tag path fails the job rather than stamping `0.0.0`, so no `:latest` is published instead of a
mislabelled one. The tag path never calls `describe`, so a release cut is unaffected either way.

### The last exemption, and how it closed

`ci-image.yml`'s checkout was the one left unset, and the reason was mechanical rather than a
judgement that it was safe: any commit touching `.gitea/workflows/ci-image.yml` re-pointed
`ci-image-pin`'s `expected` at itself and reddened that blocking job, and the file's unfiltered
`push` trigger published a toolchain image from whatever branch the edit landed on. Both are gone
(ersatztv#744): the trigger is `branches: [main]` and the file is no longer one of `expected`'s
paths, so the flag went on in the same change. **What this costs the guard is nothing, and that is
the point** — ersatztv#835 was deliberately held back rather than shipped with a one-entry
`ci-image.yml` exemption, because such an entry outlives its reason silently: once the flag went on
there the exemption would still have passed and the guard would have been permanently blind to the
file it was written for.

## Per-job declarations (`env.CI_EXECUTION_CLASS`, `env.CI_JOB_ROLE` — ersatztv#786/#789)

**Every job in every tracked workflow declares `CI_JOB_ROLE`; the jobs in `docker-build.yml` also
declare `CI_EXECUTION_CLASS`** (the scopes differ — see the table). They are metadata, not runtime
inputs: nothing reads them at execution time. They exist so that a guard comparing sets of jobs
derives its expected set from the workflow instead of from a literal list in the checker.

| Marker | Values | Scope | Derived by |
| --- | --- | --- | --- |
| `CI_EXECUTION_CLASS` | `toolchain`, `bare-runner` | `docker-build.yml` only — the sole workflow using the toolchain image | `scripts/tests/test_ci_image_pin_population.py` |
| `CI_JOB_ROLE` | `guard`, `report-only`, `none` | every tracked workflow | `scripts/tests/test_workflow_job_guards.py` |

**A missing marker and an unrecognised value both fail the suite.** Neither defaults: a scheme whose
absent value reads as some safe class stops applying the moment a job is added and nobody notices.

**`CI_JOB_ROLE` is decided by what the job PRODUCES.** A `guard` job's output is a **verdict**; a
`none` job's output is an **artifact**, and a red there means the build did not work rather than that
an invariant was violated. `docker-build.yml::build` publishes an image and smoke-tests it, and is
`none`. Which jobs are `none` is read off the markers rather than listed here — a list in prose is a
second copy of the workflow that rots on the next job added. Every `guard`/`report-only` job holds a
row in `docs/guard-inventory.md` → "Workflow-job guards", cross-checked both ways.

`report-only` is the third value: a verdict the job **cannot enforce**, because every check step
carries `continue-on-error: true`.

**A rejected alternative, so it is not re-adopted:** "a guard enforces an invariant about the
*repository*, so a job exercising the *product* is `none`" reads as more principled and is wrong —
it puts `test` and `migrations` outside the population, and those are the two **required** status
contexts on `main`, i.e. precisely where a failure to fire is fail-open against branch protection.

**Why not a literal list**, since one stood here until #789 with a real argument behind it: set
equality between two *derived* sets is blind to a member leaving both at once, so a job that loses
its `container:` block leaves the declared and the pinned set together and the comparison stays
balanced. That needs an anchor that does not move with the block — and a reviewed literal was the
only one available. The marker is the same anchor, better placed: it lives in the job it describes,
so it is reachable by anyone reading that job and it moves with the job when the job is renamed.
(Not "in the same hunk" — measured, the marker sits 6–7 lines from the `container:` block in four
jobs and **58** in `migrations`, where the `services:` block sits between them. Proximity is a
convenience; the property the guard rests on is that the marker does not move when the block does.)

**Its cost is proximity**, so the scheme does not stop there. A literal in a distant file survives a
careless workflow edit; a marker a handful of lines from the `container:` block can go with it. So
`test_no_job_needs_the_toolchain_without_declaring_it` derives the requirement a third time — from
each job's own step bodies, looking for tools that exist only in the CI image (`dotnet`,
`dotnet-ef`, `node`, `npm`, `npx`, `ffmpeg`, `ffprobe`, `playwright`, `reportgenerator` — the list
lives in `TOOLCHAIN_ONLY_TOOLS`, and this sentence is a summary of it, not a second copy to keep in
sync). `git`/`jq`/`python3`/`curl`/`tar` are **excluded** because they do not discriminate: the
bare-runner lanes provide or provision them too. That check depends on neither the marker nor the
block, and it is the only one that catches a .NET step **moved** into a bare-runner job, where no
set changes at all.

It is a **necessary condition only**: a job whose *only* toolchain use sat behind a script would be
invisible to it. Measured — the blind-spot set is **empty today**: all five declared toolchain jobs
are detected directly, `functional-e2e` included (it runs `dotnet restore` and `npm ci` in its own
steps before handing off to `scripts/e2e-local.sh`). The converse is still not asserted, because a
future job could be written that way and a red on a correct tree is how a guard gets deleted.

**The residual, stated because it is the one that matters.** All three checks go BLIND together
(none of them reddens) under one plausible edit: drop the `container:` block, flip the marker, *and* move the invocation into a
script. Both set comparisons then stay balanced and the step scan sees no token. What bounds this is
the failure **mode**, not its likelihood — the job runs a missing binary and dies with
`dotnet: command not found`, loud and immediate, where ersatztv#774's original defect sent a
*required* check green on the bare runner. Trading a silent pass for a noisy crash is the point;
it is not a claim the hole is closed.

Full rationale: `docs/decisions/records/testing/workflow-declares-its-own-job-metadata.md`.


## The workflow: `.gitea/workflows/docker-build.yml`

Single workflow. Gating jobs `test` + `migrations` run in parallel and gate `build`; a
non-blocking `docs-reminder` job runs on PRs only (see below). Prod deploy is **not** a CI
job — it's Komodo Global Auto Update off the `:prod` tag (see "Cutting a release").

### Triggers & tags

| Trigger | `test` job | `build` job | Image tags pushed |
|---------|:----------:|:-----------:|-------------------|
| `pull_request` | ✅ | — (skipped) | none |
| push to `main` | ✅ | ✅ | `:latest` + `:<short-sha>` |
| push tag `v*` | ✅ | ✅ | `:prod` + `:<version>` + `:<short-sha>` |
| `workflow_dispatch` | ✅ | ✅ | only if ref is `main`/`v*`, else build-only (no push) |

A **docs-only** change (see "Docs-only skip" below) reduces every ✅ above to a seconds-long no-op
that still reports its status.

`:latest` is the **test/dev** channel (every `main` commit). Prod's compose **follows the
floating `:prod`** tag (reverted from the 2026-07-07 version pin on 2026-07-11) — never
`:latest`. Both `:prod` and `:<version>` are produced by pushing a `v*` tag; prod tracks
`:prod` and is redeployed by Komodo Global Auto Update (see "Cutting a release"). The
immutable `:<version>` tags remain for reproducible rollback (`docker run …:26.6.0`).

Concurrency is scoped **per event+ref** (`group: ersatztv-build-${{ github.event_name }}-${{ github.ref }}`,
`cancel-in-progress` for PRs): PR runs parallelize across PRs, a new sync auto-cancels
its superseded run, and image builds still serialize within their own ref. Do NOT push
`main` and a `v*` tag simultaneously — those are separate groups but share the
`:buildcache` tag and the smoke container name; tag only after the main build is green.
(History: originally one global group serializing ALL runs for the single runner —
with three runners that starved the queue; changed 2026-07-11, server-management#574.)

**Four** runners serve the fork (server-management#570/#574/#639):

| Runner | Host | Label | Slots | Per-job cap |
|---|---|---|---|---|
| `ci-runner` | VM 127 (pve4) — no prod workload | `ubuntu-latest` | 4 | `--cpus=4 --memory=10g` |
| `bumblebee-runner` | bumblebee — prod media | `ubuntu-latest` | 2 | `--cpus=4 --memory=10g --cpu-shares=256` |
| `small-runner` | bumblebee — prod media | **`small`** | 2 | `--cpus=1 --memory=1g --cpu-shares=256` |
| `jazz-small-runner` | jazz — prod media (#633) | **`small`** | 2 | `--cpus=1 --memory=1g --cpu-shares=128` |

The `small` lane exists because Gitea dispatches a job as a runner task **even when its
`if` skips it**, and those skip-tasks used to wait behind long builds (observed 31 min),
stalling every PR run. `--cpu-shares` below the default 1024 is what makes a runner on a
prod media host acceptable: under contention CI loses to the transcoders (ersatztv 1536 /
jellyfin), which are the reason those hosts exist.

That same "dispatched even when `if` skips it" behavior is why the git-only PR gates live in
their own **[PR gates workflow](#pr-gates-workflow)** (`pr-checks.yml`, `on: pull_request`)
rather than in `docker-build.yml` — see that section (ersatztv#535).

**`small` is git-only, and that is load-bearing (server-management#639).** Everything in
the lane is a checkout plus a `git diff`: `decisions-guard`, `ci-image-pin`,
`docs-reminder` — plus `prove-fix` (ersatztv#794), the heaviest member: per commit carrying a
`Proves:` trailer it makes two `git worktree add`s and runs an arbitrary pytest selection twice,
so a PR claiming many proofs costs proportionally more than the rest of the lane combined — plus
`script-tests`, which is a checkout plus a `pytest` run needing only
`pytest` and `pyyaml` (ersatztv#631; it is NOT stdlib-only — that assumption is what turned the
job red on its first CI run, see below) — plus **`scan`** (ersatztv#767), the same
lightweight-Python shape, and **`toolchain-preflight`** (ersatztv#772), a checkout plus one `curl`.
Those last two are the lane members that live in `docker-build.yml` rather than `pr-checks.yml`, so
they are the ones that also run on a **tag push**. `scan` is still the one to think hardest about
before changing anything here: its failure does not merely redden a status but **skips `build`** —
an OOM or a wedge there yields no release image at all. `toolchain-preflight` is a `needs:` of nothing, by
design — it does not gate the jobs it diagnoses. It is not consequence-free either: like any red
job it lands in the PR's combined status, which the merge gate reads (see "When the pinned tag
disappears").
Nothing there runs a compiler or a `docker build`, which is why the lane
can be capped at 1 GiB per job. The lightweight-Python jobs are the deliberate edge of the
"git-only" rule, not an exception to it: `setup-python` + `pip install pytest` + a suite whose
heaviest allocation is a handful of temp-dir git repos stays far under the cap. Route a heavy job here and it will OOM — give it
`ubuntu-latest`, or its own label on `ci-runner`, the only host with no prod workload.

**Lane assignment (ersatztv#390).** *Slot counts below are as-of 2026-07-17; the table above is
current.* At the time, the `ubuntu-latest` lane had **4 slots** (2 + 2) and the
`small` lane 4. A 2026-07-17 audit of the Actions API found the `ubuntu-latest` lane
saturated and the `small` lane idle — **queue wait exceeded every job's runtime**:

| Job | Runtime | Queue wait | Lane |
|---|---|---|---|
| `test` | 354s | 1363s | ubuntu-latest |
| `migrations` | 639s | 1428s | ubuntu-latest |
| `functional-e2e` | 520s | 1447s | ubuntu-latest |
| `api-docs` | 5s | **1722s** | ubuntu-latest → `small` → **reverted to ubuntu-latest** (#406) |
| `format` | 37s → **~0.5s** (#469) | **1731s** | ubuntu-latest → `small` → **reverted to ubuntu-latest** (#406) |
| `docs-reminder` / `decisions-guard` | 10s | **5s** | small |

`api-docs` and `format` moved to `small` because the queue wait dwarfed their runtime. Both lanes
run the *identical* `runner-images:ubuntu-latest` base, so `small` was a label with spare capacity,
not a different capability — a move only possible because those jobs now run in the **CI toolchain
image** (below) and no longer need the runner image to supply .NET/Node.

**Reverted 2026-07-17 (ersatztv#406 / server-management#604).** #390's own caveat — *"on an
API-touching PR `api-docs` does a full `dotnet build`, so it is not always small"* — turned out to
be the deciding factor, and "capacity 4 absorbs that" held only because **nothing enforces the sum**
of the lanes' per-job caps. Each job container is correctly capped (`--memory=10g`), but 6 slots ×
10 GiB = **60 GiB on a 25 GiB host** that also runs prod media; on 2026-07-17 bumblebee hit load
340 with 21 GiB swapped. These were not small jobs — a live `docker stats` caught the `format` job
container at **3.95 GiB**, which the re-sized 2 GiB `small` lane would OOM-kill outright. #604 fixes
the queue at the source instead (`ubuntu-latest` grown to 5 slots: a 48 GiB ci-runner at capacity 4
plus a bumblebee overflow slot), so the `small` lane can be reserved for genuinely-tiny shell jobs.
(The `format` half of this is now **moot**: **ersatztv#469** moved it to `dotnet format whitespace .
--folder`, which loads no Roslyn workspace — the job's 3.95 GiB heap and multi-minute runtime are
gone, so it is no longer a reason to keep the lane large. `api-docs` on an API-touching PR still is.)

Queue wait is still a dominant cost and capacity is server-management's boundary — tracked in
**server-management#604**. The redundant triple-build behind those runtimes is **ersatztv#398**.

**The other failure mode: setup-phase starvation (server-management#639, 2026-07-20).** The
table above measures *queue wait* — time before a job is dispatched. A saturated lane also
produces a second, much more confusing symptom: a job that *is* dispatched, sits `in_progress`
for >10 minutes, writes **no log file at all** (`OpenLogs … .log.zst: file does not exist`),
and then fails — wedged in act's job-**setup** phase, before Checkout. Same-config siblings
that started 90s earlier finished in seconds; a concurrent job's log showed a normally-fast
compile taking a 7-minute gap between projects. This is the origin of the "`decisions.md` is a
known flake, just rerun it" folklore: the rerun succeeds only because it lands after load
clears, so the guard's logic gets blamed for a capacity problem.

The fix was **not** more capacity for its own sake. `small` was stuck at one slot because it
still held two heavy jobs — `docker-build.yml`'s image build and `ci-image.yml`'s toolchain
buildx (the latter reads as lightweight because it is "docker-only", but it is the heaviest
thing that ran in the lane) — and their 10 GiB requirement set the lane's per-job cap, which
on a 25 GiB host permits exactly one slot. Moving both to `ubuntu-latest` made the lane
genuinely tiny, so it could widen to **4 slots across two hosts while committing less RAM to
CI than the single slot did**. `docker-build.yml`'s `build` does not re-create #574's
skip-task queueing, because `needs: [test, migrations, scan]` means it cannot be dispatched until
the lane it would queue behind has already drained.

### CI build memory: no persistent compiler servers (ersatztv#406)

Roslyn's `VBCSCompiler` is a **persistent** compiler server — it outlives the `dotnet build` that
started it and keeps its heap warm for the next one. Locally that is a genuine speedup; in CI it
buys nothing, because each job container is torn down at the end of the run and there is never a
"next build" to warm. It was measured at **7.8 GB RSS** on bumblebee — the single largest consumer
on the host, and the reason each job needed a 10 GiB cap in the first place.

So the workflow's top-level `env:` disables the servers for every runner-side dotnet job:

| Variable | Effect |
|---|---|
| `UseSharedCompilation=false` | no persistent `VBCSCompiler`; `csc` runs per project and exits |
| `DOTNET_CLI_USE_MSBUILD_SERVER=0` | no persistent MSBuild server process |
| `MSBUILDDISABLENODEREUSE=1` | MSBuild worker nodes exit with the build instead of lingering |

These are MSBuild properties set as environment variables so they apply to every `dotnet`
invocation without touching each call site (MSBuild surfaces env vars as properties, and
`UseSharedCompilation` is only defaulted to `true` when empty, so the env var wins).

**What this does and does not shrink.** It helps the jobs that *compile* — `test`, `migrations`,
`api-docs` on an API-touching PR, and the in-Docker `build`. It never helped `format`: `dotnet
format` loaded Roslyn in-process via MSBuildWorkspace and never spawned `csc`, so the compiler-server
env vars left its measured 3.95 GiB untouched. (Moot since **ersatztv#469** switched `format` to
`dotnet format whitespace . --folder`, which skips the MSBuild/Roslyn workspace entirely — the job is
now a ~0.5 s, low-memory whitespace/BOM check with no Roslyn heap. See *Static analysis &
formatting → Formatting* below.)

The same three are repeated in `dependency-scan.yml`; workflow `env:` does not cross workflow files.
That one is lower-stakes (restore/list are MSBuild-driven, so it's lingering worker nodes rather
than a 7.8 GB VBCSCompiler) but it runs unattended on a cron against the prod media host.

**The workflow `env:` does not reach the `build` job's compilation**, which happens inside
`docker build` — the same three are set as `ENV` in the SDK stage of `docker/Dockerfile`. That is
the job server-management#570 measured pegging 5.999/6 GiB, so it is the one that most needs this.
Build-stage only; the final image is `FROM runtime-base`, so nothing lands in the shipped image.

The `test` job samples its own memory and reports it every run. Two steps (`continue-on-error`, so
they never fail a build), driven by `scripts/ci-peak-anon.sh` (ersatztv#412):

- **Start peak-anon sampler** (before the dotnet Build/Test/Coverage steps) launches a detached
  background poller that tracks the high-water mark of the cgroup's **anon** memory every 2 s.
- **Report peak container memory** (the job's last step) stops the sampler and prints the sampled
  **peak anon** — the headline number — alongside `memory.peak` and the end-of-job `anon`/`file`
  split, to the log and the job step summary.

Read the **peak anon** off a recent run to size a cap — not `memory.peak`, and here is why:

> ⚠️ **`memory.peak` is not "peak RSS".** It is the high-water mark of `memory.current`, which
> charges **page cache** to the cgroup as well as anonymous memory. Proven on bumblebee: a
> container with `anon=0` that merely reads an 800 MB file reports `memory.peak=826 MiB`, of which
> `file=800 MiB`.
>
> This matters because the naive reading **inverts the decision**: page cache is *reclaimed* under a
> tighter cap, not OOM-killed, so a big peak that is mostly `file` is **not** evidence that the cap
> must stay high. `anon` is the part that actually forces an OOM. **Size caps on peak `anon`, not on
> `peak`.**
>
> Why a *sampler* and not just the end-of-job split: the kernel exposes `memory.peak` (peak of
> anon+cache) but has **no peak-anon counter**, and the end-of-job `anon` is the composition *then*,
> not at the peak instant — a job that peaks mid-`dotnet test` and then frees reports a misleadingly
> low `anon`. The 2 s background sampler catches the true peak-anon instant; `memory.peak` and the
> end-of-job split stay in the report as a cache-inflated ceiling and a reference. (Before #412 the
> instrument printed only `memory.peak` + the end-of-job split — see #411.)

**The compiler-server A/B (ersatztv#412).** Measured on bumblebee in the CI toolchain image,
swap-off (`--memory-swap == --memory`), full-solution `dotnet build --no-incremental`, peak anon
sampled by *this* instrument, servers shut down between arms (n=2 each, interleaved):

| Arm | shared-compilation env | peak anon (2 runs) | resident after build |
|---|---|---|---|
| **OFF** (the CI config) | disabled | **5818 / 5854 MiB** (~5.84 GiB, tight) | none |
| **ON** (dotnet defaults) | enabled | **6305 / 7604 MiB** (~7.0 GiB, noisy) | ~3 GiB `VBCSCompiler` |

Two things are solid: OFF is **consistently ~5.84 GiB** and ON is **always higher** (mean delta
~1.1 GiB, up to ~1.8 GiB), so disabling the servers is worth it; and ON leaves a ~3 GiB
`VBCSCompiler` **resident** after the build — the host-between-jobs cost #406 removed. Don't read a
*precise* delta into the ON peak: it is noisy because a parallel build's peak depends on how many
`csc`/project compilations overlap at the peak instant.

**#406's premise — *"if disabling shared compilation brings peak RSS well under 6 GiB, the whole
budget loosens"* — is NOT supported.** OFF sits at ~5.84 GiB for the *build phase alone* — right at
the 6 GiB line, not well under it — and the `test` job adds `dotnet test` + coverlet +
reportgenerator on top. Disabling the compiler servers stays right (consistent reduction, no 3 GiB
resident server) but **do not bank a looser cap budget on it**: size the cap off the **live
test-job peak anon this instrument now reports** (build + test + coverage), not off this build-only
A/B.

> The earlier PR #411 probe (`peak 9457 / anon 7134 / file 421 MiB`) read *higher* than these
> swap-off sampled numbers. Swap settings and read-method (end-of-job snapshot vs sampled peak) move
> these figures by >1 GiB (#406), so treat the committed instrument's sampled peak-anon as
> authoritative and that probe as superseded.

What *is* established: no persistent compiler server survives a build, `migrations` is green with
mysql capped at 2g swap-off, and the `test` job now self-reports a true peak-anon every run.

### `services:` containers are capped explicitly (ersatztv#406)

A runner's `container.options` (`--cpus=4 --memory=10g`) applies to the **job container only, not
to `services:`**. Verified on a live `migrations` job: the job container reported
`HostConfig.Memory=10737418240`; its `mysql:8.4` service reported `mem=0 nanocpus=0` — unbounded.
So `migrations` runs added an uncapped MySQL to an already-tight host.

The mysql service now sets `--memory=2g --memory-swap=2g --cpus=2`.

**`--memory-swap` is the part that matters, and it is easy to get wrong.** Docker defaults an unset
`--memory-swap` to **twice** `--memory`, so `--memory=2g` *alone* grants 2g RAM **plus 2g of swap**.
Verified on bumblebee:

| Options | `memory.max` | `memory.swap.max` |
|---|---|---|
| `--memory=2g` | 2147483648 | **2147483648** ← 2 GiB of swap |
| `--memory=2g --memory-swap=2g` | 2147483648 | **0** ← swap disabled |

Setting `--memory-swap` equal to `--memory` disables swap for the container. On this host that is
the whole point: swap thrash is what took prod down, and a swapping mysqld mid-DDL is precisely the
pathology behind the known `Command Timeout expired` migrations flake. **Prefer a loud OOM over
silent swapping** — an OOM is a clear signal to raise the cap; swapping just degrades everything.

⚠️ **The same 2× applies to the runners' `container.options: --memory=10g`** — each job slot is
really 10 GiB RAM *plus* 10 GiB swap. The "6 slots × 10 GiB = 60 GiB on a 25 GiB host" framing
understates the promise by 2×, and it is a plausible direct mechanism for the incident's 21 GiB of
swap. Fixing that is server-management#604's call (reported there).

**On the 2g figure, honestly:** a `mysql:8.4` container with this exact env peaked at **543 MiB**
during init and settled at 481 MiB idle (probed on bumblebee, 2026-07-17) — but that is init+idle,
**not** the 787-migration replay, which grows table/definition caches idle never touches. So 2g is a
measured *floor* plus headroom, not a measured ceiling; the `migrations` job going green is what
validates it. `--cpus=2` has no measurement behind it at all — 787 sequential DDL statements on one
connection are ~1-core-bound, so it is judgement; revisit if the apply step's tail latency grows.

**Any new `services:` container needs its own explicit cap** — it will not inherit one, and it needs
`--memory-swap` set alongside `--memory` or it silently gets 2× in swap.

### `test` job

`dotnet restore` → strip the Scanner project ref (`sed -i '/Scanner/d'`, matching the
Docker build) → `dotnet build -c Release` → `dotnet test -c Release --no-build`. Gates
the image build.

- **Code coverage** (ersatztv#15): `dotnet test` runs with `--collect:"XPlat Code Coverage"
  --settings coverlet.runsettings --results-directory ./coverage`, so `coverlet.collector`
  (referenced by every `*.Tests` project) emits a Cobertura report per project. A follow-up
  **Coverage summary** step merges them with ReportGenerator (`TextSummary` to the log,
  `MarkdownSummaryGithub` to the job step summary). No floor is enforced yet ("decide on a
  floor later" — #15); the step is `continue-on-error: true`, so a missing report or a
  transient tool install never blocks a build.
  - **`coverlet.runsettings` excludes generated EF migration code** (`**/Migrations/*.cs`,
    ~2.59M generated lines vs ~200k authored). Instrumenting it OOM-killed the shared `test`
    job (exit 137); excluding it cuts the instrumented surface ~126× (2.5M→20k coverable
    lines in the whole-solution `Architecture.Tests` process) and makes the percentage reflect
    authored code.
- **Shallow checkout**: `fetch-depth: 2` — this job never runs `git describe`/`git log`, and only
  `build` needs full history/tags for version computation, so `test`, `migrations` and
  `functional-e2e` all check out shallow. It was 1 (ersatztv#190) until ersatztv#420's revalidate
  step needed `HEAD^2` to resolve on a main merge commit. `build`'s checkout stays `fetch-depth: 0`.
- **NuGet package cache**: both `test` and `migrations` cache `~/.nuget/packages` via
  `actions/cache@v4`, keyed on `hashFiles('Directory.Packages.props', 'global.json')` with
  a `restore-keys` OS-level fallback (ersatztv#190). Avoids a from-scratch `dotnet restore`
  on every run; the key only changes when the central package manifest or SDK pin changes.

### `build` job

1. Compute `INFO_VERSION` (`git describe` + short sha on `main`; tag version on `v*`).
2. `docker/setup-buildx-action` with `buildkitd-config-inline` setting `http = true`
   for `192.168.1.95:3000` — **BuildKit does not inherit the host daemon's
   `insecure-registries`**, so without this, cache/base-image/push over the HTTP
   registry fails (`http: server gave HTTP response to HTTPS client`).
3. `docker/login-action` with repo secrets `REGISTRY_USER` / `REGISTRY_PASSWORD`. **`build` is the
   ONLY job in this workflow that names them** (ersatztv#885,
   `ci.pr-route-carries-no-stored-credential`): it is gated `if: github.event_name != 'pull_request'`,
   and every job that IS on the PR route now reads the registry and the status API anonymously,
   because a `pull_request` run executes head-supplied YAML.
   **`REGISTRY_PASSWORD` is a scoped PAT (`write:package` + `read:repository`), not an account
   password** — deliberately, so head-resolved PR code cannot use it to forge a commit status
   (`ci.actions-credential-scoping`, ersatztv#697). If a job ever fails with `token does not have at
   least one of required scope(s)`, the fix is to narrow what the job does, **never** to widen the
   token to `write:repository` or to put the admin password back. Note what the scope still reaches:
   `write:package` covers `ersatztv:prod` (the tag prod's stack follows) and `ersatztv-ci:<sha>` (the
   toolchain image five `container:` jobs execute), so this is the deployment supply chain, not an
   inert endpoint — see `ci.actions-credential-scoping`.
4. `docker/build-push-action@v6`: amd64-only, `docker/Dockerfile`, `INFO_VERSION`
   build-arg, registry layer cache (`type=registry,ref=…:buildcache`,
   `cache-to … ignore-error=true`).
5. **Smoke + IPTV E2E test**: pull the just-pushed `:<sha>`, run it, poll for HTTP
   readiness (`docker exec … python3` → `http://localhost:8409/`), then assert the real
   Jellyfin-facing surfaces on the freshly built image (ersatztv#16): `/iptv/channels.m3u`
   returns 2xx containing `#EXTM3U`, and `/iptv/xmltv.xml` returns 2xx containing a `<tv`
   root. `xmltv.xml` needs `channels.xml` (written by the scheduler a few seconds after
   boot), so each endpoint is polled with a deadline. Unique container name + `trap … EXIT`
   cleanup; dumps container logs on failure. Catches routing / base-URL (#1) / migration
   regressions that leave the app "up" but serving broken output.

### `functional-e2e` job (advisory; PR + main)

Boots the app **from source** and drives the manual live-E2E curl flows sessions have historically
re-run by hand, turning them into a CI regression net (ersatztv#299). It is the automatable half of
`docs/e2e-local.md`; every script it chains runs identically locally and in CI:

1. `npm ci` + `npm run build` (SPA), `dotnet build ErsatzTV.sln -c Release`. (ffmpeg comes from the
   CI toolchain image; the old `apt-get install ffmpeg` step is gone — see below.)
2. `ETV_BUILD_CONFIG=Release scripts/e2e-local.sh <fresh-config>` — copies `wwwroot`, launches
   `dotnet ErsatzTV.dll` in the background (logging to a file so the launch step returns once the app
   is ready), prints `PID`/`CONFIG_DIR`.
3. `scripts/e2e-functional.sh http://localhost:8409 <config>` — asserts. Mostly curl-only: the
   **legacy→SPA redirect sweep** (+ the `/api`, `/artwork` never-redirect exemption), the
   **auth/CSRF/security-stamp** flow (setup-claim → read-gate 401/200 → re-claim 409 → CSRF 403 →
   login 401/200 → logout 403/204 → post-logout stamp-revocation 401), the **library-scan status
   contract** (404 unknown / 202 queued / `scan-status` 200), and the **If-Match/412** round-trip on
   `rerun-collections`. Since ersatztv#363 (extended by #444) it also asserts three **lock-contention
   409s** that aren't curl-only — see below. A `trap` kills the instance on step exit.
4. `ETV_UI_PORT=8410 scripts/e2e-ui.sh` — boots a second, **fresh** instance and runs the headless
   Playwright UI flows (ersatztv#445). See "UI-E2E step" below.

**Lock-contention 409s (ersatztv#363).** The harness now seeds rows the API can't create — a
`LibraryPath` and a Jellyfin media-source — directly into the running instance's SQLite DB (via
python3's stdlib `sqlite3`, whose busy-timeout retry serializes behind the app's writer), and synthesizes media with the
image's ffmpeg, to exercise two `IEntityLocker` contracts **deterministically** (it only fires the
racing request once the lock is *provably* held, never a sleep-and-hope): (a) the **library-scan
"already scanning" 409** — seed ~60 tiny clips into the built-in Shows library so the scanner
subprocess runs a few seconds, poll `GET /libraries/scan-status` until the library shows active (that
window is a strict subset of the scan lock's held window), then a second `POST .../scan` is a 409
(deterministic bar a tiny residual TOCTOU gap the multi-second scan covers); (b) the
**external-collections "already scanning" 409** — the per-family lock is taken
synchronously *before* the 202, so the 202 proves it held, and pointing the seeded source at a
non-routable address keeps the background sync hung so the window stays open; and (c) the
**playout-build "build in progress" 409 + `isLocked` projection** (#215/#444) — a build is enqueued
onto the single-consumer `WorkerService` and the trigger returns *before* the handler locks, so poll
`GET /playouts/{id}` until `isLocked:true` (not the accepted trigger), then fire. Seed a Classic Flood
schedule over a few short episodes and crank `PlayoutDaysToBuild` (`playout.days_to_build`) so one
build is wide enough to observe (~5 days ≈ 43k items ≈ ~1s locally, wider on slower CI); assert `PUT
/playouts/{id}` 409, `POST .../playout/reset` 409, the list `isLocked:true`, then post-build the same
`PUT` 200s. No new CI step or dependency: ffmpeg ships in the toolchain image, and python3 was already
a harness dependency (`json_field`). The scan + build flows self-skip (advisory) if ffmpeg is ever
absent, and the build flow also self-skips if the build is never observed locked (never asserts an
unproven race).

**Advisory, by design** (the issue's "keep it a separate job so a functional-E2E flake can't block the
unit-test gate"): it is **not** a `needs:` of `build` and **not (yet) a required check**, so a flake
blocks nothing. Promote it to a required check / `build` dependency once it's proven reliable — the
same staged rollout the `migrations` job used. SQLite is the default provider, so unlike `migrations`
it needs **no** DB service container. Runs on PRs and on `main` (regression net); skipped for `v*` tag
builds. The playout-build lock 409 + `isLocked` projection (#215) landed in #444.

**UI-E2E step (ersatztv#445).** The last step of this same job (item 4 in the summary above; the
`Run UI-E2E Playwright flows (headless)` entry in the YAML) runs the headless-browser flows the curl
harness structurally cannot express — client-side form validation, `AuthGate`'s rendered states, the
session cookie authenticating the **SPA's own** `/api` XHRs, and sign-out via the `UserMenu`:

```
ETV_BUILD_CONFIG=Release ETV_UI_PORT=8410 scripts/e2e-ui.sh
```

`e2e-ui.sh` owns the lifecycle (fresh config dir → boot → `playwright test` → always kill the server)
and exits with Playwright's status. Three deliberate choices:
- **This job, not a new one.** The dominant cost here is `npm ci` + the Release build, both already
  done; a separate job would duplicate them to add ~5s of browser work. The browser is baked into the
  toolchain image, so the step installs nothing.
- **Its own fresh instance on port 8410.** The first spec asserts the one-shot **Setup** gate, which
  step 3's auth section has already claimed on its config dir; a separate port also keeps this step
  independent of step 3's teardown timing.
- **`retries: 0`, `serial`, single worker.** #445 asked for deterministic flows and a retry would let a
  flaky flow merge looking green (measured: 4 consecutive clean runs, ~2s each). Full rationale and the
  rule for extending the suite: `docs/e2e-local.md` → "UI-E2E harness".

### Docs-only skip (ersatztv#416)

A change that touches **only** `docs/**` or `*.md` (anywhere: `README.md`, `CLAUDE.md`, handoff
files) has nothing for the heavy jobs to validate. Before this, such a change ran the entire matrix
— `test`, `migrations` (with its `mysql:8.4` service), `functional-e2e`, `format`, `api-docs` —
~9 min of warm CI for a Markdown edit.

**The mechanism, and why it is shaped this way.** Each heavy job (`test`, `migrations`,
`functional-e2e`, `build`) runs `scripts/ci-detect-docs-only.sh` as its first post-checkout step
(`id: detect`), which emits `docs_only=true|false` to `$GITHUB_OUTPUT`. Every real step in the job
is gated `if: steps.detect.outputs.docs_only != 'true'`. On a docs-only change the job runs only
checkout + detect and **reports `success` in seconds**.

**Shallow-checkout safe (the change-set diff).** `test`, `migrations` and `functional-e2e` all check
out shallow (`fetch-depth: 2`), and
a shallow clone has **no `origin/<base>` tracking ref and no merge-base** — so a three-dot
`origin/main...HEAD` diff *errors*, the fail-safe returns `docs_only=false`, and the skip silently
never fires (this bug shipped once — every docs-only PR still ran the full matrix; caught by
#416's "verify on a real PR" box). The script therefore fetches the base and diffs against
**`FETCH_HEAD`** (always written by `git fetch`, resolves in a shallow clone) with a **two-dot** tree
diff (`git diff --no-renames FETCH_HEAD HEAD`) — no merge-base required. (`api-docs`/`format` avoided
the bug only because they check out `fetch-depth: 0`.)

The jobs are **not** `if:`-skipped. That is deliberate and it is the whole trap of this issue:

- `main`'s branch protection requires two checks **by name** — `Build ErsatzTV Image / Build & test
  (.NET) (pull_request)` and `Build ErsatzTV Image / EF migration integrity (SQLite + MySql)
  (pull_request)`. If a docs-only PR produced **no run** for those (a workflow-level `paths-ignore`,
  or an `if:`-skipped job), those contexts would never report and the PR could **never merge** — the
  naive fix *bricks* docs PRs rather than speeding them up.
- On Gitea **1.25.4** an `if:`-skipped job reports commit-status state **`skipped`**, a distinct
  state (verified with a throwaway probe, PR #418) — not `success`. **Still true on 1.27.1**
  (re-confirmed 2026-08-28, ersatztv#747, from production PR heads rather than a new probe PR). We do
  **not** rely on how branch protection treats a `skipped` **required** context. Keeping the job running and gating its *steps*
  makes the required context report `success` unconditionally, which is safe by construction.
- Non-required jobs may skip freely: production already proves a `skipped` **non-required** context
  does not block merge (`build` is `skipped` on every PR). So `build` skips its image steps on a
  **docs-only push to `main`** (docs are not in the image, so there is nothing to rebuild); tag
  builds force `docs_only=false` in the script so a release is never skipped.

The detection **biases toward running more**: `docs_only=true` only when *every* changed path is
docs; any code path, a tag build, a non-merge push, or an undeterminable diff resolves to `false`
(run the full matrix). A false `true` would skip real tests on a code change — a correctness bug —
so every ambiguous case runs everything. The `migrations` job's `mysql` service still starts on a
docs-only run (a `services:` container starts with the job regardless of step `if:`), but the
expensive 787-migration replay is skipped; the service is capped and idle for seconds.

`api-docs` and `format` already short-circuit on docs-only changes via their own path detection (no
API path / no `.cs` changed → they pass in ~5s), so they needed no change. `docs-reminder`,
`decisions-guard`, `ci-image-pin` and `script-tests` keep running on docs-only changes — the first
two are *about* docs and must, and `script-tests` is unconditional by design (ersatztv#631).

Not in scope: the within-run triple `dotnet build` (ersatztv#398; measured and rejected as
build-once — see `docs/decisions.md`). The separate redundancy of running the **whole matrix on a
PR and again on the merge-to-`main`** over identical code (ersatztv#420) is addressed below.

### Cross-run tree-identity skip (ersatztv#420)

A merge to `main` re-runs the entire matrix over code the PR's last run already validated — the
same redundancy as docs-only, but for identical code rather than docs. On a `push` to `main` that
is a real merge commit, `test`/`migrations`/`functional-e2e` each run
`scripts/ci-detect-already-validated.sh` as a **second** detect step (`id: revalidate`, right after
the docs-only detect), and every heavy step gains an added `&& steps.revalidate.outputs.skip !=
'true'` to its existing `if:`.

**Skip condition — all four required, else fail-safe `skip=false`:**

- the event is a push to `refs/heads/main`;
- `HEAD` has a second parent `HEAD^2` (a real merge commit — the PR head CI already validated;
  squash, rebase, fast-forward, or a direct push have no `HEAD^2`, so they run);
- `git rev-parse HEAD^{tree}` equals `HEAD^2^{tree}` — main did not advance since the PR's last run,
  a byte-identical tree;
- `HEAD^2` has a **green Gitea combined commit status**, queried via the API **anonymously**
  (ersatztv#885: the three jobs that call `scripts/ci-detect-already-validated.sh` also run on the
  `pull_request` route, so they hold no stored secret; `timothy/ersatztv` is public and answers the
  combined-status GET without a credential — measured 2026-09-04). Trusting the aggregate `.state`
  is sound: a `skipped` context does **not** drag
  the combined state below `success` (verified live against this instance — a real merge commit with
  four `skipped` PR-only contexts still reported `.state == success`), and the two required jobs
  never report `skipped` (they always run and report a real `success`/`failure`), so
  `.state == success` implies they were green.

Those three jobs check out `fetch-depth: 2` so `HEAD^2` and its tree resolve.

**Why it's safe: `build` is not gated.** The three heavy jobs skip their *steps* (same
required-context reasoning as docs-only — they still run and report `success` in seconds), but
**`build` always runs on `main`, ungated**, building and pushing the image from that identical,
already-validated tree. No image ships from unvalidated source. The required contexts are
unchanged (`Build & test (.NET)`, `EF migration integrity (SQLite + MySql)`) — no branch-protection
change.

**Fail-safe bias.** Any uncertainty — not a main push, no `HEAD^2`, a differing tree, a
missing/failing/non-`success` status, missing auth — resolves to `skip=false` and runs the full
matrix. A false skip could ship an under-validated image, so every ambiguous case runs everything.

**Honest limitation — this fires rarely here, by design.** The tree is identical only on a
*fast-forward-equivalent* merge: main did not advance since the PR's last green run **and** the PR
head was not rebased at merge time. Two routine patterns defeat it in this repo: (1) under parallel
merges main usually advances; and (2) — the bigger one — the standard workflow **rebases a PR
before merging** to resolve the `docs/decisions.md` lifecycle conflict (see MEMORY: the
"decisions.md conflict treadmill"), which mints a new head SHA whose tree was never itself
CI-validated, so the tree-match check correctly declines. So the skip is a genuine but *occasional*
win (clean, up-to-date, un-rebased merges in quiet periods) — correct-but-conservative by
construction, not a general dedup. It never fires unsafely; when in doubt it runs the full matrix.

### Dropped-step guard on the required jobs (ersatztv#756)

`test` and `migrations` write the only two `docker-build.yml` contexts branch protection requires on
`main`. A step the runner declines to interpolate is **dropped, and the job still concludes
`success`** (ersatztv#751), so in these two jobs that failure is **fail-OPEN**: a required check
reports green having done no work. In `review-verdict.yml` the same drop is fail-closed — the status
is simply absent and the merge is blocked — which is why #751 fixed the safe direction first.

Two independent mechanisms hold it, and neither is redundant:

- **A static ban on expression delimiters** in any `run:` body of `test`, `migrations` **and
  `build`**. The drop mechanism *requires* an opener in the scalar, so this makes the class
  unreachable rather than merely detected — and it is the raw `${{` opener that is banned, not a
  well-formed pair, because an unclosed one triggers the same rewrite. When a step genuinely needs a
  value, pass it through the step's `env:` block, which is interpolated **per value**, so a bad
  payload there cannot take the body with it.

  **Why `build` is in the ban although it is not a required context.** Its one delimiter-bearing body
  was `Smoke + IPTV E2E`, which runs *after* `Build and push` — so on a `v*` tag the image is already
  in the registry as the release candidate and that step is what decides whether the candidate was
  ever booted. A drop there publishes an unsmoked candidate, reports green, and `DeployStack
  jazz-media` promotes exactly that image. Its two payloads moved into the step's `env:`, so the ban
  cost nothing.

  **The ban is re-checked on the release path itself (ersatztv#767).** It used to be enforced only by
  the `script-tests` job, which lives in `pr-checks.yml` (`on: pull_request`) and is **not** a
  required context — a *review-time* check on the PR that would introduce a delimiter, not a gate on
  the release. `pr-checks.yml` does not run on a `v*` tag push at all, so a delimiter that ever
  reached `main` would still drop `Smoke` on the tag build and go green; `main` being PR-only (#743)
  meant such a change had to pass through a PR where `script-tests` reddens, but a red on a
  non-required check does not block the merge server-side.

  There is now a **`scan` job** (`Delimiter ban (release path)`) that runs the PyYAML-based ban test,
  and **`build` lists it in `needs:`**. That single edge is the fail-closed property: a red `scan`
  means `build` is skipped outright, so the image is never built, let alone pushed.

  **Why a job and not a step inside `build`.** A step cannot protect the job it lives in. `build` is
  what publishes, so a guard step there fails **open** if the runner drops it — and the defence
  ("the guard's own body has no opener, so it cannot be dropped") is circular when the only thing
  enforcing that property is the same PR-only test being backstopped. That was the first design, and
  it was rejected for exactly that.

  **Why it runs the real pytest and not a bespoke scanner.** The same first cut hand-parsed the
  workflow YAML in stdlib Python, to avoid provisioning PyYAML on `build`'s bare runner. That parser
  yielded ~10 **false negatives** in one round — flow mappings (`{run: …}`), a quoted
  `"run":` key, aliases, multiline quoted scalars — making it strictly *weaker* than the check it
  backstopped, in the only direction that matters for a security gate. Running the existing test
  needs no second definition of "what is a `run:` body", so it has no drift surface at all. `scan`
  runs on `small` and provisions Python the same way `script-tests` does.

  The wiring is held by `scripts/tests/test_ci_release_path_scan_job.py` — `build` depends on it, it
  carries **no job-level `if:`** (one that excluded the tag push would restore the hole; one that
  skipped the job would skip `build` too), no step is `continue-on-error`, and every one of its own
  `run:` bodies is delimiter-free. Its steps also carry #756 markers and a trailing assert, so a drop
  *inside this job* is caught as well.

  The gate's guarantee is one **execution probe**, not a set of assertions about the command: it runs
  the scan step's real `run:` body — with the workflow, job and step `env:` tiers layered as the
  runner would — against a copy of the repo whose `Smoke` body carries an injected delimiter, and
  requires a non-zero exit, with a clean-tree negative control. Shape assertions were tried twice and
  lost twice (nine disarms across two rounds, from `echo`ing the command to `PYTEST_ADDOPTS` one env
  tier up), so **do not replace the probe with cheaper checks about the command.** One tier is
  unreachable by any static reconstruction — a step writing to `$GITHUB_ENV` injects into later steps
  at runtime — and is therefore **banned** in this job rather than modelled.

  **And one check deliberately is not a test at all.** The tests above prove the ban *holds*; they
  cannot prove it would *notice*, because pytest's own configuration outranks them — a repo-root
  `pytest.ini` with `addopts = -k "not delimiter_banned"`, or a `conftest.py` with
  `pytest_collection_modifyitems`, deselects the ban test and every test guarding it, leaving the
  whole suite green with a delimiter sitting in `Smoke`. So `scan`'s last working step runs
  **`scripts/ci-prove-ban-detects.sh`**, which poisons the checked-out workflow, re-runs the ban
  test, and requires **the ban test's `build` parametrisation to be the failure**. Not "the same
  command" (the step also runs the wiring tests; the script runs only the ban file) and not merely
  "fails if it passes": any other outcome — exit 5 because everything was deselected, exit 2 on a
  collection error, or a failure in some unrelated test — is a **refusal**, not a pass. That
  exhaustiveness is the load-bearing part, because each weaker reading was a live bug here:
  "non-zero means it noticed" vouched for the gate on a collection error, and the deselection disarm this script
  exists to catch exits **5**, not 1.

  It runs in the **real checkout** — an isolated copy does not inherit the repo-root config a disarm
  would live in, so a copy-based run reports healthy while the real invocation is deselected — which
  has happened here. The workflow file is
  restored by an `EXIT` trap, and the step carries a `selfcheck` marker so dropping it reddens the
  job.

  Two further tests exist because review disarmed the gate in ways an obvious test misses. **The
  pytest invocation is checked as a command, not as a substring**: the ban test's path must appear as
  a bare token on the real `python3 -m pytest` line, that line may carry none of
  `-k`/`-m`/`--deselect`/`--ignore`/`--collect-only`, and its exit status may not be swallowed by
  `||`/`&&`/`;`/`|`. Commenting the invocation out and echoing instead, or appending
  `-k 'not delimiter_banned'`, each left the suite green with the release path unguarded. And **the
  ban's SCOPE is pinned** — `DELIMITER_BAN_JOBS` must still contain `build`, because dropping that
  one literal makes `scan` pass while checking nothing about the job that publishes the image.

  Two operational notes. `scan` is a `small`-lane job on the tag path, which ersatztv#535 moved away
  from after such jobs wedged in act's setup phase on the v26.12.0 tag; that is safe again only
  because server-management#639 took `build` off the lane. If it ever returns, move `scan` to
  `ubuntu-latest` — never weaken the `needs:` edge. And `scan` installs from PyPI, so a network blip
  now sits between a `v*` tag and its image; fail-closed and loud, but real.

  What this does **not** claim: that no step can ever fail to run for a reason other than the
  interpolation drop. It moves the terminal assumption — to fail open you must now drop the pytest
  step **and** the assert step, rather than either one alone.

  Measuring a guard on this path does **not** require cutting a release: `build` runs on every
  push to `main` (`if: github.event_name != 'pull_request'`), and a `workflow_dispatch` on any
  other ref runs the
  job while `Build and push` publishes nothing (its `push:` is gated on `main`/`v*`). That is how
  #767 was verified — see the decision record for the run ids.

  `functional-e2e` is delimiter-free too but is deliberately **not** banned: it is
  advisory by declaration, and the rule is "ban where a drop is consequential", not "ban wherever it
  is currently free". `api-docs` and `format` keep one `github.base_ref` each in a detect step and
  gate nothing that ships.
- **Runtime per-step markers**, for a step that fails to run for any *other* reason. Every `run:`
  step that is not `continue-on-error: true` calls
  `"$GITHUB_WORKSPACE/scripts/ci-step-ran.sh" mark <key>` as its **first act**, and the job's last
  step calls `ci-step-ran.sh assert --always … --gated …`, which fails the job when an expected key
  was never recorded.

**Per step, not per job.** A marker written by the first step only proves the job *began*, which was
never in doubt. The drop that costs something is `Test`, `Build` or a migration replay — all well
past step one — so a job-level marker would have been a guard that cannot see the case it exists for.

**The guard carries no `if:`, and that is deliberate.** The #751 guard uses `if: always()` because its
job has one real step. These have a dozen, and a genuine failure in an early step legitimately skips
every later one — an `always()` guard would then announce a false *"these steps never executed:
typecheck web-test build dotnet-test"* on top of every ordinary red build, and a guard that cries wolf
gets deleted. (`migrations` is smaller — six marked steps — but the same argument applies, and its
guard comment is worded for its own keys rather than copied from `test`'s.) The default `if:` is `success()`, which is the wanted condition, and the invariant that
makes relying on it safe rather than lucky is: **the guard is skipped only when an earlier step
failed, and that already fails the job**. So *guard skipped ⇒ job red*, and every path to a green job
runs the guard. A dropped step is invisible precisely *because* it concludes `success` — which keeps
the job green and therefore reaches the guard.

That invariant has one path where it could plausibly be false and where being wrong would be silent:
a step marked `continue-on-error: true` that FAILS. If that flipped `success()`, the guard would be
skipped on a job that still concluded green — the guard rendered a no-op by exactly the failure mode
it exists to catch, with no signal. The `test` job has three `continue-on-error` steps and two of them sit immediately before the guard,
so this is a live path, not a theoretical one. **Measured** (scratch PR #766, run 1913 job 8075): the last
advisory step was made to `exit 1`, the log carries `❌ Failure - Main Report peak container memory`,
and the guard **still ran**, reported `All 12 expected step(s) executed`, and the job concluded
`success`. A failing `continue-on-error` step does not flip `success()` on this runner, so the
invariant holds where it mattered most. That run is also the `test` job's full twelve-key positive
control on the build lane.

**Adding a step to either job?** Mark it, and add its key to that job's guard list in the right
bucket (`--always` for the two detect steps, `--gated` for anything carrying the docs-only /
already-validated `if:`). `scripts/tests/test_ci_dropped_step_guard.py` derives the expected set from
the workflow, so an unmarked step or a bucket mismatch is a red — it does not rely on anyone
remembering. One caveat, since this section is careful about it elsewhere: that red is `script-tests`,
the same non-required, PR-only check discussed above. For the delimiter ban on `test`/`migrations`
that hardly matters, because the runtime guard is the fail-closed backstop — but a **newly added,
unmarked** step is caught by the static test *alone*, since the runtime guard cannot expect a key
nobody declared.

**The marker path is keyed on job + run id + attempt** — and be precise about why, because the
obvious justification is a #751 measurement that does *not* transfer. #751 found `RUNNER_TEMP` to be
`/tmp` and called it "not a private per-job directory", but that was taken on `review-verdict.yml`,
which runs *without* a `container:`. These two jobs run **inside** the CI toolchain image, so their
`/tmp` is the job container's own and starts empty. The fresh container is therefore what actually
rules out a stale marker here; the keying is defence in depth against a lane change nobody would
think to re-check this against. `GITHUB_JOB` and `GITHUB_RUN_ID` are *measured* present and the
script refuses without them rather than falling back to a name other runs share.
`GITHUB_RUN_ATTEMPT` is required too — but **how** that was established is the part worth keeping,
because the first two attempts at it were both worthless. Grepping a job log for the variable *name*
proves nothing: logs do not dump the environment. Inferring it from the *absence* of the script's
"not set" warning proves nothing either, because that warning goes to **stderr**, and whether step
stderr reaches a job log here was itself never established — the control offered for that turned out
to be an `::error::` line this script writes to *stdout*. So the script was made to **report its
resolved identity on stdout**, where capture is not in question, and the answer was simply read off
this change's own run: `Marker identity: job=test run=1916 attempt=1 (from the runner)`, and the same
for `migrations`. Both required jobs, on the lane that matters.

That measurement is what promoted it from warn-and-default to required, and it is what rules out a
rerun inheriting attempt 1's markers. The
identity line stays, as the standing evidence a future reader checks first if the keying is ever
doubted again.

**The premise was re-measured on the build lane.** The whole thing rests on the runner still executing
a later step after dropping an earlier one. #751 established that on the `small` lane; these jobs run
in a `container:` on `ubuntu-latest`, so it was measured there rather than assumed — scratch PR #765
(Gitea 1.27.1, 2026-08-10) reintroduced the exact #751 defect in the `test` job's `revalidate` step.
Recorded outcome (job `test`, run 1910, 20:03:49→20:13:31Z — a full 9m42s heavy run, so `Build` and
`Test` really executed):

- `Unable to interpolate expression 'format('# PROBE ONLY … {0}\n…', pr number)'` at 20:04:06 — the
  step was **dropped**, exactly as #751 describes, and it reported conclusion `success`.
- **Every other marked step still ran** — eleven markers were recorded, ten of them AFTER the drop
  (`restore npm-ci check-api lint typecheck web-test web-build strip-scanner build dotnet-test`),
  `detect` being the eleventh and earlier. The premise holds on this lane.
- The guard ran at 20:13:29, reported `These steps of job 'test' never executed: revalidate`, and was
  the **only** ❌ in the entire job log — every other step succeeded. Without it this run would have
  concluded `success` having never executed that step, which is precisely the fail-open being closed.
- Incidental but kept: the dropped step's output arrived as `ETV_REVALIDATE_SKIP:` **empty**, not
  `false` — the case the guard must read as "widen what is required", never as a skip.

**The positive control is the same run's `migrations` job**, which the probe did not touch: it marked
all six steps, the guard reported `All 6 expected step(s) executed: detect revalidate restore build
sqlite mysql`, and the job concluded **success**. So one run demonstrates both directions on the build
lane — a drop caught and reddened, and a clean job passing. The twelve-step `test` positive control is
this change's own CI run.

Full rationale: `docs/decisions/records/ci/required-job-step-execution-markers.md`.

### `docs-reminder` job (non-blocking, PR-only — in `pr-checks.yml`)

Two lightweight nudges, both `::warning::`-only. Neither can fail the build — they are reminders,
not gates; prose-doc gates get gamed with token edits.

**1. The parity-doc reminder** enforces the CLAUDE.md "docs-update is part of done" rule for the
one case that's easy to forget and easy to detect: a PR that touches a SPA screen
(`web/src/screens/*.tsx`) or `ErsatzTV/LegacyUiRedirects.cs` but **does not** update
`docs/blazor-route-parity.md`. It diffs the PR against its base branch and emits a
`::warning::` annotation.

**2. The session-narrative reminder** (ersatztv#784) runs `scripts/check-doc-narrative.py --diff`
over the lines this PR **adds** to `docs/**/*.md` (minus `docs/decisions/**`, exempt wholesale),
root-level `*.md`, and — since ersatztv#876 — every tracked file under `.claude/`, `.gitea/`, `.husky/`
and `scripts/` regardless of extension (minus the detector and its own test, whose fixtures are the
phrasings), flagging text that narrates the artifact's own revision history —
`docs.no-session-narrative`. It is advisory **by design and permanently**: a narrative detector is a
string predicate over prose, the class `docs/defect-shapes-773.md` §4 argues must never be
load-bearing, so the script exits 0 on every path including a bad argument or an unresolvable base
ref. It does NOT rest on prose alone, and it does not rest wholly on tests either: the record's
COVERAGE BOUNDARY names which clauses carry a mutation proof and which are defensive and unproven
(the unhandled-exception arm among them). Every argument shape is asserted per shape in
`scripts/tests/test_check_doc_narrative.py` rather than only in prose. When it cannot resolve the
base it prints `SCANNED NOTHING` instead of a clean-looking line, because a silent zero-file scan is
indistinguishable from a clean one — the same failure `ci.required-job-step-execution-markers`
exists for.

Because this step needs an interpreter, the job DOES carry `actions/setup-python` — the one
exception to the no-setup-actions note below, since `python3` is not guaranteed on the bare `small`
lane. It, the script step and the parity-doc step above it all carry `continue-on-error: true`: a script
that exits 0 does not by itself keep the JOB green, a setup-action download failure would redden an
advisory check just as effectively as a hit would, and the parity step's base-ref fetch fails loudly
(ersatztv#746) — which must reach the log without reddening a warn-only job. `Checkout` is the one
step left blocking.

The job deliberately has **no** `setup-dotnet`/`setup-node` (and
thus no `actions/cache`), so it can't hit the cache-save hangs seen on the VM-127 runner
(server-management#570). It does not cover the remaining doc obligations in the CLAUDE.md table
(domain-model, spa-conventions) — those stay on the author. (The API contract is mechanized by the
blocking `api-docs` job, and `docs/decisions.md` by the blocking `decisions-guard` job below.)

### `decisions-guard` job (`decisions lifecycle`, blocking, PR-only — in `pr-checks.yml`)

Enforces decision-record lifecycle invariants (ersatztv#521, supersedes the ersatztv#303 H9
append-only mechanic): well-formed 5-field metadata, exactly one `active` record per `key`,
reciprocal `supersedes`/`superseded-by` links, no record vanishing from the active set without an
archive copy, no rationale-prose rewrite without a `Decisions-Edit: yes` trailer on a non-merge commit
in the range (ersatztv#609), a structural per-path check that every `*.md` under
`docs/decisions/records/**` and `docs/decisions/archive/**` parses to **exactly one keyed
record** (ersatztv#621 — without it, a file the dependency-free frontmatter reader cannot parse,
such as one using a YAML block scalar, yields `[]` and vanishes from the corpus with every check
still reporting green; a file directly in `archive/` is exempt only when it really is a stripped index — one keyless
record with a known generated heading — never merely by its location; the single further exemption,
`archive/README.md`, is by exact relative path, never by basename, which would otherwise exempt the
same filename in the active wing),
and the generated active catalog (`docs/decisions/README.md`) in sync with source. Two steps:
`scripts/decisions_validate.py --base origin/<base> --head HEAD` (the merge-base diff checks, which
need a base/head range — CI-only) and `scripts/build_decisions_catalog.py --check` (catalog drift).
The **same validator** backs the Husky `pre-commit` hook (`.claude/hooks/decisions-guard.sh`, no
base/head there — structural checks only, over the working tree), so local and CI enforcement can't
drift on the rules that don't need a range. `python3` isn't guaranteed on the bare `small` lane, so
the job adds `actions/setup-python@v5` before invoking it; that install is lightweight (no
compiler/docker build), so it doesn't violate the "small is git-only" lane rule. Like
`docs-reminder`, otherwise a seconds-long `git diff` + parse with no dotnet/node setup
(`runs-on: small`).

### `prove-fix` job (`Fix proofs (Proves trailers)`, PR-only — in `pr-checks.yml`)

Runs `scripts/prove-fix.sh` for **every commit in the PR that carries a `Proves: <pytest selector>`
trailer**, and fails the PR if a claimed proof does not hold. The rule and its rationale are
`testing.fix-ships-a-witnessed-red-test`; this section is the CI-side contract.

**Opt-in by trailer, enforced when present.** Requiring `Proves:` on every commit would block docs,
CI and refactor commits that have no code side to revert, and a gate that blocks ordinary work gets
switched off — which is how a check ends up running nowhere (ersatztv#631). So the trailer is the
author's *claim* and this job checks claims. **When a PR carries none, the job emits a `::notice::`
saying it asserted nothing** — its green must not be read as fix coverage.

**It needs full history** (`fetch-depth: 0`): the script reverts each commit against its PARENT, and
a shallow clone cannot resolve `<sha>^`.

Three refusals worth knowing before you write a trailer:
- a **merge** commit is rejected — several parents means "the code before this change" is ambiguous;
  put the trailer on the commit carrying the fix;
- **more than one** `Proves:` trailer is rejected — only the first would be checked, so the rest
  would read as proven without ever running (the job checks this itself, because passing the
  selector explicitly bypasses the script's own guard);
- a **test-only** commit is rejected — there is no code side to revert.

**Only pytest exit 1 counts as red.** 2 is an interrupted collection, 3 internal, 4 usage, 5 nothing
collected, 143 a signal; `--continue-on-collection-errors` converts a genuine collection failure to
1 so add-a-file fixes stay provable. The job inherits that. Note the direction: a wobble DOES redden
this job (`prove-fix.sh` exits 5 on a harness/git failure or a signal, and the job turns that into an
error), so what the exit-code discipline buys is the other way round — a **green** here means a claim
was witnessed, never that a run was cancelled or broke.

### `script-tests` job (`Script lint and tests (ruff + pytest)`, PR-only — in `pr-checks.yml`)

> Reddens the run on failure, but like the other `pr-checks.yml` gates it is **not** one of the
> three required status checks on `main` (`Build & test (.NET)`, `EF migration integrity`,
> `review-verdict/h10`). Promoting it to required is a branch-protection change, tracked separately.

Runs the repository's Python test suite: `PYTHONPATH=. python3 -m pytest scripts/tests -q`
(773 tests at `706674272`, ~4.5 min; the suite grows fast — it was ~190 tests / ~10s when this job landed in #631 — so treat the figure as a dated snapshot, not a budget). It covers the decision-corpus parser/validator/catalog builder, the ersatztv#610
migration-equivalence harness, the merge-consent exemption logic and the ersatztv#622 review-verdict
poster.

**Until ersatztv#631, nothing ran these tests.** No workflow and no Husky hook invoked `pytest`.
`decisions-guard` executes `decisions_validate.py` and `build_decisions_catalog.py` directly — it
exercises that *code* but never its *tests* — and the `test` job is `dotnet test` only. The suite
guarding our merge-gating machinery was therefore local-only, and a test added "for CI enforcement"
was decorative.

**Why it is its own job, not a step inside `decisions-guard`.** `decisions-guard` is covered by the
standing `ci.decisions-lifecycle-flake` rule: a lone `decisions lifecycle` red is a known infra
flake and sessions are instructed *not to investigate it*. Adding the suite there would make a
genuine pytest regression surface as precisely the red everyone is told to wave through — the same
"reports success while doing nothing" failure mode ersatztv#631 exists to close. A distinct job
name keeps a real failure unambiguous.

**Why it runs unconditionally** rather than behind a `scripts/**` path filter: the suite's true
input set spans AT LEAST three top-level directories. `test_post_review_verdict.py` and
`test_merge_consent_exemption.py` execute the real `scripts/post-review-verdict.sh` and
`.claude/hooks/pretooluse-merge-consent.sh`, and since ersatztv#845 the former also READS
`.gitea/workflows/review-verdict.yml`, because the writer derives the H10 allow-list from it and
editing that literal changes the suite's outcome. So a `scripts/**` filter would silently miss a
`.claude/hooks/**` or `.gitea/workflows/**` edit. The hedge is deliberate: the corpus tests read
`docs/` too, so the real figure is higher, and an exact count here is the kind that goes stale.
The reason is the input set, not the cost: the suite was ~10s when that was decided and is now
minutes, and it would still be wrong to filter on `scripts/**`.

**It also lints (ersatztv#780).** Early in the job it installs a **pinned** `ruff==0.12.11` and runs
`ruff check` and `ruff format --check` against the repo-root `ruff.toml`. Five things are deliberate:

- The config is **committed**. Without it ruff falls back to whatever `~/.config/ruff/ruff.toml` the
  operator's machine has, so a second machine lints this repo differently or not at all.
- The version is **pinned** — an unpinned install makes the verdict a function of when the job ran,
  the same divergence one layer up, and the same argument as the `jq` pin below. `pytest`/`pyyaml`
  stay unpinned on purpose: a pytest release does not add assertions to your suite, a ruff release
  adds rules to your lint.
- Lint runs **before the jq preflight**, and after the `git` one. `Preflight jq version` is a hard
  `--expect` tripwire; a lint sitting behind it goes dark for as long as the jq contract is broken,
  under a red that says "jq". `Preflight external tools` stays ahead, because the lint steps consume
  `git` — without it a missing git reaches them as an empty population and they blame the glob.
- Neither step is `ruff check .`. Both pass an **explicit population** from
  `git ls-files -z '*.py' '*.pyi' '*.ipynb'` with `--no-force-exclude`, and fail if that list is
  empty. Discovery-based invocation is silently emptied by an `exclude` in the right config scope —
  top level empties both commands, `[lint]` empties `check`, `[format]` empties `format --check`
  (and `[format]` is where an appended line lands) — and
  `ruff check .` over zero files exits **0** with only a stderr warning, so the failure mode is a
  green gate. The measured matrix is in `ci.python-lint-ruff-config-committed`.
- `RUF100` is selected, so a `# noqa` that no longer matches anything is itself a finding.

`pyright` is not gated; the reasoning and the exemption list are in
`ci.python-lint-ruff-config-committed`.

**Dependencies: `pytest` and `pyyaml`** — the complete third-party set across `scripts/`, established
by an AST import scan rather than by reading the files that looked relevant. PyYAML does **not**
contradict the dependency-free decisions *read* path: `decisions_lib._read_frontmatter` is
hand-written exactly so validation runs where nothing is installed, but the one-shot *write* path
`migrate_decisions_split.py` uses PyYAML by design, and `test_migration_equivalence.py` imports that
module. (The first cut of this job claimed "pure stdlib + pytest", passed locally on a machine that
happened to have PyYAML installed, and went red in CI on a `ModuleNotFoundError` at collection —
which is itself a small demonstration of why the suite needed to run in CI at all.) Like the other
`small`-lane Python jobs it adds `actions/setup-python@v5` first. Checkout is at default depth: every `git` call in the suite runs
against a temp repo it creates itself, never this repository's history.

Two **preflight steps** run before the suite. The first asserts `git` is on PATH; the second runs
`scripts/jq-preflight.sh --expect 1.6`, which checks jq's **version**, not merely its presence (see
"The jq contract" below). Those two tests exec the real shell scripts, which shell out to `jq` ~26
times; the tests shim `curl` on PATH but not `jq`, so a runner image without it would surface as ~20
opaque assertion failures instead of one diagnosis. Both deliberately **check** rather than install —
ersatztv#390 removed run-time `apt-get` from CI; the fix for a genuine miss is to bake the tool into
the runner image.

### The jq contract (ersatztv#648)

> Full rationale: `docs/decisions/records/ci/jq-version-contract.md`.

Every shell gate in this repo — `decisions-guard`, `script-tests`'s own harness,
`pretooluse-merge-consent.sh`, `review-verdict.yml`, `scripts/pr-changed-files.sh` — is authored and
tested on a developer Mac shipping **jq 1.8.x**. The CI runner ships **jq 1.6**. Author to the
1.6-compatible subset; three concrete constructs diverge between the two and each one produced a real
bug when it hit CI for the first time:

- **`jq -e` over EMPTY input.** Exits 4 on jq >= 1.7, but **0** on jq 1.6. A guard that infers
  "transport failure" from that exit status silently passes an empty/failed page on 1.6.
- **`` contains("\u0000") `` (or any NUL literal).** The NUL escape truncates to `""` on jq 1.6, so
  the containment test is vacuously true for **every** string, not just ones containing a NUL. Use
  `explode | index(0)` instead — it is version-stable.
- **Parse-error exit code.** `jq empty` exits 5 on jq >= 1.7 but **4** on jq 1.6 — the same code 1.6
  uses for "no output produced". Reading that exit code as a specific failure mode conflates garbage
  input with an empty-but-valid response.

`scripts/jq-preflight.sh` makes the running version **observable** in every gate job's log (it prints
the parsed version and asserts a floor of 1.6) so a future divergence can be diagnosed from the log
alone instead of guessing at the runner image.

**Pin vs floor is deliberately asymmetric.** `scripts/jq-preflight.sh --expect 1.6` additionally pins
the version and fails loudly if it drifts, but that mode is used **only** by `script-tests`
(`.gitea/workflows/pr-checks.yml`) — advisory, not a required check. `review-verdict.yml` runs the
no-args floor-only mode and never pins, because that workflow writes `review-verdict/h10`, the
branch-protection-**required** status check on `main`: a hard pin there would mean the day the
runner's jq version changes (a base-image bump, a host reimage — nothing this repo controls), every
PR on `main` stops merging until someone notices and re-pins. A required merge gate cannot fail
because an upstream package manager did its job. The narrower pin on `script-tests` exists precisely
because that job is the suite's only 1.6 coverage — if the runner's jq silently changed, that coverage
would evaporate with no signal, so failing loudly there forces a human decision instead.

Baking a pinned jq into `docker/ci/Dockerfile` was considered and rejected: `review-verdict.yml` is
`runs-on: small` with no toolchain-image pin, and per `ci.small-lane-git-only` the small lane is
git-only, so it gets the **host's** jq regardless of what the toolchain image contains — a pin in the
image provably cannot reach the gate that broke. This was checked against the running binary, not
assumed.

## PR gates workflow

**File:** `.gitea/workflows/pr-checks.yml` — `on: pull_request` only.

The five PR gates — `ci-image-pin`, `docs-reminder`, `decisions-guard`, `script-tests`, and `prove-fix` (the one member that is NOT merely checkout + `git diff`: it installs pytest and creates worktrees)
(all described above) — live here, **not** in `docker-build.yml`, and that separation is the fix
for **ersatztv#535**.

**Why they are split out.** All three are pure `checkout + git diff` gates on the `small` lane
(no `container:`) and are PR-only (`if: github.event_name == 'pull_request'`). While they lived in
`docker-build.yml` — which also triggers on push to `main` and on `v*` tags — Gitea still
**dispatched them as runner tasks on every such push to evaluate the skip** (the `small`-lane
behavior documented above: *a job is dispatched even when its `if` skips it*). On the **v26.12.0**
release tag those dispatched skip-tasks wedged in act's setup phase and were killed by a runner
restart mid-setup, so they reported `failure` (no logs) and reddened the tag's overall commit
status even though the release built, scanned, and deployed fine. The two PR-only jobs on
`ubuntu-latest` (`api-docs`, `format`) carry the **identical** `if:` and skipped cleanly on the
same tag — the job logic was never the problem; the kill lands in the dispatch window, before any
step or `if:`-skip runs, so tweaking the `if:`/step logic could not fix it.

**Why a separate workflow fixes it.** Gitea evaluates a workflow's **trigger** before creating any
job, so a `pull_request`-only workflow produces **zero** jobs on a tag/main push: no dispatch, no
kill, no spurious red — for the whole class, permanently. The per-job `if:` guards are kept as
belt-and-suspenders (they also encode "these steps need a PR `base_ref`").

**What stays put and why.** These three carry **no** CI toolchain image pin, so `ci-image-pin`'s
grep of `docker-build.yml` still validates the five pin-bearing jobs
(`test`/`migrations`/`functional-e2e`/`api-docs`/`format`) that remain there. `api-docs` and
`format` stay in `docker-build.yml` because they carry the shared-image `container:` + pin and run
on the healthy `ubuntu-latest` lane (where they skipped correctly). None of the three moved jobs is
a **required** check — branch protection requires `Build & test (.NET)`, `EF migration integrity`
and `review-verdict/h10` (next section) — so relocating them (their status-context prefix changes
from `Build ErsatzTV Image / …` to `PR Gates / …`) does not affect merges. The file declares
`defaults: run: shell: bash` because `ci-image-pin` uses `mapfile`/`set -o pipefail`.

## Review-verdict gate (`review-verdict/h10`, required — `.gitea/workflows/review-verdict.yml`)

**A required status check named `review-verdict/h10`, written per-sha, is what actually stops an
unreviewed commit from merging** (ersatztv#622). It is not produced by a job's success/failure; it
is a commit status that `scripts/post-review-verdict.sh` POSTs onto one specific sha.

**`main` is PR-only AND admin-override-proof, and it takes both to make the check load-bearing**
(ersatztv#743, `release.main-direct-push-disabled`). Gitea evaluates `status_check_contexts` when it
**merges a PR** — a direct `git push origin HEAD:main` never consults them. So until 2026-08-05 the
entire gate was skippable with no forgery at all, which was cheaper than every route enumerated in
#697. `main` now carries **two** fields, and citing either alone is a mistake:

- `enable_push: false` — a direct push is refused server-side at pre-receive (`Not allowed to push to
  protected branch main`), for every account including a site admin. The contents API is refused too
  — measured, HTTP 403 `user cannot commit to repo`. The web editor, upload, apply-patch, revert and
  cherry-pick paths share that same `CanUserPush` predicate and are therefore expected to refuse as
  well, but were not probed (source-attested only).
- `block_admin_merge_override: true` — without it (the default is `false`), a repo admin could
  `POST /pulls/{n}/merge` with `force_merge: true` and merge straight past a missing or red
  `review-verdict/h10`. Disabling push alone just moves the bypass from the push path to the merge
  path, since `timothy` is admin and is the identity every session already uses. **Measured
  2026-08-28 (ersatztv#747)**, replacing the earlier source-attested-only note. Two throwaway scratch
  bases, each with its own rule, head and PR — differing by construction, not by mutating one rule in
  place: on `scratch/747a-base` with the field `false`, an admin `force_merge` past an ABSENT required
  context MERGED (HTTP 200); on `scratch/747b-base`, whose rule was identical but for the field set
  `true`, the same call was refused (HTTP 405 `Not all required status checks successful`). The pair is
  the evidence — arm B's 405 has no attribution control of its own. It never required merging an
  unreviewed PR into `main`: a scratch BASE branch carries its own rule, so the experiment is
  disposable. Full arm table, including the ordinary-merge arm and its control:
  `release.main-direct-push-disabled`. `enable_bypass_allowlist` is **not** a
  substitute in the one configuration tested: enabled with an EMPTY username list (and
  `block_admin_merge_override: false`) the admin force-merge still returned HTTP 200. A non-empty
  list was not tested.

**Operator recovery when a required context gets stuck.** `block_admin_merge_override: true` removes
the "Merge (admin)" / `force_merge: true` escape that used to unstick a PR whose required context was
absent or wrongly red — a recurring situation here (a killed run overwriting a newer green, an
advisory red counted into the combined status, a gate workflow that cannot post). That escape is gone
*by design*: it was also the bypass. The supported recovery is to fix the status
(re-run the job, or re-post the verdict with `scripts/post-review-verdict.sh`); the last resort is to
`PATCH .../branch_protections/main` setting `block_admin_merge_override: false`, merge, and set it
straight back. Do the last one deliberately and say so in the PR — it is the one action that
re-opens the hole this section exists to close.

Practical consequences: **every** change to `main` goes through a PR, including a one-line docs fix;
and the client-side Husky guards (H6/H11/H13) remain useful friction but were never the control —
they are fail-open and `--no-verify` bypasses them. Tag pushes go through a separate mechanism and
are meant to stay unaffected for the release operator: since ersatztv#885 `tag_protections` carries
one rule, `v*` whitelisted to `timothy` (`release.tag-protection-v-star`), which is intended to leave
the release cut in "Cutting a release" working unchanged while refusing a `v*` tag from the `renovate`
bot — a push that would otherwise have published a `:prod` image. Both halves of that are EXPECTED,
UNVERIFIED: only the `timothy` credential exists here, so neither a real release cut nor a refused bot
push has been exercised. `release.tag-protection-v-star` carries the same caveat and waits on the
next cut.

**A fourth arm: guard-scope freshness (ersatztv#787).** `scripts/tests/test_ci_dropped_step_guard.py`
decides which CI jobs must carry per-step execution markers by DERIVING that scope from
`.gitea/required-status-contexts.json`, the committed mirror of `main`'s required status checks. It
derives rather than repeating a literal because its own CI job, `pr-checks.yml::script-tests`, checks
out with `persist-credentials: false` and cannot ask Gitea — which leaves the snapshot as the one
hand-maintained input in the chain. So the merge hook reconciles it: it reads
`branch_protections`, hands the payload to `scripts/check-required-contexts.sh`, and **asks** (never
denies) when the live list no longer matches. The comparison is against the EFFECTIVE required set:
a rule with `enable_status_check: false` requires nothing whatever its `status_check_contexts`
still holds, so reading the list without the flag would certify disabled protection as current.

Two consequences worth knowing before you meet them. **The immediate merge path now reads a
repo-admin-scoped endpoint**: before #787 `branch_protections` was fetched only when arming a
scheduled auto-merge, so a credential without repo-admin scope kept its auto-grant on the common
path and now gets a prompt there instead. That is the fail-closed direction and it is loud, but it
is a real change in what the gate needs. **The arm runs only for the repo the snapshot names**
(its `repo` field): the merge tool is called with whatever owner/repo is being merged, and judging
another repo by this one's mirror would report a confident finding about a repo the mirror does not
describe.

And what it does not reach: a docs/process-only PR exits at the hook's docs-only passthrough before
this arm, so a PR editing only `.gitea/required-status-contexts.json` — the snapshot-narrowing
direction — never triggers it. That passthrough is a human prompt rather than an auto-grant, and the
arm could not have judged such a PR anyway, since it compares the live server against the snapshot in
the local checkout rather than the one the PR proposes. Ask rather than deny because drift does not make that
merge unsafe — Gitea enforces the live required set server-side — while a stale guard scope means a
required context may have no dropped-step guard at all. The arm runs **last**, so it can only
downgrade an otherwise-satisfied auto-grant and never masks a more serious refusal, and it runs on
**both** merge paths: the pre-existing branch-protection read sits inside the
`merge_when_checks_succeed` branch and never fires on an immediate merge. Rule selection is shared,
not copied — both the hook and the checker load `scripts/lib/branch-rule-classifier.jq`.

**The hole it closes.** `pretooluse-merge-consent.sh` proves its three consent conditions at the
moment the merge tool is called. Pass `merge_when_checks_succeed=true` and Gitea performs the merge
*later*, against whatever head is green then — while the Done-when and review-verdict checks were
proven against the head at **scheduling** time. Every commit pushed in between merges unreviewed.

This was demonstrated as a controlled A/B rather than inferred (`ci/fake` stands in for a slow CI
check so Gitea waits, as it really does): review head A → post its verdict → schedule auto-merge →
push an unreviewed commit B → CI greens on B. **Without** the required verdict context, B merged.
**With** it, the same sequence was refused, and merged only once B itself was reviewed.

Note the motivating anecdote in ersatztv#622 — "PR #619 merged 263 insertions with no verdict" —
is **wrong**: #619 does carry `Review-verdict: MERGEABLE @ 02c82b35`, posted six seconds before the
merge, explicitly re-reviewing the follow-up commits. It was filed from an API read that lagged.
The gap is real anyway, and structural: nothing *forced* that re-review inside the 45-minute window
where Gitea would have merged whatever went green. This turns diligence into construction.

**Why a commit status fixes it and a smarter hook cannot.** A status belongs to exactly one sha, so
a new commit *cannot inherit it*: the required context is simply absent on the new head, Gitea's
merge-requirement check reads that as not-passing, and the scheduled auto-merge refuses to fire.
The invariant self-invalidates — nothing has to notice the push. It also covers merge paths the
hook never sees (Gitea UI, raw API, another agent's session).

**Posting a verdict.** After reviewing a PR's *current* head:

```bash
ETV_GITEA_BASICAUTH=user:pass scripts/post-review-verdict.sh <pr> MERGEABLE [note...]
```

That posts both the `Review-verdict: … @ <sha>` comment (the human-readable artifact, and the
hook's condition (c)) and the `review-verdict/h10` status on the same sha. `BLOCKED` /
`NOT-MERGEABLE` post a `failure` status instead. The script re-reads the head after commenting: if
a commit landed mid-flight it writes **no** status and exits non-zero rather than retargeting your
verdict at a commit you never read.

**The verdict words are declared once** in `scripts/lib/review-verdict-vocabulary.sh`, and both
sides derive from it: `post-review-verdict.sh` classifies through `etv_verdict_class`, and
`check-review-verdict.sh` builds its `POS_RE`/`NEG_RE` from `etv_verdict_alternation`. Adding a word
is a one-line edit to that file and it reaches both sides by construction (ersatztv#788,
`release.verdict-vocabulary-shared`). Only the WORD SET is shared — the grammar the classifier
applies around it (the column-0 anchor, the whole-word boundary, the anchored `@ <sha>` field) stays
in `check-review-verdict.sh`, where every #629 false-open actually lived. Run
`scripts/post-review-verdict.sh` with no arguments to print the live vocabulary.

Two things follow that are easy to get wrong. **Do not reintroduce a parity test** that compares the
two scripts' sources: #774 shipped one and withdrew it after six review rounds, each finding another
shell construction that escaped the extractor or reddened it on a correct tree. And a vocabulary
that cannot be loaded or does not validate yields **no words on either side** — the reader exits 2
(which callers treat as an unreadable input and fail closed on), the writer exits 1 having posted
nothing. Words are interpolated into a regex, so a `.*` in the positive list would classify an
explicit `BLOCKED` as `positive`; validation restricts every word to `[a-z][a-z-]*`.

**What enforces that is a data dependency, not the `if` around the validator** — worth knowing
before editing either script, because the obvious reading is wrong and the first version of #788
shipped the hole. Under `set -u`, an unbound-variable error inside a function aborts the function
but not the script, and in an `if ! validate` form neither branch runs, so the `exit 2` is skipped.
The reader has no `set -e`, so unvalidated words reached the regex and an explicit `BLOCKED @ head`
classified `positive`, exit 0. Validation now sets a sentinel on its last line and the derived views
refuse without it. Likewise a top-level `exit` in a sourced library terminates the sourcing script
where it stands, so both consumers probe-source the library in a subshell and require an
end-of-file marker before trusting it.

The status description also records the base branch — `Review-verdict: MERGEABLE @ abc1234 (base:
main)` — and the merge-consent hook denies when that no longer matches the PR's live `base.ref`
(ersatztv#632). Retargeting a PR changes the effective diff without moving the head sha, so the
per-sha binding alone cannot see it. This is **detection on the hook path only**: a commit status
carries no base of its own, so a merge driven through the Gitea UI or API is unaffected. The
comparator is the base *branch*, never its tip sha — a base that merely advances is ordinary churn,
and comparing tips would invalidate every open verdict on every unrelated merge to `main`.

**Exemptions** are handled by `review-verdict.yml` on every `pull_request_target` event, which posts the
status as `success` for **Renovate-authored** PRs (it uses `platformAutomerge: true`, so a required
verdict with no exemption would stall every dependency bump) and for **docs-only** PRs, and as
`pending` for everything else so the block has a visible reason. Both exemptions are **void when the
PR touches `.claude/`, `.codex/`, `.gitea/`, `.husky/`, `scripts/` or `docker/ci/`** — a PR that can weaken the
gate must not be able to exempt itself from the gate. That includes Renovate's `docker/ci` base
bumps, which already need the manual publish-then-pin two-step anyway.

The Renovate exemption additionally requires **every** changed path to be a dependency manifest —
`Directory.Packages.props` or `.config/dotnet-tools.json`, and only those (ersatztv#698). The npm
manifests are deliberately excluded: `renovate.json` enables only `nuget`/`github-actions`/`dockerfile`,
so npm is unmanaged here, while `package.json` `scripts` are executed by CI (`npm ci`, `npm run build`)
— exempting it would put a code-execution path inside the allow-list for no benefit. An author match alone is not enough, because `pull_request.user.login` is the PR's
*immutable creator* while its head is not: pushing application code onto an open Renovate branch
leaves the PR still "authored by renovate" and, previously, still exempt. A Renovate PR touching
anything else — a `.csproj`, a source file — is not blocked, it just needs a real verdict. **If a
dependency PR is unexpectedly asking for a verdict, this is why**; the status description says so.

The two exemptions are evaluated as **independent predicates**, never as an `elif` chain: a Renovate
PR touching only `docs/` still gets the docs-only exemption on its own merits.

An existing `review-verdict/h10` **`success`** on the head is **only** left alone when it is positively
identifiable as a verdict from an approved reviewer — a `.creator.login` in the workflow's
`H10_REVIEWERS` allow-list **and** a `Review-verdict:` description, which is what
`post-review-verdict.sh` writes. Anything else, including any shape the workflow does not recognise,
is **re-derived** rather than inherited. An existing **`failure`** is left alone on a weaker test —
attributable (`Review-verdict:`, a non-null creator, and a base matching this PR's) but **not**
necessarily allow-listed — see the state asymmetry below. (Measured: a status POSTed with a user credential carries a creator; one POSTed by an
Actions job carries `"creator": null`.) Without this, an exemption obtained once was accepted
unchanged on every later run.

**The allow-list replaced a "non-null creator" test (ersatztv#742), and how wide that old test
really was is the point.**
Non-null proved only that *some* user credential POSTed the status, and this repo stores credentials
for accounts that are not reviewers. `RENOVATE_TOKEN` is a `write:repository` PAT of the `renovate`
bot; unlike the registry credential it **cannot be scoped down** (Renovate needs repo write to push
branches and open PRs), and secrets are a per-repo *store* rather than a per-workflow grant, so a
PR-added workflow can reference it and POST a `Review-verdict:` success that reads back
`creator: renovate` — non-null, and therefore inheritable. (Inferred, not observed: `renovate` has
never actually posted a commit status in this repo — every row on all six Renovate-authored PR heads
carries `creator: null`. The user-credential-vs-Actions-token asymmetry underneath it IS measured.)
Membership makes every non-reviewer account's
**`success`** re-derive exactly as an Actions-token status does. The list is a **literal in the workflow**, not a
repo `vars:` entry or a secret: those are settable without a PR, i.e. without this gate, while
`.gitea/` is on the PROTECTED path list so the edit cannot ride a docs-only exemption either.

**The allow-list governs an inherited `success`. It does NOT govern a `failure`**, and the asymmetry is
deliberate: an off-list `failure` is left alone, because inheriting a rejection can only ever withhold
an exemption while re-deriving one can turn it green on an exempt PR. The worst a forged `failure`
achieves is a stall, and `post-review-verdict.sh` POSTs unconditionally, so any human clears it in one
command.

**Beyond that, two other readers of `.creator` stay broad, and the split is load-bearing.** `read_existing_verdict` carries two flags: `ex_attributable` (an attributable
`Review-verdict:` row for this base, whoever wrote it) and `ex_human` (`ex_attributable` **plus**
membership).

- The **inheritance short-circuit** reads `ex_human`. Narrowing is safe there: an unrecognised account
  falls through to re-derivation.
- The **last-moment re-read** reads `ex_attributable`, because it asks the opposite question — "did a
  reviewer post a verdict while we were classifying". Narrowing it makes the job stop abstaining and
  post its exemption over the row, and the post-write repair would not cover that — it is skipped
  whenever the high-water mark could not be established, which since ersatztv#849 is also the case in
  which no exemption is posted at all. It additionally requires the
  state/creator/description triple to have *changed* since the first read, because the two calls no
  longer compute an identical predicate and "changed" can no longer be inferred from "fired".
- The **post-write raced check** stays broad on `creator != null`. Not because narrowing it would let
  a rejection go green — a real reviewer is on the list by construction — but for the
  misconfiguration case, a second human account added to the repo and forgotten from the literal.
  Its reach is narrower than it looks: it covers such a verdict landing inside the **write window**
  and nothing else, since a verdict already on the head sits below the high-water mark and is
  invisible to it. Within that window the errors are asymmetric — a repair to `pending` is
  recoverable, a wrongly-standing `success` is not — so it stays broad. The cost: any verdict-shaped
  row landing in that window trips the sticky repair sentinel and costs that head its exemption until
  a human clears it.

Pinned by `test_the_MID_RUN_abstain_still_fires_for_a_NON_allowlisted_creator` and
`test_the_RACED_verdict_check_still_counts_a_NON_allowlisted_creator`.

**If a reviewer's verdicts stop being honoured, this is why.** Whenever the job declines to inherit an
existing status it logs the creator, the allow-list and the description together, and says to add the
account to `H10_REVIEWERS` in `.gitea/workflows/review-verdict.yml`. (A separate `::warning::`
annotation carrying the same three values was tried and withdrawn: its only delta was a severity whose
rendering could not be probed, and it needed a guard of its own to stop contradicting the line above
it.) `post-review-verdict.sh` posts
with whatever account owns `ETV_GITEA_TOKEN`/`ETV_GITEA_BASICAUTH`, and since ersatztv#845 it
**checks that it is an allow-listed one** — at the terminal that made the mistake, not one PR event
later. It reads the status back after posting, identifies its own write by state and description, and
refuses **before writing the comment** unless `.creator.login` is a member; a `creator: null`
(Actions-token) status gets its own message, since there is no account to name. The allow-list is
DERIVED from the `H10_REVIEWERS` literal in `review-verdict.yml` by `scripts/lib/h10-reviewers.sh`,
not restated beside it, so there is one declaration rather than two plus a parity test.

**The membership requirement is `success`-only, mirroring the gate's own asymmetry.** An existing
`failure` is inherited on ATTRIBUTABILITY alone, whichever account recorded it, so requiring
membership for a `BLOCKED` verdict would refuse one the gate honours perfectly well — and would
leave an off-list reviewer no supported way to record a rejection, since the comment would be
withheld too. A negative verdict from any attributable account is written and commented normally.

What that changes is the *diagnostic*, not the safety direction: a non-allow-listed POSITIVE verdict
was always a stall, never a fail-open. Before, a changed writer credential reported success and each
*subsequent* PR event re-derived the status — to `pending` on a PR that needs a verdict, or to an
exemption **`success`** on a docs-only or Renovate-manifest one, which is the quieter and worse
half — so the symptom was a PR that kept losing its verdict rather than an immediate refusal. Now it
is an immediate refusal naming the account and the allow-list.

**The residual, stated precisely, because "the `ask` half-state" describes only half of it.** The
check runs after the POST, so a refused verdict leaves a `review-verdict/h10` **standing on the
head** with no comment beside it. On the *hook* path that is the safe half-state
`release.verdict-writes-status-before-comment` designates — nothing for condition (c) to classify,
so it asks rather than granting. On the *server* path it is weaker: branch protection binds the
context NAME and not its issuer, so a `success` posted by any account satisfies the required check
until the gate re-derives it on the next PR event. **A refusal from this tool is not a reason to
trust the green it leaves behind.** That is not a regression — it was equally true before #845, when
the tool reported success and said nothing — and it is deliberately not repaired here, because a
second corrective write is the sticky-sentinel mechanism ersatztv#849 is separately designing.

**One false ACCEPT, named because the direction matters.** The writer reads the workflow in the local
checkout while the gate runs the base-resolved one, so on a branch that ADDS a reviewer the writer
accepts an account the live gate will not yet inherit from — the #845 stall, reached through the
check meant to catch it, on the one branch where someone is editing the allow-list. Accepted:
closing it means a live remote read on the merge-consent write path, whose own read failure would
have to resolve either to refusing every verdict or to passing. The window is one PR long and the
residual is a stall, never an unreviewed merge.

This remains a *provenance* check, not an authentication one — an **allow-listed** collaborator's own
write-scoped token can still POST a verdict they did not form, branch protection binds the context
*name* rather than its issuer, and no repo-side change closes that (ersatztv#697). That provenance
asymmetry is *why* the credential scoping in `ci.actions-credential-scoping` mattered: a forgery
through an allow-listed **user** credential inherits, while one through a job's `GITEA_TOKEN` carries
`creator: null` and is re-derived, so it must win a race. Note also that re-derivation is **not** a
race the attacker can lose: it fires only on the trigger's `types`, and posting a status is not one of
them, so a POST timed after the last PR event stands until the next one.

Deciding either exemption requires the PR's **complete** changed-file list, which the workflow does
not compute itself: it calls `scripts/pr-changed-files.sh`, the single shared implementation also
used by the advisory hook `.claude/hooks/pretooluse-merge-consent.sh` (ersatztv#649). The workflow
reads that script's **exit status** — a non-zero exit means "could not tell" and withholds the
exemption; its stdout is meaningless on any failure path and is never consumed.

**Never write a classification guard as `producer | grep -q…` here.** Under `set -o pipefail`, `grep -q`
exits at its first match, the producer takes SIGPIPE (141), and a MATCH is reported as a failed
pipeline — inverting the guard for any PR whose path list exceeds the pipe buffer. That let a large PR
be classified docs-only, and let one editing `.gitea/` skip the protected-path check entirely. A
here-string is **also** wrong (bash spills a large one to temp storage, which fails the same way when
temp is full). **Count** instead — `grep -c` drains stdin over an ordinary pipe — evaluate the counts
once at top level rather than inline in an `if`, and fail closed on a non-numeric result. Full detail:
`ci.grep-q-pipefail-inversion`.

That script takes the expected base branch as a **required 5th argument** and refuses to enumerate when
the PR's live base does not match it, checked both before and after paging (ersatztv#698).
`/pulls/{n}/files` diffs against the PR's *live* base, so retargeting changes the answer without moving
the head sha — a PR opened into `main` and retargeted mid-run was granted a docs-only exemption while
its diff against `main` carried a C# file. The workflow passes the base from the `pull_request_target`
payload, which a retarget cannot rewrite, and `edited` is in `types:` so a retarget reclassifies.
`edited` gives **detection, not atomicity**: runs are not serialized, so a stale run could still post
`success` after the reclassifying run posted `pending`.

**That residual is now fenced — NARROWED, not resolved (ersatztv#706; correction ersatztv#849).** The pre-write fence does not re-count *after* the POST, so a retarget landing between its final pre-write count and the write used to yield a **permanent** forged green: the successor run consumes the `edited` event and exits on the existing status, and the stale run then posts last with nothing left to correct it. Since ersatztv#849 a SEPARATE post-POST re-count (below) makes that green transient rather than permanent. Runs are still not serialized — instead a run that was
overtaken *declines to write its classification*. The job counts `change_target_branch` events on the PR's issue timeline
at start and again immediately before its POST, and posts **nothing** if the count moved. The count is
the key precisely because the branch *name* is ABA-vulnerable: `main → scratch → main` reads `main` at
both ends, which is how the forged exemption was obtained in the first place. Abstaining never strands
a PR, because every retarget fires `edited` — the event that makes one run abstain has already queued
its successor.

If the count can't be established (unreadable timeline, an empty FIRST page, or paging that reached
the 20-page cap on a NON-empty one — since #870 an empty page before the cap no longer ends the walk),
only the exemption `success` is withheld; `pending` still posts, since it blocks the merge
immediately and withholding it would strand ordinary PRs whenever the timeline is unreadable. That is
the right trade but not a free one — a generic `pending` still masks a rejection landing in its own
write window. What made that durable is gone: since ersatztv#849 the post-write check runs after
**every** write, so such a write is repaired to the sticky sentinel instead of being re-derived into
an exemption by a later run. **If an exempt PR is unexpectedly missing its status after a retarget,
this is why** — the job log names the counts.

**The count is also re-taken AFTER the POST, on the exemption path only** (ersatztv#849,
`ci.verdict-unverified-write-sentinel`). The pre-write fence covers *writing while overtaken*; it never
covered *being overtaken after writing*, which was the worse of the two — a retarget landing after the
final pre-write count left a stale `success` with the `edited` event already consumed by a successor
that short-circuited, so nothing remained to reclassify. It makes that green **transient rather than
permanent**, not absent: the `success` is live between its POST and the repair, so branch protection
or a scheduled auto-merge can still observe it. Closing that window needs a compare-and-set the API
does not offer. The retarget axis only: a push after the POST
moves the head, so the status no longer gates that PR, while a retarget changes the effective diff with
the sha unchanged.

Worth knowing before reaching for the obvious alternative: **a concurrency group does not work here**,
measured rather than assumed. Gitea 1.25.4 auto-cancels superseded `push` runs on a branch, but *not*
`pull_request_target` runs — two runs for one PR genuinely overlap, and adding
`concurrency: {…, cancel-in-progress: false}` changed nothing (probe runs still overlapped by 36s).
`cancel-in-progress: true` is deliberately untried, because a cancelled run leaves an exempt PR
statusless with nothing left to re-trigger it. Full measurements and the two surviving residuals:
`ci.verdict-write-retarget-fence`.

Since ersatztv#803 that fence counts **two** event types on the one timeline walk:
`change_target_branch` for the base alias (`main -> S -> main`) and `pull_push` for the HEAD alias
(a force-push `H1 -> H2 -> H1` spanning `pr-changed-files.sh`'s paging, which leaves every sha
comparison equal). Either count moving withholds the write. The run log line is
`Mutation fence: N retarget event(s) and M push event(s) … (trusted=…)`.

The `push`-supersession half of that claim stays **1.25.4-dated on purpose** (ersatztv#747,
2026-08-28). It is not unobservable, but it is no longer reproducible ON DEMAND. No workflow triggers
on a push to a *non-`main`* branch — `docker-build.yml` filters its `push` trigger to `main` plus `v*`
tags, `ci-image.yml` to `main` and `docker/ci/**` with deliberately no `tags:` key
(`ci.toolchain-image-publish-is-a-dispatch`), and the other four have no `push` trigger at all — and
`main` itself refuses direct pushes since `release.main-direct-push-disabled`. A PR merge IS a push to
`main`, so two merges landing close together remain a route; a `v*` tag push triggers `docker-build.yml`
but is **not** one, because supersession is per-ref and a release tag is pushed once. What is gone is the
cheap disposable probe — the scratch-branch route the other measurements used. Re-running it would mean
arranging two real merges, or adding a throwaway trigger; both cost more than the claim is worth,
because nothing branches on it. The `pull_request_target` half the fence actually relies on was **not**
re-measured either and is 1.25.4-dated too.

The two facts this paragraph rests on — that no workflow triggers on a push to a non-`main` branch,
and that `main` refuses direct pushes — were **re-checked on 2026-09-02 (ersatztv#869)** and both still
hold. Note what that does and does not cover: it re-checks the REASON, not the claims. The
`push`-supersession and `pull_request_target` measurements above remain 1.25.4-dated and were not
re-run by #869 either. Re-checking the reason matters on its own, because a reason decays exactly like
the claim it excuses. One limit is worth stating so this is not over-read: the enumeration covers
`push:` triggers only. Four of the six workflows carry `workflow_dispatch:`, so a disposable run can
still be raised on demand — what is gone is the cheap scratch-BRANCH push probe, not every route to a
throwaway run.

Separately, after posting **any** status the job re-reads the per-POST status history and, if a human
`Review-verdict:` row appeared during the write window, overwrites its own status with `pending` and
logs an error. It ran only for a `success` write until ersatztv#849, on the claim that a `pending`
cannot turn a rejection green — false, since a generic `pending` masks a rejection just as well and,
carrying no marker, is re-derived into an exemption by the next run with the human's row now below
THAT run's high-water mark. The repair is `pending`, never a copy of the human's verdict, which would
attribute a human decision to the job.

**Which sentinel the repair writes is part of the answer.** There are two, and they are not
interchangeable (`ci.verdict-unverified-write-sentinel`):

- `Human verdict raced this exemption write — re-post the verdict` **asserts** that a verdict existed
  and a write buried it. Only the arm that actually *counted* such a row may claim it, and only a
  human re-posting clears it.
- `Status write could not be verified — re-post the verdict` states only what was established —
  that the write could not be checked. Every "could not check" arm writes this one: an unreadable or
  over-cap history, an impossible empty history, an unusable count, no high-water mark at all, a
  post-POST retarget, and an unreadable combined-status read. It is equally sticky, and additionally
  **reconcilable**: a later run pages `/statuses/{sha}` in full and either finds a verdict underneath
  it — upgrading to the repair sentinel — or finds none and clears it, so a transient API failure does
  not cost a head its exemption permanently.

**No high-water mark means the write becomes the sentinel** — every re-derivable state, not only the
exemption — withheld before the POST rather than posted and repaired: the defect is known in advance,
and publishing a green to take it back opens a window branch protection, and an already-scheduled
auto-merge, can see. A generic `pending` is included because it is exactly what a later run
re-derives; `$REPAIR_DESC` is not, being the stronger fact.

**Every path that cannot establish what the head carries REPLACES the unknown state** rather than
merely declining to write: the combined read (retried once first), all four page-2 completeness
refusals, a row or consumed field whose type the schema does not allow, the path-predicate failure,
the fence branch that cannot trust its retarget count while holding a derived `success`, and the two
OBSERVED-mutation arms when the head carries a row the run did not inherit. Declining protects a real verdict and leaves a *forged* one standing, which is what an
off-list `success` is; the job went red on a status branch protection does not read, and the fence
branch was reached only after the classification had DECLINED to inherit the very row it then left
current. Nothing is destroyed — `/statuses/{sha}` keeps one row per POST, so the next run's
reconciliation finds a masked verdict and tells the reviewer to re-post. Writing the sentinel and
failing the job are separate decisions: the read refusals were already non-zero exits and stay red,
the fence branch exited 0 and still does.

**Reconciliation needs a witness.** It may clear the sentinel only over a complete history that
CONTAINS the sentinel's own row, matched by the `id` the combined read reported — degrading to a
DESCRIPTION match, which identifies a row with the same text rather than that row, only where the
server omits `id`. `ex_unverified` means the combined endpoint just returned that row
and `/statuses/{sha}` keeps one row per POST, so a complete-but-empty history contradicts a write
that demonstrably happened — and `page_statuses` accepts an empty page 1 as complete, which is what
made the shape reachable. The cost lands where the replacement actually cost something: on an
APPROVED PR, reconciliation finds the masked verdict and upgrades to the repair sentinel, so the
recovery is "you are asked to re-post", not "it clears itself".

**Every `/statuses/{sha}` read is PAGED** (ersatztv#763). The history is read twice on an ordinary
run — before the write for the high-water mark, after it for the race check — and a THIRD time on a
head carrying the unverified sentinel, for the reconciliation (ersatztv#849). `limit` clamps to
`MAX_RESPONSE_ITEMS` (measured 50), so a single read of a busy head returned a partial list.

**Be precise about what that cost** — ersatztv#763's framing of it is too strong. Under the server
default (`created_unix DESC`) page 1 holds the *newest* rows and ids are monotonic with `created_at`,
so page 1 already carried both the true maximum id and every row newer than the mark — the only rows
the post-write check selects on. A single-page read therefore missed a raced verdict only if **more
than 50 rows were created inside the write window**, not merely on "a head with more than 50 rows".
What actually removed PR #761's stall is the deletion of the page-2 probe below, not the paging.

The walk earns its place for a different reason: it removes the gate's one fail-toward-SUCCESS path
from depending on an **undocumented ordering** the server honours only coarsely (page 1 came back
`114,112,113,111,110` — newest-first, but not strictly, and nothing in the API contract promises even
that). Both reads walk to a **validated empty page**, never terminating on a short one, retrying each
page once. The termination rule is page-size independent; the *capacity* is not — 20 requests of 50
validate at most 950 rows, since the twentieth must be the empty terminator. The terminator is measured
per endpoint and they differ: `/statuses/{sha}` returns `[]`, `/issues/{n}/timeline` a bare `null`,
`/commits/{sha}/status` an object with `statuses: null`.

**The two walks terminate by OPPOSITE rules, and that asymmetry is deliberate** (ersatztv#870, settled
by ersatztv#893 on 2026-09-02). `count_pr_mutations` reads the timeline to its cap and trusts the
counts only when the LAST page came back empty; `page_statuses` still returns on its FIRST empty page.
They differ because the ENDPOINTS differ, established from the v1.27.1 source: `getCommitStatuses`
serializes every row the paged query returns — an unconditional `append` loop, the only filter being a
SQL `WHERE` the database evaluates in the same query as the `LIMIT`/`OFFSET` — so a serialized page is
exactly as long as its database page, and an empty page really is the end. `ListIssueCommentsAndTimeline` instead drops
`CommentTypeCode` rows AFTER paging, which is what makes a fully filtered page byte-identical to the
end of the list. The same source explains the header: `/statuses/{sha}` sets `X-Total-Count` from a SQL
COUNT, so it is a true total (page 1 of 50 on a head reporting 63, measured 2026-09-02), where the
timeline reports its post-filter page length. **Do not "tidy" the two walks into agreement.**

**The page-2 probe is gone.** ersatztv#751 treated "there are rows I did not read" as "assume raced".
That fired on Renovate PR #761: a head that grew past one page over ordinary CI re-runs had its
exemption repaired away, with an `::error::` asserting a human verdict was overwritten when the head
carried none, and the sticky sentinel then refused re-exemption on every later run.

The two directions are **not** symmetric, and the asymmetry is deliberate:

- **Post-write, uncertainty fails closed.** An unreadable history, an over-cap history, or a count that
  does not come back as a number all repair to `pending`. Previously an unreadable history warned and
  left the exemption green while the page-2 probe repaired on the same uncertainty — one check
  disagreeing with itself.
- **Pre-write, a partial list still yields a mark.** The mark gates the post-write check entirely, so
  refusing one *skips* the check and is itself a fail-open. This rests on the DESC default above: the
  newest row, carrying the maximum id, is on page 1, so a walk that fails later still saw it. A
  *validated* empty history is a different case and is not abandoned: it yields a mark of 0, which is
  correct, since every later row is newer — that is the normal first run. What abandons the mark is a
  read that both **failed** and returned nothing, the pre-existing gap tracked as ersatztv#849 and
  unchanged here. A non-empty history carrying no numeric id is reported as unusable rather than
  collapsed to 0.

Two id comparisons are **numeric-only**, closing the *string* half of a fail-open (a type guard, not
a value bound — a corrupt but genuinely numeric id would still inflate the mark, which is not
attacker-controllable since ids are server-assigned): jq orders strings above every number, so one
`"id": "99999"` would inflate the high-water mark until nothing looked newer, and
`.id > $since` would read any string id as newer than any mark — making a pre-existing base-mismatched
verdict look raced on every run, a permanent stall. `.creator` is likewise type-tested before it is
indexed, because `.creator.login` on a non-object exits jq 5 and, under `set -e`, took the step down
after the green was posted and before the repair.

**Rejected alternatives**, recorded so they are not re-adopted:

- **A currency witness** (require a row above the mark, proving the read reflects the POST just made).
  Counting *any* row above the mark does not witness this job's write, so a stale-but-valid snapshot
  carrying an unrelated newer row still passes; and a schema-valid stale read is not retried, so one
  such response turns a transient anomaly into a permanent sentinel. Gitea here is a single instance
  with no read replicas, so the hazard has no mechanism behind it.
- **`sort=highestindex`** (index ASC, measured), which would close the mid-walk-insert gap below. ASC
  puts the *oldest* rows on page 1, which inverts the partial-mark fallback: the incomplete walk's
  maximum becomes the max of the oldest rows, leaving every pre-existing row above the mark and read as
  raced. That is the #761 failure, re-introduced to close a smaller one. A test asserts the request
  carries no sort parameter.

**Accepted residual: a row inserted mid-walk can be missed.** These are independent offset-paginated
GETs with no snapshot token, so under DESC a new row lands at position 0 — on a page already read —
while everything else shifts down. It is bounded: a row arriving after this job's POST is not one the
job overwrote, and being newest it wins on the combined endpoint branch protection reads.

**An empty post-write history is rejected.** The walk terminates on an empty page — correct before the
write, impossible after it, since this job has just POSTed and the endpoint returns one row per POST. A
well-formed "no statuses exist" is not retried (the body is valid), so accepting it would conclude
`raced=0` from a list that cannot be real, silently. This is *not* the withdrawn currency witness: that
asked whether ANY row sat above the mark, which an unrelated newer row satisfied; this asks only whether
the list is empty, which no unrelated row can produce.

The `::error::` names its own cause, and there are three: a verdict actually **found**; a read that
could not be **completed**; and a read that completed but returned an **impossible** answer. The third
is not a variety of the second — an operator holding a sticky sentinel needs to know whether to go
looking for an API failure that never happened. The sentinel *description* stays generic, because the
classification recognises it as a fixed point.

Three properties of this workflow are security-relevant and are **structurally** asserted by tests in
`scripts/tests/test_pr_changed_files.py` — those tests pin the workflow's shape, which is not the same
as establishing that the gate cannot be forged (see the residual below, and ersatztv#697/#698):

- **The trigger is `pull_request_target`, scoped to `branches: [main]`** — never plain
  `pull_request` (ersatztv#672). Gitea resolves a `pull_request` workflow *definition* from the PR's
  own head, so under that trigger a PR editing `review-verdict.yml` ran its own rewritten copy and
  could post `review-verdict/h10=success` for itself. The base-ref checkout below binds the scripts
  this job runs; only the trigger binds the definition. The `branches` filter is half the fix, not a
  refinement of it: base resolution means the *base branch* supplies the gate, so an unfiltered
  trigger merely moves the rewrite to an attacker-pushed base — and a status forged there is
  inherited by any later PR carrying the same head sha (ersatztv#663). `pull_request_target` is safe
  here **only** because this job never checks out or executes head-supplied code. Verified on this
  instance with four scratch PRs rather than inferred from GitHub; full rationale in
  `docs/decisions/records/ci/gate-trigger-base-resolved.md`. **This closes the rewrite route through
  this workflow, not the class:** `docker-build.yml` is also head-resolved and must stay on
  `pull_request` because it builds the PR's code, so it got the read-only status identity instead —
  the `ETV_STATUS_AUTH` it passed was a PAT scoped `write:package` + `read:repository`, which the
  status endpoint refuses (`ci.actions-credential-scoping`, ersatztv#697). Since ersatztv#885 it
  passes no credential there at all: the status read is anonymous, so the PR route materialises
  nothing to refuse. The inventory was never that one
  workflow, though: Gitea injects a write-capable `GITEA_TOKEN` into every job and branch protection
  binds the *context*, not its issuer. Gitea >=1.26 with the Actions default set to **Restricted**
  (server-management#714) binds the injected token, but does not close the class either — not against
  a personal token, and not against `RENOVATE_TOKEN` (ersatztv#742). **And none of it was necessary:
  direct pushes to `main` were server-side permitted, so the gate could be skipped without any forgery
  (ersatztv#743). That is now CLOSED — `main` carries `enable_push: false` **and**
  `block_admin_merge_override: true`, so it is reachable only through the PR merge path, the one path
  on which Gitea evaluates `status_check_contexts`, and an admin cannot `force_merge` past them
  (`release.main-direct-push-disabled` — neither field is citable alone).** Note the fix is *disabling* push, not whitelisting it: a
  push whitelist naming `timothy` was measured to still admit the push, and `timothy` is the identity
  every session, PAT and injected `GITEA_TOKEN` already acts as, so the whitelist form would have
  closed nothing. The block binds a site admin at pre-receive but not a credential that can first
  PATCH branch protection off — an accepted residual, recorded in that decision. The
  exemption path has separate defects of its own (ersatztv#698). One operational
  consequence of the trigger change: a PR whose base is not `main` now gets **no**
  `review-verdict/h10` at all. That is fail-closed. `edited` **is** now among the trigger's `types`
  (ersatztv#698), so a PR retargeted onto `main` reclassifies instead of staying statusless until its
  next push — but note that only gives *detection*: runs are not serialized, so a stale run can still
  post `success` after the reclassifying run posts `pending` (ersatztv#706).
- **The checkout takes the PR's BASE ref**, `ref: ${{ github.event.pull_request.base.sha }}` with
  `persist-credentials: false` — never the head. This job judges the PR, so the PR must not supply
  the code that judges it; a head checkout would let a PR edit the enumeration to return an empty
  list and exempt itself.
- **`scripts/jq-preflight.sh` runs in floor-only mode**, never `--expect`. This job writes a
  branch-protection-**required** status, so an exact version pin would turn any jq upgrade on the
  runner into a repo-wide merge deadlock.

A PR whose base predates ersatztv#658 has no such script on its base ref; that case posts `pending`
with the reason rather than dying with no status at all.

⚠️ **Changing `review-verdict.yml` itself: it is not exercised by its own PR.** Base resolution cuts
both ways — the PR editing this workflow runs the version already on `main`, so an edit goes live
**only on merge**, repo-wide, having never run. A broken edit merges green and then breaks the gate
for every subsequent PR, and the PR that would repair it is gated by the same broken workflow. Do not
trust the editing PR's own checks. Verify the way ersatztv#672 did:

1. Push a scratch **base** branch carrying the candidate workflow.
2. Open a throwaway PR from a scratch head *into that base*, so the candidate is the definition that
   runs. Have it post a **probe-named** context (e.g. `review-verdict/h10-PROBE`), never the real
   `review-verdict/h10` — a probe must not be able to forge the gate it is testing.
3. Read the resulting commit statuses to see which definition actually ran, then delete both
   branches.

The same shape is what makes a `branches:`/`types:` change verifiable at all, since neither can be
observed from the editing PR. Note step 2 requires the scratch **base**'s own `branches:` filter to
name that base — the definition comes from the base, so a base the filter does not admit produces no
run at all.

⚠️ **Never write an expression delimiter inside a `run:` body here — a comment is NOT inert**
(ersatztv#751, `ci.workflow-run-body-no-expressions`). A `run:` body is not shell when the runner
reads it. The runner scans the whole scalar for the expression opener and, on finding one, rewrites
the **entire** body into a single `format(...)` call so the result can be spliced back in. That
rewrite is all-or-nothing: a payload that does not evaluate fails the interpolation of the whole
scalar, and **the runner then drops the step and concludes the job `success`**.

That is not hypothetical. From 8f6d4f443 (2026-08-03) to 2026-08-06 the classify step **never ran**.
The #706 note above, explaining why a concurrency group does not work here, quoted a `concurrency:`
snippet containing a PR-number expression *as an illustration*, in a shell comment. `pr number` is not
a valid expression. So `review-verdict/h10` was posted by nothing but a human hand for three days,
both exemption classes silently stopped working, and every run reported success. The prose documenting
a fix disabled the fix.

**The silent green is the real defect.** An absent required status reads as "not reviewed yet", which
is indistinguishable from the correct pending state — so an ordinary PR looked ordinary while the gate
was dead, and the cost landed only where no human was in the loop. PR #739 (docs-only) merged
2026-08-05 with **zero** commit statuses on its head, and got in only because admin force-merge was
still enabled; ersatztv#743 removed that escape the next day, so a docs-only or Renovate-manifest PR
arriving after that would simply have been stuck with no bypass. The two Renovate PRs in the window
escaped by timing, merging minutes before the bad commit.

Three things now hold the line, and they are deliberately different in kind:

- **The prose names expressions instead of quoting them** — write "a
  `github.event.pull_request.number` expression", not the delimiters. Pass values in through the
  step's `env:` block, which is interpolated per value, so a bad payload there cannot take the body
  with it.
- **A start-marker guard turns a dropped step RED.** The classifier writes a marker as its first act
  and an `if: always()` step fails the job when it is missing. It asserts execution *started*, never
  that it completed — the classifier has several legitimate `exit 0` abstention paths. The guard's own
  body must stay expression-free, or the mechanism it guards against can delete the guard too, and
  that absence would be silent as well.
- **Two static guards**, in `scripts/tests/test_pr_changed_files.py`: no delimiter in *any* `run:`
  body of this file (absolute — a dropped step here is a dead merge gate, and its bodies are ~700
  lines of prose), and repo-wide, every expression payload's **head token** must name a context or
  function the runner can resolve (permissive, because the other workflows interpolate into `run:`
  legitimately — 5 occurrences today, in `ci-image.yml`, `docker-build.yml`'s `api-docs`/`format`
  and `pr-checks.yml`'s two git-diff gates; #756 removed `build`'s two and banned that job as
  well, so the ban now covers `test`, `migrations` and `build`). Be precise about the second
  one's reach: it catches the
  historical defect (`pr number`) and a nonexistent context, but **not** a syntactically invalid
  payload whose tokens are all known (`${{ github.ref == }}` passes), nor a renamed output
  (`steps.metadata.outputs.shortsha` passes — every token after the first is preceded by `.` and is
  skipped), nor an unclosed opener. Catching those needs an expression parser — it does **not**
  catch "a payload that cannot evaluate, wherever it sits".

Worth knowing why nothing caught this for three days: every *other* workflow-shape test in that file
reads `_code_lines()`, which strips comments. That is correct for what it was for, but it encodes the
assumption this bug falsifies. The strict test reads the raw scalar, and must never adopt
`_code_lines`.

⚠️ **A page past the end of `/issues/{n}/timeline` is JSON `null`, not `[]`** — and this instance is
not consistent between endpoints (`/issues/{n}/comments` returns `[]` when empty). The retarget
fence's timeline walk (`count_retargets`, renamed `count_pr_mutations` by #803) gated on
`type == "array"`, so it read the real terminator as *unreadable*:
the walk never reached a validated empty page, `rt_ok` was never `yes` for **any** PR, and the fence
therefore withheld **every** exemption `success`. Renovate and docs-only PRs got no status at all —
the same user-visible outcome as the dropped step above, by a completely unrelated route. So fixing
the interpolation alone would not have restored the exemptions.

Two things kept it invisible, and both are worth generalising:

- It shipped in the **same commit** (8f6d4f443) that stopped the step executing, so the fence had
  never once run in production. A guard's first real execution is not the same event as its merge.
- The **test double asserted the wrong shape while claiming measured fidelity.** Its comment read
  "Real shapes, measured on this instance and deliberately mirrored" and it printed `[]` for a page
  past the end. Every fence test was green against a response the server never produces, so the
  `array`-only gate was never exercised by the suite either. With the double corrected and the old
  gate restored, **most of the fence suite fails** — 18 tests when first measured at `c710db4a1`, 21
  once three more fence-dependent tests existed. The invariant is the point, not the count: they had
  all been passing for the wrong reason. (Given as a range on purpose: a bare count goes stale within
  a couple of commits, inside a section about stale claims.) When a double claims
  fidelity, that claim is a test assertion and needs re-measuring like any other.

The type is now read as a value (`case` over `jq -r 'type'`) rather than through `jq -e`, whose
exit-status semantics already bit this workflow once at jq 1.6, and both `null` and `[]` are read as
the same EMPTY shape. The regression test is parameterised over both because both are live on this
server. An empty page — `null` OR `[]`, which take the SAME rule since ersatztv#803 — is **no longer
accepted as exhaustion before the page cap** (ersatztv#870, 2026-08-30). Gitea pages this endpoint at the database level and
filters afterwards, so a page whose 50 rows are all inline-code comments serializes as exactly the
value the walk used to trust as the end of the list, while later pages still hold events — 50 comments,
which a PR author can create on their own PR, bought silence on both axes. The walk now SKIPS an empty
page before the cap, reads every page up to its 20-page cap, and trusts the counts only when the LAST
page it read came back empty — the same rule the status walk above states for the CAP, though not for termination: that walk still returns on its first empty page, and whether it needs the same change is tracked at ersatztv#893 rather than assumed (measured 2026-08-30, `/statuses/{sha}` returns a TRUE `X-Total-Count`, which this endpoint does not). **That
moves the inference to the cap rather than removing it, and the residual is real**: pages 1-19 cover
offsets 0..949, so a timeline of more than 1000 rows whose offsets 950..999 are all filtered still
returns an empty page 20, and the walk publishes a too-low count as trusted. Both re-reads share the
blind spot, so this is the same defeat at about 10x the price — the 50-row filtered block is the
same, but the timeline it must sit in grows from ~100 rows to over 1000, and on their own PR the
padding is the attacker's to supply — not a closed hole. It is NOT the status walk's over-cap residual: `/statuses/{sha}` does not
filter after paging, so only here can a page-20 response be empty while later pages hold rows. An empty
FIRST page still ends the walk untrusted: every real PR's first page carries events (spot-checked
non-empty across #752/#753/#749/#739/#717; the counts are deliberately not recorded here because
timelines grow and such figures go stale within days), so an empty first page is ANOMALOUS rather than
EXHAUSTED, and the walk should not certify "no retarget happened" from a response it cannot explain.
`X-Total-Count` cannot shortcut this: on this endpoint it is the post-filter length of the PAGE, not a
total (`?limit=1` returns 1 on a 14-row timeline), unlike `/activities/feeds`, which does return one.

**The same nil-slice shape bites `/commits/{sha}/status`** — a third instance. A head with no statuses yet returns
`{"state":"pending","total_count":0,"statuses":null}` (measured on PR #739's head). `read_existing_verdict`
gated on `.statuses | type == "array"`, so it hit its `exit 1` and posted nothing at all — fail-closed,
same user-visible outcome. `null` is now accepted there only when `total_count` is 0, so a body that
merely lost its array is still refused and an existing verdict is still protected from a transient
error. `scripts/pr-changed-files.sh` was swept and is unaffected (`pulls/{n}/files` returns `[]`).
**The generalisable rule: a nil Go slice serialises to `null`, so every list-shaped field on this API
is suspect and only a per-endpoint measurement settles it.**

**Establishing that "no verdict exists" needs a second page, and both arithmetic guards for it are
no-ops here.** `read_existing_verdict` concluding absence is what licenses posting an exemption over a
verdict the job cannot see, so that conclusion has to be earned. Two obvious checks were tried and both
proved empty:

- **`.statuses | length` vs `.total_count`** — `total_count` is the count for the **page returned**, not
  for the commit. Measured at 1.27.1 on `3aed43c6` (6 contexts): `?limit=1` returns
  `len=1, total_count=1`, `?limit=3` returns `len=3, total_count=3`. Equal by construction, so the check
  reads as a completeness proof while proving nothing.
- **"refuse when the page comes back full at the requested `limit=100`"** — this instance caps `limit`
  at the server-wide `MAX_RESPONSE_ITEMS`, **measured at 50** (`/issues?limit=100` returns 50). A
  response can therefore never carry 100 rows, and the comparison was **dead code**. The repo already
  documented that cap in `scripts/pr-changed-files.sh`, two test files and `ci.script-tests-job`; the
  guard was written against 100 anyway. Hardcoding 50 instead would
  re-break the day the setting changes.

So the job **asks the server, and only when it matters**: if the `review-verdict/h10` row is on page 1
there is nothing further to learn (this endpoint returns the latest status per *context*, and a context
cannot recur on a later page). When the row is absent it reads **page 2** — any rows there mean the list
runs longer than one page and a verdict could be beyond it, so it refuses instead of concluding absence.
Cap-independent by construction. Paging is real here: measured `?limit=3&page=2` returning three further
rows, and `page=9` returning the same `statuses: null` terminator.

The `total_count` zero-check also requires the JSON **type** to be a number: `jq -r` renders `0` and
`"0"` identically, so a text compare would accept a schema-corrupted `"total_count": "0"` as "no
statuses".

**The repo-wide expression guard scans PARSED scalars, not raw file text.** A delimiter in an ordinary
top-level YAML comment is inert — the runner never evaluates it — so redding on it is a false positive,
and this file has now produced that false red twice. PyYAML drops those comments. A `run:` body is
itself a scalar and keeps its *shell* comments, which is the point: inside a `run:` scalar a comment is
not inert. Verified both directions by mutation — an inert top-level comment passes; the same payload
in a run-body comment still reds.

**`CLAUDE.md` and `AGENTS.md` are now PROTECTED paths.** `DOCS_ONLY` matched them, so the documents
that *define* the completion protocol, the merge-consent convention and the H10 rule were themselves
docs-only-exemptible while `.claude/` was protected — the same self-exemption the gate rules out, one
directory over. Driving the real classify body with a lone `CLAUDE.md` change produced
`review-verdict/h10=success`. It is fixed here rather than deferred because restoring the exemptions is
what makes it reachable: no exemption `success` was writable at all while the classify step was
dropped. `README.md` is deliberately not listed — ordinary prose, no enforcement. For the same reason,
#706's known residual returns with the working fence: while `rt_ok` was never `yes`, route 1 was closed
by accident.

**That gap is now closed** — `docker-build.yml`'s `test` and `migrations` jobs are also required
contexts, and there a dropped step is **fail-OPEN**: the required check goes green having done no work,
which is strictly worse than an absent status (compare #684). ersatztv#756 gave those two jobs
per-**step** execution markers and extended the delimiter ban to them; see
"Dropped-step guard on the required jobs" above.

It lives in its **own workflow file** on purpose: `pr-checks.yml` sets `cancel-in-progress: true`,
and a cancelled run there would leave an exempt PR with no status and no further push to
re-trigger it. Its own job context (`Review verdict / Set review-verdict status`) is **not** the
required check — a workflow must not satisfy the gate merely by running successfully.

Full rationale: `docs/decisions/records/release/verdict-status-check.md` and
`docs/decisions/records/ci/shared-pr-file-enumeration.md`.

## CI toolchain image (`docker/ci/Dockerfile`, `.gitea/workflows/ci-image.yml`)

The jobs that need a toolchain — `test`, `migrations`, `functional-e2e`, `api-docs`, `format` —
run **inside** a shared image via `container:` instead of installing their toolchain per run
(ersatztv#390). They therefore carry **no** `setup-dotnet`, **no** `setup-node`, **no** `apt-get`,
and **no** `dotnet tool install`.

**What it ships:** .NET 10 SDK, Node 22, prod-identical ffmpeg/ffprobe, `git`/`python3`/`jq`/`zstd`,
the `dotnet-ef` + `dotnet-reportgenerator-globaltool` global tools (which the `migrations` and
`test` jobs used to install on every run — **bump those versions in the Dockerfile, not the
workflow**), and **headless Chromium** for the UI-E2E flows (below). Project dependencies (NuGet/npm)
are deliberately **not** baked in — they change per commit and stay on `actions/cache`
(`~/.nuget/packages`, `~/.npm`).

**Headless Chromium for UI-E2E (ersatztv#445).** `PLAYWRIGHT_BROWSERS_PATH=/ms-playwright` holds
`chromium-headless-shell`, installed with `--with-deps` at image build time so the `functional-e2e`
job installs no browser per run. Measured on this exact base: the headless shell is **267M** where
full `chromium` is **656M**, and `chromium.launch()` resolves to the shell anyway because
`web/playwright.config.ts` never asks for headed — the accepted tradeoff being that a *headed* run
inside this image would fail. Also verified on the real base rather than assumed: Chromium launches
as **root inside a container with no `--no-sandbox`/`chromiumSandbox:false` opt-out**, so the config
carries no sandbox workaround. The Dockerfile's build-time smoke test actually *launches* the browser,
so a missing system library fails the image build rather than a CI run.

⚠️ **`ARG PLAYWRIGHT_VERSION` must equal `web/package.json`'s `@playwright/test` pin**, which is
deliberately **EXACT** (no caret): Playwright ties a browser *revision* to the package version, so a
mismatch leaves no usable browser. Renovate bumps the npm pin but cannot know about this ARG — when it
does, bump the ARG, publish the new `:<sha>` (step 1 below), then update all five container pins.
`scripts/e2e-ui.sh` guards the drift by *launching* a browser up front and failing with exactly that
instruction (it probes by launch, not by path, because `chromium.executablePath()` reports the
full-chromium path that a headless-shell-only image deliberately lacks).

**How it's layered:** `FROM ersatztv-ffmpeg:8.1.2` + `COPY --from=mcr.microsoft.com/dotnet/sdk:10.0-noble-amd64
/usr/share/dotnet` — the same pattern `docker/Dockerfile` uses for the prod image. Our ffmpeg base is
`ghcr.io/linuxserver/baseimage-ubuntu:noble`, the same Ubuntu release as the SDK image, so the copied
SDK matches the base's glibc/ICU. Keep the ffmpeg tag on that `FROM` equal to the one
`docker/Dockerfile` pins, so CI's ffmpeg stays prod-identical — that fidelity is what the
ersatztv#299 seeded-media/scanner E2E follow-ups will need.

**Bumping the toolchain is a deliberate two-step.** The jobs pin an immutable `:<sha>`, never
`:latest`, so a bad toolchain push cannot break every job at once. Both steps land in the **same
PR** — a PR that changes `docker/ci/**` without moving the pin turns `ci-image-pin` red, and the
merge-consent hook reads the combined status, so it will not auto-grant:

1. **Publish.** Push the branch with the `docker/ci/Dockerfile` change as HEAD, then dispatch
   `ci-image.yml` on that branch; it publishes `ersatztv-ci:<sha>`. (`:latest` moves only from
   `main` — a human pointer; **jobs must never consume it**.) Since ersatztv#744 a branch *push*
   does not publish; see "Publishing from a branch is a dispatch, not a push" below.
2. **Pin.** Update the pin in `docker-build.yml` — **all five jobs together** — on top. That PR's own
   CI is what proves the new image works. The pin is repeated per job because
   `jobs.<id>.container.image` cannot read the workflow `env` context.

**The tag is exactly 7 hex chars — get the length right, not just the commit** (ersatztv#594).
`ci-image.yml` tags with `git rev-parse --short HEAD` under `fetch-depth: 1`, and that shallow clone
holds few enough objects that git always abbreviates to **7**. A full local clone abbreviates to **8**,
so the natural command prints one character too many:

```bash
git rev-parse --short HEAD     # 8 chars in a full clone — WRONG, no such registry tag
git rev-parse --short=7 HEAD   # 7 chars — what ci-image.yml publishes. Use this.
```

An 8-char pin names the *right commit* but *no existing image*: it satisfies a resolve-and-compare
check, then every `container:` job dies at image-pull with `manifest unknown`, which reads like a
registry outage rather than a one-character pin error. `ci-image-pin` therefore checks the pin's
**length** as an invariant separate from its correctness, and prints the exact tag to use.

> **Caveat worth knowing before you trust the 7:** `ci-image.yml` still tags with a plain `--short`,
> whose length git *auto-scales* to the object count. 7 is therefore an empirical property of today's
> shallow clone, not an enforced invariant — if that count ever crosses git's threshold, the publisher
> emits 8, the correct pin becomes 8, and the gate's hardcoded 7 goes permanently red demanding a tag
> with no image behind it. Making the publisher emit `--short=7` is tracked as ersatztv#597.

`ci-image.yml` triggers on pushes **to `main`** touching `docker/ci/**`, `workflow_dispatch` on any
branch, and a weekly Monday 05:00 UTC cron (base-image security updates; Gitea registers `schedule`
only from `main`). It runs on
`ubuntu-latest` — it was on `small` until server-management#639, where "docker-only" was found to be
a poor proxy for "small": this is a full buildx of the .NET toolchain image, the heaviest job in that
lane. Like `docker-build.yml`, it needs BuildKit's inline `http = true` for the HTTP
registry. Renovate tracks the Dockerfile's image pins (`dockerfile` manager, see `renovate.json`).

**Three container-specific gotchas** — worth knowing if you add a job or a step:
- **`sh`, not bash, is the default shell inside a `container:`.** act_runner runs steps as
  `sh -e {0}` (dash) because it can't assume bash exists in an arbitrary image — even though ours
  has it. Every bashism (`set -o pipefail`, arrays, `shopt`, `mapfile`) then dies instantly with
  `set: Illegal option -o pipefail`. `docker-build.yml` therefore declares a workflow-level
  `defaults: run: shell: bash`. **If you add a workflow with containerized jobs, do the same** —
  outside a container the shell defaults to bash, so this failure only appears once you containerize
  and it looks nothing like a shell problem (it surfaced as the `migrations` job dying in 0.13s).
- `actions/checkout` clones as root into a mounted workspace, which trips git's *"detected dubious
  ownership"* guard and breaks every `git` call in a step. Fixed in the Dockerfile with
  `git config --global --add safe.directory '*'`.
- *(defensive, not load-bearing)* The ffmpeg base sets `ENTRYPOINT ["ffmpeg"]` because it ships as an
  ffmpeg CLI, so the Dockerfile resets `ENTRYPOINT`/`CMD`. act overrides the entrypoint anyway
  (`entrypoint=["/bin/sleep" "10800"]`), so this is belt-and-braces for anyone running the image by
  hand — unlike the two above, which are real.

⚠️ **A REBASE invalidates the pin.** The pin must equal the short sha of the commit that touched
`docker/ci/**`, and a rebase **rewrites that commit's sha** — so `ci-image-pin` goes red on a branch
that was green before, with a pin that still resolves to a real (now-orphaned) commit and an image
that still exists in the registry.

### Publishing from a branch is a dispatch, not a push (ersatztv#744)

**A branch push of this file's committed form no longer publishes anything.** `ci-image.yml`'s
`push` trigger is `branches: [main]`, because Gitea resolves a `push` workflow's definition from the
pushed branch — so an unfiltered trigger meant any branch push touching those paths executed that
branch's own YAML, unreviewed, on a docker-capable runner holding the credential that writes
`ersatztv:prod` and the toolchain image five `container:` jobs run. Credential scoping
(`ci.actions-credential-scoping`) bounds what such a job may *do*; it never bounded whether it runs.

**Read that heading exactly as narrow as it is written.** The filter is loaded from the pushed ref
like every other line in the file, so a branch that *deletes* it re-enables the route in one edit.
What is closed is the **drive-by** case: an ordinary push of a legitimate `docker/ci` change
publishing an image nobody asked for, with no deliberate act anywhere in the loop. Against someone
who intends to run their own YAML this is not a boundary at all, and citing it as one is the mistake
to avoid — see #853, which settled its DISPATCH row; the `v*` tag-push and `pull_request:` rows are
live in #885.

**What this does not close — beginning with this file's own filter, then sorted by REF CLASS
rather than by which keyword a trigger carries.**
`docker-build.yml` has `branches: [main]`, so a keyword test passes it; it also has `tags: ['v*']`
and `pull_request:`, both head-resolved and both reaching the same registry credential. A tag push
is outside `release.main-direct-push-disabled`, so `git push origin <commit>:refs/tags/v99.9.9` runs
that commit's `docker-build.yml` and can publish `ersatztv:prod` — no dispatch needed. Separately,
`ci-image.yml`, `docker-build.yml`, `dependency-scan.yml` and `renovate.yml` all carry an
unrestricted `workflow_dispatch`, which loads its definition from the selected ref. The filter here
removes the *drive-by* route, where merely pushing a branch was enough; it is not a boundary against
a malicious writer.

**That remaining class was probed and ACCEPTED in #853 — read this before proposing a fix for it.**
Gitea 1.27.1 has no way to restrict `workflow_dispatch` by ref (the API takes `ref` as a required
free-form string) and no protected-environment concept to gate a secret behind — zero of its 308
documented API paths mention "environment", and the config file the running server names in its own
`--config` sets only `ENABLED` and `DEFAULT_ACTIONS_URL` under `[actions]`. More to the point,
restricting dispatch would close nothing: the `pull_request:` row above is cheaper than any
dispatch, because it runs **attacker-authored YAML**, which can name any secret in the repo store —
not only the ones the committed workflows reference. That is what makes it cover `renovate.yml`'s
`RENOVATE_TOKEN` without dispatching `renovate.yml` at all. (`dependency-scan.yml` references no
secrets at all.) Full reasoning and the probe: `ci.workflow-dispatch-ref-unrestricted`.

Six jobs in `docker-build.yml` used to hold `REGISTRY_PASSWORD` on that route, two of them required
contexts. Since ersatztv#885 **no job on the `pull_request` route names any stored secret**
(`ci.pr-route-carries-no-stored-credential`): the toolchain image is pulled through the registry's
anonymous bearer-token flow and the commit-status API is read unauthenticated, leaving the PAT to
`build`, which the route cannot reach. The invariant is held by
`scripts/tests/test_workflow_persist_credentials.py::test_no_PULL_REQUEST_route_job_names_a_STORED_secret`,
which derives the population from the git index by the predicate *every job of a
`pull_request`-triggered workflow that names a `secrets.*`* — never a list, and never "every
`container:` job" either, which names five of the six because `toolchain-preflight` is
container-free. The `v*` tag-push row is closed against the bot by `release.tag-protection-v-star`.

What that does **not** close, and must not be read as closing: `REGISTRY_PASSWORD` is still in the
repo's Actions store, and head-supplied YAML can still name it, `RENOVATE_TOKEN` or
`SERVERMGMT_DEPLOY_KEY` in a job of its own. "Push a branch, open a PR" still reaches every secret in
the store with **no act outside the ordinary contribution flow**. Bounding the store needs
per-environment secret scoping, which Gitea 1.27.1 does not have. What is removed is the routine
materialisation of a write-capable credential into six PR-run environments — blast radius, not the
route.

**A `v*` tag push does not fire `ci-image.yml`.** There is no `tags:` key, and a `branches:` filter
is compared against a branch ref — so a release cut no longer republishes the toolchain image as a
side effect. The exact matcher semantics are unprobed; the observable claim is the one stated.

The capability removed is real and is **not** assumed away — it is replaced by `workflow_dispatch`,
which Gitea loads from the branch you select. The constraint to plan around is that `ci-image.yml`
tags `git rev-parse --short HEAD`, i.e. **the branch HEAD at dispatch time**, not the commit that
touched `docker/ci`. So the branch-side two-step is:

1. Push the branch with the `docker/ci/**` change as **HEAD**. Anything that changes how the image is
   built — a `ci-image.yml` edit to build-args, Dockerfile path or platforms — must be **in that
   commit or below it**, because the dispatch builds HEAD's whole tree. Commit it afterwards
   and the published image was built by the *old* recipe while the pin and `expected` both name the
   `docker/ci` commit, so `ci-image-pin` is green over an image the merged recipe would not produce.
2. Dispatch `ci-image.yml` on that branch (Actions → *Build CI Toolchain Image* → *Run workflow* →
   pick the branch). It publishes `:<short sha of that HEAD>`; `:latest` stays put, because only
   `main` may move it.
3. Commit the pin bump in `docker-build.yml` on top. HEAD moves, but nothing needs republishing —
   `expected` is the `docker/ci` commit, which the pin now names.

After a rebase, redo all three: the sha is new, so the image must be republished under it. That is the
loop ersatztv#445 paid twice when `main` moved mid-review.

> **Probed on this instance 2026-08-27**, because the whole branch-side path rests on it. Dispatched
> `ci-image.yml` on `fix/744-ci-image-branch-filter` at `43b1e45` (run 2340, `conclusion: success`).
> Gitea bound the job to `head_branch: fix/744-ci-image-branch-filter` / `head_sha: 43b1e45b5554…`,
> so it loads both the definition and the tree from the selected non-default ref, and the run
> published `ersatztv-ci:43b1e45` — it tags the branch HEAD at dispatch time, as described above.
> `:latest` still resolved to `sha256:7baf403c…`, the digest it carried before the dispatch. Note the
> limit of that last check: `docker/ci` was unchanged, so the rebuild was content-identical and the
> new `:<sha>` carries the same digest as `:latest`. It therefore shows `:latest` did not CHANGE, and
> cannot by itself distinguish "not pushed" from "pushed the same digest" — the `refs/heads/main`
> guard in the workflow is what makes that decision.

The cheapest way to avoid the rebase loop entirely is still to **land a toolchain-image change on its
own, before** the work that consumes it, so the consuming branch never carries the `docker/ci` commit
through a rebase.

> **`docker/ci/**` has no cheap-edit path, comments included.** Editing any byte there re-points
> `ci-image-pin`'s `expected` at the editing commit, so the five `container:` pins go stale and a
> blocking job fails — a one-line comment fix costs a full publish-and-repin, and a rebase charges it
> again. Ride prose fixes in that directory with the next genuine toolchain bump. The TWO stale
> publish instructions still in `docker/ci/Dockerfile` — the `PLAYWRIGHT_VERSION` bump comment, and
> the rebase warning's "re-dispatching does not help", which is now the opposite of the documented
> recovery — are #854, deferred on exactly these grounds.

**A change that lives only in `ci-image.yml` publishes nothing.** That file is no longer one of its
own trigger paths, nor one of `ci-image-pin`'s `expected` paths. Those two lists must otherwise name
the same sources, and `scripts/tests/test_ci_image_paths_pin_agreement.py` holds them to it — it
derives both from the workflow documents and compares them for set equality, so editing one alone
reddens `script-tests`. That job is not a required context, so the red does not block server-side; it
denies the merge-consent auto-grant through the combined status. **Read that guard before taking the
reversal described next**: it models a `<dir>/**`-against-`<dir>` pair only, so putting
`ci-image.yml` back into both lists is a file-against-file pair it refuses, and would need a file arm
added there first (ersatztv#855). That pairing is a **decided
tradeoff, not a necessity** — keeping the file in `expected` still works, since the branch dispatch
can publish the ci-image.yml commit itself and the pin then matches. What it costs is the reason it
went: that route charges a full ~2 GB publish and a five-pin bump for *every* edit to the file, a
comment included, and a rebase charges it again. That price reverted the first attempt (#697) and is
why this change was landed alone — also a priced decision, not a rule. The chosen side pays the
opposite risk instead, below.
The consequence to live with: if you change **how** the image is built (build-args, Dockerfile path,
platforms) without touching `docker/ci/**`, CI keeps running an image built by the previous recipe
and nothing goes red. **Make the change alongside a `docker/ci/**` edit** — a comment bump in the
Dockerfile is enough. That is not one option of two: publishing after the merge and then pinning
does not work, because `expected` is the last `docker/ci` commit, so a pin naming the workflow-only
commit fails `ci-image-pin` and leaving the pin where it is keeps consuming the old image. The
`docker/ci` touch is what gives the new recipe a commit the pin is allowed to name.

### When the pinned tag disappears

⚠️ **An immutable PIN is a promise about what we consume, not about what the registry keeps.** It
means the jobs never follow a floating tag like `:latest`, so a fresh toolchain push cannot change
what today's CI runs. It does not promise the tag will still EXIST — nor, strictly, that the tag's
content is frozen: `ci-image.yml` tags `git rev-parse --short HEAD`, so a `workflow_dispatch` or a
weekly `no-cache` run at the same HEAD republishes that same `:<sha>` from a rebuilt image. Those are
three different claims, and existence is the one that is not ours to make: the registry belongs to
server-management, and an owner-level Gitea *package cleanup rule* there (`keep_count` 15,
`remove_days` 1, `remove_pattern` `.*`, and a `keep_pattern` that no 7-hex sha can match) deletes any
sha tag once 15 newer versions of the package exist. `ci-image.yml` publishes a new `:<sha>` weekly
and on every push to `main` touching `docker/ci/**` (plus every branch dispatch), while the pin only
moves when a human bumps it — so a pin ages toward eviction on its own. That is what happened between 2026-08-11 and
2026-08-13 (ersatztv#772): the tag vanished, and every `container:` job — **both required contexts
included** — died at image pull, before running a step, with

```
Error response from daemon: failed to resolve reference ".../ersatztv-ci:<pin>": not found
```

buried in each job's log. Nothing said "your toolchain image is gone", so the natural first reading
was "my diff broke the build", and that is where the review time went. The durable fix is
registry-side and is tracked in **timothy/server-management#842**; until it lands, assume any pin
older than a couple of weeks can evaporate.

**How firm that cause is, since it decides whether you go looking further.** The rule and its nightly
execution are directly observed; the specific deletion is not, because Gitea hard-deletes package
versions with no audit row. What ties them is the same rule's fingerprint on the sibling `ersatztv`
package — every `:<sha>` older than the 15-slot window gone, every `keep_pattern` tag kept back to
`26.3.1`. Reproduce both halves on the Gitea host (LXC 119, `192.168.1.95`):

```bash
# on the Gitea host: the rule itself
sqlite3 /var/lib/gitea/data/gitea.db 'select * from package_cleanup_rule;'
# from anywhere: that the cleanup task is scheduled and has been running (schedule/prev/exec_times)
curl -s -u user:pass 'http://192.168.1.95:3000/api/v1/admin/cron?limit=50' \
  | jq '.[] | select(.name == "cleanup_packages")'
```

(Do **not** reach for `journalctl -u gitea | grep ExecuteCleanupRules` — that identifier reaches the
log only via slow-query warnings, so an empty grep on a healthy host would read as "the rule never
ran", which is the inverse of what it means.)

If a pin disappears again *after* #842 changes that rule, treat this cause as refuted rather than
re-applying it — something else is deleting tags.

**Detection.** `docker-build.yml::toolchain-preflight` (`scripts/ci-toolchain-image-resolves.sh`)
resolves every pin in `docker-build.yml` against the registry on every run and fails with a message
that names the tag. It is container-free by necessity — a job consuming the missing image could not
run to report it — and deliberately **not** a `needs:` of the five jobs it diagnoses: the container
jobs already fail fast, so gating them would tax every green run to speed up a rare red one.

**Everything it cannot establish is a FAILURE, not a warning**, and the arms are worth knowing
because they send you to different places:

| Answer | Job | Message says |
|---|---|---|
| HTTP 200 with a manifest body | green | resolves |
| HTTP 404 | **red** | `IS GONE` — rebuild the tag (recovery above) |
| HTTP 200, body is not a manifest | **red** | something is answering for the registry (proxy, login page) |
| 401 / 403 that survives a bearer the run actually obtained | **red** | the registry `refused an ANONYMOUS read ... even after a Bearer token was obtained` — about the PACKAGE: check that the `ersatztv-ci` package is still PUBLIC. It is linked to no repository (measured 2026-09-05: every version reports `"repository": null`), so this is the package's own visibility and NOT `timothy/ersatztv`'s |
| 401 whose token leg ANSWERED and yielded no bearer — no `Www-Authenticate` at all, a challenge naming no realm, or a token endpoint answering with no token | **red**, asked once | `could NOT OBTAIN an anonymous pull token ... after 1 token-leg attempt(s)` — about the registry's TOKEN ENDPOINT, not the package's visibility. The endpoint said something, so asking again cannot change it |
| 401 whose token endpoint could not be REACHED, or answered 5xx | **red** after `ETV_CI_ATTEMPTS` token-leg tries | the same message, naming the attempts it actually made. An endpoint that said nothing is the same transport blip a flaky manifest read gets to survive — the two legs of one read must not have opposite flake tolerances when a red here denies a merge |
| 403 on the FIRST read, before any token leg — `probe` enters it on a `401` only | **red** | `refused an ANONYMOUS read ... and NO TOKEN WAS EVER REQUESTED` — it reports what the RUN did (never followed this answer as a challenge), not what the registry sent. About ACCESS to the registry: visibility, or something answering for it |
| anything else (5xx, unreachable, no `curl`) | **red** after `ETV_CI_ATTEMPTS` tries | `could NOT VERIFY` — check the registry's health, NOT the pin |

**The refusal rows say only what actually ran**, and that is the point of wording them apart
rather than a nicety: a message naming a step the run skipped is evidence for a diagnosis nobody
performed. So the arm branches on `token` first — a bearer was obtained — then on whether the token
leg was attempted at all, and only then reports the never-asked case; and the token-leg message
carries its own attempt count rather than leaving the reader to infer one from the retry policy.

Those refusal rows and the `could NOT VERIFY` row are the ones worth defending, because warning
on them and exiting 0 is the natural way to write this check and it is wrong: a missing `curl`, a
moved registry and a DNS change all land in the last of them, and a green-with-a-warning job is
indistinguishable from a healthy pin forever after. The
unknown arm retries first (`ETV_CI_ATTEMPTS`, default 3, `ETV_CI_RETRY_SECONDS` apart) so an ordinary
registry blip does not redden a PR — that pause is what makes failing on unknown affordable, and
shortening it silently trades this guard for flake. **An unreachable token endpoint is in that
class, not in the refusal class**, and is retried on the same budget: the split is by what the
endpoint SAID, never by which leg of the read it happened on.

**It is not a `needs:` of anything, but it is not consequence-free either.** The merge-consent hook
reads the PR's **combined** status and denies on a non-`success` combined state (a `skipped` context
counts as green, ersatztv#593; an advisory red does not, ersatztv#598), so a red preflight blocks the
merge exactly like any other red job. "Advisory" would be the wrong word for
it — what it does not do is *skip* the jobs it diagnoses.

**Recovery, without needing CI to be healthy.** The tag names a commit, and that commit still builds
the same image, so the fastest fix is to republish the *same* tag by hand — no PR, no pin bump, no
green CI required, and every open branch recovers at once. Run this on a host with docker and this
registry in `insecure-registries` (bumblebee `192.168.1.99` or jazz `192.168.1.29`):

```bash
repo=$(pwd)                                  # keep the CURRENT checkout: the pin commit predates
                                             # the preflight script and the verify step below
pin=$(grep -oE 'ersatztv-ci:[0-9a-f]+' .gitea/workflows/docker-build.yml | cut -d: -f2 | sort -u)
git worktree add /tmp/etv-toolchain "$pin"   # the pin IS the commit's short sha
cd /tmp/etv-toolchain

# The registry is HTTP-only and BuildKit does NOT inherit the daemon's insecure-registries, so the
# `docker-container` driver (anything created by `docker buildx create`) will try HTTPS and fail.
# Either build on the default `docker` driver — `docker buildx use default` — or give the container
# driver the same inline config ci-image.yml passes it:
#   [registry."192.168.1.95:3000"]
#     http = true
prev_builder=$(docker buildx inspect 2>/dev/null | awk '/^Name:/{print $2; exit}')
docker buildx use default                    # needs the containerd image store to --push;
                                             # both named hosts have it (checked 2026-08-22)
docker login 192.168.1.95:3000 -u timothy
docker buildx build --platform linux/amd64 --provenance=false \
  -f docker/ci/Dockerfile -t "192.168.1.95:3000/timothy/ersatztv-ci:$pin" --push .

cd "$repo" && git worktree remove /tmp/etv-toolchain
[ -n "$prev_builder" ] && docker buildx use "$prev_builder"   # leave the builder as you found it
```

Then confirm the tag resolves before re-running anything — the preflight script does exactly this
check and takes no arguments. Run it from the CURRENT checkout, not the pin worktree, which is why
`$repo` is kept above:

```bash
scripts/ci-toolchain-image-resolves.sh
```

**What this rebuild does and does not restore.** It restores a *working* toolchain at that tag, built
from that commit's `docker/ci` — not a bit-identical copy of what was deleted: the base image tags
and the apt/NodeSource packages the Dockerfile pulls are mutable, so a rebuild picks up whatever they
point at today. That is the same exposure the weekly `no-cache` cron has by design. Prefer this over
the two-step above whenever the pin is *missing* rather than *stale*: the two-step exists to move the
pin to a NEW image, and running it here would leave the repo pinning a different sha for no reason.

The push path was exercised against this registry on 2026-08-22 — a throwaway `docker push` of a
13 MB image to `timothy/etv-772-recovery-probe:probe1` from bumblebee, `HEAD /v2/.../manifests/probe1`
→ `200`, then `DELETE /api/v1/packages/timothy/container/etv-772-recovery-probe/probe1` → `204` and
the manifest read back `404`. Re-run that shape against a scratch package name to re-establish it;
what it establishes is the auth + HTTP-registry push path, not the toolchain build itself.

**Bumping the pin is enforced, not remembered.** The `ci-image-pin` job (blocking, PR-only; defined
in `pr-checks.yml`, but it greps `docker-build.yml` where the pins live) fails if
`docker-build.yml`'s pin isn't the short sha of the last commit to touch `docker/ci/**` (since
ersatztv#744 that path alone — `ci-image.yml` is deliberately not part of it), if that pin isn't
exactly 7 chars long (see above), or if the five jobs ever pin different tags. This exists because **Renovate manages
`docker/ci/Dockerfile`'s base pins but cannot bump an opaque `:<sha>`** in `container.image` — so a
Renovate base bump would otherwise publish a new image, test the *old* one, and merge with the
Dockerfile disagreeing with the pin. A red `ci-image-pin` means: dispatch `ci-image.yml` on the
branch to publish the new
`:<sha>`, then update all five pins to it.

**What it is and isn't worth.** Measured honestly (ersatztv#390): the image saves ~15–40s per job
(`setup-dotnet` is 8–19s, `setup-node` 2–5s cached, the two tool installs ~9s) plus the 110s
apt-ffmpeg step — roughly 3–8% of runtime. It is **not** where CI time goes; see the lane table above
(queue wait, server-management#604) and ersatztv#398 (742s of redundant compilation). Its durable
value is prod-identical ffmpeg, a pinned/consistent toolchain, and making jobs runner-agnostic — the
last is what allowed the lane rebalance.

## Dockerfile notes (`docker/Dockerfile`)

- Base image: **`192.168.1.95:3000/timothy/ersatztv-ffmpeg:8.1.2`** (our Gitea fork of
  the archived `ghcr.io/ersatztv/ersatztv-ffmpeg`). FFmpeg 8 base image work landed in
  ersatztv-ffmpeg#4; app-side compatibility work landed in ersatztv#9.
- Copies `Directory.Build.props`, `Directory.Build.targets`, `Directory.Packages.props`,
  `global.json`, `.editorconfig` before `dotnet restore` so the image build uses the same
  MSBuild config, central package versions, SDK pin, and analyzer severities as local/CI
  builds (it previously copied only `*.sln`). `Directory.Packages.props` is **required**
  here: under Central Package Management the csproj carry no inline versions, so the
  image's restore fails (`NU1015`) without the central manifest.
- amd64-only (the runner/build host is x86_64). No arm32/arm64, no DMG/exe artifacts, no GHCR/DockerHub.
- **openapi-generator jar layer ordering** (ersatztv#190): the `wget` for the
  openapi-generator-cli jar runs *before* the `COPY` of `ErsatzTV/wwwroot/openapi/`, so the
  ~30MB download layer is cached independently of the openapi spec. Previously the jar was
  downloaded after that `COPY`, so any PR touching the spec (e.g. `v1.json`) busted the
  download layer too and re-fetched the jar on every such change. Codegen itself still runs
  after the spec `COPY`, since it needs both the jar and the spec files.

## Dependency management (Central Package Management + scans)

**Central Package Management (CPM)** — package versions live in a single repo-root
`Directory.Packages.props` (`ManagePackageVersionsCentrally=true`); the per-project
csproj reference packages by name only (no `Version=`). One source of truth, atomic
one-line bumps, and cross-project version drift is structurally impossible. To add or
change a dependency, edit the `<PackageVersion>` entry centrally — never put a `Version=`
back on a `<PackageReference>` (that trips `NU1008`). The Docker build must copy this file
before restore (see Dockerfile notes). The `.mcp/` vendored tool (gitignored, not in the
solution) keeps inline versions via a **local-only** `.mcp/Directory.Packages.props`
opt-out (`ManagePackageVersionsCentrally=false`). (ersatztv#14)

**NuGet audit** — .NET 10 runs NuGet audit on restore. Several projects set
`TreatWarningsAsErrors=true`, so vulnerable transitive packages failed the build.
`Directory.Build.props` demotes low/moderate/high advisories (NU1901-1903) to warnings
and promotes NU1904 (critical) to an error in **every** project via `WarningsAsErrors`.
The advisories that prompted this were resolved in ersatztv#8 (NCalcSync→6.x; SQLitePCLRaw
bundle 3.x) and ersatztv#314 (**Microsoft.OpenApi 2.0.0→2.7.5**, GHSA-v5pm-xwqc-g5wc High —
direct-pinned in `ErsatzTV.csproj` over the 2.0.0 that `Microsoft.AspNetCore.OpenApi` +
`Scalar.AspNetCore` pull transitively; the SQLitePCLRaw override pattern; regenerates the
OpenAPI doc byte-identically). The NU1901-1903 demotion is **kept by design**: criticals (NU1904)
still hard-block, while low/moderate/high advisories surface as warnings + via the weekly scan and
Renovate security PRs, rather than breaking unrelated PRs the moment a new transitive
advisory drops.

**Scheduled vulnerability scan** — `.gitea/workflows/dependency-scan.yml` runs weekly
(cron `0 6 * * 1`) + on `workflow_dispatch`: `dotnet list package --vulnerable
--include-transitive` over the **full** solution (incl. Scanner, which the image build
strips). `dotnet list` exits 0 even with findings, so the step (`bash -euo pipefail`)
greps for the "has the following vulnerable packages" marker and fails the run if present.
Detection only — it surfaces advisories on a schedule, a Gitea-native stand-in for
Dependabot; it does **not** open update PRs (that's Renovate — server-management#484).
Gitea registers `schedule` triggers only from the default branch, so the cron starts
after merge to `main`; use `workflow_dispatch` to run on demand. It went **green** once
ersatztv#8 cleared the NCalcSync/SQLitePCLRaw advisories — a red run now means a **new**
advisory has appeared. (ersatztv#14, ersatztv#8)

**Renovate (automated update PRs)** — `.gitea/workflows/renovate.yml` runs self-hosted
[Renovate](https://docs.renovatebot.com) weekly (cron `0 3 * * 1`) + on `workflow_dispatch`,
as a `renovate/renovate:43` container job on the shared act_runner. This is the *proposing*
layer the scan above deliberately omits: it opens grouped dependency-update PRs and
OSV-driven vulnerability-fix PRs against `main`, and maintains a **Dependency Dashboard**
issue listing the full backlog. Config is the repo-root `renovate.json` — managers `nuget`
(via CPM), `github-actions`, and `dockerfile` (scoped to the built `docker/Dockerfile`; it reads the
HTTP-only Gitea registry for the `ersatztv-ffmpeg` base via a `RENOVATE_HOST_RULES` host rule —
`insecureRegistry` + registry read creds, set in the workflow env, not the committed config). The
docker-compose manager is unused (repo compose files are `build:`-only). Auth: a dedicated
**`renovate` Gitea bot** (Write
collaborator) via repo Actions secrets `RENOVATE_TOKEN` (bot PAT) + `GH_COM_TOKEN` (no-scope
github.com PAT for changelogs — named `GH_`, **not** `GITHUB_`, a prefix Gitea reserves).
**Patch** bumps to test/dev-only packages (NUnit\*, NSubstitute, Shouldly, coverlet,
`Microsoft.NET.Test.Sdk`, `Testably.Abstractions*`, threading analyzer) **auto-merge** once
the `Build & test (.NET)` check passes — branch protection on `main` requires that context;
everything else is manual review (ersatztv is prod-bearing). Range-pinned packages (e.g. EF
Core `[9.0.x,10)`) are respected — no v10 jump. PR volume is throttled (`prConcurrentLimit`
5 + `config:recommended`'s `prHourlyLimit` 2); tick a dashboard checkbox or raise the limits
to drain faster. `workflow_dispatch` defaults to a safe **dry run**. Cross-repo rollout
tracked in server-management#484. (server-management#484)

## Security scanning — black-box DAST + SAST (`scripts/security-scan.sh`, ersatztv#314)

Every other security check we run is **in-ecosystem / white-box** — SonarAnalyzer, NetArchTest, the
adversarial fork + Codex review passes, the api-docs/format/decisions CI gates, `dotnet list package
--vulnerable` — so they share our blind spots. `scripts/security-scan.sh` is the **out-of-ecosystem,
black-box** complement and a **#197 exit criterion** (HARD GATE before remote exposure): it drives the
running product from outside our C#/review stack.

- **What it does.** Boots a **throwaway** container from the image under test (fresh empty config volume;
  never the deployed prod/test container — the authenticated *active* scan sends attack payloads to write
  endpoints), reads the generated machine key, and runs an **authenticated OWASP ZAP** API scan
  (`zap-api-scan.py`) that imports the static `/openapi/v1.json` so it exercises **every** declared
  `/api/v1` operation, injecting `X-Api-Key` on every request via a ZAP replacer rule so it reaches the
  `[RequiresAuthentication]` + `RequireKeyForReads` surface (not just the `/app` shell an unauthenticated
  spider sees). Then a **semgrep** SAST cross-check (`p/security-audit` + `p/secrets` + `p/csharp`). The
  container is torn down on exit.
- **Where/when.** Runs on the **docker host** (jazz — the Mac has no docker), like
  `migration-smoke.sh`: `scripts/security-scan.sh [IMAGE] [PORT]` (defaults `…:latest` / `8411`). It is a
  **manual release-gate**, deliberately **not** a per-PR CI job — it needs docker + a booted image, takes
  several minutes, and is noisy (expect to tune, not take raw). The continuous layer is the per-PR
  white-box gates + the weekly `dependency-scan`; this is the per-release black-box pass. Re-run it each
  release and before any change to the exposure posture.
- **Exit-code contract (ersatztv#338).** `zap-api-scan.py`'s raw exit code is NOT a simple pass/fail — it
  conflates a clean run with a warnings-only run unless you know its wrapper contract: **0** clean (no
  FAIL or WARN alerts), **2** WARN-only (triage required, but **not** release-blocking), **1** FAIL (at
  least one FAIL-level alert — release-blocking), **124** the script's own `timeout` wrapper killed a
  hung post-scan cleanup (the report written before the hang is still usable — triage it), any other
  code means the scanner/tool itself errored (not a scan result at all). `scripts/security-scan.sh`
  encodes this in `classify_zap_exit()` and prints an unambiguous `==> ZAP result: <PASS|WARN|FAIL|
  TIMEOUT|TOOL ERROR> ...` line; the script's own exit status reflects that classification (0 for
  clean/WARN, 1 for FAIL/timeout/tool-error) rather than ZAP's raw code, so a warnings-only run no longer
  reads as a failed scan. Found when the v26.8.0 release scan (#335) returned raw exit 2 for a report with
  `FAIL-NEW: 0` and two known/expected warning classes — the shell result looked like a failure though the
  release gate had actually passed. Run `scripts/security-scan.sh --selftest` for a docker-free regression
  check of the classification logic.
- **Triage.** Triage each WARN/FAIL finding false-positive vs real. Real, in-scope, go-live-blocking
  findings get fixed (e.g. the security headers from the #319 baseline; the Microsoft.OpenApi pin above);
  LAN-expected noise (Private-IP disclosure) is revisited only for genuine remote exposure. nuclei
  (template-based CVE fingerprinting) is an optional third pass — deferred while its template fetch is
  blocked in the runner env (pre-seed a template volume to add it); ZAP covers the DAST baseline and
  semgrep the SAST, so it is not on the critical path.

## Static analysis & formatting

**Analyzers** — `Directory.Build.props` enables the SDK analyzers at `latest-All` and turns on
`Microsoft.VisualStudio.Threading.Analyzers` for every centrally managed project.
`Directory.Build.targets` also references
**Roslynator**, **SonarAnalyzer.CSharp**, **Meziantou.Analyzer**, and **AsyncFixer** repo-wide (versions
central via CPM). All analyzer package references are guarded on `ManagePackageVersionsCentrally`, so the
gitignored `.mcp` tool—which deliberately uses inline package versions—does not inherit versionless
references.
They are introduced **incrementally** (ersatztv#15). `eng/analyzers/sdk-all-suggestion.globalconfig`
enumerates the .NET 10 SDK `All` inventory at `suggestion`; this exact-ID baseline is necessary because
the SDK's generated `latest-All` severities outrank `.editorconfig` bulk settings. `.editorconfig` keeps
the threading and curated-pack baselines at `suggestion`. Diagnostics remain visible to IDEs and
`dotnet format analyzers`, but do not create a wall of failures (a direct `latest-All` trial activated
455 existing errors in the TWAE projects).

**Promotion is the enforcement** — set a reviewed rule to `warning` in `.editorconfig` and append its ID
to the central `WarningsAsErrors` list in `Directory.Build.props`. The explicit list makes the rule block
in every project, including test projects that do not otherwise use TWAE. On a major SDK upgrade,
regenerate the checked-in SDK baseline from `analysislevel_<major>_all.globalconfig`, preserve SDK `none`
entries, and review newly introduced rules before accepting the snapshot.

Promoted rules are recorded here so the blocking subset stays intentional and reviewable:
- **Sonar `S3981` — `warning` + `WarningsAsErrors`** (ersatztv#15): rejects collection-count comparisons that are constant
  regardless of collection size. Its first finding exposed `Workers.Count >= 0`, which permanently
  classified scheduled memory releases as busy and skipped the intended aggressive idle collection.

- **StyleCop.Analyzers is intentionally excluded**: its latest stable (1.1.118) crashes (`AD0001`) on
  C# `record` declarations, and its rules overlap the existing `.editorconfig`/Roslynator. Revisit via
  the record-compatible `1.2.0-beta` only if specifically wanted.
- **The former Blazor `.razor` caveat is retired**: Blazor removal deleted the Razor sources and their
  temporary Sonar `NoWarn` list. The `.razor`/`.cshtml` suggestion scopes remain in `.editorconfig` only
  as a defensive default if server-rendered view code is ever reintroduced.

**Formatting** — the inherited tree still contains legacy UTF-8 BOM/whitespace debt, so the standing
policy is **format as you touch**, not a mass rewrite (ersatztv#311). The Husky pre-commit hook and
the blocking `format` CI job both run `dotnet format whitespace . --folder --verify-no-changes
--include <changed .cs>` — scoped to the files the commit/PR touches. Untouched legacy files remain
outside the gate; `.gitattributes` pins line endings. A one-time full-tree normalization remains a
separate, unmade decision.

*Why `whitespace . --folder`, not the full `dotnet format <sln>` (ersatztv#469)* — the gate only
needs to enforce `.editorconfig` **whitespace** (indent/EOL/trailing/final-newline) and **charset**
(no UTF-8 BOM). The old recipe (`dotnet format ErsatzTV.sln --no-restore --verify-no-changes
--include`) loaded the entire ~10-project MSBuild workspace and built a Roslyn compilation per
project *before* checking a single file — `--include` narrows *which* files are checked, never what
gets loaded. Measured whole-solution `dotnet format` ran **~480s locally**; folder mode runs in
**~0.5s** and needs no `dotnet restore` (the NuGet-cache + Restore steps were removed from the job).
`--folder` treats the tree as a plain folder of files, skipping MSBuild/Roslyn entirely, and still
reads `.editorconfig`. Verified **non-vacuous**: it exits non-zero on an injected trailing-whitespace
line (`error WHITESPACE`) and on a prepended UTF-8 BOM (`error CHARSET`), and exits 0 on a clean file.
**No coverage was lost**: the full `dotnet format` gate did **not** enforce the style/analyzer pass
either — a probe injecting a `warning`-severity naming violation (`local_constants` not `ALL_UPPER`)
passed the *full* solution format (exit 0): the only `.editorconfig` rule above `:suggestion`/`:none`
severity is that one naming rule, and naming violations have no `dotnet format` batch code-fixer, so
`--verify-no-changes` reports no change regardless of severity. The analyzers that **must** block
(`NU1904`, `S3981`) are enforced at *compile time* via
`WarningsAsErrors` in `Directory.Build.props`, not by this job. Devs fix a violation with `dotnet
format whitespace . --folder --include <files>` (the full `dotnet format ErsatzTV.sln --include
<files>` is a superset and also works).

## Migration integrity (EF Core, both providers)

`TvContext` (`ErsatzTV.Infrastructure/Data/TvContext.cs`) has **two** migration sets — one per
provider project: `ErsatzTV.Infrastructure.Sqlite/Migrations` and
`ErsatzTV.Infrastructure.MySql/Migrations`, each with its own `TvContextModelSnapshot`. **A model
change needs a migration in BOTH.** Add them with `scripts/add-migration.sh <Name>` (runs the EF CLI
for each provider). The EF CLI pattern (provider selected by the post-`--` arg, which `Startup`
reads as the `provider` config key):

```
dotnet ef <cmd> --context TvContext --startup-project ErsatzTV \
  --project ErsatzTV.Infrastructure.{Sqlite|MySql} -- --provider {Sqlite|MySql}
```

The **`migrations` job** in `docker-build.yml` runs on every push/PR and, for **each** provider:

1. `dotnet ef migrations has-pending-model-changes` — fails if an entity changed without a matching
   migration (**model drift**), so a forgotten migration can't merge.
2. `dotnet ef database update` against a **fresh empty DB** — applies *all* migrations in order and
   fails on any broken/un-orderable one.

- **SQLite** (the prod provider) uses a throwaway file (`ETV_CONFIG_FOLDER=$(mktemp -d)`); no service
  needed. Validated: 787 migrations → 139 tables.
- **MySql** uses `ServerVersion.AutoDetect`, which **connects at config time**, so the job needs a
  reachable server — provided by a `services: mysql:8.4` container (the act_runner uses Docker
  execution with an auto-created per-job network — service reachable as `mysql:3306` (the old
  bumblebee runner pinned network `downloadswarm`; relocated in server-management#570)). Connection
  string via `MySql__ConnectionString` (→ config key `MySql:ConnectionString`). Validated: 305
  migrations → 137 tables. It's an **independent gate** (not yet a `needs:` of the image build) so
  the new MySql-service dependency can't block image builds until it's proven; promote it to a
  required check once stable.

**Caveat — non-transactional operations**: some migrations (e.g. SQLite `PRAGMA foreign_keys`) run
outside a transaction and warn at startup; they can't be rolled back mid-migration, so review such
migrations carefully (this is part of what motivated the apply-to-fresh check before the prod
cutover, server-management#481).

**Resilience — the MySql apply is retried (concurrent-runner contention, not a model bug)**: both
runners (`ci-runner` VM 127 + `bumblebee-runner`) serve `ubuntu-latest`, and when two migration jobs
land on the **same host at once** (common when several PRs push together), each spins its own
`mysql:8.4` service container and they starve each other — producing intermittent `Command Timeout
expired` **or** mid-replay `MySqlEndOfStreamException` (dropped connection) on the MySql
apply-to-fresh-DB step. This is pure infra flakiness — `has-pending-model-changes` (the actual model
check) still passes, and the *same commit* passes on a quieter host. The job hardens against it two
ways: the connection string sets `DefaultCommandTimeout=300` (up from MySqlConnector's 30s default),
and the apply is wrapped in a **3× retry** that resumes from `__EFMigrationsHistory` (EF commits each
migration in its own transaction, so an interrupted one rolls back and the retry continues). A *real*
migration failure fails deterministically on every attempt, so the retry never masks it. If a run
still flakes past the retry, re-trigger (Gitea has no rerun API on this version — push, or the run
drains); don't treat a lone MySql-apply red as a code problem without checking the failure mode.

### Migration-on-prod-copy smoke — release path (`scripts/migration-smoke.sh`, ersatztv#315)

The `migrations` job above only proves a migration is **well-formed against a fresh, empty DB**. It
can't prove it applies cleanly to the **accumulated prod SQLite** — real row volume, historical values,
and the post-migration data steps ErsatzTV runs on startup: `DatabaseMigratorService` (a
`BackgroundService`) applies pending migrations, then `DbInitializer.Initialize` + `PopulatePathHashes`
(an `UPDATE` over the real `MediaFile` table). A migration green on a fresh DB can still fail or corrupt
on prod, and today you'd only find out mid-deploy after the container recreates.

`scripts/migration-smoke.sh` rehearses it on a **throwaway copy** of the latest prod backup — it never
touches the live DB:

```
scripts/migration-smoke.sh --image <ref-about-to-be-promoted> [--db <backup.sqlite3>] [--timeout 180]
```

It copies the backup into a temp config dir, boots the new image against it (`ETV_CONFIG_FOLDER`), and
**gates PASS on the `Done applying database migrations` log line** — not merely on HTTP readiness, since
the migrator runs concurrently with Kestrel, so the web server can serve before/while migrations run.
FAIL = the container exits before finishing, a migration exception appears in the logs, migrations
don't finish within `--timeout`, or the app won't serve `/iptv/channels.m3u` afterwards. The smoke
container, the DB copy, and the temp dir are always torn down on exit (the ErsatzTV image runs as root,
so cleanup deletes its root-owned config files from inside a throwaway root container — otherwise each
run would leak the multi-hundred-MB copy). Exit 0 = clean, 1 = migration/boot failure, 2 = usage error.

- **`--db` default**: the newest `ersatztv.sqlite3` under `$ETV_BACKUP_DIR`
  (default `~/downloadswarm/ersatztv-backups` — where the host-side pre-deploy backup hook writes
  timestamped snapshots). Pass `--image` = the version tag about to be promoted.
- **Where it runs**: it's meant to run **on the docker host** as a **Komodo pre-deploy step** (which
  already produces the backup — see "Cutting a release" and the #553 pre-deploy backup caveat), so a
  bad migration aborts the promote before the live container recreates. Wiring it into that hook is a
  server-management concern (cross-repo — this repo owns the script + docs, server-management owns the
  Komodo hook). Until wired, run it by hand before cutting a migration-bearing release.
- Validated live 2026-07-12: `:latest` against a copy of the 283 MB prod backup → migrations applied
  cleanly, app booted and served, temp dir removed.

## Pre-commit hooks (web/)

The repo uses **husky** git hooks (installed via `web/`'s **lint-staged** + npm) to catch
lint/format/type/API-drift errors locally, before they reach CI. Because the git root and
the npm project dir differ (monorepo: no root `package.json`, the JS/TS project lives
entirely in `web/`), the wiring is:

- `husky` + `lint-staged` are devDependencies of `web/package.json` (not a root package —
  there isn't one).
- The committed hook scripts live at the repo root: `.husky/pre-commit`, `.husky/pre-push`,
  `.husky/commit-msg`.
- `web/package.json`'s `prepare` script (`cd .. && husky`) runs on `npm install` inside
  `web/` and points git at the repo-root `.husky` dir (`git config core.hooksPath
  .husky/_` — the `_` subdir is husky's generated internal dir, gitignored via its own
  `.husky/_/.gitignore`; only the hook scripts themselves are committed). This works
  because npm keeps `web/node_modules/.bin` on `PATH` for the `prepare` script even after
  it `cd ..`s to the repo root (which husky's init requires — it hard-checks for `.git`
  in the *current* directory).

**The four hooks:**

1. **`pre-commit`** — (a) `cd web && npx lint-staged`: runs `eslint --fix` on staged
   `web/src/**/*.{ts,tsx}` files, then a project-wide `npm run typecheck` (`tsc -b` isn't
   file-scoped, so it runs the full check, but only when a `.ts`/`.tsx` file is staged);
   (b) back at the repo root, if any **`*.cs`** files are staged, `dotnet format
   ErsatzTV.sln --verify-no-changes --include <staged .cs>` — a formatting violation
   blocks the commit. The .cs step is **skipped entirely when no .cs is staged**, so
   web-only commits don't pay the sln-load cost; when it does run it's scoped to the staged
   files (~6-7s wall in practice, dominated by the workspace load); (c) **H3** (ersatztv#303) —
   refuses a staged **root-level `*.png`** (`git diff --cached --name-only | grep -E '^[^/]+\.png$'`),
   belt-and-suspenders with the `.gitignore` screenshot rule so a forced `git add -f` still can't land
   a review/debug screenshot at the repo root. Nested `*.png` (real assets) pass; (d) **decision
   lifecycle validator** (ersatztv#521, supersedes the ersatztv#303 H9 append-only mechanic) — runs
   `.claude/hooks/decisions-guard.sh` (no args; a fail-open shim around
   `scripts/decisions_validate.py`), the structural checks over the working tree (metadata
   well-formedness, one active record per key, reciprocal links). It has no base/head here, so the
   body-diff/no-vanish checks it also knows about are skipped locally and only run in the CI
   `decisions lifecycle` job, which has a PR base to diff against.
2. **`pre-push`** — CI-parity gate: `cd web && npm run check:api && npm run lint && npm run
   typecheck && npm run build`. `check:api` guards generated-OpenAPI drift
   (`ErsatzTV/wwwroot/openapi/v1.json` → `web/src/api/generated/v1.d.ts`); the full
   lint/typecheck/build catch a staged change that breaks an *unstaged* file (lint-staged
   only sees staged files). Any failure blocks the push.
3. **`commit-msg`** — enforces the CLAUDE.md protocol: the message must carry a
   `Co-Authored-By:` trailer, else the commit is rejected (merge commits are exempt, detected
   via `git rev-parse --verify MERGE_HEAD`). The decision-lifecycle check lives in `pre-commit`
   (above), not here — the `Decisions-Edit:` trailer is read from the commit message, but only by the CI
   `decisions lifecycle` job's body-diff step (`range` mode over the PR's merge-base diff), which is
   the only place a base/head range exists to diff against.

- **Worktree/subdir gotcha**: git exports `GIT_DIR` (and friends) while running hooks. In a
  worktree or any subdir, an explicit `GIT_DIR` makes nested `git` commands mislocate the
  working tree — `pre-push`'s `check:api` (`git diff --exit-code`, run from `web/`) then
  silently reports "no diff" and lets drift through. `pre-push` therefore `unset`s
  `GIT_DIR GIT_WORK_TREE GIT_INDEX_FILE` first. (`pre-commit`'s `.cs` collection uses
  `git diff --cached`, index-vs-HEAD, which needs only `GIT_DIR` and is unaffected.)
- **Practical effect**: a fresh `web/` `npm install` (after cloning or pulling this change)
  installs all four hooks automatically — no separate setup step. Commits that touch only
  non-`web/`, non-`.cs` files skip linting/formatting (lint-staged no-ops with nothing to
  run, the `.cs` step is skipped).

## Registry

Gitea Packages, HTTP-only at `192.168.1.95:3000`. the `ci-runner` VM's Docker daemon (192.168.1.127) has it as an
insecure-registry (server-management#172; runner relocated off bumblebee in #570). Images: `192.168.1.95:3000/timothy/ersatztv:<tag>`.

## Test / prod environments

Container/compose wiring lives in **server-management** (project boundary): test
`ersatztv-test` on 8410 (`:latest`), prod `ersatztv` on 8409 (`:prod`). See
server-management#481 for the full spec (registry pull on the docker host, volumes, Jellyfin
isolation for test, Watchtower/manual promotion).

## Retired upstream workflows

The upstream `.github/workflows/` (`ci.yml`, `docker.yml`, `artifacts.yml`,
`release.yml`, `pr.yml`, `issue-stale.yml`) were removed — they targeted
GHCR/DockerHub + Azure/Apple signing and called reusable workflows at dead
`ersatztv/ersatztv@main` paths, and ran as noise (incl. a daily stale-issue cron) on
the Gitea runner. Upstream is archived, so there are no future merges to preserve them
for. The dead `.github/dependabot.yml` and `FUNDING.yml` (upstream-pointed) were also
removed.

## Known follow-ups

- Pin third-party actions to commit SHAs (currently floating major tags cloned from
  github.com at runtime) — low priority for a homelab; tracked informally.