feat(acceptance): fail-closed fidelity gate — cold WHAT-conformance auditor + PR-body CI gate (acceptance contour Phase A)

[artyhoo]

Summary

Phase A of the acceptance contour: the factory can now be held to what the kickoff asked for, mechanically. A cold agent protocol (agents/fidelity-auditor.md) produces a WHAT-conformance verdict at every stage-PR boundary, and a deterministic CI gate makes that verdict's presence non-optional on base=staging PRs. Design altitude only — code quality stays with the existing review path. Phase B (skill/CLAUDE.md choreography wiring) ships separately.

Design: docs/superpowers/specs/2026-07-23-acceptance-contour-design.md (ACCEPTED r3 after two rounds of cold two-altitude review). Plan: docs/superpowers/plans/2026-07-23-acceptance-contour.md §Phase A.

Changes

• agents/fidelity-auditor.md — cold WHAT-conformance auditor: inputs are kickoff/spec + 3-dot diff only (never chat or implementation logs); reports missing/extra/diverged drift with file:line; emits the machine-consumed grammar FIDELITY: GO|REVISE|STOP + Basis/Round/Audited-SHA/Evidence; KICKOFF-AMBIGUOUS routes a broken kickoff to re-design instead of burning rework rounds.
• packages/core/hooks/checks/pr-body-fidelity.ts (+ .test.ts, -bin.ts) — deterministic PR-body validator: FIDELITY: GO requires Basis + Round + Audited-SHA equal to the PR head (staleness guard: a commit pushed after the audit reddens the check) + ≥1 file:line evidence; FIDELITY: skipped — <rationale ≥20 chars> is the non-stage-PR path, and it is REJECTED on a stage PR (one whose ## Provenance declares a substrate) — without that the gate was opt-in for exactly the population it covers. Any heading closes the section, so evidence cannot be borrowed from a neighbouring ### §1.7 block; fenced code blocks do not truncate it. 39 tests, key guards mutation-verified.
• .github/workflows/pr-body-fidelity.yml — unfiltered pull_request events (the pr-body-prior-art.yml feat(prior-art): PR-body Prior-art gate — trailers survive squash (incident #1094→#1097) #1098 model; a paths: filter would un-fail-close tier-0/docs PRs), job scoped if: base.ref == 'staging'. Promote PRs (base=main) never run it; the primary exemption is registering the check as required only in the staging protection set.
• .github/pull_request_template.md — acceptance-package sections (Provenance / Review findings / Fidelity verdict / Parked questions), placed before the §1.7 block so the §1.7 awk substance-capture boundary is unchanged.
• docs/meta-factory/prior-art-evaluations.md — SSOT entry 228 (BUILD).
• tests/install-sh/baselines/* — regenerated: the new shipped agents/*.md file shifts install fingerprints (SNAPSHOT_MODE=capture, then 13/13 compare green).

Prior-art consult

Prior-art: prior-art-evaluations.md#228 (PR-body fidelity acceptance gate, verdict BUILD — upstream PR-body validators check form only and cannot host the semantic half, which is a session-bound cold agent under no-paid-llm-in-ci; the deterministic arm reuses the in-repo #1098 pr-body gate pattern).

• Capability commits in this PR carry a Prior-art: trailer. Reclassified during implementation: the plan called this a non-capability commit at 57 LOC and recorded a forward-guard — «if the file grows ≥80 LOC, the trailer requirement becomes real». It did: closing the code-review MAJORs grew pr-body-fidelity.ts to 165 lines and pr-body-fidelity.test.ts to 219, both over the 80-LOC threshold (packages/core/hooks/checks/prior-art.ts:134). The CI gate caught the stale classification; the guard fired exactly as written.
• New capability area: SSOT entry 228 added in the same PR, with Verdict / Rationale / Trigger to revisit.
• Existing entries matched: none superseded.
• Research: upstream PR-body validators (danger-js class) check form only and cannot host the semantic half — a session-bound cold agent under no-paid-llm-in-ci.md; the deterministic arm reuses the in-repo feat(prior-art): PR-body Prior-art gate — trailers survive squash (incident #1094→#1097) #1098 pattern. Recorded in entry 228.

Test plan

• npm test --workspace=@rules-as-tests/core --run -- pr-body-fidelity → 39/39 (RED first: module absent, then GREEN). Mutation-checked: removing the non-GO branch fails 2 tests, reverting the section boundary fails 2, disabling the stage detector fails 1, first-match-only Provenance fails 1.
• npm test --workspace=@rules-as-tests/core --run -- 09-doc-authority → 37/37 (new agent auto-enumerated, header accepted).
• Bin smoke, three directions: valid skipped → OK rc=0; empty body → FAIL rc=1 with the missing-section error; BASE_REF=main → out-of-scope pass rc=0.
• actionlint .github/workflows/pr-body-fidelity.yml clean; zizmor no findings.
• npm run --prefix packages/core test:principles → 335/335; make self-audit green.
• bash scripts/run-local-ci-sweep.sh → all gates PASS (actionlint, md-ci-only, meta-all-wired, format-check, typecheck, byte-identical, agnosticism).
• CI green on this PR.

⚠️ Read first — one operator action is required, and CI cannot enforce it. Add fidelity-verdict-in-pr-body to Settings → Branches → staging → Require status checks (the context is registrable once this PR's workflow has run). Register it on staging only — never on main, or the promote flow breaks (CLAUDE.md Harness gates).

Why this is a manual step rather than a gate — measured, not assumed: a cold code review flagged that resting the contour's fail-closed property on a manual click is #hope-as-gate, so this PR extended the existing workflow-integrity.yml staging assertion with a REQUIRED_CONTEXTS list. Running it live proved that assertion cannot work: reading branches/*/protection needs admin rights and GitHub exposes no workflow permission scope that grants them (administration is not a valid scope — actionlint rejects it). Observed on this very PR: protection exists with contexts ["ci-success"], and the job printed «No branch protection or active rulesets» and exited 0. So the pre-existing R11 assertion has been silently warn-passing all along, and the new context inherits that. The assertion is kept (it is correct wherever it can read state) but is documented as best-effort — its green does not prove registration. Escalation paths, both operator decisions and neither taken here: a PAT secret with admin scope, or a local pre-push probe using your own gh token. Recorded in the spec's Known limitations.

Provenance

• Basis: docs/superpowers/plans/2026-07-23-acceptance-contour.md §Phase A (spec D2/D3/D4)
• Substrate: in-session (top-tier design + implementation; discipline-bearing surfaces are deliberately not factory-dispatched — spec D9)
• Base: origin/staging; 18 commits, 21 files
• Fidelity Round: 13 (R1 GO → R2 REVISE → R4-R6 GO → R7-R9 REVISE → R10-R13 GO; see Review findings)

Review findings

• Design spec: cold two-altitude review, 2 rounds. Round 1 — both seats REVISE (0 BLOCKER, 5 + 2 MAJOR); round 2 — bottom-up GO, top-down REVISE with one MAJOR closed by code evidence (aif plan artifact readable via GET /tasks/:id → plan: packages/runtime-bridge/src/cli/aifHttp.ts:15,51, consumer questions.ts:53,89). All findings applied (spec r2/r3).
• Plan: cold two-altitude review. Bottom-up GO (all old_string anchors verified verbatim; 8 test cases traced against the regexes). Top-down REVISE → all 5 findings applied → confirmation round GO.
• Independent cold full review of the branch (operator-requested): REVISE — 1 BLOCKER, 6 MAJOR, 7 MINOR, all closed. The blocker was this PR failing its own gate on a stale Audited-SHA. The MAJORs that mattered: FIDELITY: skipped was unconditional, so the gate was opt-in for exactly the population it exists to cover; the section boundary let Evidence: be satisfied by a neighbouring ### §1.7 block; D6's first amendment made the rework cap unfalsifiable; the plan reproduced already-fixed defects on re-execution; §7 recorded 3 audit rounds when 10 had run. A follow-up audit then found a decoy ## Provenance heading bypassing the new stage detector — closed, mutation-verified.
• Code review of this diff (T19 own cold-QA): REVISE — 4 MAJOR, 11 MINOR, 0 BLOCKER, all applied in 89c33e321. The four that mattered: (1) the gate's fail-closed property rested on a manual branch-protection edit tracked only as a prose operator note — #hope-as-gate on the load-bearing precondition, now an executable assertion in workflow-integrity.yml (REQUIRED_CONTEXTS); (2) the bin failed open — an unset/renamed BASE_REF or HEAD_SHA exited 0, turning the gate into a permanently-green no-op; (3) multi-round semantics were undefined, giving both a false-pass (an appended skipped neutralised a recorded REVISE) and a false-fail (an appended round-2 GO was shadowed by the round-1 REVISE); (4) the REVISE test was mutation-survivable — deleting the entire non-GO branch left it green. Verified after the fix: deleting that branch now fails 2 tests.
• Phase B entry condition (from the round-7 audit): the routing exception in Phase B (D1 — Tier-2 /arch kickoffs run the executor-tier pipeline) is the first thing that genuinely depends on the registration being real. It must not ship until the registration is either mechanically verifiable (PAT secret / local pre-push probe) or converted into an explicit operator attestation. Nothing in Phase A rests on it.
• Dogfood fidelity audit, 13 rounds: R1 GO with one MINOR (SSOT entry 228 sat outside the §4 table → fixed f55591a83); R2 REVISE on a real MAJOR — the branch-protection assertion checked the job id while GitHub reports the job's name:, so it could never be satisfied, and registering the id would have deadlocked every staging PR on a context that never reports (fixed 229983652 by pinning name == id, the ci-success precedent at audit-self.yml:1354); R4 GO, drift lists clean.
  The auditor re-anchored its own verdict when the head advanced mid-audit, noting that a block naming the older SHA would fail the staleness guard this PR ships — the guard proving itself on its own author.

Fidelity verdict

FIDELITY: GO
Basis: docs/superpowers/plans/2026-07-23-acceptance-contour.md#phase-a
Round: 13
Audited-SHA: 38b2bd4
Evidence: packages/core/hooks/checks/pr-body-fidelity.ts:22-29; docs/superpowers/specs/2026-07-23-acceptance-contour-design.md:52
Findings: none — drift lists missing/extra/diverged all empty

Parked questions

none

§1.7 Self-discipline check (REQUIRED if PR touches discipline-bearing files)

§1.7 Forward-check applied

This PR introduces a discipline-bearing agent + CI gate, so it is checked against every active discipline it touches. attention-is-not-a-mechanism.md §1: the detection layer is a NAMED cold-agent protocol with structured output (agents/fidelity-auditor.md:47-57), never bare attention; the CI job is the fail-closed transport, not the judge (.github/workflows/pr-body-fidelity.yml:36-40). no-paid-llm-in-ci.md §1: the gate is pure string validation — no API-billed call anywhere in the workflow (packages/core/hooks/checks/pr-body-fidelity.ts:11-17 are the only decision points); the semantic half runs session-bound under the operator's own subscription. ci-tool-pinning.md §1: both actions SHA-pinned and npm ci --prefix used, not npm install (.github/workflows/pr-body-fidelity.yml:26,29,33). doc-authority-hierarchy.md §2-§3: the new agent carries Class + Authoritative-for + NOT-authoritative-for (agents/fidelity-auditor.md:3-15), verified by principle 09's dynamic enumerateFlatRequiredDocs (37/37). build-first-reuse-default.md §3: BUILD verdict recorded with rationale in SSOT 228 (docs/meta-factory/prior-art-evaluations.md:301) — reuse of the in-repo #1098 gate pattern, BUILD only for the unserved semantic-half seam. language-discipline.md §1: all shipped artefacts English.

§1.7 Backward-check applied

Class of this change = a PR-body section whose presence is enforced by a deterministic CI job on pull_request events. Complete enumeration of that class in the repo (grep -l "github.event.pull_request.body" .github/workflows/*.yml + the §1.7 arm): (1) .github/workflows/pr-body-prior-art.yml:20-47 — SWEPT-CLEAN: unfiltered events, no base filter, distinct section (Prior-art: line), no interaction with the new job; this PR copies its event set deliberately and does not modify it. (2) .github/workflows/discipline-self-check.yml:14-24,56-66 — SWEPT-CLEAN with an action taken: it is paths:-filtered (so it is NOT the model for a fail-closed every-PR gate — that is why the new job follows #1098 instead), and its backward-check awk captures from ### §1.7 Backward-check applied until the next /^###/; the four new template H2 sections are therefore placed BEFORE the §1.7 block (.github/pull_request_template.md:25,29,33,42 all precede :50,55,58) so the capture boundary is provably unchanged. (3) .github/workflows/audit-self.yml §7 real-commit trailer backstop — SWEPT-CLEAN: it is scoped to base_ref==main && head_ref==staging (the promote flow); the new job never runs on base=main, and its required-check registration is staging-only, so the fragile promote path (CLAUDE.md Harness gates) is untouched. (4) .github/pull_request_template.md — extended, not superseded: existing sections keep their order and content. No sibling gate is superseded, duplicated, or left inconsistent by this change.