docs: close S-E round-4 fidelity watch-list W-5 (park-record contract) + W-10 (P2b withdrawal in the umbrella kickoff)

[artyhoo]

Summary

Closes the two loose ends left by the round-4 fidelity audit of PR #1237 (merged as 1cb8fc8641). Two atomic commits, both on planning-session-owned surfaces.

W-5 — KICKOFF-AMBIGUOUS, decided and encoded. A stage kickoff pairs a ## §2 Permitted files allowlist with the §3a/§4c park-don't-guess contract, but no §2 revision enumerated an artefact where a fired PARK — or the propagation of its correction — could be recorded, so the S-E session edited both its own kickoff and the design spec, outside the allowlist. Verified across all 6 unique kickoff revisions: none lists .claude/orchestrator-prompts/* or docs/superpowers/specs/*.

The decision is REUSE of channels that already ship, not new machinery:

What	Where it is recorded	Already exists at
The fired PARK itself	park payload — manualReviewRequired / blocked_external + blockedReason stating «Option A → consequence X / Option B → consequence Y»	packages/core/../cli/questions.ts:85-93 (detection)
The git-durable record	the stage PR body's ## Parked questions section	.github/pull_request_template.md:53; acceptance-package section per spec D4 (docs/superpowers/specs/2026-07-23-acceptance-contour-design.md:67); routed by .claude/skills/dispatcher/SKILL.md:280
The PARK's correction propagated into a spec / ADR / kickoff	a separate owner commit on that artefact, never the stage PR	spec → precedent PR #1252; kickoff → .claude/rules/kickoff-staging-placement.md:28

So ## §2 Permitted files needs no new entry — it needs to say that recording a PARK is not a file write. Encoded in /pipeline §5 (.claude/skills/pipeline/SKILL.md:315), which already owns the kickoff-side park precondition, together with the canonical §2 clause for kickoff authors to copy and a new anti-pattern #park-record-outside-the-allowlist with its falsifier.

Honest note on the encoding target. There is no stage-kickoff template file in this repo — Permitted files appears in zero files outside .claude/orchestrator-prompts/ (grep -rln "Permitted files" | grep -v orchestrator-prompts → empty), and .claude/skills/pipeline/templates/ holds only meta-kickoff.template.md + state.md.template. The §2/§3a shape propagates by copy between kickoffs. Minting a template now would fork a second source for ~20 already-authored kickoffs (#sync-by-copy-paste), so the skill that already states the kickoff contract is the encoding target.

W-10 — umbrella bookkeeping. .claude/orchestrator-prompts/arch-v2-context-pipeline/kickoff.md:260 still listed the withdrawn P2b deliverable ("local-shadow pre-push check"). Annotated with the same strikethrough + dated-WITHDRAWN + do-not-rebuild shape the sibling spec-owner commit #1252 used, and carrying forward the unclaimed inverse hazard (a local file adding excludes to hide always-on rules). Per spec §3 row P4 this is a planning-session-owned surface closed by an umbrella bookkeeping commit, not by S-E.

Provenance

n/a — not a stage PR. No kickoff governs this change and nothing was factory-dispatched; it is a planning-session change on two planning-session-owned surfaces. Base origin/staging 3df8fb07d1; inputs are the PR #1237 round-4 fidelity watch-list items W-5 and W-10.

Review findings

Own cold-QA pass over the branch diff (T19), no sub-agent dispatched (none requested). Verification run on this host:

• npx vitest run packages/core/principles → 37 files, 352 passed | 1 skipped.
• make self-audit → green; link-check 96 OK / 0 errors; always-on 48671 B within ceiling 54000 B.
• bash scripts/run-local-ci-sweep.sh → 3 gates passed (md-ci-only, format-check, byte-identical).
• In-flight probe before authoring: gh pr list --state open → no PR touches either file; aif queue → no in-flight task on this umbrella.

Fidelity verdict

FIDELITY: skipped — not a stage PR: no kickoff governs this change, so no Basis artefact exists for agents/fidelity-auditor.md to audit the diff against

Parked questions

none — no fork fired while authoring. (Dogfood note: this PR is itself the conformant shape the W-5 decision prescribes — the umbrella-kickoff correction rides a separate owner commit rather than a stage PR.)

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

§1.7 Forward-check applied

• .claude/rules/build-first-reuse-default.md:1 — verdict REUSE, not BUILD: both record channels already ship (.github/pull_request_template.md:53, .claude/skills/dispatcher/SKILL.md:280). No new artefact, no dependency, no module → not a capability commit per CLAUDE.md «What is a capability commit?»; both commits carry the Prior-art: skipped — … hatch with ≥20-char rationale.
• .claude/rules/rule-enforcement-channel-selection.md:31 — detectability axis: «did the author state the park-record clause in §2» is judgment at authoring time, so the delivery is prose in the skill that authors kickoffs, not a gate (#gate-where-judgment-needed). The mechanically-checkable half is captured as the anti-pattern's falsifier for review time.
• .claude/rules/attention-is-not-a-mechanism.md:26 — honest scope: this closes the ambiguity (§2 said nothing), not the detection (nothing blocks a stage session that edits outside its allowlist). Stated as a falsifier rather than claimed as a gate.
• .claude/rules/no-paid-llm-in-ci.md — zero CI mechanism added, zero API-billed calls.
• .claude/rules/doc-authority-hierarchy.md:21 — .claude/skills/pipeline/SKILL.md:21 already declares «§0 invocation through §11 failures»; the new §5 prose falls inside it, so no header amendment. Kickoffs stay tracked-but-header-exempt per that rule's §2.
• .claude/rules/language-discipline.md:9 — both edits are repo artefacts, authored in English.
• .claude/rules/kickoff-staging-placement.md:28 — the umbrella-kickoff edit rides its own staging PR before any further dispatch, which is the rule's binding sequence.

§1.7 Backward-check applied

Class of this change = statements of the stage-kickoff §2/park contract, and statements of the withdrawn P2b deliverable. Enumerated by grep -rn "Parked questions" and grep -rn "P2b\|local-shadow" over the tracked tree (worktrees excluded):

• .claude/skills/dispatcher/SKILL.md:280 — SWEPT-CLEAN. Already names the ## Parked questions PR channel; the new clause points at it rather than re-describing it, so no #two-prompts-drift.
• .github/pull_request_template.md:53 — SWEPT-CLEAN. The section stub already exists; no edit needed, which is what makes this REUSE rather than BUILD.
• .claude/skills/arch/SKILL.md:101 — SWEPT-CLEAN. The factory-bound row already routes umbrellas to /pipeline; adding a duplicate clause there would be #sync-by-copy-paste.
• .claude/rules/kickoff-staging-placement.md:32 — SWEPT-CLEAN. Same paths: scope, orthogonal concern (merge timing, not §2 content); cited as the kickoff-correction route, not amended.
• docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — SWEPT-CLEAN, no edit. All four P2b statement sites were struck through by the sibling owner commit spec(token-economy): P2b withdrawal recorded at every statement site (owner commit per S-E fidelity r4 W-9) #1252; re-touching them here would duplicate that commit.
• .claude/orchestrator-prompts/arch-v2-context-pipeline-s-e/kickoff.md:55,60,122,124 — GAP-FOUND, deliberately NOT actioned. It still carries P2b as a live deliverable and the overturned replace-per-key overlay model. S-E has merged (PR arch-v2 S-E: L1 budget gate + config-assertion asserts + InstructionsLoaded verification (fidelity GO — round 4) #1237), and .claude/orchestrator-prompts/README.md declares prompts append-only audit-trail artefacts («do not edit after the umbrella PR is open»). Retro-editing a closed dispatch input would rewrite the record the fidelity audit was run against. Surfaced, not silently swept.
• Un-dispatched sibling stage kickoffs (S-D′, S-F, S-I) — forward-going, not retrofitted. The clause applies to the next kickoff authored; retrofitting live dispatch inputs is a separate, operator-invited change (CLAUDE.md PR strategy).

Self-application (T15). The decision is dogfooded by this PR: the umbrella-kickoff correction is a separate owner commit rather than a stage-PR rider, and ## Parked questions above is filled for real. The W-5 clause was derived by reading the shipped channels (questions.ts, the PR template, dispatcher/SKILL.md, spec D4) rather than from recall — source-before-shape.md:1 reuse surface.