Align clean-result-critic Lens 2 result-heading rule with the v4 SPEC contract (rubric is strictly stronger than SPEC; 5 invented-rule FAILs across 3 tasks)

Goal

Reconcile the clean-result-critic Lens 2 result-heading bullet with the v4 SPEC contract, so the review rubric is not strictly stronger than the spec it enforces.

Problem

.claude/rules/clean-result-critic-lens-reference.md (Lens 2, the ## Results block, around lines 252-257) requires that each ### <result> heading "states the result WITH the number in the heading". That phrasing has no basis in the v4 contract:

• The v4 contract (.claude/skills/clean-results/SPEC.md around 422 / 573 / 586) requires only a claim-stating, standalone ### <result> H3 per result. It contains no heading-number clause.
• The only "with the number in the heading" text in SPEC is the v3 template placeholder (SPEC.md around 1229).
• The rubric bullet's own enumerated FAIL class is bare outline labels (### Headline result, ### Subset checks, ### Sample completions, ### Plan deviations, ### Methodology corrections) — a different shape from a claim-stating heading that happens to carry no number.

Impact (measured, not hypothetical)

The divergence has produced the identical Codex invented-rule FAIL five times across three tasks: #2223 round 3, #2333 twice, and #2479 rounds 1 and 4. On #2479 alone it fired twice, and the second firing came through a binding round-1 reconciler rejection of the same demand — the round-4 re-raise landed under a new concern id (result-headings-omit-outcome-numbers vs the round-1 numbered-result-headings) against byte-unchanged headings, and cost a full extra Codex dispatch, a second binding reconcile round, and a defer-concern record.

Two second-order costs worth naming: a rubric strictly stronger than SPEC makes every Claude-side PASS on that lens look like an under-application (it is not), and re-raising a rejected finding under a fresh concern id defeats ledger-based dedup.

Proposed fix (pick ONE deliberately — SPEC is the source of truth)

(a) Align the rubric down to SPEC v4 — the presumption, since SPEC is canonical:

Each ### <result> heading states the result as a claim (an outcome number in the heading is encouraged, not required); FAIL only on bare outline labels: ### Headline result / ### Subset checks / ### Sample completions / ### Plan deviations / ### Methodology corrections.

(b) Or amend SPEC's v4 contract deliberately to make the heading number binding, and keep the rubric as-is. If this is chosen, note that some results are irreducibly multi-number (on #2479 the retrieval-failure result carries four statistics and the mediators result carries three), so any binding form needs an explicit carve-out for them rather than forcing a misleading single number.

Whichever is chosen, the rubric and SPEC must end up saying the same thing.

Acceptance criteria

1. clean-result-critic-lens-reference.md Lens 2 and SPEC.md's v4 contract agree on whether a heading number is required.
2. The chosen wording states the FAIL class explicitly, so a reviewer cannot read a preference as a categorical requirement.
3. If (b) is chosen, the multi-number carve-out is written into both surfaces.
4. Consider a dedup note for the re-raise channel: a finding a binding reconciler has rejected on a given task should not re-enter that task's ledger under a new concern id against unchanged content. (Scope this only if it is a small edit to the same surfaces; otherwise name it as a follow-up rather than widening this task.)

Provenance

workflow_fix_target: .claude/rules/clean-result-critic-lens-reference.md

Raised as a workflow-fix-candidate v1 by the binding reconciler during #2479's Step 9a-bis clean-result gate, round 4 (epm:review-reconcile row v11, sentinel v4, 2026-08-24T17:29:51Z). Evidence: #2479 concerns.jsonl rows numbered-result-headings (raised r1, rejected + addressed) and result-headings-omit-outcome-numbers (raised r4, reconciler-deferred 17:28:44Z); epm:review-reconcile rows v1 (14:20:03Z, finding 9) and v11 (17:29:51Z); reconciler agent-memory DISCARD classes Z and AD.