workflow-fix: verify_task_body check for bare #K refs in Methodology/Results

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced on task #841 (emitting agent: clean-result-critic, round 2).

Goal

Add a verify_task_body.py check that FAILs bare #<digits> issue references inside the ## Methodology / ## Results section spans of a v4 body when not inside a markdown link target, not in a table row, and not in the footer/frontmatter.

Workflow gap

• Bug observed: bare #779 references (~8×) in #841's ## Methodology prose survived TWO clean-result-critique ensemble rounds: in round 1 BOTH reviewers PASSed Lens 2 — the Claude critic on a mistaken claim that the canonical v4-657 exemplar carries prior-issue refs in its Results (it does not; verified against the exemplar in round 2), the Codex critic flagged it but the finding was waived on the Claude critic's mistaken recalibration. Only a round-2 re-check against the exemplar caught it.
• Why it is a workflow gap: the Lens-2 standalone-sections rule (## Takeaways / ## Methodology / ## Results carry no bare #K; prior-issue links live only in the Goal context slot + footer) is spec text with NO mechanical enforcement in verify_task_body.py — an LM-judgment-only rule that two independent reviewers can (and did) mis-apply in the same direction. The critic explicitly tagged the finding mechanizable: yes.
• Confidence (emitter): high (live two-round miss on #841; the exemplar check that settles it is deterministic).

Proposed change (candidate diff sketch — refine in planning)

• New check in scripts/verify_task_body.py (v4 bodies only): scan the ## Methodology and ## Results section spans (and ## Takeaways) for #\d+ tokens, excluding:
  • matches inside markdown link syntax [...](...) (both label and target),
  • lines that are table rows (| ... | — the Training-table Source column is a sanctioned grounding convention),
  • the **Repro:** / **Context:** footer block and YAML frontmatter,
  • HTML comments.
• FAIL (or WARN for grandfathered bodies) with the offending line numbers; v3/v2/legacy bodies exempt (forward-only rule).
• Regression tests: the #841 round-2 shape (bare #779 in Methodology prose → FAIL; the same reference in a Source-column table row → PASS; [#779](url) in the Goal slot → PASS).
• Mirror one line into the SPEC.md mechanical-checks list so the lens and the verifier stay in sync.

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Also update: .claude/skills/clean-results/SPEC.md (mechanical-checks list), tests/test_verify_task_body.py.
• Grep before editing: grep -n "check_" scripts/verify_task_body.py for numbering conventions; verify the v4-657 exemplar and a sample of recent v4 bodies pass the new check (no false positives on sanctioned forms).

Constraints / invariants

• Workflow-surface only. Forward-only: never newly hard-FAIL a grandfathered v3/v2/legacy body.
• Existing verifier tests stay green; SPEC.md is the source of truth — update it alongside per the standing rule.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: scripts/verify_task_body.py
• fingerprint: 7a967742b4bb

Surfaced prose (verbatim, from the #841 round-2 clean-result-critic report): "mechanizable: yes — a verify_task_body.py check could FAIL on #\d+ inside the Methodology/Results section spans when not inside a ](...) link, not in a | … | table row, and not in the footer/frontmatter. I did NOT emit a formal workflow-fix-candidate block (running under AUTO_REVIEW_DISABLED=1), but flagging it for you: this is the exact gap that let round 1 miss it."