workflow-fix: WARN-tier plan-conditions coverage check (silent condition drop)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced by the clean-result-critic on task #1774 (round 1, Lens 13 finding: a silent scope drop that passed every mechanical check).

Goal

Add a WARN-tier plan-conditions coverage check to verify_task_body.py that parses the plan §5 conditions-table config-slug column and WARNs when a slug appears neither in the body nor in an explicit skip/descope record.

Workflow gap

• Bug observed: Task #1774's plan-committed pretrained-reads robustness condition (cell_pre_own, plan §5 conditions table + §4 substrate) was silently dropped — no artifact, no descope marker, aggregate .skipped empty, zero body mentions — and passed every mechanical check (verify_task_body.py 61 checks + audit_clean_results_body_discipline.py); only the LM critic's plan read caught it.
• Why it is a workflow gap: Lens 13's silent-drop check (plan §5 conditions vs body mentions) is currently LM-only; the planned-vs-actual mechanical surface (check 11b) keys on conditions the body NAMES, so a condition dropped from the body entirely is invisible to it. A slug-level WARN closes the gap cheaply.
• Confidence (emitter): medium (false-positive risk from paraphrased plain-English names — hence WARN tier + slug-level matching, per the emitting critic's own caveat)
• verified-at-filing: grep -cn 'conditions table\|§5\|plan_conditions' scripts/verify_task_body.py → 0 hits (no plan-§5-conditions coverage check exists in the target — absence claim, 0-hit in-target result IS the evidence); incident cross-check: grep -c 'pre_own' tasks/reviewing/1774/body.md → 0 while plan §5 names cell_pre_own (2026-07-29)

Proposed change (candidate diff sketch — refine in planning)

+ # Check N (WARN): plan §5 conditions coverage — parse the plan conditions
+ # table's config-slug column (plans/plan.md, when resolvable); for each slug,
+ # WARN unless it appears in the body OR in an explicit skip/descope phrase
+ # ("not run", "descoped", "named deviation") within +-1 sentence of the slug
+ # or its plain-English row name. WARN tier only — plain-English paraphrase
+ # makes a FAIL tier too false-positive-prone.

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Consider whether the check belongs beside check 11b (planned-vs-actual) or as a new numbered check; the plan path resolves via the task folder (plans/plan.md symlink). Grep the workflow surface for prior art (grep -rn 'planned-vs-actual\|check 11b' scripts/ .claude/rules/) and keep clean-result-critic Lens 13's prose in sync if the mechanical tier lands.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; WARN tier only (a FAIL tier is explicitly out of scope for this filing).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: scripts/verify_task_body.py
• fingerprint: 227f48f2d53f

Verbatim surfaced prose (clean-result-critic r1 on #1774): "Lens 13's silent-drop check (plan §5 conditions vs body mentions) is currently LM-only. A WARN-tier check in scripts/verify_task_body.py (or verify_plan.py-adjacent tooling) could parse the plan §5 conditions table's config-slug column and WARN when a slug (or its plain-English name) appears neither in the body nor in an aggregate skipped/deviation record — slug-level matching with WARN severity to bound the paraphrase false-positive risk. Evidence: this round's cell_pre_own drop passed every mechanical check and was only caught by the plan read. mechanizable: yes (with the stated false-positive caveat)."