workflow-fix: verify_plan check — fail-loud acceptance claims need committed tests

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up raised on task #913 (emitting agent: codex-critic statistics twin, S5; surfaced by the reconciler's "Observed but not raised").

Goal

Add a verify_plan.py check requiring plans whose acceptance criteria assert fail-loud / no-silent-swallow behavior to name a committed test backing the claim, not only run-book grep commands.

Workflow gap

• Bug observed: Task #913's plan certified its fail-loud acceptance criterion only via run-book grep gates; the Codex statistics critic showed a differently-worded re-swallow would ship green past all committed tests
• Why it is a workflow gap: verify_plan.py (and the Statistics critic lens) has no check that a fail-loud acceptance claim is backed by a committed test — run-book grep gates verify the invariant once at review time, not durably.
• Confidence (emitter): low

Proposed change (candidate diff sketch — refine in planning)

(none — synthesized from prose follow-up)

• verify_plan.py: new WARN-level check — when a plan's acceptance criteria / success criteria contain fail-loud vocabulary ("fail loud", "no silent", "not swallowed", "raises", "no except") coupled to a run-book-only verification (a grep gate with no named test), WARN asking for a committed test naming.
• Optionally mirror as a Statistics & Measurement lens bullet in .claude/rules/critic-lens-reference.md.

Scope / surfaces

• Primary target: scripts/verify_plan.py
• Grep the workflow surface for the pattern before editing (grep -rln 'fail.loud\|grep gate' .claude/ scripts/verify_plan.py) and update every hit; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; keep the check WARN-level (advisory), never a hard FAIL on grandfathered plans.
• 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_plan.py
• fingerprint: a326765ae05d

Verbatim surfaced prose (codex-critic statistics twin, #913 round 1, S5): "the acceptance items certified only by run-book checks are 'no except Exception remains' and the broader 'no warning-and-continue path remains.' For this fix, that should be a committed AST/source test, not only a grep command. This is a recurring workflow-surface verifier candidate: fail-loud acceptance claims should be backed by committed tests when the defect is silent swallowing."