workflow-fix: verify_plan WARN when a named regression anchor is neither run nor gate-selected

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up raised on task #1536 (emitting agent: Statistics & Measurement critic, adversarial-planner Phase 2 round 1).

Goal

Add a verify_plan.py WARN check that each plan-named regression-anchor test file either appears in the plan's exact test command or is returned by select_step9c_tests.py --map-files for the plan's declared touched files (or direct-imports a touched module).

Workflow gap

• Bug observed: plan-named regression-anchor tests can be neither auto-selected by Step 9c (the selector's import map is deliberately one-hop) nor present in the plan's exact test command, so the anchor claim ships unexecuted. Live instance: #1536 plan v2 §4 claimed tests/test_issue906_marker_mix_budget.py + tests/test_issue906_tiny_real_e2e.py "sit on the Step-9c mapped-scan path for sft.py edits" — verified false (select_step9c_tests.py --map-files src/explore_persona_space/train/sft.py returns empty; both files import train_lora only transitively via scripts/issue906_phase1_pilot). Caught only by the round-1 Statistics critic (Must-Fix 1).
• Why it is a workflow gap: verify_plan.py mechanically checks plan structure but has no check cross-referencing named anchor test files against the Step-9c selection surface; the shape is likely to recur on infra plans that lean on "the gate will run it".
• Confidence (emitter): medium
• verified-at-filing: grep -n 'select_step9c_tests' scripts/verify_plan.py → 0 hits in the named target (absence-of-guard claim; 0-hit in-target IS the evidence — the file's only 'anchor' hits are the unrelated success/kill-anchor internals at L224/247/814/981/1041); git log --oneline --since='7 days ago' -- scripts/verify_plan.py → checked at filing (2026-07-19; output recorded in filing-session transcript — no landed anchor-cross-check commit)

Proposed change (candidate diff sketch — refine in planning)

(none — synthesized from prose follow-up)

Sketch for the planner: a new WARN-severity check (e.g. c40_regression_anchor_executed): parse the plan for test-file paths named as regression anchors / in a "must stay green" clause; parse the plan's exact test command(s) for explicitly-run test files; for each named anchor not explicitly run, check whether select_step9c_tests.py --map-files <plan's declared touched files> selects it (or a direct-import grep links it to a touched module); WARN naming each anchor that is neither explicitly run nor auto-selected.

Scope / surfaces

• Primary target: scripts/verify_plan.py
• Secondary: tests/test_verify_plan.py (pin the new check), possibly scripts/select_step9c_tests.py (only if a library entry point is needed for the map query — prefer reusing the existing --map-files CLI).
• Grep the workflow surface for the pattern before editing (grep -rln 'select_step9c_tests' .claude/ CLAUDE.md scripts/) and update every hit the plan touches; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• WARN severity, never FAIL (an anchor claim can be legitimately satisfied by prose the parser cannot see; fail-open).
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; tests/test_verify_plan.py stays green.
• 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, .claude/rules/workflow-fix-on-bug.md § Recursion guard).

Provenance

• workflow_fix_target: scripts/verify_plan.py
• fingerprint: 7ee32b4d1a65

Verbatim surfaced prose (task #1536, Statistics critic round-1 report): "Follow-ups (orchestrator should consider): a verify_plan.py check that cross-references plan-named regression-anchor test files against select_step9c_tests.py --map-files output + a direct-import grep for the plan's declared touched files, WARNing when a named anchor is neither auto-selected nor present in the plan's exact test command (the Must-Fix-1 shape is likely to recur on infra plans that lean on 'the gate will run it')."