workflow-fix: verify_plan.py check for false numeric containment claims

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1315 (emitting agents: statistics-critic + critic/Alternatives lens, both flagged mechanizable: yes independently in the same round).

Goal

Add a verify_plan.py check that a plan's explicit numeric containment claims ("N inside/within the A–B spread/band/range") are arithmetically true — FAIL (or WARN) when N ∉ [A, B].

Workflow gap

• Bug observed: plan v4 for task #1315 asserted "Tier-2 0.724 — … inside the siblings' realized 0.737–0.820 spread", which is arithmetically false (0.724 < 0.737); the verify_plan.py mechanical pre-pass returned PASS n_fail=0 n_warn=0 and two independent critics caught the false numeric by hand (Statistics lens finding 1; Alternatives lens finding 2).
• Why it is a workflow gap: verify_plan.py mechanically checks many prose contracts but has NO check that explicit numeric containment prose is arithmetically consistent, so a false "inside the X–Y spread" claim rides into a registered plan and can only be caught by an LM reviewer.
• Confidence (emitter): medium
• verified-at-filing: grep -niE 'spread|containment|inside the' scripts/verify_plan.py → 1 hit (line 3430, an unrelated comment about paren placement inside quotes); this is an ABSENCE-of-guard claim, so the 0 relevant in-target hits ARE the evidence (2026-07-16)

Proposed change (candidate diff sketch — refine in planning)

+ def c36_numeric_containment_claims(text):
+     # match: "<num> ... (inside|within) ... <a>-<b> (spread|band|range)"
+     # tolerate en-dash/hyphen ranges, "the registered"/"the realized" fillers
+     for m in re.finditer(NUM_CONTAINMENT_RE, text):
+         n, lo, hi = float(m['n']), float(m['lo']), float(m['hi'])
+         if lo > hi: lo, hi = hi, lo
+         if not (lo <= n <= hi):
+             fail(f"numeric containment claim false: {n} not in [{lo}, {hi}]")

Design notes for the planner: keep the matcher conservative (only fire when a single number, a containment verb, and an explicit numeric range co-occur in one sentence) to avoid false positives on ranges that describe something other than the claimed value; a WARN tier may be safer than FAIL for v1. Add the canonical N/A escape phrase if a legitimate non-containment reading needs an exemption, and pin with tests in tests/test_verify_plan.py.

Scope / surfaces

• Primary target: scripts/verify_plan.py
• Secondary: tests/test_verify_plan.py (pin the new check + the false-claim fixture from #1315 plan v4), .claude/skills/adversarial-planner/SKILL.md (add the new check id + N/A escape phrase to the canonical list) if a new escape phrase is introduced.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• 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: ff49ea6ff46e

Surfaced prose (statistics-critic, task #1315 amendment round, 2026-07-16): "§4.1 (plan line 66) claims Tier-2 0.724 is 'inside the siblings' realized 0.737–0.820 spread' — it is not: 0.724 < 0.737. […] mechanizable: yes — assert min(sibling spread) ≤ claimed-inside value ≤ max whenever a plan asserts 'inside … spread' with explicit numerics (verify_plan.py c27-family check)." Independently (critic, Alternatives lens, same round): "mechanizable: yes — assert committed 0.724 ∉ [0.737, 0.820] against the plan's prose claim."