workflow-fix: verify_task_body per-result multi-figure (non-pair) check

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a mechanizable clean-result-critic finding raised on task #1769 (emitting agent: clean-result-critic, fu1 re-gate round, 2026-07-30).

Goal

Add a verify_task_body.py check flagging a single ### <result> H3 under ## Results that carries MORE THAN ONE inline ![...] figure when the figures are not a declared raw+processed (aggregate + per-unit) pair — mechanizing the lens-9 one-result-one-figure rule's multi-figure side.

Workflow gap

• Bug observed: task #1769's fu1 fold placed two distinct-analysis figures (fig_dose_ladder.png + fig_alpha3_lattice.png) inside one ### <result> section; verify_task_body.py PASSed (the only flag was an unrelated word-count WARN) and the violation was caught only by the LM clean-result-critic (fu1 re-gate REVISE, lens 9, mechanizable: yes).
• Why it is a workflow gap: the v4 spec's one-result-one-figure three-beat allows >1 figure per ### <result> only for a raw-alongside-processed pair, but no mechanical check counts inline figures per result section — the existing figure checks are per-figure (caption shape, panel drift, count claims), never per-section cardinality — so a two-analysis section passes silently and burns an LM-critic round.
• Confidence (emitter): high (the critic tagged it mechanizable: yes and named the fix).
• verified-at-filing: grep -n "one-result-one-figure\|one_figure" scripts/verify_task_body.py → 11 hits, ALL per-figure helper functions (_panel_drift_for_one_figure, _prose_numerics_for_one_figure, _beat_claims_for_one_figure, _count_claims_for_one_figure) — no per-### figure-count rule; git log --oneline --since='7 days ago' -- scripts/verify_task_body.py → 1 commit (218a6d0425, #1493 N/A-verdict for infra/batch/survey — unrelated) (2026-07-30)

Proposed change (candidate diff sketch — refine in planning)

+ def check_result_sections_single_figure(...):
+     for each `### <result>` under `## Results` (v4-sentinel bodies only):
+         n = count of inline `![alt](url)` lines in the section
+         if n > 1 and not a declared raw+processed pair (heuristic: one
+             figure's alt/caption names a per-unit/per-question companion of
+             the other, or the prose declares the pair):
+             WARN (or FAIL for v4) "### '<h3>' carries N inline figures that
+                   are not a raw+processed pair (one-result-one-figure)"
+ # WARN-vs-FAIL and the pair heuristic are planning decisions; the
+ # forward-only rule keeps v3/v2/legacy bodies exempt.

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Add paired rows in tests/test_verify_task_body.py; keep v3/v2/legacy bodies exempt (v4-sentinel-gated per the forward-only rule); update .claude/skills/clean-results/SPEC.md mechanical-checks list alongside.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• Never a new hard FAIL on grandfathered (non-v4) shapes; verify_task_body.py stays consistent with .claude/skills/clean-results/SPEC.md.
• 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

Surfaced finding (verbatim, clean-result-critic fu1 re-gate on #1769): "Lens 9 — ### The dose ladder places the CJK collapse between α=2 and 3… carries TWO figures (fig_dose_ladder.png + fig_alpha3_lattice.png) that are two distinct analyses, not a raw+processed pair — the only sanctioned >1-figure exception. [...] mechanizable: yes — extend verify check 21/38 to FAIL >1 non-paired inline figure per ### <result>."

• fingerprint: a7bcd9d11aea