workflow-fix: verify_task_body.py — mechanize the clean-result-critic three-beat-ordering + arrow-range-drift checks

Goal

Close two mechanizable gaps in scripts/verify_task_body.py that the clean-result-critic caught by hand on #2356, each consuming review rounds the mechanical gate should have pre-empted.

Gap 1 — check 21 (check_v4_results_beat) misses a multi-figure result ending on a caption

The Lens-3 three-beat rubric (.claude/rules/clean-result-critic-lens-reference.md, SPEC one-narrative-unit clause) hard-FAILs a ### <result> whose declared multi-figure unit ends on an image/blockquote caption instead of an interpretation beat ("what-is-plotted above the pair, interpretation below the pair"). check_v4_results_beat PASSes that shape. On #2356 this slipped past BOTH the mechanical gate and Claude's round-3 lens, surfaced only by the Codex twin in round 3, then recurred in round 4 — two review rounds spent on a mechanizable defect.

Fix: extend check 21 — for a ### <result> with >1 inline figure, WARN/FAIL when the section's LAST non-empty block (before the next H2/H3 or the footer rule) is an image (![…](…)) or a blockquote caption (> **Figure.** …) rather than non-caption interpretation prose.

Gap 2 — the plotted-value-drift check does not cover arrow-form numeric ranges

On #2356 Result 3, the parenthetical "(0.13→0.29 harmful, 0.56→0.61 over-refusal)" quotes only real artifact values, but the arrow semantics differ across arms: armA reads min→max across cells (0.132→0.291) while armB's 0.61 endpoint is the pooled-3b value (0.6097), not the max fold (0.634). Direction and the gate claim are unaffected (no overclaim — the true top cell is higher than quoted), so round-5 Claude flagged it non-blocking. But it is mechanizable.

Fix: extend the plotted-value-drift check to parse X→Y pairs adjacent to a figure reference and verify both endpoints appear among (or bound) the figure sidecar's plotted values; WARN when the max plotted value exceeds Y by more than one rounding unit.

Provenance

workflow_fix_target: scripts/verify_task_body.py

Surfaced during /issue 2356 clean-result-critic ensemble (rounds 3–5). Both gaps are in scripts/verify_task_body.py; fixes should add fixtures reproducing each shape and update any affected lens-reference wording.