Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1092 (emitting agent: clean-result-critic, 9a-bis round 1).

Goal

Capture rendered figure text (title/suptitle/axes titles/legend labels/series names) in savefig_paper's .meta.json sidecar and extend verify_task_body.py's figure-text checks to scan it.

Workflow gap

• Bug observed: Two figure defects in #1092 (bare cell slugs as read4c panel titles; a "both input arms" / "one bar per re-fit item" beat contradicting the plotted series/bar structure) passed the mechanical figure-text checks because the sidecar captures only per-point data (commit/figsize/points/n_series/total_points) — no rendered text fields — so the opaque-config-code and panel/series-drift checks were blind to them.
• Why it is a workflow gap: the sidecar is the verifier's only window into rendered figure text; without title/legend capture, the existing opaque-code and drift checks structurally cannot see the reader-facing labels they are specified to gate.
• Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

• paper_plots.savefig_paper: meta["text"] = {"suptitle": ..., "axes_titles": [...],

•                                          "legend_labels": [...], "series": [...]}
  

• verify_task_body.py check "figure text opaque config codes": also scan meta["text"] fields

• verify_task_body.py check "panel/series drift": parse beat phrases r"one bar per (\w+)" /

• r"both input arms" and compare against sidecar n_series/total_points/series labels

Scope / surfaces

• Primary targets: scripts/verify_task_body.py, .claude/skills/paper-plots/SKILL.md
• The sidecar writer lives in src/explore_persona_space/analysis/paper_plots.py — src touch for the planner to judge (analysis helper consumed by the workflow verifier; candidate emitter flagged it explicitly).

Constraints / invariants

• scripts/workflow_lint.py default run passes; ruff clean on touched files; existing sidecars (no text field) must not FAIL retroactively — the new checks fire only when meta["text"] is present (forward-only).

Provenance

• workflow_fix_target: scripts/verify_task_body.py, .claude/skills/paper-plots/SKILL.md
• fingerprint: pending-wrapper-computed

Verbatim candidate block preserved in the epm:clean-result-critique v1 marker on task #1092.