paper_plots sidecar: keep exact-(0,0) singleton collections and propagate per-panel series labels

Goal

Fix two silent under-recording defects in the figure sidecar capture used by savefig_paper, both of which make a CORRECT figure look wrong to every downstream reviewer and verifier. Each has now produced a false finding in a #2329 report-gate round, three rounds running.

The two defects

(a) _extract_scatters drops genuine exact-(0,0) single-point collections

A scatter collection holding exactly one point at the origin is discarded by the extractor, so the sidecar under-counts the figure's real points.

Observed: q35_ladder_decay_transfer — the sidecar captured 14 points while the PNG renders 16. The two missing points are the genuine all-zero (0,0) install-prefix-end cells: real data at the origin, not padding. A reviewer comparing sidecar to caption sees a 14-vs-16 discrepancy and has to open the PNG to discover the figure was right all along.

Suspected mechanism: an origin-valued singleton is indistinguishable from an empty/placeholder collection under a falsy-or-emptiness test. The fix must distinguish "collection with one point whose coordinates happen to be (0, 0)" from "collection with no points".

(b) Multi-panel sidecars record series: None for every panel after the first

On a multi-axes figure the sidecar keeps series labels only for the first panel; subsequent panels get series: None.

Observed: q35_ladder_decay_decay_raw renders BOTH model panels (q25 and q35), but its sidecar's q35-panel series names are None. This directly caused a wrong review finding: the #2329 methodology-critic read the sidecar and ruled the figure "q25 only", a claim the PNG refuted. The report text was correct; the sidecar was not.

Why this is worth fixing rather than working around

The sidecar exists so a reviewer can check a figure WITHOUT loading the PNG. Both defects make it lie in the conservative direction — under-reporting real content — so they generate FALSE FAILs, which cost a full revision round each time and train reviewers to distrust the sidecar. Three consecutive #2329 gate rounds paid this cost (the 14-vs-16 count, the "q25 only" misread, and the follow-on re-verification of both).

Acceptance criteria

1. A figure with a scatter collection holding exactly one point at (0, 0) has that point present in its sidecar; a genuinely EMPTY collection is still omitted. Both directions covered by tests.
2. A multi-panel figure records per-panel series labels for EVERY panel, not only the first — reproduce with a 2-panel figure whose panels carry distinct labels and assert both appear.
3. Regression fixture reproducing the #2329 shapes: a transfer-like scatter whose point set includes exact-(0,0) singletons (assert the sidecar count equals the rendered count), and a decay_raw-like 2-panel figure (assert no panel's series is None when labels exist).
4. Existing sidecars stay readable — no schema change that breaks a consumer (scripts/verify_report.py, report-verifier, the plotter). If the schema must change, state the migration and keep the old shape parseable.
5. Tests failing before / passing after; no new red in the no-flags workflow_lint.py run or the mapped-test selection.

Notes for the implementer

• Do NOT retroactively regenerate #2329's committed sidecars as part of this task — its figures are pinned by SHA in a report body under review, and a re-render would force a re-pin of every image URL. This task fixes the capture path forward; #2329 recorded both defects in its verification marker.
• The consumer side already has a correct habit worth preserving: when sidecar and PNG disagree, the PNG is authoritative. That is a workaround, not a fix.

Provenance

Surfaced as a prose mechanizable-class recommendation by report-verifier during #2329 round q35_ladder_decay report verification (round 1 FAIL, 2026-08-21), after the same extractor had produced a false finding in each of the three preceding gate rounds. The orchestrator verified both shapes against the committed sidecars + PNGs before filing. Evidence: #2329 events.jsonl epm:report-verified round 1; figures/issue_2329/q35_ladder_decay/ sidecars for q35_ladder_decay_transfer (14 vs 16) and q35_ladder_decay_decay_raw (q35-panel series: None) on branch issue-2329-q35-ladder-decay.

• target_file: src/explore_persona_space/analysis/paper_plots.py
• fingerprint: sidecar-origin-singleton-drop-and-multipanel-series-none
• confidence: high — both shapes observed on committed artifacts, each having produced a false review finding