verify_report.py: assert per-stage cap-hit disclosure + declared-cap parity (catches stale cap_hit_report residue)

Goal

Mechanize the CLAUDE.md max_new_tokens rule's per-stage reporting duty ("Every generation stage REPORTS its realized cap-hit fraction") as a scripts/verify_report.py check, and catch stale/contradictory cap_hit_report_*.json residue in the same pass.

The two failure shapes (both realized in #2329 q35_ladder_decay)

A round produced TWO generation stages, each with its own cap-hit report:

• cap_hit/cap_hit_report_grid.json — grid stage, 29/2160 = 1.343% at cap 4096. Disclosed.
• cap_hit/cap_hit_report_anchors.json — anchors stage, realized 5/420 = 1.19% at 4096/4097 (per the staged anchors_gate_w0.jsonl). NOT disclosed in the report's G5 bullet; it survived only as a line in a figure caption.

Shape 1 — silent per-stage omission. The report disclosed the stage the author was thinking about and omitted the sibling stage. Nothing mechanical objected. A reader of the "G5 disclosure" bullet would reasonably believe it covered the round.

Shape 2 — stale residue that CONTRADICTS the report. The committed cap_hit_report_anchors.json declares max_new_tokens: 2048, realized_row_caps: [2048], partial: true — residue from an earlier --cap-scope both run — while the round actually ran the anchors stage at 4096. So the committed artifact contradicts the (correct) report text, and whichever a later reader trusts, one of them is wrong. Detected only by a human critic opening both files.

Proposed check

For each cap_hit_report_*.json under the round's eval_results/<issue>/<round>/:

1. Assert the report body discloses THAT stage's realized cap-hit fraction (stage name + number), not merely some stage's.
2. Assert the report's declared max_new_tokens for the stage equals the report JSON's max_new_tokens / realized_row_caps, and that both match the realized truncation cap observed in the stage's shard rows where those are reachable.
3. FAIL loud on a partial: true / mismatched-cap report that no report text explains — either the residue is regenerated or the report names it as known-stale.

Keep it evidence-based rather than name-based: key off the report JSONs actually present, so a round with one stage is not forced to invent a second disclosure.

Acceptance criteria

1. Reproduce shape 1 with the #2329 fixture: a report body disclosing only the grid stage while cap_hit_report_anchors.json is present ⇒ FAIL naming the anchors stage.
2. Reproduce shape 2: cap_hit_report_anchors.json at max_new_tokens: 2048 / partial: true against a report declaring a uniform 4096 ⇒ FAIL naming the contradiction.
3. A round whose every present stage IS disclosed, with matching caps, PASSes — the check must not become unsatisfiable for compliant rounds.
4. A single-stage round is not required to disclose a stage that does not exist (no false-positive on absence).
5. Tests failing before / passing after; no new red in the no-flags workflow_lint.py run or the mapped-test selection.

Provenance

Surfaced as a prose mechanizable: yes recommendation by the methodology-critic during #2329 round q35_ladder_decay report review (round 1 FAIL, item 3), 2026-08-21. The orchestrator verified both halves against the committed artifacts before filing. Evidence: #2329 events.jsonl epm:methodology-check round 1; cap_hit/cap_hit_report_anchors.json and cap_hit/cap_hit_report_grid.json on branch issue-2329-q35-ladder-decay; marker v185 (the --cap-scope both run that produced the residue).

• target_file: scripts/verify_report.py
• fingerprint: per-stage-cap-hit-disclosure-and-declared-cap-parity
• confidence: high — both shapes observed on committed artifacts in one round