verify_report.py: assert committed-under paths resolve at the pin, and assert code-SHA rows match per-phase reproducibility cards

Goal

Add two pure asserts to scripts/verify_report.py. Both caught real defects in #2162's round-1 methodology-accuracy review, both are mechanical rather than judgement calls, and both are the kind of error a human reviewer catches only by manually running a git command per claim.

Check (a) — every "committed under <path>" claim resolves at the pinned tree

A report that says an artifact set is committed under some path should be falsifiable by git ls-tree -r <pin> -- <path> returning at least one blob.

What it caught on #2162. The draft asserted that judge scores and items were committed under eval_results/issue_2162/judge/. At the pinned tree that directory held gates (7), raw (70), audits (2), items 1 of 168, judge_summary.json and pools.json — and zero files under judge/scores/, while 336 score files existed on disk untracked. The corpus genuinely lives on the HF data repo under raw_completions/judge_raw per the wave-output convention, so the claim named a git home the artifacts do not have.

Why it matters beyond tidiness: a reader who trusts that sentence will later clone at the pin and find nothing, and will not know whether the data was lost or merely lives elsewhere. The convention that wave outputs stay out of git is correct — the report just has to say so.

Check (b) — every Code-SHA row matches the per-phase reproducibility card

A multi-phase run has one reproducibility card per phase, each carrying its own git_commit. A report's code-SHA table should agree with them phase by phase.

What it caught on #2162. The table pinned the stage-1 grid to the MARGIN leg's commit (ba3485b619…) when the grid/anchors phase's own card records b4ab6ed5f9…, six commits earlier. The reviewer then diffed the full constant block of the run script between the two commits and found them byte-identical, so no hyperparameter value was wrong — but the provenance pairing was, and only an explicit comparison surfaced it. A future run where the constants DID change between phases would turn the same mistake into wrong reported hyperparameters.

Sources to compare against: the per-phase upload_done.json reproducibility cards, plus gate reports that carry a repro block.

Suggested shape

• Both as asserts in scripts/verify_report.py, so they run in the existing report-verifier gate rather than as a separate tool.
• Check (a): parse the report for "committed under <path>"-shaped claims and their pin, then git ls-tree -r <pin> -- <path>. Fail loud naming the claim and the empty path. Be careful to accept a correctly-worded HF-home claim — the goal is catching a false GIT-home claim, not discouraging the convention.
• Check (b): collect git_commit from every reproducibility card under the issue's eval_results/<issue>/**, then assert each code-SHA row in the report matches the card for the phase it names. When a run legitimately spans multiple commits, the expected output is a per-phase split, not a single SHA — so the failure message should suggest that rather than implying one is wrong.

Scope notes

• These are ADDITIVE asserts; do not weaken any existing verify_report.py check.
• #2162 is the FIRST workflow: v2 task in the repo, so this is the first report the gate has ever run against — expect more mechanizable patterns to surface as later v2 reports land, and prefer a shape that is easy to extend.
• The reviewer that surfaced these also verified a large amount of the draft as clean, so the intent is not to distrust the methodology-writer — it is to move two specific classes of error from human re-derivation into a machine check.
• Confidence is high on the value of both checks (each has a concrete, confirmed instance) and moderate on the exact parsing shape, since report phrasing is free-form. The implementing session should look at #2162's actual sections at tasks/*/2162/artifacts/issue-2162-report-sections.md for realistic input before settling on a matcher, and should prefer a conservative matcher that under-fires over one that produces false FAILs on well-formed reports.