[P1] reproduce: validate sample coverage and selected rubrics before reporting PASS/FAIL

Problem

The reproduction verdict compares three percentages but ignores the observed/published task counts. Meanwhile --rubric lenient and --rubric strict are supported options, yet missing unrequested metrics are replaced with zero and still compared. This can accept mismatched samples and reject a correct single-rubric reproduction.

Sources: verdict(), summary-to-verdict conversion.

Verified offline examples

from clawbench.eval.reproduce import verdict, PUBLISHED_V2_HERMES
verdict((50, 40, 20, 5), (50, 40, 20, 130), 2)[0]
# True: same percentages are accepted despite 5 versus 130 tasks.
verdict((54.6, 44.6, 0, 130), PUBLISHED_V2_HERMES['claude-opus-4-7'], 2)[0]
# False: a matching lenient-only result fails against the uncomputed strict column.

The second example models the exact summary.get('reward_pct_strict', 0) behavior for --rubric lenient.

Acceptance criteria

• Verify nonzero expected task coverage before declaring a full reproduction; compare task identities/revision when available, not only count.
• Compare only explicitly requested rubric columns; display omitted metrics as not evaluated rather than 0%.
• Distinguish incomplete/invalid evidence from a measured mismatch, including missing required metrics and judge-inconclusive counts.
• Test both single-rubric modes, full mode, sample-count mismatch, missing metrics, and empty input.

Suggested priority P1: scientific result validity. This is a verdict contract bug, not the historical corpus-count issue #244 or path portability #296. The numbers above exercise checked-in reference data, not a claim about today's live leaderboard.

@Perry2004 please review the definition of a valid reproduction and its exit-status semantics.