Fail closed on unsafe incremental resumes

[davanstrien]

Summary

• Fingerprint the sampled pages and each OCR column, then stamp every comparison and checkpoint with per-pair provenance.
• Include source repo, config and split, raw judge backend identity, sampling, prompt, preprocessing, token cap, and minimum text threshold in resume identity.
• Reuse unchanged old-model pairs when a model is added, but require --full-rejudge for legacy unstamped or mismatched current-grid rows.
• Make stamped checkpoint rows authoritative over stale completed-run metadata after an interrupted full rejudge.
• Publish only one-way judge-spec hashes so endpoint credentials and private URLs are not exposed.

Compatibility

Legacy result repositories fail closed once and can be upgraded with --full-rejudge. Historical comparisons outside the current rankable grid remain preserved.

Stack

This is stacked on #56 because source split is part of the provenance contract. Change the base to main after #56 merges.

Validation

• 632 tests passed
• Ruff passed on src and tests
• ty passed on src
• git diff --check passed