Add opt-in targeted adaptive judging and size-aware stopping

[davanstrien]

Summary

The current adaptive loop checks for convergence every five pages, but keeps judging the full all-pairs grid until every adjacent CI separates. On a 14-model run that means up to 91 pairs per page batch, so persistent near-ties can turn “adaptive” mode into almost the full benchmark.

This PR adds an opt-in targeted strategy while preserving the historical balanced default:

ocr-bench judge <output-repo> \
  --adaptive-strategy targeted \
  --size-tie-ratio 3

Targeted allocation

• starts with enough balanced all-pairs evidence to compute an initial board
• then tops up only adjacent pairs whose 95% CIs remain unresolved
• recomputes adjacency after every batch, so ranking changes can activate new pairs
• records adaptive_strategy in results metadata
• leaves balanced as the default while the sparse strategy is validated
• documents that conditioning later sampling on interim ranks means ordinary bootstrap CIs may be somewhat optimistic

Size-aware practical stopping

• --size-tie-ratio RATIO treats an overlapping adjacent pair as a practical smaller-model preference when one model has at least that parameter advantage
• applies to either adaptive strategy: it can satisfy the whole-run stopping condition in balanced, while targeted also stops topping up the pair
• --size-tie-min-samples N requires direct pair evidence first (default: 10)
• requires both marginal CIs to exist; missing CI/size data fails safe as unresolved
• adds a ★ annotation to CLI, dataset rows/cards, and the viewer
• explicitly labels this as a deployment preference: it does not change ELO/rank or claim statistical equivalence

Resume, history, and compatibility hardening

• reconstructs the targeted pair set from checkpointed comparisons on resume instead of filling the intentionally sparse grid
• detects new/incomplete model-grid warm-ups and continues balanced allocation until every current pair has direct evidence
• excludes subset/retired/failed-model comparisons from the current ELO fit while preserving them in both checkpoints and final Hub publication
• fails closed when an existing comparisons/metadata config has files but cannot be loaded, preventing transient/schema errors from becoming destructive replacement pushes
• makes equal-ELO ordering deterministic by model name, since adjacency now controls allocation
• stores strategy/size-rule provenance in the append-compatible metadata schema; old viewers safely ignore the new leaderboard annotation column

bench forwards the new options.

Validation

• pytest tests/ -q — 594 passed
• ruff check src/ tests/
• ty check src/
• independent review loop covered targeted resume, equal-ELO determinism, model-grid expansion/interruption, history preservation, fail-closed Hub reads, and missing-CI semantics

Closes #35.

[davanstrien]

Addressed the review in ad72d86:

• History preservation: out-of-grid rows (subset/retired configs and currently failed models) are excluded from the current ELO fit but retained in both checkpoint payloads and final comparisons publication. Existing comparisons/metadata loads now fail closed when config files exist but cannot be loaded, so a transient/schema read failure cannot become a destructive replacement push.
• Strategy semantics: CLI help and README now state that --size-tie-ratio applies to balanced and targeted stopping; targeted additionally stops topping up that pair.
• Missing CIs: practical preference now requires both marginal CIs to exist.
• Statistical caveat: README notes that targeted allocation conditions sampling on interim ranks, so ordinary bootstrap CIs may be optimistic.
• Kept rankings_resolved as a documented public statistical-only helper and short-circuited preference recomputation when no size rule is enabled.

Added regressions for retired/subset and all-sentinel history preservation, checkpoint preservation, fail-closed Hub reads, and missing-CI behavior.

Validation: 594 tests, Ruff, ty, and CI all green.