feat(reeval): cap per-trial POVs and add reeval merge script

[acorn421]

Summary

• Bound per-trial POV verification load via discovery-time stratified sampling (CRSBENCH_REEVAL_POV_SAMPLE_SIZE, default 100; 0 disables).
• Apply the cap on both sync (_reeval_bug_finding) and async (_enqueue_trial_povs) verification paths so cloud workers see the same sampled subset.
• Add scripts/merge_reeval_into_source.py to overlay a reeval result tree onto the source experiment tree with a surgical pov_store.json merge that preserves original CRS timing while adopting reeval verdicts.
• Update docs/design/distributed/cloud-reeval.md to document the per-trial POV cap.
• Add tests/test_reeval.py covering sampling/env parsing behavior.

Description

Re-evaluation of large CRS runs was bottlenecked by trials with thousands of POVs. The cap enforces a deterministic per-trial budget by sorting POVs by mtime (a proxy for CRS discovery time), partitioning into equal time buckets, and randomly choosing one per bucket — seeded by the trial path so distributed workers agree on the subset. The merge script lets reeval results be folded back into the original experiment tree without losing original-run timing or downgrading POVs that reeval intentionally skipped (sampling) or never saw.

Commits:

• d4e23beb feat(reeval): cap per-trial POVs via stratified discovery-time sampling
• 05e4c563 chore(reeval): lower default POV sample cap to 100
• 8b3068db feat: add reeval merge script