feat: clean-restart attempt isolation — CCRM retry-contamination fix (#972)

[tcconnally]

Summary

Implements #972 — clean-restart attempt isolation (CCRM arXiv:2605.08563 borrow), the retry-contamination fix for multi-attempt workloads.

• src/perseus/retry_isolation.py — deterministic, stdlib-only:
  • transactional checkpoints: digest-sealed context snapshots at attempt boundaries (snapshot_context/verify_snapshot);
  • fence on failure: build_retry_context restores the pre-attempt snapshot + injects a bounded structured failure summary (attempt, step, kind, error — hard token cap); failed turns are quarantined and the builder fails closed if any leak into the restored context;
  • contamination events — digest-sealed, replayable (verify_isolation_event);
  • attempt-budget allocation — the paper's closed form T* = sqrt(B·log(1/(1−ε1))/log(1/(1−ε0))) with derivation inputs for audit;
  • seeded CCRM simulation — IID vs contaminated cascade vs clean restart at the ~7.1x cascade ratio: IID overestimates pass@3 by 10.63pp, clean restart recovers all of it (clean-restart dominance in simulation). The paper's SWE-bench numbers: 17.4pp overestimate, cascade ratio 7.1.
• benchmark/retry-isolation/ — offline seeded gate: overestimate ≥ 8pp, recovery ≥ 8pp, closed-form exactness, fence demo with quarantine verification. All gates PASS.
• CI — new retry-isolation-bench job. Docs + CHANGELOG included.

Composes with #968: fenced failure summaries feed TRACE attribution as explicit dissatisfaction signals.

Test Plan

• tests/test_retry_isolation.py — 17 tests (snapshots, quarantine, budget cap, closed form, seeded simulation, replay)
• Rebased onto main with feat(context): trajectory-mined context-source failure attribution (TRACE 2608.09153) #968–feat(context): commitment-preserving verifiable compression — Context Codec borrow (2605.17304) #971 merged: all six modules coexist (build --check clean)
• Retry-isolation benchmark gate PASS on the merged tree
• CI (Linux matrix + Windows + retry-isolation-bench)

Closes #972