feat(context): slice 2 — structural-v1 policy, path index, replay-lite

[akougkas]

Part of #179. Depends on slice 1.

path-index.ts (paths from tool-call args, fileEntry entries, write/edit results; no fingerprints), protect.ts, policies/structural.ts with rules in order: stale_after_mutation, superseded_read, failure_resolved, listing_consumed, thinking_turn_closed, age_horizon (last rung, newest-safe-first, batched to target). replay/ over Clio ledgers only: loader, reference graph (file_reread, file_discovery, file_rewrite), oracle and random controls, metrics (retention, eviction_precision, tokens_evicted, churn), JSON + markdown report. CLI clio-coder context replay --sessions <dir>.

Acceptance

• Frozen fixture under tests/fixtures/context-replay/ produces the same contextEviction sequence through the replay runner and the live engine.
• On local Clio ledgers, structural-v1 retention >= age-horizon and eviction precision >= random; table committed under benchmarks/results/context-replay/.
• Only then: default policy flips to structural-v1.
• Supersession is range-aware (a range read supersedes only an identical or containing range).

[akougkas]

Landed on v0.3.4 via PR #183, range 8a1c8304..9073e395 (merge f404745f).

Against the acceptance list:

• The frozen fixtures under tests/fixtures/context-replay/ produce the same contextEviction sequence through the replay runner and the live engine (working-set-scenarios-replay.test.ts, working-set-replay-claude-code.test.ts).
• Retention and precision: local Clio ledgers were too short to measure anything (129 sessions, 123 under 8 turns), so the corpus is 165 Claude Code transcripts of this repository through a loader that normalizes them into Clio-shaped entries. At the shipped protectLastTurns: 6, structural-v1 retention (mean) is at or above age-horizon at every budget (0.480/0.590/0.812 vs 0.479/0.582/0.788) and precision is above random at 32k and 64k (0.953/0.962 vs 0.947/0.960); at 128k it is 0.979 against random's 0.980. That one cell misses the precision >= random clause by 0.001 and is recorded in benchmarks/results/context-replay/README.md with the full grid rather than rewritten around.
• Default policy flipped to structural-v1 on that basis (44f7e94f, numbers refreshed in c56030f3).
• Supersession is range-aware (working-set-structural.test.ts).

Replay surface changes on review: the churn column was 1 − precision and is gone; context replay with nothing kept exits 1; the loader merges Claude Code's per-block assistant records into one message (ca3f49b6), which superseded the earlier tables.