fix(hygiene): keep a carried trim through interrupted and busy boundaries; review-wave-1 fixes and the findings reports

[alpertarhan]

Review wave 1: verified findings on the #76–#79 code, plus the findings reports

Four external reviews landed under docs/findings/ (first commit). Every claim acted on here was re-verified against the code and Pi 0.87.1 first; three claims were found wrong and left alone (noted below).

Fixed (second commit)

1. Carried trim dropped on non-committing boundaries (claude B1, deepseek P1-2, muse P0-2). turn_end nulled applied at its top and restored it only on a contested boundary. Pi persists an aborted/failed response with a fresh timestamp (agent-session.js:333), so after such a turn the next request looked warm, went out untrimmed and rewrote the prefix the interrupted request had just cached. applied now lives until a boundary commits it, unchangedSince fails, the session changes, or contextHygieneEnabled is turned off. Tests append the interrupted response before turn_end (aborted and error), cover a busy boundary (canAutoTrim false) and cleanup turned off.
2. Cache lifetime from the last response only (claude B2). A fully cached (cacheWrite = 0) or interrupted response after a 1h write made the prefix look 5 minutes old. cachedPrefix() dates the entry by the last response and takes the lifetime from the last response that wrote cache — the host-cache-ledger rule.
3. Automatic trims need a reachable smart_context (deepseek P1-9, muse P0-5): canAutoTrim checks toolExposure.reachable("history") like artifact offload; ARCHITECTURE claimed it, nothing enforced it. Smoke through src/index.ts wiring with a fake host: eager/lazy trim (1 context_edit), off does not. Readiness & details names the inactive state.
4. Silent drop of a queued change on an incomplete turn (glm B2): the usual "Context operation not applied" note now appears. Rewind attributed as trim in the ledger (glm B1): staged.kind, new ContextEditKind rewind.
5. Small: explicit source for noteForeignCompaction instead of startsWith("Another") (claude B4); redundant before_switch/before_fork ledger resets removed (claude B5; session_start resets); write-only markedAt removed; break-even docstring counts the trimmed request (claude B7); veto derivation comment; bench case price a held trim (trimTokens, p95 3.3 ms) beside plan trim; package.json files excludes docs/findings (claude C2; packed count unchanged at 269).

Claims checked and not applied

• Veto math (deepseek P1-3 = muse P0-3): the code is right. Pi's missCost = (w − r)·P; the warm path still reads X at r, the cold+trim path neither reads nor writes X, so the avoided cost is missCost − w·X. The comment now states the derivation.
• Conflict notice truncated (claude B3): measured 360 chars with two conflicts, cap 400; only 3+ conflicts truncate.
• Warming stop is sticky (muse P1-1): Pi re-decides per request (cache-warmer.d.ts:47).
• Per-turn planning cost (claude P1): measured 4.9 ms at 560 entries / 2.9 MB, 17.8 ms at 2,800 entries / 14.4 MB — real, small; the bench now pins pricing too.
• lifecycle-e2e flake (deepseek P0-9): 4 full-suite runs green.

Not in this PR

#75-area defects confirmed by re-verification (artifact-storage chunk views — reproduced with a 1.2 MB line; tombstone pruning; Hindsight fail-open ledger; native compaction budget/estimate/backup; bridge lock; Mnemopi readiness; Home settings writeConfig; dashboard in RPC; settled watchdog; receipts; UI batch) are queued for separate PRs.

[alpertarhan]

Receipts (local, env -i, private HOME/TMPDIR) at ca70613:

• bun run typecheck: clean (src, scripts, tests, bench).
• bun test: 1568 pass / 0 fail across 120 files (4 earlier full runs on main also green; the reported lifecycle-e2e flake did not reproduce).
• bun run gate: 376 / 0, EESV adversarial release gate passed (26 suites).
• bun run bench: exit 0 — plan trim over 120 archived reads p95 2.3 ms, new price a held trim over 120 archived reads p95 3.3 ms (limit 25 ms).
• bun run build: pass; bun run release:audit: pass (269 packed files; docs/findings excluded).
• dist/index.js sha256 8de115bff66d9ac1454369ede1480ea628a66785c4fcdbcb2746699e7dc6744f.
• Mutation checks (each reverted after): dropping applied at turn_end top → 4 tests fail; lifetime from the last message only → the 1h test fails; staged.kind hardcoded → the rewind test fails; no incomplete-turn note → the aborted table row fails; keeping a carried trim with cleanup off → its test fails. A faithful pre-fix pending variant fails exactly the 3 new retention tests.
• Reachability smoke through src/index.ts with a fake host that activates registered tools: eager → 1 context_edit, lazy → 1, off → 0.