Reliability: android reaper meta key, evidence-walk cost, atomic writes

[ElbertePlinio]

From the 2026-07-19 complementary swarm review (read-only, orthogonal to #44).

P0 — reliability

• Android session create/destroy never sets REAPER_CLEANUP_PENDING_META_KEY (packages/android/src/session.ts:106-119,137-153), unlike desktop/browser. A leaked emulator can become permanently untracked by the reaper. Fix: set/clear the pending-cleanup meta key with the same protocol the other runtimes use.

P1 — performance & reliability

• appendAction does a full recursive readdir+lstat walk of the run directory on every evidence action (packages/core/src/evidence.ts:925-957,1069). Cache or incrementally track size instead of rescanning.
• MCP run-log resource reads entire log file into memory with no size cap (packages/mcp-server/src/resources.ts:589 — screenshots are guarded at :528, logs are not), then 9 regex redactSecrets passes. Add the same size guard as screenshots + stream/tail semantics.
• Browser supervisor child-shim buffers Chrome stderr unbounded when no \n arrives (packages/browser/src/supervisor.ts). Bound the buffer.

P1 — simplify

• Atomic write (tmp + rename + counter) hand-rolled 5× (core/session.ts:87-122, run.ts:61-92, config.ts:105-125, evidence.ts:66,744-751,1244-1251); 3 copies leak the tmp file on rename failure. Correct impl exists at agent-installers/src/atomicFile.ts. Hoist writeFileAtomic into core and use everywhere (do NOT touch evidence.ts's concurrency-critical claim/lock paths beyond the two safe call sites).

Validation

• bun test per touched package; typecheck; verify reaper picks up an artificially leaked android session.

[ElbertePlinio]

Opened #47 implementing all checklist items:

• P0: packages/android/src/session.ts create/destroy now set/clear REAPER_CLEANUP_PENDING_META_KEY with the same protocol as desktop-linux/browser. Added packages/android/test/reaper.test.ts proving a leaked android session (failed destroy) stays reaper-trackable and gets reaped.
• P1 perf: appendAction no longer full-walks the run dir per action — cached artifact byte total, invalidated via a cheap entry-name snapshot (immune to journal-lock/truncation-sentinel churn); journal's own size read live off the open handle. Orphaned on-disk artifacts still counted (existing cumulative-bytes tests pass unchanged, plus a new cache-hit test).
• P1 perf: MCP run-log resource now caps memory with tail semantics (mirrors the screenshot guard), instead of reading the whole file before redaction.
• P1 perf: browser supervisor child-shim stderr buffer bounded to last 64KB.
• P1 simplify: hoisted writeFileAtomic into packages/core/src/paths.ts; used in core/session.ts, run.ts, config.ts (fixes the tmp-file leak on rename failure in all three), agent-installers' call sites, and the two safe evidence.ts call sites — claim/lock code paths in evidence.ts untouched otherwise.

Validation: tsc --noEmit clean, full vitest run — 78 files / 950 passed, 2 skipped. No sudo usage introduced; no overlap with #44.