Bounded GC for crash-abandoned policy-lock claims and stale secure-temp roots

[fitchmultz]

Context: follow-up from PR #108 / v0.4.1 (https://github.com/fitchmultz/pi-agent-browser-native/releases/tag/v0.4.1, immutable head d6cde09).

Follow-up: cheap GC for crash-abandoned policy-lock claims and stale secure-temp roots

Observed (2026-08-11, mega-dogfood gate cycle)

• Killed/aborted Pi test runs (and by extension real Pi crashes) abandon .pi-agent-browser-policy-*.lock-v2 / .lock-v3.claim-* dirs under /tmp/pi-agent-browser-policy-<uid>.
• 54 accumulated stale claim dirs measurably degraded later policy-lock acquisitions: two contention-window tests (serializes caller-owned live verification, revalidates daemon policy after cross-instance lock contention) failed deterministically in a 7-file suite run and passed 91/0 after quarantining the stale dirs. Healthy runs leave zero claims.
• Reclaim exists (PID/start-identity probes reclaim proven-dead current-protocol claims) but pays per-candidate process probes at acquisition time; abandoned v2 owners fail closed by design, so buildup taxes every later same-identity acquisition.
• Separately, TMPDIR pi-agent-browser-* secure-temp roots accumulate from aborted runs (observed 1600+ in one day of heavy testing); stale-root ownership checks handle them but a cheap opportunistic sweep would bound the scan set.

Proposal (post-merge, separate PR)

• Opportunistic bounded GC at coordination-directory ensure time: reclaim/remove proven-dead current-protocol claims and v2 dirs older than a floor (e.g. 30 min) with dead owners, capped per call (e.g. 8 entries) so acquisition latency stays bounded.
• Same bounded sweep idea for marked secure-temp stale roots (already ownership-checked; just add an age floor + per-call cap).
• Keep fail-closed semantics for ambiguous/foreign/legacy entries exactly as today; GC only provably-dead current-protocol artifacts.
• Tests: unit coverage for cap/floor/dead-proof gating; regression that stale-claim buildup no longer delays an unrelated-identity acquisition.

Non-goals

• No change to fail-closed abandoned-v2 semantics, lock protocol, or claim formats.
• Not a substitute for the cross-process global recording claim follow-up (separate file).

[fitchmultz]

Reproduced abandoned v3 claim accumulation and an unbounded stale-temp sweep on current main under Termux. Draft PR #174 contains the bounded/rotating cleanup, before/after reproduction results, broken-then-fixed regression evidence, and explicit MacBook validation gaps. The historical contention-test timeout itself was not reproduced. No release or merge performed.