Cache and back off Web Inspector socket discovery

[shaun0927]

Tracked by #713.

Background

Web Inspector socket discovery can run lsof, ps, and filesystem scans repeatedly during proxy startup. This is valid work, but repeated discovery increases startup latency.

Objective

Reduce repeated socket discovery cost while preserving correctness for booted simulator selection.

Scope

• Add short TTL cache for UDID-to-socket path results.
• Invalidate cache after failed socket probe.
• Use exponential or staged backoff during wait loops.
• Keep existing lsof and mtime fallback behavior.

Target files and paths

• src/simulator/socket-finder.ts
• src/simulator/proxy.ts
• tests/unit/socket-finder.test.ts
• tests/unit/proxy-timing.test.ts if proxy wait behavior changes

Implementation steps

1. Add a cache keyed by target UDID and no-target lookup separately.
2. Store only socket paths that pass probeSocket.
3. Invalidate cached path on failed probe.
4. Replace fixed 500ms retry with staged backoff.
5. Add tests for cache hit, stale cache invalidation, and no-target behavior.

Acceptance criteria

• Repeated discovery for the same UDID reuses a recently probed live socket.
• Stale cached sockets are invalidated and rediscovered.
• No-target discovery does not incorrectly return a socket for a requested UDID.
• Existing fallback paths remain available.

Verification

• npm run lint -- --quiet
• npm test -- --runInBand tests/unit/socket-finder.test.ts
• Manual macOS smoke: boot simulator, start proxy twice, confirm second discovery is faster or avoids repeated expensive commands in debug logs.

Dependencies / recommended order

Run after proxy readiness split if both touch startup timing. This issue is independent of HTTP/security work.

Risks and cautions

• Do not return unprobed socket paths from cache.
• Do not cache failures for so long that newly booted simulators are missed.
• Do not delete stale sockets unless they were explicitly probed stale.

Non-goals

• Do not change proxy port allocation.
• Do not replace lsof fallback entirely.
• Do not introduce platform support beyond macOS/Xcode simulator needs.

Ambiguity reducers

• “Cache hit” requires a live probe success within the TTL.
• “Backoff” must still respect the configured overall timeout.
• Target-specific and target-agnostic cache entries are separate.

Validity review

Proceed as a focused speed issue. It supports startup responsiveness without changing automation semantics.

Implementation detail refinements

• Make TTL configurable for tests.
• Keep cache in module or finder instance with a reset hook for tests.
• Expose debug counters only in tests or verbose logs.

Verification checklist reinforcement

• Live socket cache hit avoids lsof in unit test mock.
• Stale cache triggers rediscovery.
• Backoff loop respects total timeout.
• Existing mtime fallback test still passes.

[shaun0927]

검증 완료(2026-05-16 KST): 최신 origin/develop / OpenSafari v0.6.0 기준으로 해당 이슈의 실행 가능한 체크리스트를 확인했습니다.

Evidence:

• PR-linked implementation is merged into develop.
• npm test -- --runTestsByPath ... targeted suites: 10 suites / 103 tests passed for HTTP auth, high-risk HTTP gating, auth persistence, audit logging, network interception, zombie cleanup, and device_boot WebKit retry.
• Broader targeted suites: 23 suites / 557 tests passed for lazy tool registration, proxy readiness, WebKit performance/refactors, socket discovery cache, simulator lifecycle/refactors, native input backends, DOM input scripts, coordinate-space, and AX diagnostics.
• npm run lint -- --quiet ... passed for the changed/validated paths.
• npm run build passed for v0.6.0.
• Real OpenSafari HTTP smoke: serve --http bound to 127.0.0.1; /health returned 200; unauthenticated /mcp returned 401; authenticated initialize succeeded; disallowed CORS origin returned 403.
• Real OpenSafari device smoke for Recover device_boot when WebKit client registration fails #756: device_boot on iPhone 16 iOS 26.4 returned a booted device and proxy; subsequent diagnose retained the simulator in device even while WebKit was unavailable.

체크리스트 중 외부 앱/장기 수동 검증 성격이 아닌 항목은 모두 체크 가능하다고 판단하여 완료로 닫습니다.