feat(reliability): isolated gauntlet, warm Runtime replay, and lifecycle fixes

[kitlangton]

Why

Browser Control's hostile-page gauntlet could only run against the user's live browser and relay, so regressions were hard to reproduce and timing evidence was not trustworthy. Behind it sat real reliability bugs. Adopting a busy tab could hang. Reconnecting a canonical CDP client always paid a ~3 s Runtime reset. Rotated credentials and OAuth fragments leaked into captures. Concurrent same-ID session lifecycle calls could acknowledge uncommitted state.

What Changes

Isolated regression loop. pnpm gauntlet:isolated builds this checkout outside the source tree. It launches a disposable Chromium profile with the real MV3 extension and a private relay/HOME/ports, and owns every process group. Leaks, cleanup failures and budget overruns can never be relabelled as expected failures. Correctness failures are never retried into green. A PR lane runs six real-extension cases; a nightly lane repeats the full suite three times.

Faster warm reconnect. After its frame-tree response, a canonical client that reconnects to a live target receives validated cached execution contexts. It no longer waits for a native event and then resets Runtime.

Warm reconnect + evaluate (1 warmup, 7 measured)	Median
Before	3013 ms
After	7.7 ms

Correctness fixes

Area	Before	After
Tab adoption	Waited for a responsive renderer; could hang	Binds exact target identity immediately; the first execute initializes lazily with a bounded timeout and reports session-page/adopted-initialization-timeout without running user code or dropping the tab
Title reads	Ambiguous hang	session-page/context-read-timeout; operation=page.title without page replacement
Capture redaction	Rotated values and OAuth #access_token= fragments could appear in artifacts	Historical values redact to stable ${BC_SECRET_N} refs; profiles store only current values
Session lifecycle	Same-ID create/ensure/delete/reset could interleave and acknowledge uncommitted state	Serialized publication and rollback, with atomic identity claim
Handoff start	Early human completion could leave a pending action unbounded	Deadline and target cancellation stay active until the action settles
Tab detach while staged	Relay per-tab state survived	Cleared, along with staged descendants
Snapshots	Native roles and document baselines could drift	Correct native roles; navigation retires baselines and refs

BROWSER_CONTROL_AUTOSTART=false lets a supervisor make calls fail instead of spawning a replacement relay.

Runtime Context Replay

sequenceDiagram
  participant C as Canonical client
  participant R as Relay (CdpRuntime)
  participant X as Extension / Chrome
  C->>R: Page.getFrameTree
  R->>X: forward
  X-->>R: frame tree
  R-->>C: response (records main frame id)
  C->>R: Runtime.enable
  R->>X: forward
  X-->>R: ack
  R-->>C: replay cached contexts for this requester (deduped, validated)
  Note over R: target, generation, ownership or detach changes invalidate the cache;<br/>misses and aliases keep the bounded reset path

Loading

The client pool is the only thing that retires requester state, for announced and silent detaches alike. Cache eviction relies on Chrome's destroyed/cleared events rather than Page.frameNavigated, matching Playwright's Chromium path.

Dependencies

• Effect 4.0.0-rc.112 → rc.118: effect/unstable/* imports become effect/*, and the Config and CLI constructors are renamed to PascalCase.
• Playwright 1.63, ws 8.22, Vitest 5, Knip 6.38 and current tooling.
• CLI --help output is byte-identical across all 39 commands.
• MCP tool failures now follow Effect's updated server behavior (noted in the changeset).

Scope

This PR owns the isolated harness, the fixes above, the Runtime replay and the dependency update. Follow-ups are tracked separately:

• alias Runtime event routing
• staged-tab cleanup on extension reconnect
• serializing execute/adopt auto-created sessions behind the identity gate
• cross-tool benchmarks

Verification

pnpm typecheck && pnpm check:locals && pnpm check:unused
pnpm test
pnpm audit:duplicates
pnpm build:cli --outdir "$TMPDIR/bc-build" && pnpm build:extension
pnpm gauntlet:isolated

• 970 tests pass across 69 files; typecheck, locals and Knip are clean; 0 duplicate clones.
• Full isolated gauntlet on the final tree (Chrome 153.0.8010.12, macOS): 10/10 genuine passes, 0 xfail/fail/budget overruns, cleanupVerified: true, 143 s total; runtime replay connect + evaluate took 10.4 ms.
• Linux CI real-extension lane (Chrome 153, ubuntu-latest): 6/6 passes.