fix: make browser session starts recoverable

[iuyo5678]

Problem

Fixes #245.

The DSH plugin previously recorded ownership only after a successful CLI reply. If the CLI was cancelled, timed out, or terminated after window creation, that reply could be lost and the Agent Window became unreachable from the plugin. Initial navigation/emulation failure could also discard ownership even when the compensating close failed.

Change

Represent each managed start as a recoverable request whose identity exists before any browser side effect:

• Persist the request and conversation ownership before dispatch; prepare, start, then claim only after initialization succeeds.
• Deduplicate retries by request token. Cancellation is terminal, including cancellation before a delayed start arrives. An unprepared start is rejected after a daemon restart.
• Keep ownership until window closure is confirmed. Wait for delayed creation to settle, retain failed startup compensation in the extension, and retry failed cleanup.
• Recover abandoned plugin journals; cancel pending and completed starts on archive/unload. Unclaimed requests expire through the existing daemon reaper.
• Keep pending cleanup visible and counted against capacity. Verify request ownership when short session IDs are reused, and release transfer resources on confirmed stop.

Ordinary CLI starts keep their existing behavior. Managed starts require matching CLI/daemon support; update the extension too for retryable startup-compensation failures. No new package dependencies.

Lifecycle and recovery contract

Validation completed on macOS

Suite	Passed
CLI/daemon Rust unit tests	352
Protocol tests	154
Session IPC, cancellation, queue and user-interrupt integration tests	50
DSH plugin tests	268
Extension session-manager and session-tool tests	59

Rust formatting, Clippy with warnings denied, TypeScript checks, and targeted Biome checks passed. The plugin build also passed.

The integration tests use a real daemon, IPC/WebSocket connections, and a controlled extension peer. One test kills an actual CLI child after dispatch and before its successful reply, then recovers and closes the session through a new CLI process. Additional cases cover delayed creation, duplicate requests, failed close/retry, failed startup compensation, archive/unload, journal recovery, capacity, and foreign-session isolation.

Windows/DSH/Chrome GUI validation is still pending. The results above do not claim that coverage.

Windows verification requested

Use CLI, daemon, extension, and DSH plugin builds from this branch; ensure the DSH profile points to that CLI and the daemon is also running this build. Keep the same working directory and journal location when testing host restart recovery.

1. Start, navigate, list, and stop normally. Expect one Agent Window and consistent current-session/list state.
2. Cancel a start around window creation, and separately terminate only that test's CLI process while the start is pending. Expect eventual closure; a temporarily unresolved outcome must remain recoverable instead of disappearing.
3. Trigger an initialization failure, then retry cleanup if the initial close fails. Expect the owned resource or pendingCleanup to remain visible until closure is confirmed.
4. Archive the conversation or unload/reload the plugin during a pending start. Expect no late successful start to become usable; cleanup must continue or recover on reload.
5. Terminate and restart the test DSH host. Expect abandoned requests to be recovered and cancelled. Keep a second live DSH instance/session open to verify it is unaffected.
6. Repeat interrupted starts and cleanup. Expect no growing window count, no bypass of maxSessions, and no closure of unrelated Agent Windows.

Focused automated checks to run on Windows:

cargo test -p bsk --lib --locked session_requests
cargo test -p bsk --test sessions_ipc --locked recoverable_starts
pnpm --filter @wxg-prc-cpg/browser-skill-dsh-plugin test

Please include Windows/Chrome or Edge/DSH versions, commit SHA, reproduction steps, and the before/after window and session counts with test results. Reviewer attention is especially useful on lifecycle races, journal recovery ownership, and cancellation/cleanup boundaries.