fix(server): Cursor adapter turn-lifecycle hardening (races, missing terminal events, lock retention)

Summary

While building the omp ACP provider (#8985 / #9038, modeled on the Cursor driver) we verified three latent issues in apps/server/src/provider/Layers/CursorAdapter.ts. They are inherited patterns, not regressions — filing separately so they can be fixed on their own merits. The omp adapter shipped with the fixes; Cursor still has the originals.

1. sendTurn binds activeTurnId too late

ctx.promptsInFlight += 1 (CursorAdapter.ts ~L927) and ctx.activeTurnId = turnId (~L948, after effectful session configuration) leave an interleaving window: a second concurrent sendTurn reads the previous turn's activeTurnId and steers/attributes events onto the wrong turn. Fix (as in #9038): bind activeTurnId in the same synchronous stretch as the increment.

2. No terminal event when a started turn fails

turn.started is emitted before attachment reads and acp.prompt; failures from either propagate with no turn.completed, relying entirely on ProviderCommandReactor.handleTurnStartFailure to settle the session. RuntimeTurnState already supports "failed" with errorMessage — emitting the terminal event adapter-side (guarded on !ctx.stopped and the last-in-flight rule) makes the runtime event stream self-consistent.

3. Cancellation during turn preparation is ignored

interruptTurn only calls ctx.acp.cancel, which is a no-op before acp.prompt starts. A turn cancelled during applyRequestedSessionConfiguration / attachment reads still gets prompted. Fix (as in #9038): record the cancelled turn id in a session set, check it before turn.started and again before prompting, settle as cancelled.

4. threadLocksRef grows forever

Every thread that passes through getThreadSemaphore retains its semaphore for the adapter's lifetime — stopSession/stopAll/failed startups never remove it. Fix: delete the per-thread lock on session teardown and startup-failure paths (careful: the startSession replacement path must keep the lock it holds).

All four were verified against Cursor's code by reading; fixes 1-3 exist and are tested on the omp side (#9038, tests in OmpAdapter.test.ts) so the port is mechanical.

Start in apps/server/src/provider/Layers/CursorAdapter.ts, focusing on sendTurn, interruptTurn, getThreadSemaphore, and session teardown or startup-failure paths. Read the corresponding fixes and tests in OmpAdapter.test.ts, then verify the Cursor adapter covers turn binding races, failed-turn terminal events, preparation-time cancellation, and per-thread lock cleanup.