test(computer-use): synchronize session cancellation

[liugddx]

Summary

• make the session-clear test wait for a same-pipe round trip before cancelling the silent request
• have the mock executor acknowledge cancellation and await that causal response before reading its log
• remove two fixed 80ms sleeps while retaining state, generation, and cancel-log assertions

Why

clearSession() synchronously sends $/cancel; the CI flake came from reading the child-process log after a fixed delay that did not guarantee the child had consumed stdin and appended the record. The mock records cancel before sending its response, so awaiting that response is a deterministic acknowledgement instead of a scheduling guess.

Validation

• npm --workspace @maka/computer-use run build
• target test repeated 30 times
• node --test packages/computer-use/dist/__tests__/maka-cu-service.test.js (17 pass)
• Biome and git diff --check

[Astro-Han]

Thanks — the barrier-based synchronization is genuinely causal, not another timing guess: the stdin FIFO guarantees window.list is answered only after the mock consumed the observe line, and await call waits for the cancel reply (and its cleanup of the kill timer). 30 runs including a CPU-saturated pass all green, and the old test was arguably hollow (mock never answered cancel, so the child was doomed to SIGKILL at +2s anyway — the assertions passed right before the behavior they claimed to verify broke). The new test strengthens rather than compromises the semantics.

One follow-up suggestion (not blocking, P3): the sibling file maka-cu-backend.test.ts:1165-1168 still waits a fixed delay(120) for the fire-and-forget session.end — the exact CI-flake mechanism this PR's description names. The same barrier trick (a round-trip await after clearSession) would make it deterministic too.

Merging.