fix(coding-agent): cancel the final kernel snapshot before disposal cleanup

[snimu]

What was wrong

When a kernel was being disposed, its final state snapshot was raced against a 5-second guard that couldn't cancel anything: if the snapshot was slow, the guard "won", disposal proceeded to close the ZMQ sockets, and the still-running snapshot then failed against dead sockets.

The fix

The final snapshot now runs through the kernel's existing cancellable execution path (interrupt, then force-abort after a grace period). Disposal waits for the snapshot to settle — completed or cleanly aborted — before any socket cleanup. The separate dispose-timeout constant and the uncancellable race are deleted; there is one timeout policy for snapshot execution, applied in one place. Both teardown paths (dispose and shutdown-with-snapshot) share the fix.

Net −13 lines.

How it's verified

The existing timeout test was updated to exercise the real flow: dispose() directly, the timeout starting only after queued work, and cleanup deferred until the aborted snapshot settles. tsgo/biome/suite green, re-validated after each stack merge. Two-model implement/review loop, approved first pass.

Stacked on #1705 (test the whole stack at the leaf; merge base-first).

Note: intentionally no Linear ticket for this cleanup stack, so that check stays red.

---

Note

Medium Risk
Changes kernel teardown ordering and snapshot behavior on dispose/shutdown; mistakes could hang disposal or skip needed state persistence, but scope is bounded to ReplKernelManager graceful cleanup.

Overview
Fixes a race where graceful kernel disposal could close ZMQ sockets while a final namespace snapshot was still running, leaving that work failing against dead sockets.

flushSnapshotForDispose no longer races an uncancellable snapshotState() against a separate dispose guard. On dispose/shutdown-with-snapshot it interrupts any active execution, waits for the executionQueue to settle within SNAPSHOT_EXECUTION_TIMEOUT_MS, then runs captureSnapshot through the normal cancellable execution path (or skips the final snapshot if the queue never settles). SNAPSHOT_DISPOSE_TIMEOUT_MS is removed so snapshot timing uses one policy in one place.

Tests in repl-kernel-abort.test.ts cover dispose waiting until a timed-out snapshot aborts before cleanupResources, and tearing down when a hung execution blocks the queue without starting a snapshot.

^{Reviewed by Cursor Bugbot for commit 35d4433. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Cancel hung final kernel snapshot before disposal cleanup in ReplKernelManager

Rewrites ReplKernelManager.flushSnapshotForDispose to interrupt any active execution and wait for the executionQueue to settle within SNAPSHOT_EXECUTION_TIMEOUT_MS before taking a final snapshot. If the queue does not settle in time, the snapshot is skipped so teardown proceeds without hanging.

• Removes the now-unused SNAPSHOT_DISPOSE_TIMEOUT_MS constant from shared.ts
• Adds tests in repl-kernel-abort.test.ts covering both the interrupted-snapshot path and the permanently-hung execution path
• Behavioral Change: during dispose(), if an active execution exists it is interrupted immediately, and the final snapshot is only captured when the queue settles within the timeout; previously the snapshot was raced against a separate dispose guard timeout

^{Macroscope summarized 35d4433.}

Linear: ENG-5650

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 35d4433. Configure here.}

[cursor]

Dispose snapshot wait can hang

Medium Severity

flushSnapshotForDispose only bounds the wait for the queue tail captured at entry, then awaits captureSnapshot with no outer deadline. A second execute (or other enqueue) that joins during that wait becomes prev for the snapshot and has no timeout, so dispose/shutdown can block indefinitely and delay /reload and signal teardown.

 

^{Reviewed by Cursor Bugbot for commit 35d4433. Configure here.}

[snimu]

Superseded by #1816: GitHub's stack lock prevents retargeting stacked PRs, so this PR was recreated as a plain PR against main (same branch, same content, all review threads here are resolved). Please review/merge #1816 instead.