test(runtime-host): validate Windows crash recovery

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!

[liugddx]

Summary

• enable the killed Runtime Host recovery gate on Windows
• use an explicit fixture IPC shutdown request for portable graceful Host teardown
• close execution and task-ledger stores before releasing root leases so Windows can remove SQLite files
• refresh the Windows skip inventory

Validation

• npm --workspace @maka/runtime-host run build
• npm --workspace @maka/runtime-host run typecheck
• node --test packages/runtime-host/dist/__tests__/execution-host-queue.test.js (9/9 passed)
• shared execution-host fixture suites run serially (36/38 passed before the two remaining same-root-cause lease leaks were fixed)
• the two corrected execution-host cases passed after the fix
• final post-rebase Windows gate set passed (3/3)
• npm run windows:inventory
• git diff --check

The full runtime-host suite still exposes pre-existing Windows baseline failures outside this change, primarily SQLite file removal and live-root rename semantics tracked by #2142.

Part of #2142

[jackwener]

Review decision

Problem real: Windows could not exercise Host crash recovery; fixture held SQLite handles across lease release (EBUSY), and SIGTERM-based graceful stop is not a portable Host teardown contract.

Fix correct:

1. Enable killed-Host recovery test on Windows
2. IPC → for graceful stop; SIGKILL path waits on
3. / task-ledger before owner/reader release

Landing via rebased PR after inventory conflict resolution.

[jackwener]

Landing via rebased maintainer PR. Thanks @liugddx!

[Astro-Han]

Thanks — the recovery harness changes look well scoped. Keeping the real crash path as SIGKILL, registering the close listener before termination, using an explicit fixture IPC shutdown for graceful teardown, and closing SQLite stores before releasing the root lease are all appropriate Windows boundaries.

The recovery scenario itself is valuable: it verifies recovery before successor readiness, app_restarted, one interaction closure, one terminal event, and no repeated recovery after a third Host starts. I do not think these assertions should be reduced.

The remaining work is integration with current main:

Please rebase and regenerate the Windows inventory

The PR currently conflicts only in docs/windows-test-inventory.md. #2475 has already moved current main to 20 windows-backend-gap entries / 55 total declarations. Removing this Runtime Host skip should produce 19 / 54.

Please rebase and run npm run windows:inventory rather than resolving the generated counts manually. The four Runtime Host test/fixture files otherwise merge cleanly.

P3 — Add focused Windows CI evidence for the new contract

The existing Windows baseline builds the repository and runs inventory, scripts, smoke, and Storage tests, but it does not execute Runtime Host tests. The target recovery case currently runs automatically only on Linux; the Windows 3/3 result in the PR description is local evidence.

A narrow Windows step for the single recovery case would protect the new platform contract without running the full Runtime Host suite, for example:

node.exe--test --test-name-pattern="a killed Host is recovered exactly once" packages/runtime-host/dist/__tests__/execution-host-queue.test.js

This can remain part of the current non-blocking Windows baseline. It is a coverage improvement rather than a reason to redesign the harness.

After the rebase, regenerated inventory, and fresh CI run, this should be ready for another quick review. Thanks!