Follow-ups from the 0.9.21 test-suite batch #263

Closed
opened 2026-10-27 22:18:10 -05:00 by ethan-hann · 3 comments
Owner

Smaller items from the 0.9.21 review and gap analysis (D70):

  • BackendWorkspace's world cache retries only OperationCanceledException (BackendWorkspace.cs ~132), so a cancelled capture that surfaces as an xunit assertion is not retried.
  • The gate's count check has two implementations (a regex in check-test-counts.sh, XML parsing in the .ps1); give them one shared listing format or a cross-check test.
  • No test covers the one-process App mode's thread cap; no test runs the real memory reading, and an unknown reading picks the heavier lanes layout.
  • The memory check reads /proc/meminfo, which shows the host's memory inside a container; read the cgroup limit when there is one.
  • The 7 GB layout rule means a small Linux VM always runs one App process, so it no longer mirrors Linux CI's lanes.
  • The stale-lock cleanup in build/vm-common.sh may end leftover build servers after every normal VM run; check on the VM.
  • Mutation-check the shared-window pill sweep and the merged tab walk.
  • Log at session end whether the game was seen under its process name, so a missed exit can be told apart from a wrong name (#255, #256).
  • The guides should say to render a screen before a change as well as after, since a fresh worktree has no previous render.
  • Copy the memory fallback note into the gate.ps1 header.
  • Record the measured CI time in D70 after the first PR gate with the lanes.
Smaller items from the 0.9.21 review and gap analysis (D70): - BackendWorkspace's world cache retries only OperationCanceledException (BackendWorkspace.cs ~132), so a cancelled capture that surfaces as an xunit assertion is not retried. - The gate's count check has two implementations (a regex in check-test-counts.sh, XML parsing in the .ps1); give them one shared listing format or a cross-check test. - No test covers the one-process App mode's thread cap; no test runs the real memory reading, and an unknown reading picks the heavier lanes layout. - The memory check reads /proc/meminfo, which shows the host's memory inside a container; read the cgroup limit when there is one. - The 7 GB layout rule means a small Linux VM always runs one App process, so it no longer mirrors Linux CI's lanes. - The stale-lock cleanup in build/vm-common.sh may end leftover build servers after every normal VM run; check on the VM. - Mutation-check the shared-window pill sweep and the merged tab walk. - Log at session end whether the game was seen under its process name, so a missed exit can be told apart from a wrong name (#255, #256). - The guides should say to render a screen before a change as well as after, since a fresh worktree has no previous render. - Copy the memory fallback note into the gate.ps1 header. - Record the measured CI time in D70 after the first PR gate with the lanes.
ethan-hann added this to the Code Cleanup milestone 2026-10-27 22:18:10 -05:00
Author
Owner

Also: after several VM gates, ~/gate/tree/.tools/tmp/linux on the Linux VM held 22,368 leftover test folders (59 MB). Some test classes do not remove their temp trees; find which (the folder names say which helper made them) and make them clean up on dispose, and have the gate clear .tools/tmp/ at the start of a run.

Also: after several VM gates, ~/gate/tree/.tools/tmp/linux on the Linux VM held 22,368 leftover test folders (59 MB). Some test classes do not remove their temp trees; find which (the folder names say which helper made them) and make them clean up on dispose, and have the gate clear .tools/tmp/<os> at the start of a run.
Author
Owner

Also: RealSessionTests.A_game_that_exits_at_once_and_is_started_again_by_another_program_is_followed_to_the_end_of_the_new_copy only exercises the relaunch-stub path when the watcher lists the stub under its name before its 0.6 s exit, so it passes or fails by timing. Make it deterministic by having the stub wait for the watcher to see it before it exits.

Also: RealSessionTests.A_game_that_exits_at_once_and_is_started_again_by_another_program_is_followed_to_the_end_of_the_new_copy only exercises the relaunch-stub path when the watcher lists the stub under its name before its 0.6 s exit, so it passes or fails by timing. Make it deterministic by having the stub wait for the watcher to see it before it exits.
Author
Owner

Items 1 to 10 shipped in SavePoint 0.9.22. Item 6: the VM check found that gate-linux.sh and package-linux.sh start the gate with the lock descriptor closed (8>&-), so build servers never hold the lock and the stale-lock cleanup cannot end them; no change needed. Left: item 11. The first PR gate with the lanes (PR #274, run 197) took 34.6 minutes of wall clock; that figure goes into D70 in the next batch.

Items 1 to 10 shipped in [SavePoint 0.9.22](https://git.tortal.tech/ethan-hann/SavePoint/releases/tag/v0.9.22). Item 6: the VM check found that gate-linux.sh and package-linux.sh start the gate with the lock descriptor closed (`8>&-`), so build servers never hold the lock and the stale-lock cleanup cannot end them; no change needed. Left: item 11. The first PR gate with the lanes (PR #274, run 197) took 34.6 minutes of wall clock; that figure goes into D70 in the next batch.
Sign in to join this conversation.
No milestone
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
ethan-hann/SavePoint#263
No description provided.