test(desktop): scope slash-menu mutation watch to the menu and read once by yunaremaia · Pull Request #3737 · apache/maka

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.

yunaremaia · 2026-08-24T20:15:11Z

Summary

Implements the direction suggested in #3727 for the slash-command-menu flake:

The mutation watch now observes the menu itself (subtree) plus its parent, instead of document.body with subtree: true. Overlays elsewhere in the document (toasts, tooltips, other popups) can no longer contribute removals to the counter.
The observers disconnect once the refresh rounds have settled (visibility wait + one frame pair), and the count is read exactly once. The old expect.poll on a monotonically increasing counter could never retry into success - it only added 3s of latency before failing.
No retries are added; the config's fail-loudly stance (retries: 0) is untouched.

Why this is safe

The guarded property (#2667) is "a same-content projection refresh must not tear down the skills group or the listbox". The new watch counts exactly two things during the refresh window: a [role="group"] removed inside the menu, or the listbox element itself removed. Anything else was noise under the old predicate, which counted any removed element containing a listbox/group anywhere in the document.

Verified locally under xvfb (Electron 43):

Spec green: 5/5 tests, six consecutive runs of the target test all passed.
Regression check: temporarily injecting a group teardown right after arming makes the rewritten test fail with removals = 1 - the narrowing does not mute the failure mode the test exists to catch.

Note on artifacts

The trace/video artifacts from the five failing runs linked in the issue had already expired, so I could not diff them directly; the code-path analysis above is the basis for the window-scope reading.

Fixes#3727

The watch armed on document.body with subtree:true for the whole poll window and counted any removed element containing a listbox or group, so an unrelated overlay unmounting within 3s reported the apache#2667 regression. Because the counter only increases, expect.poll could never retry into success - it added latency, not tolerance. Scope it to the guarded property instead: one observer on the listbox (subtree) catches a group torn down inside it, one on its parent catches the whole listbox being replaced. Disconnect after the refresh rounds settle and read the count once. Verified locally (xvfb): spec green 5/5 across six runs; an injected group teardown during the window still fails the test (removals=1), so the narrowing does not mute the regression it guards. Fixesapache#3727

Astro-Han

I reviewed this head and found no blocking issues. No P0-P2.

Checks on 7dfa534 are test: success.

简体中文

该头未发现阻断问题。

Astro-Han

APPROVE — 7dfa534, test-only fix for #3727, no P0-P2, observer correctly narrowed. See prior COMMENT 5014208436.