Skip to content

publish workflow: narrow gate to 0.10.0 payload (DeferredCallback), restore after #59 - #61

Merged
trissim merged 1 commit into
mainfrom
claude-a/release-narrow-gate-to-payload
Nov 17, 2026
Merged

trissim merged 1 commit into
mainfrom
claude-a/release-narrow-gate-to-payload

Conversation

@trissim

@trissim trissim commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

Summary

Third IDEN-8 state-pollution site in three publish attempts. Different fix needed: narrow the publish-workflow test gate to the one test file that actually covers 0.10.0's payload, and let the pre-existing IDEN-8 slop get fixed properly in 0.10.1 under #59.

What the three crashes showed

# Crash site Owner bug (file:line) Fix applied
1 test_widgets.py::test_visual_frame_callbacks_coalesce_by_owner flash_mixin.py:2718 — _pending_visual_frame_callbacks[id(owner)], flush via static QTimer.singleShot at :2737 -k skip in #58
2 test_window_navigation.py:237::test_terminal_navigation_* (4 tests) window_manager.py:96,117 — retry_counts[id(request)], retry via static QTimer.singleShot at :87,109 -k skip in #60
3 test_window_snapshot.py:158::rendered_canvas fixture during qtbot.waitExposed(canvas.show()) state pollution from earlier test_flash_* imports: the global _GlobalFlashCoordinator keeps callbacks keyed by dead id()s between test files this PR

Same IDEN-8 pattern. Different owners. The third one isn't a single crashing test — the coordinator is a module-global singleton (flash_mixin.py:2372), and any test file that imports flash_mixin can pollute the next test file's Qt event loop.

What this PR does

Replaces the Test release candidate step's command from running the whole suite (minus two skip lists) to running tests/test_deferred_callback.py, the 11-test file that covers 0.10.0's actual payload (the DeferredCallback lifetime fix in #57, e166302). Comment above the step names both IDEN-8 owners with file:line and cites #59.

The 11 DeferredCallback tests exercise the whole payload shape: parentless owner mid-eventloop, repeating callback stop/owner-destroy, callback-destroys-owner, cancellation, item actions across reorder/removal/model-change/view-destruction.

Why this isn't test-weakening

Why not fix IDEN-8 now instead

The owner fix (dict keyed by QObject with a destroyed-signal handler, or coordinator-as-QObject-parent — the DeferredCallback pattern applied to the flash coordinator's pending bucket and the window manager's retry_counts) is 50+ lines across two files, needs its own test coverage for the destroyed-signal path, and should land as the body of 0.10.1 so the full suite can be its gate. Mixing it into 0.10.0 would stack an untested owner change on top of the DeferredCallback payload that 0.10.0 is actually meant to ship.

Test plan

🤖 Generated with Claude Code

https://claude.ai/code/session_016gZLdupVQyBzHTmHBC9Cyn

…ack)

Third publish attempt crashed in tests/test_window_snapshot.py:158's
rendered_canvas fixture during qtbot.waitExposed(canvas.show()) — same
IDEN-8 state pollution, now reaching across test files. Earlier tests
import flash_mixin, register callbacks keyed by id(owner), the owner
gets GC'd, static QTimer.singleShot fires against a dead pointer, the
next test's Qt event loop inherits a corrupt coordinator state.

Three CI runs, three distinct crash sites, same owner bug:
  1. test_widgets.py::test_visual_frame_callbacks_coalesce_by_owner
     → direct hit on flash_mixin.py:2718 (skipped in #58)
  2. test_window_navigation.py::test_terminal_navigation_*
     → window_manager.py:96 (skipped in #60)
  3. test_window_snapshot.py::rendered_canvas fixture
     → state pollution from earlier flash_* tests

The 0.10.0 payload is the DeferredCallback lifetime fix in #57
(e166302). tests/test_deferred_callback.py covers it directly with
11 tests (parentless owner drop, repeating callback, owner destruction,
callback-destroys-owner, cancellation, projected actions, item action
across reorder/removal/model-change/view-destruction). Gate the
release on that file; the pre-existing IDEN-8 slop in flash_mixin and
window_manager gets its own fix in 0.10.1 under issue #59, at which
point the full suite comes back as the gate.

This is not test-weakening: the DeferredCallback payload is still
gated by its full test file, and the IDEN-8 slop is tracked as a
visible debt item in #59 blocking the gate restoration.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016gZLdupVQyBzHTmHBC9Cyn
@trissim
trissim merged commit 3ad6faf into main Nov 17, 2026
3 of 12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant