[Roman Lutz (romanlutz)]

Description

Remove two remaining accidental speed requirements from backend correctness tests using the existing finite TEST_WATCHDOG_SECONDS backstop:

• _settle_send_async now defaults to the shared 300-second watchdog rather than ten seconds. It still polls through finalization to a semantic terminal state, and expiration still leaves the actual send running.
• The concurrent first-request analytics singleton test uses the same watchdog for report, quick-query, and queue budgets instead of thirty seconds. The backend analytics fixture also references the shared constant instead of repeating its value.
• Regression checks guard the helper default and the singleton controller's actual budgets. An event-controlled service test expires only the observer with timeout=0, checks that the send has no cancellation request, then releases it and verifies successful completion and reservation cleanup. The existing failure-teardown test also uses deterministic zero-timeout expiry and checks that only lifecycle shutdown cancels the send.

Production deadlines, intentional timeout coverage, and the async helper implementation are unchanged. This completes remaining adoption after #3203, #3268, and #3278. Open #3443 is complementary: its analytics changes cover explicit-expiry checkpoint and cleanup observation, not singleton-controller correctness budgets. This PR does not depend on it.

Merged upstream main ddaa6af996ac6fc1debf2828d0eb050ea509b7b8 without rebasing. Conflict resolution preserved both the combined async helper imports and main's newer message-send ownership tests, plus both adjacent documentation sections.

Tests and Documentation

Validated after merging main on Windows with CPython 3.14.4:

Exact command	Result
uv sync --frozen --extra all	Passed; synced merged dependency manifests.
uv run --no-sync pytest -q -n 4 --dist=loadfile tests\unit\test_async_utils.py tests\unit\backend\test_message_send_service.py tests\unit\backend\test_conversation_editor.py tests\unit\backend\test_temporary_components.py tests\unit\backend\test_analytics_routes.py tests\unit\backend\test_analytics_lifecycle.py tests\unit\analytics\test_execution.py	458 passed, 79 warnings in 140.60 seconds. Includes helper callers and analytics deadline/cancellation coverage.
uv run --no-sync ruff check tests\unit\backend\mocks.py tests\unit\backend\conftest.py tests\unit\backend\test_analytics_routes.py tests\unit\backend\test_message_send_service.py	Passed.
uv run --no-sync ruff format --check tests\unit\backend\mocks.py tests\unit\backend\conftest.py tests\unit\backend\test_analytics_routes.py tests\unit\backend\test_message_send_service.py	Passed; four files already formatted.
uv run --no-sync ty check --output-format concise tests\unit\backend\mocks.py tests\unit\backend\conftest.py tests\unit\backend\test_analytics_routes.py tests\unit\backend\test_message_send_service.py	Passed.
uv run --no-sync python -m build_scripts.validate_docs	Passed; 168 TOC references validated, no orphaned documentation.
git diff --cached --check	Passed before the merge commit.

Applicable merge-commit pre-commit hooks passed. A worktree-local pre-commit cache was used to avoid contention on the shared cache lock.

Updated doc/contributing/4_running_tests.md to explain the manual-send watchdog and non-cancelling observer ownership. JupyText is not applicable: no notebooks or executable documentation samples changed. Full-suite and macOS/Linux CI were not run locally. The overloaded-runner failure was not reproduced with real ten/thirty-second delays; regression guards verify the budgets directly without sleeps or global clock patches.

[Richard Lundeen (richlundeen)]

Follow-up: make send cleanup ownership consistent across the helper's callers.

tests/unit/backend/mocks.py:20-28 correctly times out the observer without cancelling the send. But the normal message_send_service and real_send_context fixtures (test_message_send_service.py:92-93,118-139) do not shut down accepted sends. For example, test_accepts_before_target_finishes_and_status_cancellation_does_not_cancel_send_async (lines 3114-3151) holds the provider behind release, then waits and asserts without a finally block or lifecycle owner. An assertion failure before release.set() leaves the send blocked as the provider patch and database fixture close. A completion-watchdog failure also leaves the send running beyond those scopes. The new ownership tests use a lifecycle context explicitly, so they do not cover this normal caller path.

Please apply the existing message_send_lifecycle_async ownership pattern to these callers, with shutdown inside the relevant patch scopes and gate release in finally. A fixture can provide a final safety net, but teardown after a test-local patch exits is too late. Add one controlled-failure check through a normal caller to confirm that sends and reservations are drained before mocks and memory close. Keep the observer non-cancelling; it is not the send's owner.

Review authored with GitHub Copilot.

[Richard Lundeen (richlundeen)]

Follow-up: remove the remaining short correctness deadlines in the same manual-send tests, not only the shared completion helper.

Current main still has these limits:

• test_message_send_service.py:3948-3952: five seconds for one conversation to finish finalization while sibling sends remain held.
• test_conversation_editor.py:164: five seconds for the provider to signal entry.
• test_message_send_service.py:1627,1777-1779,1837,1842: a ten-second concurrency-test class limit, a three-second independent-send limit, and thirty-second readiness waits.

These tests check ownership and independent progress, not a response-time requirement. A busy worker can still fail them after this PR raises the final completion budget. Open #3491 fixes analytics and other watchdogs, but does not cover these cases.

Please handle these together: use shared correctness watchdogs, make the whole-test backstop leave room for cleanup, and observe the owned task as well as its readiness signal so early failure is reported at once. For the per-conversation finalization case, retain polling or a scoped completion signal for that conversation: waiting for the entire send would deadlock against the deliberately held siblings. Keep the existing conflict, held-peer, and independent-progress assertions. Larger backstops need not add sleeps or increase successful test duration.

Review authored with GitHub Copilot.