Verify completed-task cancel race (#140) is resolved on main; close superseded PR #180

[millsmillsymills]

Summary

Verify the completed-task cancel race (#140) is actually resolved on current main, then close the now-superseded draft PR #180.

Background

Draft PR #180 fixes the race in tests/mandatory/protocol/test_tasks_cancel_method.py, but that file no longer exists on main — the tests/mandatory/ tree was restructured into tests/compatibility/ (commit 2bb20c2), which is why the PR is CONFLICTING. The race appears already handled by the replacement coverage in tests/compatibility/core_operations/test_task_lifecycle.py:

• TestCancelTask.test_cancel_task_returns_updated_state (CORE-CANCEL-001) tolerates the terminal-state race.
• TestCancelTask.test_cancel_terminal_task_returns_error (CORE-CANCEL-002) asserts -32002 TaskNotCancelableError on terminal tasks.

Suggested work

• Run the new TestCancelTask cases against a synchronous SUT and confirm the completed-task race no longer produces a false failure.
• Also check the restructured equivalent of test_a2a_v030_new_methods.py::TestMethodMappingCompliance::test_core_method_mapping_compliance (named in test_tasks_cancel_valid: race condition causes false failures on synchronous agents #140) for the same immediate-cancel race.
• If confirmed resolved, close test_tasks_cancel_valid: race condition causes false failures on synchronous agents #140 and the superseded draft fix(tests): tolerate completed task cancel race #180. If it still reproduces, fix in test_task_lifecycle.py (not the deleted path).

Follow-up from review of #180.

[millsmillsymills]

Re-verified against main — structural facts hold, but two overclaims in this issue need correcting before anything is closed:

Confirmed:

• tests/mandatory/protocol/test_tasks_cancel_method.py is gone (mandatory tree removed in 2bb20c2); that's why PR fix(tests): tolerate completed task cancel race #180 is CONFLICTING — it edits a deleted path and is genuinely superseded.
• TestCancelTask.test_cancel_task_returns_updated_state (CORE-CANCEL-001) and test_cancel_terminal_task_returns_error (CORE-CANCEL-002) exist in tests/compatibility/core_operations/test_task_lifecycle.py under an @must class.

Corrections:

• CORE-CANCEL-002 does not assert -32002. Its body only does if response.success: errors.append(...) — it asserts that cancel on a terminal task returns some error, never inspecting the code. TaskNotCancelableError/-32002 appears only in the docstring. A SUT returning a wrong-but-present error code still passes.
• CORE-CANCEL-001 does not "tolerate the terminal-state race." It sidesteps the race by requesting a tck-input-required (non-terminal) scenario via create_working_task() and pytest.skip-ing if the task isn't in input_required; its body still flags a failed cancel as an error. So a synchronous SUT that immediately completes the task gets a skip — silent non-coverage — not a tolerated pass.
• The second site named in test_tasks_cancel_valid: race condition causes false failures on synchronous agents #140 (test_a2a_v030_new_methods.py::TestMethodMappingCompliance::test_core_method_mapping_compliance) no longer exists on main, so that re-check step is moot.

Net: PR #180 can be closed as superseded, and the original crash/false-failure path is gone. But the cancel-success behaviour for synchronous/immediate-completion SUTs is avoided-by-skip, not proven-resolved — so the right gate before closing #140 is exactly this issue's own suggestion: run CORE-CANCEL-001/002 against a synchronous SUT and confirm coverage isn't silently skipped. Keeping open until that run happens.

[millsmillsymills]

Verification result — synchronous-SUT run

Ran the upstream/main TestCancelTask methods (verbatim — test_task_lifecycle.py and _task_helpers.py confirmed identical to main) and create_working_task/create_completed_task against two simulated SUTs: a scenario-aware client (honors tck-input-required) and a synchronous client (completes every task, ignores the scenario prefix). Driven natively under pytest so skip/pass/fail is pytest's own verdict.

Test	Scenario-aware SUT	Synchronous SUT
CORE-CANCEL-001 (test_cancel_task_returns_updated_state)	PASS	SKIP
CORE-CANCEL-002 (test_cancel_terminal_task_returns_error)	PASS	PASS

The CANCEL-001 skip is the load-bearing result, with the exact reason:

SKIPPED tests/compatibility/_task_helpers.py:98:
  Expected task state 'TASK_STATE_INPUT_REQUIRED' but got 'TASK_STATE_COMPLETED'

Confirmed: on a synchronous/immediate-completion SUT, CORE-CANCEL-001 is silently skipped — create_working_task requests tck-input-required, the task completes instead, validate_task_state fails the precondition, and the helper pytest.skips. The cancel-success path is never asserted. This is avoidance-by-skip, not race-tolerance.

Also confirmed (CANCEL-002 weakness): a synchronous SUT whose terminal-task cancel returns a wrong error code (-32603 instead of -32002) still PASSES, because the test only checks if response.success and never inspects the code.

Conclusion / actions

• Do not close test_tasks_cancel_valid: race condition causes false failures on synchronous agents #140 as resolved on the strength of the current tests. They do not cover the synchronous-agent cancel path — they skip it. Track a follow-up to give CORE-CANCEL-001 a synchronous-path assertion (cancel a task created via the default-complete scenario; accept either updated-state success or -32002).
• PR fix(tests): tolerate completed task cancel race #180 can still be closed — it edits the removed tests/mandatory/... path and the original crash is gone.
• Strengthen CORE-CANCEL-002 to assert the -32002 code specifically.

Method note: the bundled sut/a2a-python is scenario-aware (it implements tck-input-required), and a synchronous build needs an SDK checkout not present locally — so the synchronous case was exercised by driving the real test+helper code with simulated transport responses. #190's logic is entirely in the TCK test layer, which is run verbatim here; the SUT only supplies the task-state responses the table varies.