test(npu_pipeline): test_dispatcher.py's async mock_post breaks dispatcher.py's sync-call contract #13004

What

Discovered while verifying issue #12979 batch 9 (unrelated scope — services/npu_pipeline/dispatcher.py's HTTP session is a documented RAW carve-out in that batch, not converted). Confirmed pre-existing via a cp-swapped baseline against origin/Dev_new_gui — identical failures with the unmodified file.

autobot-backend/services/npu_pipeline/test_dispatcher.py fails 5/5 of its tests when run standalone:

FAILED test_happy_path_three_workers
FAILED test_worker_drop_mid_pass_failover
FAILED test_single_worker_produces_tokens
FAILED test_worker_drop_no_peer_raises
FAILED test_context_manager_closes_session

Root cause

• services/npu_pipeline/test_dispatcher.py:78, :175, :198 — mock_post is declared async def mock_post(url, json=None): ... and assigned as session_mock.post = mock_post. Production code (dispatcher.py:151, :193) calls it synchronously — async with self._session.post(url, json=payload) as resp: — expecting .post(...) to return an async context manager directly, not a coroutine. Calling an async def function returns a coroutine object without executing its body, so:
  • test_happy_path_three_workers, test_single_worker_produces_tokens: TypeError: 'async for' requires an object with __aiter__ method, got coroutine (the coroutine is passed to async with/async for unexecuted).
  • test_worker_drop_no_peer_raises: the raise inside mock_post never executes (coroutine body doesn't run until awaited), so the expected WorkerDropError chain doesn't fire as designed; observed failure is TypeError: Expected a BaseException type, but got 'MagicMock' in the pytest.raises(WorkerDropError, ...) context — WorkerDropError itself resolves to a MagicMock, which also implicates the module-loading indirection below.
• services/npu_pipeline/test_dispatcher.py:22-30 (_load_module) — the file bypasses sys.modules stubs the root conftest installs for services.npu_pipeline.*, but the reuse guard (if name in sys.modules and hasattr(sys.modules[name], "__file__")) can pick up a stub module rather than the freshly-loaded one depending on import order, which would explain WorkerDropError binding to a MagicMock attribute instead of the real exception class in test_context_manager_closes_session / test_worker_drop_no_peer_raises.

Evidence

$ python3 -m pytest autobot-backend/services/npu_pipeline/test_dispatcher.py -q
5 failed in 1.00s

Same 5 failures reproduce against origin/Dev_new_gui's unmodified dispatcher.py (cp-swap baseline, no diff).

Suggested fix

Change mock_post in each of the 3 affected tests from async def to a plain def returning the pre-built resp (or an AsyncMock with side_effect), matching the synchronous-call / async-context-manager-return contract HTTPClientManager-style clients use. Separately, audit _load_module's stub-reuse guard against the root conftest's services.npu_pipeline.* stubs.

Model Used

Sonnet (claude-sonnet-5)