fix(testing): fail fast on invalid handler output

[Thanatos9404]

Issue #, if available: #670

Description of changes:

The local runner sent every invalid invocation output through _retry_invocation: up to five attempts, five seconds apart, then a failure with ErrorObject.from_exception(e). As #670 describes, the service fails the execution on the first invalid output and retries only PENDING with no pending operations, up to three attempts. This implements the fix proposed there.

• Executor._finish_invocation: a validation failure now fails the execution at once with {ErrorType: "InvalidParameterValueException", ErrorMessage: <validation message>}. That covers FAILED with a Result, SUCCEEDED with an Error, and a status the runner does not accept (such as RETRY). PENDING with no pending operations still goes through _retry_invocation, now capped at MAX_CONSECUTIVE_INVALID_PENDING_ATTEMPTS = 3, and fails with the same error shape.
• A missing or unknown Status, or an output that is not a JSON object, never reached _finish_invocation: DurableExecutionInvocationOutput.from_dict raised inside the invoker, and the executor retried that as a failed invocation. Both invokers now parse through new InvokeResponse.from_output and InvokeResponse.from_payload factories, which raise a new InvalidInvocationOutputError (a DurableFunctionsTestError). _invoke_handler records the invocation and fails the execution with the same error shape. A stop or timeout that landed during the invocation still stands. LambdaInvoker parses the payload after its try, so a payload that is not JSON is no longer reported as "Unexpected error during Lambda invocation".
• _retry_invocation takes an optional max_attempts. Handler exceptions, invoke errors and invocation timeouts keep MAX_CONSECUTIVE_FAILED_ATTEMPTS = 5.
• docs/architecture.md: the retry budget section describes the split.

The PENDING retry shares the existing consecutive_failed_invocation_attempts counter, so earlier consecutive failed attempts count toward its cap of 3 (two handler errors followed by an invalid PENDING fail at once). I can give it its own counter if the service counts them separately.

Testing:

• New tests/e2e/invalid_invocation_output_test.py: six invalid outputs (FAILED with a Result, SUCCEEDED with an Error, no Status, unknown Status, RETRY, a JSON array) each fail the execution after one handler call with the InvalidParameterValueException error. On main all six time out while the retries run.
• executor_checkpoint_test.py: PENDING with no pending operations fails on the third attempt. PENDING once, then SUCCEEDED, completes and resets the counter.
• executor_test.py: the four test_should_retry_when_* tests for outputs that now fail fast are now test_should_fail_when_*. Two new tests cover the unparseable-output path, one with a stop during the invocation. test_should_retry_when_pending_response_has_no_operations now sets seq_counter; without it the test hit a TypeError in max() and retried through the generic error path, not validation.
• invoker_test.py, exceptions_test.py: both invokers raise InvalidInvocationOutputError for outputs they cannot parse.

Locally (Windows, Python 3.13, Hatch 1.16.5): hatch run test:all has no new failures (only the pre-existing Windows-only ones in the core lambda_service_test.py negative timestamps and the microvm-worker worker_test.py port tests, as on main); hatch run dev-testing:cov passes at 96.08%; hatch run dev-testing:typecheck, hatch run types:check and hatch fmt --check are clean. The asserted error messages are identical on Python 3.11 to 3.14. hatch run dev-testing:js-examples was not run, because the harness needs fcntl; CI runs it. No conformance tests were added or changed.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[zhongkechen]

/ai review

[zhongkechen]

/ai review

[github-actions]

Codex AI review

No actionable findings. The prior request-ID finding is resolved. Review was limited to static inspection; runtime behavior and test results were not independently verified.

Reviewed commit a389f5ffa979f3e058f2242c5d1d6b47eb0e290c. Workflow run

[zhongkechen]

Reviewed commit a389f5f. I reproduced two remaining gaps in invalid-output handling and an incomplete part of the request-ID fix, detailed inline. These are existing paths left incompletely fixed, rather than newly introduced regressions.

Validation: all 267 selected executor, checkpoint, invoker, exception, and invalid-output tests passed. Additional targeted reproductions covered the three store backends, undecodable response bytes, and real boto3 parsing against a local HTTP server.