fix: a body that stops mid-read is transient, and a round that dies still says so

Round 5 died on `http.client.IncompleteRead(0 bytes read)` and left no record
at all. Two defects, both ours.

The retry classified only connection-level failures. IncompleteRead and
RemoteDisconnected arrive while the *body* is being read, after the connection
succeeded, so they are not URLError -- they escaped both the retry and the
error wrapper and walked out of the round. The network is allowed to be
unreliable; losing a paid round to it is not the provider's defect.

And the exception path was wrapped around the stages that come *after* the
proposal is applied, because those are the ones with something to revert. The
stages before it had already spent money, and a round that leaves no record is
a round nobody can account for. Every outcome is journalled now, with the
exception type in the reason.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG

tests/test_learning_round_evidence.py

@@ -224,7 +224,7 @@ def test_a_stage_that_raises_is_reverted_journalled_and_re_raised(world) -> None
         _run(world, _proposal(), build)
     assert (world["configs"] / "injection" / "checklist.md").read_text() == "- old\n"
     outcomes = [(item.outcome, item.reason) for item in world["journal"].rounds()]
-    assert outcomes == [("reverted", "stage_failed")]
+    assert outcomes == [("reverted", "stage_failed: RuntimeError")]  # the exception type is part of the record
     assert (world["out"] / "rounds" / "1" / "scores.json").is_file()
 
 
@@ -406,3 +406,23 @@ def test_a_round_journals_why_the_proposer_had_nothing_to_say(world) -> None:  #
     )
     assert record.outcome == "rejected"
     assert record.reason == "no_proposal: meta_agent_failed: RuntimeError"
+
+
+def test_a_round_that_dies_before_it_applies_anything_is_still_journalled(world) -> None:  # type: ignore[no-untyped-def]
+    """Measured: round 5 died on a network read during its *first* stage and left no record at all.
+
+    The exception path was wrapped around the stages after the proposal is applied, because those are the ones that
+    have something to revert. The ones before it had already spent money."""
+
+    def build(config: FamilyConfig):  # type: ignore[no-untyped-def]
+        def scan(root: Path) -> list[StaticFinding]:
+            raise ConnectionError("the body stopped arriving")
+
+        return scan
+
+    with pytest.raises(ConnectionError):
+        _run(world, _proposal(), build)
+    outcomes = [(item.outcome, item.reason) for item in world["journal"].rounds()]
+    assert outcomes == [("reverted", "stage_failed: ConnectionError")]
+    assert (world["out"] / "rounds" / "1").is_dir()
+    assert (world["configs"] / "injection" / "checklist.md").read_text() == "- old\n"

tests/test_provider_retry.py

@@ -63,3 +63,53 @@ def test_non_retryable_value_error_propagates_immediately() -> None:
 
     with pytest.raises(OpenRouterError):
         call_with_retry(op, attempts=3, base_delay=0, sleep=_no_sleep)
+
+
+def test_a_body_that_stops_mid_read_is_transient() -> None:
+    """Measured: an evolve round died on `http.client.IncompleteRead(0 bytes read)`.
+
+    The retry classified only connection-level failures, so a response whose *body* stopped arriving escaped both
+    the retry and the error wrapper, and the exception walked out of the round. The network is allowed to be
+    unreliable; losing a paid round to it is our defect, not the provider's."""
+    import http.client
+
+    from openultrasast.provider.openrouter import OpenRouterError, _is_transient, call_with_retry
+
+    assert _is_transient(http.client.IncompleteRead(b"")) is True
+    assert _is_transient(http.client.RemoteDisconnected("closed")) is True
+    assert _is_transient(ValueError("a bug in our own code")) is False
+
+    attempts = []
+
+    def flaky() -> str:
+        attempts.append(1)
+        if len(attempts) < 3:
+            raise http.client.IncompleteRead(b"")
+        return "ok"
+
+    assert call_with_retry(flaky, attempts=3, base_delay=0.0, sleep=lambda _s: None) == "ok"
+    assert len(attempts) == 3
+    assert issubclass(OpenRouterError, Exception)
+
+
+def test_a_read_failure_is_reported_as_a_provider_error_not_a_raw_http_exception() -> None:
+    import http.client
+    import urllib.error
+
+    from openultrasast.provider.openrouter import OpenRouterChatClient, OpenRouterError
+
+    client = OpenRouterChatClient(api_key="k", base_url="https://example.invalid", max_attempts=1)
+
+    def boom(*_args: object, **_kwargs: object) -> object:
+        raise http.client.IncompleteRead(b"")
+
+    import openultrasast.provider.openrouter as module
+
+    original = module.urllib.request.urlopen
+    module.urllib.request.urlopen = boom  # type: ignore[assignment]
+    try:
+        with pytest.raises(OpenRouterError, match="IncompleteRead"):
+            client.complete_chat_raw(model="m", messages=[{"role": "user", "content": "x"}])
+    finally:
+        module.urllib.request.urlopen = original  # type: ignore[assignment]
+    assert urllib.error.URLError is not None