fix(gateway): bounded fail-open evidence assembly + PM decision transient-failure bypass

Evidence-assembly git legs (diff, changed-files, branch fetch, advisory
conventions run) ran unbounded inside claim_review / claim_doc_task /
claim_gate_review / evidence() / i_am_done's envelope build, so a slow
clone turned the whole verb into a silent 120s FlowVerbTimeout 504.
Each leg now runs through run_bounded_leg under a shared LegBudget
(evidence_assembly_timeout_seconds, 45s total): a timed-out leg — both
asyncio TimeoutError and git's own GitTimeoutError — degrades into an
evidence_gaps note on the envelope instead of hanging the verb, while
non-timeout git errors still propagate. The advisory conventions run
gets an inner-only timeout (conventions_validator_advisory_timeout_
seconds, 30s) threaded down to the subprocess so it is never orphaned
by an outer cancel; the fail-closed i_am_done/pr_pass conventions
gates keep their hardcoded 120s.

_ensure_pm_decision now reports a PmDecisionOutcome: a transient DB
failure recording the PM's decision journal (e.g. lock timeout under
load) no longer launders into a journal:decision gate rejection that
escalates and BLOCKS the task — the verb's own rationale satisfies the
gate with a structured warning, across all seven PM verbs.

Also: repo-wide ruff realignment to the lockfile-pinned ruff (8
format-only diffs, 14 UP038 isinstance conversions) that a transiently
newer venv ruff had masked.

Gate: 15474 passed, 459 skipped; ruff/mypy/xenon/vulture/bandit/
pip-audit/deptry/import-linter/foundation-check all green.
This commit is contained in:
Renn F
2026-07-31 13:56:19 +02:00
parent ef8849b0cd
commit ce4d02b3c2
41 changed files with 2160 additions and 138 deletions
@@ -989,6 +989,19 @@ async def test_escalate_up_matches_spec(role: str, status: str) -> None:
id=agent_id, role=role, team="backend", slug=None, escalation_target="main-pm"
)
task_svc.escalate.return_value = after
# escalate_up's own _ensure_pm_decision write-then-gate opens a
# session.begin_nested() savepoint unconditionally (even on the
# fresh-decision no-write path below) — an unshaped AsyncMock's
# auto-attribute return doesn't support `async with`, which orphans
# the mock's internal coroutine (AsyncMockMixin._execute_mock_call
# never awaited).
task_svc.session = MagicMock()
task_svc.session.begin_nested = MagicMock(
return_value=MagicMock(
__aenter__=AsyncMock(return_value=None),
__aexit__=AsyncMock(return_value=False),
)
)
journal_svc = AsyncMock()
# journal:decision is not on the spec — satisfy it so the verb-specific
# preflight does not surface a non-spec tracing_gap on the allowed branch.