Add robot/e2e/m5_acceptance.robot with 21 zero-mock E2E tests covering
all M5 acceptance criteria: context assembly (add/list/show/clear),
context scaling (10,000+ files without timeout), context policy
configuration (view-specific settings), budget enforcement
(max_file_size/max_total_size constraints), context analysis (ACMS
pipeline inspect/simulate), and plan execution with real LLM calls
(openai/gpt-4o-mini via plan use + plan resume).
Bug fixes discovered during E2E testing:
- Fix _save_policy_json commit bug: session.flush() -> session.commit()
so policy changes persist after session.close() (project_context.py).
Rollback wrapped in contextlib.suppress(Exception) to prevent masking
the original commit exception.
- Add session_factory DI provider to Container as providers.Singleton so
the four project context commands (set/show/inspect/simulate) resolve
a real sessionmaker instead of hitting a missing attribute. Return
type annotated as sessionmaker[Session] using future annotations and
TYPE_CHECKING guard (container.py).
- Propagate GEMINI_API_KEY in nox e2e_tests session (noxfile.py).
- Add Google/Gemini API key pattern (AIzaSy...) to secret redaction
in redaction.py.
Review feedback addressed across eight review passes:
- Budget enforcement test verifies large_file.py exclusion by
max_file_size constraint (not tautological token comparison).
- 10K scaling test validates simulation returns valid fragment_count
(non-negative; non-zero deferred to full ACMS pipeline integration).
- Plan resume test parses JSON via Extract JSON From Stdout with
TRY/EXCEPT — consistent with plan use test pattern.
- __import__("sqlalchemy") replaced with proper import in step defs.
- Context clear test verifies all three files removed (config/main/utils).
- Inspect test asserts tier_metrics and tier_budget are non-empty.
- _save_policy_json rollback wrapped in contextlib.suppress(Exception).
- Skip If No LLM Keys rewritten to use Evaluate with os.environ.get —
keys never stored in any Robot variable.
- All assertion messages follow safe pattern (no stdout/stderr embedding).
- Review finding ID prefixes removed from all comments.
- Type annotations added to all 8 _build_session_factory step functions.
- Default view verification checks exclude-path __pycache__.
- Prerequisite skip guards added to sections 2-4 (Variable Should Exist).
- Plan ID extraction moved inside TRY block.
Results: 25/25 E2E (21 M5 + 2 smoke + 2 M1), 1536/1536 Robot
integration, 383/383 Behave features (10,926 scenarios), coverage 97%.
ISSUES CLOSED: #745
hurui200320