feat(core): wire evolution hooks into execute_stream path (U2, OQ6 fix)
ConfigDrivenAgent.execute_stream() now fires on_task_complete/on_task_failed
evolution hooks in its finally block, achieving lifecycle parity with the
sync execute() path. This fixes the OQ6 gap where WebSocket-routed streaming
tasks bypassed evolution entirely.
Implementation:
- Module-level backpressure manager (_schedule_evolution / drain_pending_evolution_tasks)
with cap = max(2, max_concurrency * 2), drop + log + counter on exceed, and
shutdown drain via asyncio.gather(return_exceptions=True).
- _trigger_evolution_hooks / _evolve_safe methods on ConfigDrivenAgent: fire-and-forget
via asyncio.create_task, evolution errors swallowed (never fail the stream).
- execute_stream finally block distinguishes cancelled (CancelledError /
TaskCancelledError -> CANCELLED), failed (Exception -> FAILED), completed
(final_answer received -> COMPLETED), and early-close (no completion, no
error -> CANCELLED "stream closed before completion").
- app.py shutdown drains pending evolution tasks.
- plan_exec_engine.py / reflexion.py: doc comments noting hooks fire at the
ConfigDrivenAgent layer (single chokepoint, no double-fire).
- portal.py: verification comments at 3 execute_stream call sites (these call
react_engine.execute_stream directly, bypassing ConfigDrivenAgent - known gap
tracked separately).
Tests (8 new in test_execute_stream_hooks.py):
- Happy path: success fires COMPLETED, failure fires FAILED.
- Edge cases: cancellation fires CANCELLED, early aclose fires CANCELLED,
evolution error suppressed, backpressure cap drops + counts.
- Parity: REST on_task_complete vs execute_stream both fire COMPLETED.
- Disabled: _evolution_enabled=False fires no hooks.
dd259153fa