fix(handoff): bound escalation assessment latency

Co-Authored-By: Codex <noreply@openai.com>

.ai/HANDOFF.md

@@ -23,10 +23,12 @@ Fixes made:
 - The SSE handshake test now calls `stream_escalations()` directly and consumes only the first generator yield, avoiding HTTPX's infinite-stream buffering behavior.
 - Handoff manager/API tests stub `_generate_ai_assessment()` with an `AsyncMock`.
 - `EscalationBus` normalizes string/UUID account IDs at subscribe/publish/unsubscribe/subscriber_count boundaries, with a regression test.
+- Follow-up fix: escalation AI assessment is now latency-bounded by `ESCALATION_AI_ASSESSMENT_TIMEOUT_SECONDS` (default 5s). If it times out, handoff creation proceeds with no assessment instead of blocking on the model/network path.
 
 Verified:
 - `pytest tests/test_escalation_bus.py tests/test_handoff_manager.py tests/test_session_handoffs_api.py tests/test_flowpilot_analytics_escalations.py --override-ini=addopts= -q --durations=20` → `31 passed in 46.95s`
 - Same subset with `-n auto` → `31 passed in 17.80s`
+- After the assessment-timeout fix: same subset with `-n auto` → `32 passed in 17.77s`
 - No remaining pytest processes or `resolutionflow%test%` Postgres sessions after the run.
 
 ## Resume point
@@ -48,4 +50,4 @@ Verified:
 - Do not reintroduce `client.stream()`/ASGITransport tests for infinite SSE responses; test the generator directly or use a real server-level test.
 - `DROP SCHEMA public CASCADE` per test is still the dominant cost: DB-backed tests spend ~1.7-2.8s in setup. Use `-n auto` for focused backend loops.
 - The bus is acceptable for v1 pilot scale only because Railway is single-replica. Redis pub/sub is the obvious swap when horizontal scaling appears.
-- Synchronous `_generate_ai_assessment()` during escalation creation remains product-latency risk; tests are now isolated from it, but the UX path should be watched as the magic-moment screen is built.
+- Escalation assessment can be missing when the 5s timeout fires. The handoff-context UI must render a graceful "assessment unavailable/in progress" state rather than treating it as required.