Operator API hardening: error mapping and input validation (QA panel on #3970)

[mabry1985]

The QA panel flagged these on #3970. They're in operator routes that existed before that PR, so they were left out of it. Re-check each against current main before fixing.

Error mapping

• operator_api/routes.py:~135: _http_error returns str(exc) as the 500 detail, leaking internal file paths and library internals to the client. Log the exception and return a generic message.
• No not-found mapping, so these return 500 instead of 404:
  • DELETE and PUT /api/scheduler/jobs/{id} for a missing job
  • the task close route for a missing issue
• _runtime_status and _goal_status have no try/except guard, unlike their sibling routes.
• POST /api/chat has no structured mapping for chat() failures. /v1 maps the same failures to 429, 502 and so on.

Availability

• operator_api/routes.py:~119: the SSE event stream ends on a single malformed bus event, because per-event key access and json.dumps aren't guarded.

Input validation

• SubagentBatchRequest.tasks has no upper bound.
• GET /api/background?status=: the value isn't checked against the documented 3-value enum.
• POST /api/goals takes a bare dict instead of a Pydantic model.
• graph/goals/controller.py:~419: max_iterations and no_progress_limit from a JSON goal spec aren't type-checked, so a string such as "abc" is accepted.
• ScheduleAddRequest.schedule and .timezone aren't format-checked at the route. The scheduler now normalises errors to 400 (fix: settle ledger on restart reconcile, ValueError for bad schedules, watch clear race, partial scheduler PUT (#3943) #3969), but checking at the route would be clearer.

Nit

• graph/checkpoint_prune.py:~97: the LIKE pattern in the delete_thread cascade doesn't escape % or _ in thread_id.

[mabry1985]

More items from the QA panel's next round on #3970. They're in code that predates that PR; re-check each against main before fixing.

• server/__init__.py:~505: --headless maps to --ui console instead of --ui none.
• operator_api/routes.py:~360: DELETE /api/background/{id} and POST /api/background/{id}/cancel don't validate job_id, unlike GET.
• SubagentBatchRequest.tasks is list[dict] with no per-item model.
• Task routes:
  • registered even when tasks_store is None, so every call returns 500 (AttributeError);
  • GET /api/tasks/status reports initialized=true when there's no store;
  • PATCH /api/tasks/issues/{id} returns 500 for a missing issue, and DELETE returns 200 {deleted:false} instead of 404.
• GET /api/subagents has no error guard.
• graph/checkpoint_prune.py:
  • prune and aged-thread lookups are N+1;
  • reclaim() runs a full VACUUM on legacy DBs, which holds an exclusive lock.

[mabry1985]

From the QA panel's final round on #3970 (merged): knowledge/hybrid_store.py:~435 _drop_vectors (used by delete_forget_pending) lacks the sqlite3.DatabaseError guard the other vector-cleanup methods have, so a transient vector-table error aborts the whole forget-delete instead of degrading to chunk-only deletion.

[mabry1985]

From the QA panel on #3971 (merged): server/turn_sync.py:~245 performs the post-turn _pending_interrupt_value check outside the per-thread lock. That code predates #3971 (aed5688). Decide whether the post-turn check needs the lock.