ai: approval prediction — serving path (event listeners + prediction service)

[babltiga]

Part of #645 (approval-outcome prediction epic — read it first for the big picture).

Goal

The serving path: when a query lands in review, load the org's trained model, extract features, score, and persist the prediction row. Event-driven, fail-safe — a prediction failure must never affect workflow state.

Depends on: core.api services, the feature extractor, and the trainer/model classes.

Steps

1. ai/api/ApprovalPredictionService.java — interface: void predictForQuery(UUID queryRequestId) (plus void trainAll() used by the scheduled job in the next sub-issue). Implementation ai/internal/DefaultApprovalPredictionService.java:
   • Guards, in order (each persists a skipped=true row with a distinct skipped_reason string): feature disabled → DISABLED; no model row / serving=false → MODEL_NOT_SERVING (covers cold start and the below-AUC case).
   • Happy path: load persisted AI analysis + cost estimate + query attributes + rate counts (all via core.api lookups), extract features, model.predict(...), persist via ApprovalPredictionPersistenceService, publish ApprovalPredictionCompletedEvent.
   • Any unexpected exception → catch (specific types where possible), log ERROR, persist a failed=true sentinel row with error_message. Mirror how the AF-624 estimate path and the AI-analysis path handle failure sentinels.
2. ai/internal/ApprovalPredictionListener.java — two @ApplicationModuleListener methods:
   • on(QueryReadyForReviewEvent event) → predictForQuery(...). This event is published by workflow/internal/QueryReviewStateMachine exactly when a query enters PENDING_REVIEW and never for auto-approved/rejected/break-glass paths — which is precisely the population we want.
   • on(QueryEstimateCompletedEvent event) → the estimate is computed in parallel and may arrive after review-readiness. If a prediction row exists whose feature snapshot recorded estimate_missing=true and the query is still PENDING_REVIEW, recompute and replace the row (this is the only update path). Otherwise no-op.
3. Register nothing in routing/workflow decision paths — the prediction is advisory-only by construction (see epic).

Acceptance criteria

• DefaultApprovalPredictionServiceTest (Mockito, no Spring): disabled → skipped(DISABLED); no serving model → skipped(MODEL_NOT_SERVING); happy path persists probability + publishes event; lookup throwing → failed sentinel, no exception propagates.
• ApprovalPredictionListenerTest: ready-for-review triggers prediction; estimate-completed recomputes only in the estimate_missing + still-pending case (3 no-op branches tested).
• Integration test ApprovalPredictionIntegrationTest (Testcontainers postgres:18-alpine, model on BehaviorAnomalyDetectionIntegrationTest): seed ≥ 50 decided queries → train (call trainAll() directly) → submit a query to PENDING_REVIEW → assert an approval_predictions row with a probability exists.
• mvn test -Dtest=ApplicationModulesTest passes (ai → core via api only).