fix(predictions): circuit breaker resilience + graceful degradation

- Reset consecutiveFailures on cooldown expiry (half-open state)
  so a single retry failure doesn't immediately re-open the circuit
- Exclude AI Engine app-level 500s from circuit breaker count
  (only network/infra errors: timeout, 502, 503, 504, 429)
- Return null gracefully instead of throwing 503 when no cache exists
- Add DB fallback for non-cooldown AI Engine failures
- Remove blocking wait-and-retry that held requests for up to 20s

src/common/utils/ai-engine-client.ts

@@ -183,8 +183,11 @@ export class AiEngineClient {
     }
 
     this.logger.warn(
-      `[${this.serviceName}] AI circuit breaker cooldown elapsed, allowing a recovery attempt`,
+      `[${this.serviceName}] AI circuit breaker cooldown elapsed, allowing a recovery attempt (resetting failures from ${this.consecutiveFailures})`,
     );
+    // Half-open state: reset failures so a single retry failure doesn't
+    // immediately re-open the circuit at threshold+1
+    this.consecutiveFailures = 0;
     this.circuitOpenedAt = null;
   }
 
@@ -233,8 +236,18 @@ export class AiEngineClient {
     if (!error.response) {
       return true; // Network error, timeout, etc.
     }
+    // Only count infrastructure-level errors toward circuit breaker:
+    // - No response (network failure) → already handled above
+    // - Timeout (ECONNABORTED) → infrastructure
+    // - 429 (rate limit) → infrastructure
+    // - 502/503/504 (proxy/gateway errors) → infrastructure
+    // Do NOT count 500 (app-level crash in AI Engine) — it may be
+    // match-specific and shouldn't block all other matches.
+    if (error.code === 'ECONNABORTED') {
+      return true;
+    }
     const status = error.response.status;
-    return status >= 500 || status === 429;
+    return status === 429 || status === 502 || status === 503 || status === 504;
   }
 
   private toRequestError(error: unknown): AiEngineRequestError {