fix: add 30s cooldown on 400 errors to prevent retry loop

Since findAvailableSlot always starts from slot 0, a 400 error
without cooldown would keep hitting the same Mistral slot forever.
Now it gets a 30s cooldown so the pool moves to the next Mistral
model, then Groq if all Mistral models are on cooldown.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/server/src/providers/pool.ts

@@ -100,14 +100,16 @@ export class AIProviderPool {
         }
 
         if (error.status === 400) {
-          // 400 errors: invalid model, json_validate_failed, bad request — disable slot and try next
+          // 400 errors: invalid model, json_validate_failed, bad request
           const bodyStr = error.body ?? error.message ?? '';
           const isInvalidModel = bodyStr.includes('invalid_model') || bodyStr.includes('Invalid model');
           if (isInvalidModel) {
             slot.disabled = true;
             logger.warn(`Invalid model on ${slot.provider.name}, permanently disabled`);
           } else {
-            logger.warn(`400 error on ${slot.provider.name}, trying next provider`);
+            // Cooldown to prevent retry loop (findAvailableSlot starts from 0)
+            slot.cooldownUntil = Date.now() + 30_000;
+            logger.warn(`400 error on ${slot.provider.name}, 30s cooldown. Trying next...`);
           }
           this.currentIndex = (this.currentIndex + 1) % this.slots.length;
           continue;