fix: adjust batch size to 50 and enforce strict Lite-only policy

- Reduced BATCH_SIZE_FAST to 50 to prevent AI request timeouts.
- Hardened call_gemini_with_retry to strictly use Flash/Lite models when prefer_flash=True.
- Ensured newest model versions are prioritized within the Flash tier.

src/gemini_utils.py

@@ -293,9 +293,13 @@ async def call_gemini_with_retry(prompt: str, response_format: str = "json", max
     consecutive_429s = 0
     base_wait_time = 2.0
     
-    # 1. Smart Re-ordering
+    # 1. Smart Filtering and Re-ordering (Strict Lite-Only Policy for Tier 1)
     if prefer_flash:
-        models = sorted(models_pool, key=lambda m: 0 if "flash" in m or "lite" in m else 1)
+        # Strict filter: Only allow flash/lite models, maintaining version sorting from pool
+        models = [m for m in models_pool if "flash" in m or "lite" in m]
+        if not models:
+            # Emergency fallback: if no flash/lite discovered, use hardcoded safe defaults
+            models = ["gemini-1.5-flash", "gemini-1.5-flash-latest"]
     else:
         models = models_pool
 

src/v2_optimizer.py

@@ -319,7 +319,7 @@ class V2VisionEngine:
             analyst_results = []
 
             # 1.1 Fast-Track: Large Batches, NO GROUNDING (Fast)
-            BATCH_SIZE_FAST = 100 # Increased from 40 for optimal RPM/TPM balance
+            BATCH_SIZE_FAST = 50 # Balanced "Sweet Spot" for RPM/TPM and timeout safety (2026)
             total_fast = len(fast_track)
             for i in range(0, total_fast, BATCH_SIZE_FAST):
                 batch = fast_track[i:i+BATCH_SIZE_FAST]