promote hybrid: 8B map + 30B read + split build (VRAM) + FLUX preload + CUDA cache

arcana/llm.py

@@ -150,9 +150,10 @@ class LocalLLM:
                  timeout: float | None = None) -> str:
         from .gpu_models import llama_map, llama_read
         if json_mode:
-            # structured mapping → small dense model (reliable JSON, fast, low VRAM)
-            return llama_map(system, user, max_tokens=10000, temperature=0.4)
-        # prose readings → reading model (the 30B where configured → richer prose)
+            # mapping → 30B, think-suppressed (fast); generous tokens so the full
+            # 22-card JSON never truncates
+            return llama_map(system, user, max_tokens=12000, temperature=0.5)
+        # prose readings → 30B (think-suppressed for clean, fast output)
         return llama_read(system, user, max_tokens=self.max_tokens,
                           temperature=self.temperature)