Fix: 8192 ctx window + trim safety context to 3000 chars

modal_inference.py

@@ -65,7 +65,7 @@ class LlamaCppServer:
             [
                 "llama-server",
                 "--model",        MODEL_PATH,
-                "--ctx-size",     "4096",
+                "--ctx-size",     "8192",
                 "--n-gpu-layers", "99",
                 "--port",         "8080",
                 "--host",         "127.0.0.1",