Raise warm-call timeouts (text 8->20s, FLUX 12->30s) so the hosted Space gets real AI + comics instead of cold-start fallbacks

game/comic.py

@@ -15,7 +15,9 @@ import requests
 
 from .trace import trace
 
-FLUX_TIMEOUT = float(os.environ.get("BDS_FLUX_TIMEOUT", "12"))
+# 30s warm default: on a hosted Space a slightly-cold / loaded FLUX container
+# needs more than 12s, else panels silently no-op. Warm renders are ~3-6s.
+FLUX_TIMEOUT = float(os.environ.get("BDS_FLUX_TIMEOUT", "30"))
 FLUX_COLD_TIMEOUT = float(os.environ.get("BDS_FLUX_COLD_TIMEOUT", "90"))
 _warmed = False
 

game/llm.py

@@ -20,8 +20,10 @@ from .schemas import SCHEMA_BY_CALL_TYPE
 from .state import GameState
 from .trace import trace, trace_payload
 
-# 8s in production (TECHNICAL.md); raise via env for slow local CPU inference
-TIMEOUT_S = float(os.environ.get("BDS_LLM_TIMEOUT", "8"))
+# warm-call ceiling. 20s default so a hosted Space (cold/loaded Modal containers,
+# concurrent players) gets real AI instead of constant fallbacks; lower via env
+# (run_modal.ps1 sets 20 explicitly) — 9B crisis JSON warm-generates in ~8-12s.
+TIMEOUT_S = float(os.environ.get("BDS_LLM_TIMEOUT", "20"))
 # the FIRST call to a cold Modal container waits longer (boot + model load)
 COLD_TIMEOUT = float(os.environ.get("BDS_COLD_TIMEOUT", "55"))
 _warmed = False  # flips true after the first successful live response