Keep one container of each model warm during the demo/judging window

The vLLM coach cold start is ~90s, which made the live demo feel broken on the
first question after idle. min_containers=1 holds one warm container per service
so judges never pay that. Revert to 0 after judging to stop idle GPU burn.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

modal_app.py

@@ -104,6 +104,9 @@ hf_cache = modal.Volume.from_name("coach-cockpit-cache", create_if_missing=True)
     volumes={"/cache": hf_cache},
     timeout=20 * MINUTES,
     scaledown_window=300,
+    # Keep one container warm so judges never hit a cold start during the demo
+    # window. Set back to 0 after judging to stop the idle GPU burn.
+    min_containers=1,
 )
 class Extractor:
     @modal.enter()
@@ -261,6 +264,9 @@ class Extractor:
     volumes={"/cache": hf_cache},
     timeout=20 * MINUTES,
     scaledown_window=300,
+    # vLLM cold start is ~90s — keep one container warm so the demo never eats it.
+    # Set back to 0 after judging to stop the idle GPU burn.
+    min_containers=1,
 )
 class Coach:
     @modal.enter()