Prewarm GraphCast at FastAPI startup to avoid cold-start fallback

Mirrors the MI Chronos prewarm pattern. load_model() was called lazily
on the first _step_predict, which on a freshly-rebuilt HF Space took
~30-120s (downloading the 1.4GB checkpoint from GCS plus normalization
stats). That puts the pipeline's first weekly trigger at risk of the
same silent-fallback path that bit MI before PR #5.

Fix: spawn a daemon thread in the lifespan context that calls
load_model() at startup. JAX compile + GCS download happen in the
background while FastAPI comes up; by the time someone hits
/api/pipeline/trigger the model is resident. Failures log and move on.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/api.py

@@ -38,6 +38,21 @@ logger = logging.getLogger(__name__)
 _db_conn = None
 
 
+def _prewarm_graphcast() -> None:
+    """Load GraphCast model into memory at startup so the first pipeline
+    trigger doesn't pay the ~30-120s download/init cost. Runs in a
+    background thread — failures are logged, not fatal.
+    """
+    try:
+        from src.prediction.graphcast_inference import load_model
+        import time as _time
+        t0 = _time.time()
+        load_model()
+        logger.info("[PREWARM] GraphCast loaded at startup (%.1fs)", _time.time() - t0)
+    except Exception as exc:
+        logger.warning("[PREWARM] GraphCast prewarm threw: %s", exc)
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     global _db_conn
@@ -66,6 +81,12 @@ async def lifespan(app: FastAPI):
     except Exception as e:
         logger.warning("DB init failed (non-fatal): %s", e)
         _db_conn = None
+
+    prewarm_thread = threading.Thread(
+        target=_prewarm_graphcast, daemon=True, name="graphcast-prewarm",
+    )
+    prewarm_thread.start()
+
     scheduler = _start_scheduler()
     yield
     if scheduler: