fix(warehouse): do NOT create the live-mark matview on store init (crashlooped the cockpit)

ensure_latest_matview() in _create_schema ran on EVERY cockpit pod startup (the store is built at app
startup), and CREATE MATERIALIZED VIEW over the ~4M-row bybit_features hypertable blocked uvicorn from
binding 8080 -> readiness/liveness failed -> crashloop -> cockpit down (and image is :latest so a rollout
undo did not help). The matview is now created/refreshed ONLY by the nightly ingest (refresh_latest);
read_latest falls back to the LATERAL until it exists, so the cockpit boots instantly and is never wrong.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/fxhnt/adapters/warehouse/timescale_feature_store.py

@@ -92,9 +92,11 @@ class TimescaleFeatureStore:
         # Hypertable + compression: best-effort perf layer, guarded so it no-ops without timescaledb.
         if self._is_pg:
             self._try_make_hypertable()
-        # Live-mark matview: idempotently create it at setup so the cockpit's read_latest fast path is live as
-        # soon as the bybit_features store exists (no separate migration step). No-op off-PG/off-bybit.
-        self.ensure_latest_matview()
+        # NB: the live-mark matview is NOT created here — CREATE MATERIALIZED VIEW over the ~4M-row hypertable
+        # is far too heavy for the store-init path, which runs on EVERY cockpit pod startup (it blocked uvicorn
+        # from binding → readiness/liveness failed → crashloop). The matview is created + refreshed ONLY by the
+        # nightly ingest (refresh_latest → ensure_latest_matview); read_latest falls back to the LATERAL until
+        # then, so the cockpit boots instantly and is never wrong.
 
     # -- live-mark matview (PG + bybit_features only; perf layer, never affects correctness) ---------
     def ensure_latest_matview(self) -> None: