fix: study.db migration fallback + degenerate HMM training guard

- optimizer.get_best_params: if study_<broker>.db not found, fall back to
  legacy models/study.db so optimization runs before the per-broker split
  are not silently discarded (prevents 'using defaults' fallback)
- main.cmd_train: abort and sys.exit(1) before saving if HMM min state
  persistence < 0.70 — prevents deploying a garbage degenerate model when
  Optuna params were not loaded; prints clear remediation steps

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

main.py

@@ -172,6 +172,25 @@ def cmd_train(args):
         params = {}
 
     result, model_hmm, state_map, models_ensemble, thresholds, metrics, *_ = _train_for_tf(tf, balance, broker, params)
+
+    # Guard: refuse to save a degenerate HMM — identical state means + near-zero persistence
+    # indicate the Kalman/HMM params were bad (usually because Optuna params weren't loaded).
+    _min_persist = min(model_hmm.transmat_[i, i] for i in range(model_hmm.n_components))
+    if _min_persist < 0.70:
+        logger.critical(
+            "TRAINING ABORTED: degenerate HMM (min persistence=%.4f). "
+            "Model NOT saved. Usually caused by missing Optuna study for tf=%s broker=%s. "
+            "Ensure --mode optimize has completed and re-run --mode train.",
+            _min_persist, tf, broker,
+        )
+        print(
+            f"\nERROR: Degenerate HMM detected (min state persistence={_min_persist:.4f} < 0.70).\n"
+            f"The model was NOT saved — saving it would only produce garbage signals.\n"
+            f"\nFix: python main.py --mode optimize --tf {tf} --broker {broker} --trials 300\n"
+            f"Then: python main.py --mode train    --tf {tf} --broker {broker}\n"
+        )
+        sys.exit(1)
+
     save_hmm(model_hmm, hmm_model_path(tf, broker))
     save_xgb_ensemble(models_ensemble, thresholds, metrics, get_ensemble_path(tf, broker))
 

src/optimizer.py

@@ -454,18 +454,39 @@ def get_best_params(
     tier = _get_tier(balance)
     name = _study_name(broker=broker, tier=tier, tf=tf)
     storage = _study_db(broker)
-    try:
-        study = optuna.load_study(study_name=name, storage=storage)
-        return study.best_params
-    except Exception:
-        if tier == "small":
-            raise   # small tier is the baseline — nothing to fall back to
-        # Growth (or future tiers) — reuse the small-tier study params so live
-        # trading always benefits from Optuna thresholds.
-        fallback = _study_name(broker=broker, tier="small", tf=tf)
-        logger.info(
-            "Growth study '%s' not found — using small-tier params from '%s'.",
-            name, fallback,
-        )
-        study = optuna.load_study(study_name=fallback, storage=storage)
-        return study.best_params
+
+    def _try_load(db: str) -> dict | None:
+        try:
+            study = optuna.load_study(study_name=name, storage=db)
+            return study.best_params
+        except Exception:
+            return None
+
+    params = _try_load(storage)
+    if params is not None:
+        return params
+
+    # Legacy fallback: studies optimised before per-broker DB split were saved
+    # to models/study.db.  Try it so users don't lose their optimisation work.
+    _legacy = "sqlite:///models/study.db"
+    if storage != _legacy and os.path.exists("models/study.db"):
+        params = _try_load(_legacy)
+        if params is not None:
+            logger.info(
+                "Loaded params for study '%s' from legacy models/study.db. "
+                "Re-run --mode optimize to persist to study_%s.db.",
+                name, broker,
+            )
+            return params
+
+    if tier == "small":
+        raise KeyError(f"No Optuna study '{name}' found in {storage} or study.db")
+    # Growth (or future tiers) — reuse the small-tier study params so live
+    # trading always benefits from Optuna thresholds.
+    fallback = _study_name(broker=broker, tier="small", tf=tf)
+    logger.info(
+        "Growth study '%s' not found — using small-tier params from '%s'.",
+        name, fallback,
+    )
+    study = optuna.load_study(study_name=fallback, storage=storage)
+    return study.best_params