sweep runner: per-config isolation, counted trials = evaluated configs

sweep.py

@@ -88,12 +88,20 @@ def main() -> int:
 
     # --- full-sample net returns per config -----------------------------
     returns: dict[str, np.ndarray] = {}
+    failures: dict[str, str] = {}
     for i, (name, (fn, kw)) in enumerate(grid.items()):
-        pos = fn(bars, funding, **kw)
-        bt = run_backtest(bars, pos, funding)
-        returns[name] = bt["net_return"].to_numpy()
+        try:
+            pos = fn(bars, funding, **kw)
+            bt = run_backtest(bars, pos, funding)
+            returns[name] = bt["net_return"].to_numpy()
+        except Exception as e:  # one bad config must not discard the sweep
+            failures[name] = repr(e)
+            print(f"[sweep] CONFIG FAILED {name}: {e}", flush=True)
         if (i + 1) % 25 == 0:
             print(f"[sweep] {i+1}/{n_trials} backtests done", flush=True)
+    if failures:
+        print(f"[sweep] {len(failures)} configs failed", flush=True)
+    n_trials = len(returns)  # counted trials = configs actually evaluated
 
     sr = {n: float(np.nanmean(r) / (np.nanstd(r, ddof=1) + 1e-12))
           for n, r in returns.items()}
@@ -159,6 +167,8 @@ def main() -> int:
         "data": "Avi-L/crypto-quant-eval-data",
         "n_bars": len(bars),
         "n_trials": n_trials,
+        "n_configs_failed": len(failures),
+        "config_failures": failures,
         "pbo_splits": args.pbo_splits,
         "grid": trials,
         "best_config": best,