Bug: ADX division by zero when di_plus + di_minus == 0 — inf propagates through ewm

[bradsmithmba]

Summary

The ADX formula divides by di_plus + di_minus with no guard. At the start of a series before directional movement accumulates, both values are zero, producing inf. The subsequent ewm call propagates inf through the entire ADX series. fillna(0) does not remove inf, so it enters the feature tensor.

Evidence

src/features/regime_features.py line 128:

dx = 100 * np.abs(di_plus - di_minus) / (di_plus + di_minus)

No zero-division guard. di_plus and di_minus are both zero on the first period bars.

Fix

denom = di_plus + di_minus
dx = np.where(denom > 0, 100 * np.abs(di_plus - di_minus) / denom, 0.0)

Then proceed with pd.Series(dx).ewm(...).