Inconsistent behavior between dowhy wrapper and wrapped model

For some reason the trained T and Y models exposed by the dowhy wrapper do not accept input of the correct shape, seeming to expect twice as many X columns as they should (e.g. with 3 X columns and 5 W columns, the error message will complain that 8 columns were supplied but 11 were expected):

model_dml = LinearDML(model_y=LinearRegression(),
                      model_t=LinearRegression())

d_x = 2
d_w = 5

N_SHAPE = 1000
Y = np.random.normal(size=N_SHAPE)
T = np.random.randint(2, size=N_SHAPE)
X = np.random.normal(size=(N_SHAPE, d_x))
W = np.random.randint(2, size=(N_SHAPE, d_w))

model_dml = model_dml.fit(Y=Y, T=T, X=X, W=W, inference='auto', cache_values=False)
model_dml_dowhy = model_dml.dowhy.fit(Y=Y, T=T, X=X, W=W, inference='auto', cache_values=False)

# works
model_dml.models_y[0][0].predict(np.concatenate((X, W), axis=1))

# This will error!
model_dml_dowhy.models_y[0][0].predict(np.concatenate((X, W), axis=1))

Start by reproducing the provided LinearDML example, comparing predictions from models_y before and after accessing the dowhy wrapper. Trace the dowhy.fit path and the models_y predict entry point to identify why the wrapped model expects extra columns; done means correctly shaped X and W input behaves consistently in both cases.