Pandas join with categorical indices

I'm getting unexpected results with attempting to join Pandas DataFrame objects with categorical indices. Here is the minimum reproducible example (boiled down from my real-life use case):

shape_categories=['square', 'circle']
color_categories=['red', 'blue', 'green']

test_a = pd.DataFrame({
    'shape': pd.Categorical(['square', 'circle'], categories=shape_categories, ordered=True),
    'color': pd.Categorical(['red', 'blue'], categories=color_categories, ordered=True),
    'value_a': [1.0, 2.0]
})
test_a.set_index(['shape', 'color'], inplace=True)

test_b = pd.DataFrame({
    'shape': pd.Categorical(['square', 'square', 'circle', 'circle'], categories=shape_categories, ordered=True),
    'color': pd.Categorical(['red', 'blue', 'red', 'blue'], categories=color_categories, ordered=True),
    'value_b': [10.0, np.nan, np.nan, 40.0]
})
test_b.set_index(['shape', 'color'], inplace=True)

test_a.join(test_b, how='left')

I expect

but instead I get

What am I missing? I've tried to be careful to keep the dtypes of the categorical variables exactly the same.