Unexpected None behavior when working with computation ak.cartesian

[yimuchen]

Version of Awkward Array

2.6.6

Description and code to reproduce

When working with pair-wise computations with ak.cartesian, if a collection was reordered using a None-able array, the computation results can cause a un-used field to be computed as None.

In the case below, the array idx and either be None-able (signature 20 * var * ?int64) or not-None-able (signature 20 * var * int64) by toggling the fill_none line, even if the numerical value of the array is unchanged. The final output of the unused field col1.y will change depending on:

• Whether idx was a None-able array
• Whether we attempt to modify the collection with col1["r"].

import awkward as ak
import numpy as np

print(ak.__version__)

rng = np.random.default_rng(seed=1234)

n_events = 20

# Making the first collection with a hand ful of entries
n_col1 = rng.choice([2, 3], size=n_events)
col1 = ak.zip(
    {
        "x": rng.normal(size=ak.sum(n_col1)),
        "y": rng.normal(size=ak.sum(n_col1)),
        "z": rng.normal(size=ak.sum(n_col1)),
    },
)
col1 = ak.unflatten(col1, n_col1)

# Making second collection with few entries
n_col2 = rng.choice([0, 1], size=n_events)
col2 = ak.zip(
    {
        "x": rng.normal(size=ak.sum(n_col2)),
        "y": rng.normal(size=ak.sum(n_col2)),
        "z": rng.normal(size=ak.sum(n_col2)),
    }
)
col2 = ak.unflatten(col2, n_col2)

# Ordering the objects a None-able index array
idx = ak.pad_none(ak.local_index(col1), 4, axis=1)
idx = idx[ak.local_index(col1)]  # Making a non-able version of array
# idx = ak.fill_none(idx, 4 ) # Should crash if None actually exists
col1 = col1[idx]

print(ak.sum(col1.x != pre_x) == 0) # Showing that the array values are indeed identical be and after ordering
print(ak.to_list(col1.y), col1.y.__repr__)

# Making pair of pair computations
d = {"c1": col1, "c2": col2}
pairs = ak.cartesian(d, axis=1, nested=True)
diff_pairs = pairs.c1.x - pairs.c2.x

# Pushing results back into collection
col1["r"] = ak.min(diff_pairs, axis=2)

print(ak.to_list(col1.y), "\n", col1.y.__repr__())