bed_to_regions silently drops strand when polars Categorical

[bschilder]

Summary

_dataset/_utils.py::bed_to_regions checks bed.schema.get('strand', None) == pl.Utf8 before mapping +/- to 1/-1. When strand is a Categorical column (which is what gvl.write stores when the input BED has a repeated strand vocabulary — very common), the check returns False, and the fallback branch cols.append(pl.col('strand')) appends the raw string column without mapping.

.to_numpy() on the resulting polars frame then produces dtype=object (int columns + string strand). Downstream, the njit-compiled get_diffs_sparse and reconstruct_haplotypes_from_sparse fail with:

numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
non-precise type array(pyobject, 1d, A)

Reproduction

On GVL 0.21.4 + polars 1.40 + numba 0.65:

import polars as pl
import genvarloader as gvl

bed = pl.DataFrame({
    'chrom': ['chr17'] * 3,
    'chromStart': [100, 200, 300],
    'chromEnd': [150, 250, 350],
    'strand': pl.Series(['+', '+', '-'], dtype=pl.Categorical),
    'transcript_id': ['t1', 't1', 't2'],
    'exon_number': [1, 2, 1],
})
gvl.write('/tmp/test_gvl', bed, '<some_pgen>')
ds = gvl.Dataset.open('/tmp/test_gvl', reference='...').with_seqs('haplotypes')
ds[0, 0]  # -> numba TypingError

Fix

Two small changes in _dataset/_utils.py::bed_to_regions:

-    if bed.schema.get('strand', None) == pl.Utf8:
+    if bed.schema.get('strand', None) in (pl.Utf8, pl.String, pl.Categorical, pl.Enum):
         cols.append(
-            pl.col('strand').replace_strict({'+': 1, '-': -1}, return_dtype=pl.Int32)
+            pl.col('strand').cast(pl.String).replace_strict({'+': 1, '-': -1}, return_dtype=pl.Int32)
         )

(.cast(pl.String) is a no-op on String columns and unwraps Categorical/Enum cleanly.)

Verified locally: ds._full_regions.dtype goes from object to int32, and ds[i, j] returns the expected Ragged without tripping numba.

Happy to open a PR if helpful.

[d-laub]

Fixed by #156