[antonsoo]

map_to_global documents queries in the first table and exon annotations in the *2 arrays, but passes them to an exon-first dispatcher in the opposite order. One query spanning two exons raises a length error. With equal row counts, the call can silently return wrong genomic coordinates: query [5, 10) against exon local [0, 10) / genomic [100, 110) returns [95, 105) instead of [105, 110).

Pass the annotation table first and retain the query dtype when restoring genomic coordinates. Clarify the existing sort requirement. The native kernel and other operations are unchanged.

Validation: all 51 tests pass on Python 3.12, 3.13 and 3.14; the 15 new cases fail before the correction. An independent per-base reference on public GTF/BED exon data checks query attribution and both strands, including equal and unequal table sizes. It reproduces 12 exceptions and 12 incorrect-coordinate configurations, while 12 complete identity controls remain exact. A clean ABI3 wheel repeats these results on 3.12/3.14, and 16 public PyRanges1 operation results per profile remain unchanged.

[endrebak]

Thanks for the pr :)

@marco-mariotti You are the expert here :)

[endrebak]

This seems like a major bug that is easy to understand.

I see why we did not catch it:

We used the temporary pandas solution in pyranges1 still so this code was never called.