openvax / topiary / 33418641910

93%

push

github

Treat an empty bucket as empty, and refuse a colliding wide column (#218)

Two silent-failure bugs found while auditing for the pattern the last few
releases kept producing.

#216: SelfProteome.nearest crashed with "attempt to get argmin of an empty
sequence" when the proteome had a bucket for a peptide length but nothing
in it -- reachable from from_fasta and from_peptides whenever the sources
are shorter than the window. _build_index made an entry for every requested
length regardless, and nearest guarded on key presence rather than
emptiness, so the graceful path was skipped and argmin ran over a
zero-width axis. The same question already had a right answer by another
route: a query length with no bucket returns a clean empty row. Absent and
empty mean one thing here and now give one answer. Unfillable lengths are
no longer recorded either, so peptide_lengths stops advertising coverage
the proteome lacks.

#217: to_wide merged the passthrough annotation columns against the
generated prediction columns with pandas' default suffixes, so a name on
both sides silently became _x / _y. The canonical name then did not exist,
from_wide found no value for it, and a prediction that went in came back
NaN. That is the same silent round-trip loss as #208 and #211, one layer up
at the point where the names are assembled, and _x / _y were merge
artifacts leaking into a schema that never documented them. Now refused,
naming the column. The realistic read_lens -> to_long -> to_wide path was
never affected, since to_long consumes the prediction columns rather than
leaving them as annotations.

11 tests. Version 5.31.1.

Claude-Session: https://claude.ai/code/session_01BAUqjqDDPXaTdtUDJqapJG

9 of 9 new or added lines in 3 files covered. (100.0%)

5731 of 6290 relevant lines covered (91.11%)

2.73 hits per line