openvax / topiary / 33428957974

93%

push

github

Stop treating a missing value as the text "nan" (#223) (#225)

Two more instances of the defect behind #214, #216 and #219: two code
paths disagreeing about what "nothing" means.

Version selection compared stringified values, so a row whose
predictor_version is missing could be addressed as version "nan" --
returning it. 5.31.0 had already taught the ambiguity check and
resolve_default_versions that a missing version names no version, so the
two actively disagreed; fixing only half of it is what left this. It was
also self-inconsistent, since the "no such version" error lists available
versions from dropna() and so never offered "nan" while accepting it.
Selection now uses the same test as everything else and, like the
resolver, tolerates surrounding whitespace.

read_pvacseq's coordinate fallback concatenated stringified columns, so a
blank field became text inside the identifier: chr1-154590262-nan-A. That
looks well-formed, is wrong, and is stable, so every row sharing the gap
grouped under one fabricated id with nothing said. Those rows now get a
null variant and a warning naming the count and the incomplete columns.
An absent identifier is honest; a fabricated one cannot be told from a
real one downstream. Complete coordinates, and files carrying an Index
column, are unchanged and silent.

17 tests, parameterized over the five spellings of a missing version and
the four coordinate components.

Version 5.34.0.

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

23 of 23 new or added lines in 3 files covered. (100.0%)

5846 of 6413 relevant lines covered (91.16%)

2.73 hits per line