openvax / topiary / 33415632931

93%

push

github

Add default_versions to resolve an ambiguous predictor version (#214) (#215)

* Add default_versions to resolve an ambiguous predictor version (#214)

default_methods answers "which model, when a kind has several". Nothing
answered "which version, when a model has several", so an unqualified
reference on a multi-version frame raised with no configured way through
and a consumer's default scoring expression could not run at all.

This gap was ours, and our own fix made it reachable. 5.28.2 added the
version-ambiguity raise so two versions could no longer resolve silently
to whichever row came first -- correct, but shipped without the way
through, in the same release series whose notes described the identical
problem one level up as "a working configuration turned into a hard
error". Keeping both versions of a multi-version LENS table (#208) is
what made a frame like that reachable in the first place.

default_versions is keyed on (kind, model) since a version means nothing
without its model, and is forwarded by apply_filter/apply_sort/
evaluate_scores like the other context options.
resolve_default_versions(df, prefer="newest") orders by PEP 440 -- the
only ordering where 4.10 beats 4.9 -- and sorts non-PEP-440 versions
before everything that parses, so "newest" means a real release rather
than whichever string sorted last.

packaging becomes a declared dependency. It was already there
transitively, but the ordering depends on it and catching ImportError
alongside InvalidVersion would have quietly demoted every version to
string order.

26 tests. Version 5.31.0.

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

* Treat an unknown predictor_version as no version, not as "nan"

Frames that do not record predictor_version are ordinary -- a reader with
no version column, or a source that reports one for some rows and not
others. The ambiguity check compared versions as strings, so a missing
value became a version called "nan": one r... (continued)

83 of 92 new or added lines in 2 files covered. (90.22%)

5724 of 6283 relevant lines covered (91.1%)

2.73 hits per line