openvax / topiary / 33433872283

93%

push

github

Make fragment_from_effect and the version rule public (#226)

* Make fragment_from_effect and the version rule public

Standing preference recorded in AGENTS.md: logic that does real work
belongs in one documented public function, not a private helper. The
failure mode is concrete and has happened repeatedly across these repos
-- a consumer needs behavior it cannot import, reimplements it, and the
copy drifts.

fragment_from_effect (was _fragment_from_effect) builds a ProteinFragment
from a varcode effect: the varcode arm of the multi-source fragment
story, and the thing a caller doing its own variant annotation would
otherwise rewrite. Its docstring now states the two non-obvious rules
rather than leaving them to be rediscovered -- the window clips at the
first stop codon, and reference_sequence is populated only when the pre-
and post-mutation proteins align 1:1, since slicing the same offsets out
of a frameshifted protein would present a different piece of protein as
the comparator.

is_named_version is the sharper case, because AGENTS.md's own example is
this rule: "if str(v).strip()" admits the literal string "nan" and so
treats a row with no version as a version named "nan". That shipped in
topiary and, independently, in a consumer that had reimplemented it.
Leaving it private while writing the rule down would have been
hypocritical. The vectorized internal form is documented as the same rule
and a test asserts the two agree, so there is one rule with two shapes
rather than two rules.

23 tests. Version 5.35.0.

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

* Address review: one version rule, and fragment windows that hold

The review found the previous commit contradicting its own AGENTS.md
section in three places, plus real defects in the newly-public function.

The rule was not centralized. wide.py's _version_str, io.py's
_model_version_str and cached.py's identity check each had their own
copy with different covera... (continued)

62 of 65 new or added lines in 9 files covered. (95.38%)

5880 of 6446 relevant lines covered (91.22%)

2.74 hits per line