openvax / topiary / 33439414783

91%

master: 93%

Pull #228

github

Order the cache guards so each owns its case

A cache spanning two models also has two rows per key, because
prediction_method_name is not part of the cache key. My duplicate check
ran inside _normalize, before _unique_version_pair, so a multi-model
frame got "duplicate key" instead of the precise "this cache spans two
models" that already existed for it.

Not a functional regression -- both reject the frame -- but the wrong
guard was answering, and error quality is the whole point of a guard.
The duplicate check now runs after the single-model invariant.

Found by asking what my own change could break rather than trusting a
green suite; no test covered a multi-model cache, so nothing failed.

2 more tests. Version 5.36.0.

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

17 of 18 new or added lines in 2 files covered. (94.44%)

5901 of 6467 relevant lines covered (91.25%)

2.74 hits per line