ponder-lab / Hybridize-Functions-Refactoring / 26981614816

78%

push

github

Auto-inject `TensorSpec` and dtype constants for signature emission (#574) (#589)

When `convertToHybrid` finds no TensorFlow import, it auto-injects one and
proceeds. Previously it injected a bare `from tensorflow import function`,
leaving `TensorSpec` and the dtype constants out of scope, so input-signature
emission was skipped.

Extend the auto-inject to bring the signature's required symbols into scope:
`from tensorflow import function, TensorSpec, <dtypes>`. Track the injected name
set per file (replacing the `Set<File>` guard with a `Map<File, Set<String>>`),
so a later function in the same file emits when its dtypes are covered and is
safely skipped by the #585 gate otherwise, never a `NameError`-raising
decorator. A file-level union pre-pass to cover divergent dtypes across
functions in one file is left as a follow-up.

Also fix the emission-test helper to apply edits highest-offset-first, so an
injected import (low offset) does not shift the unapplied decorator edit's
anchor; in production LTK coordinates offsets in one change tree.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

15 of 15 new or added lines in 1 file covered. (100.0%)

1329 of 1714 relevant lines covered (77.54%)

0.78 hits per line