ponder-lab / Hybridize-Functions-Refactoring / #2850

85%

main: 85%

Pull #892

github

Pin the dict-parameter and reconfigure populations #890 left blind

Neither population had a fixture, which is why widening the container gate
to every container and letting a container verdict outrank a parameter's
own typing could leave a green suite. Both are pinned as properties rather
than as recorded defects: the current answers are the ones wanted.

`score` takes a dict of tensors, the multi-input Keras shape and the form
deep_recommenders' estimator rankers take as `call(self, features)`. It
classifies through the container check, the nested reduction models only
positionally-indexed sequences, so the form is reported unsupported and
the function hybridizes bare at P1. Worth being exact about what #890
would have moved here: a dict-only parameter already reaches the container
check, since the dict itself carries no tensor type, so only a dict
arriving beside a tensor-typed value would have been rerouted. The pin
covers the base case.

`train_step` reduces TensorFlow2.0-Examples' CNN.py: already hybrid with a
bare decorator, two tensor parameters, no Python literal, and the variable
collection read after the forward pass. Adding the inferred signature is
its own transformation there, RECONFIGURE at P4, which is what makes the
population different from the conversion path, since a withheld signature
costs the transformation rather than only the specification. Nothing
pinned a live P4 before, which is exactly the silent loss #890 risked.

Both files execute under the pinned TF 2.9.3. The suite runs 689/689.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R1emLt6bvQa993huMzTaKu

89 of 97 new or added lines in 5 files covered. (91.75%)

3400 of 4014 relevant lines covered (84.7%)

0.85 hits per line