ponder-lab / Hybridize-Functions-Refactoring / #2812

85%

push

github

Treat Keras-supplied call parameters as supplied at every call site (#885)

A call override declaring training or mask with a default receives it
from Keras's Layer.__call__ on every invocation, from outside the
analyzed program, so isSuppliedAtCallSite answers FALSE for a supplier
no call-site examination can see: the parameter was filed as omittable,
the emitted all-or-nothing input signature covered only the tensor
parameter, and the first call raised (#881, observed on
TensorFlow2.0-Examples' RPN model, with the same shape in MTCNN, FCN,
and FPN).

The two framework names on a call override of a Keras model or layer
subclass now route to the existing DEFAULTED_PARAMETER_SUPPLIED block:
conservative, and the same outcome the tool already produces when it can
see the supplying call site. Subclass recognition reuses the PyDev class
hierarchy resolution behind hasTensorContext() (AST-base fallback,
last-segment match, over-recognition only withholds an unwritable
signature). The gate is local to computeInputSignature rather than
folded into isSuppliedAtCallSite, whose other consumer (the #795
star-arg exemption) must keep treating a source-unsupplied parameter as
bound to its default.

The vendored RPN reduction pins the flip two-sided: pre-fix a signature
covering only x is emitted and the absence assertion fails; post-fix the
signature is withheld, the function stays P1, and the class runs
627/627, so no existing emission pin moved.

Fixes #881.


Claude-Session: https://claude.ai/code/session_01X8W9PDNUSQFzZRUAYJmRiE

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

10 of 10 new or added lines in 1 file covered. (100.0%)

3204 of 3809 relevant lines covered (84.12%)

0.84 hits per line