ponder-lab / Hybridize-Functions-Refactoring / #3359

88%

push

github

Read through only trampolines when resolving a training-surface call (#1041)

* Read through only trampolines when resolving a training-surface call

`dispatchTargets` replaced every target that was neither program code
nor in the TensorFlow namespace with what it calls, though its Javadoc
said each such target was a trampoline. A builtin passes the same
test. When a Keras model's `predict` was replaced with `len`, the
builtin was read through to nothing, the call fell back to its
receiver, and the TensorFlow model there blocked a call that runs
under `tf.function`.

Only targets the engine names as trampolines are read through now.
Any other target is kept, and since it is not the endpoint it does not
block. A trampoline that forwards nowhere still contributes no target,
so the receiver decides, as it does for a call site with no targets.

Fixes #1039.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Say where the builtin-replaced endpoint's pin moves

From Ariadne 0.52.105 the call also targets the model's own predict
trampoline, so calls_replaced blocks there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

2 of 2 new or added lines in 1 file covered. (100.0%)

4798 of 5455 relevant lines covered (87.96%)

0.88 hits per line