ponder-lab / Hybridize-Functions-Refactoring / #2599

84%

push

github

Add an invalid-name-argument safety precondition (#815)

* Add an invalid-name-argument safety precondition

A function that passes a non-string constant where a TensorFlow API declares its `name` parameter (e.g., `tf.sqrt(x, tf.float32)` in word2vec's `evaluate`) runs eagerly but raises under `tf.function` tracing, which opens a name scope with the value and validates it. Such functions now fail with `HAS_INVALID_NAME_ARGUMENTS` (code 17), joining the safety family after `HAS_NUMPY_CALLS_ON_PARAMETERS` and before the barren benefit signal.

Per the scope note on #814, the check is the narrow syntactic rule: a keyword `name=` argument passing a non-string constant fails for any TensorFlow-rooted callee, and a positional argument is recognized for the ops whose TensorFlow 2.9.3 signature is exactly `(x, name=None)`, generated from the `tf`, `tf.math`, `tf.linalg`, and `tf.nn` namespaces with `inspect.signature`. Flagged constants are numeric literals, boolean literals, and TensorFlow dtype constants; `None` (the declared default) and unknown values are conservatively accepted. Because the scan needs no call graph and costs almost nothing, it runs unconditionally rather than sharing the `barrenCouldDecide` gate and `alwaysCheck*` knob pattern of its sibling safety checks; the evaluator reports the verdict in a new `invalid name arguments` column of `functions.csv`.

`testInvalidNameArgumentsBlockHybridization` pins the positional form (distilling word2vec's `evaluate`) and the keyword form, with a string-name P1 control on the same op.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GcqABsog1rtb4m2CDZdcRD

* Degrade to undetermined on a traversal failure

An AST traversal failure in the syntactic scan threw, killing the whole project's refactoring run. Mirror the sibling safety checks instead: log and leave the result null, so the precondition neither blocks nor passes on it.

Co-Authored-By: Clau... (continued)

47 of 49 new or added lines in 3 files covered. (95.92%)

2444 of 2940 relevant lines covered (83.13%)

0.83 hits per line