ponder-lab / Hybridize-Functions-Refactoring / #2304

80%

push

github

Walk every node of a callee reference in the side-effect closure (#724)

`allCreationsWithinClosureInteral2` expanded each callee `MethodReference` through only the first of its call-graph nodes, assuming the calling context makes no difference. A shared synthetic summary breaks that assumption: `tensorflow.data.map` has one node per callback context, each with a different callee, so the walk descended into whichever callback the arbitrary first node carried. A creation on the unreached branch then failed the within-closure proof, and op-modeling allocations (result tensors/tuples, `read_data` stores) read as external writes—`HAS_PYTHON_SIDE_EFFECTS` misfired on pure-tensor code (`Streaming.call` and `FactorizedTopK.update_state` in deep_recommenders, both P1 on 0.52.13).

The walk now unions the successors of all of a reference's nodes. The closure stays reference-granular (and hence context-blind, as before); the union can over-widen through a shared summary's foreign contexts, which errs toward filtering—acceptable against a mod set that is itself context-sensitive.

`testPythonSideEffects68` pins the scenario: two nested callbacks piped through `dataset.map(...)`, the second performing the tensor work; red without the fix.

Closes #722.


Claude-Session: https://claude.ai/code/session_01S1U9TQRo3Wmhfu86DRcBNW

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

6 of 6 new or added lines in 1 file covered. (100.0%)

1689 of 2124 relevant lines covered (79.52%)

0.8 hits per line