ponder-lab / Hybridize-Functions-Refactoring / #2383

83%

push

github

Restore conservative call-site tainting in the numpy-flow scan (#746)

* Restore conservative call-site tainting in the numpy-flow scan

The return-taint precision refinement adopted in #743's review severed element-through-container flow: a list comprehension over a tainted parameter compiles to a synthetic callee returning a FRESH container whose elements are tainted but whose value is not, so the callee reported a clean return and the call site went untainted, letting NLPGNN's `TUDataset.cat` (`np.concatenate` over comprehension results) escape the precondition at corpus scale. Call-site definitions are now tainted conservatively whenever any argument is tainted, and the dead return-taint tracking is removed: for a safety precondition, over-tainting only declines an optimization while under-tainting ships a crasher, and value-level return-taint structurally cannot represent container-element flow. `testNumpyCallsThroughComprehension` pins the distilled `cat` shape.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

* Address review: cite #745 in the comprehension pin's Javadoc

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

---------

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

6 of 7 new or added lines in 1 file covered. (85.71%)

1913 of 2375 relevant lines covered (80.55%)

0.81 hits per line