ponder-lab / Hybridize-Functions-Refactoring / #2464

83%

push

github

Extract the numpy parameter-flow scan from `Util` into a dedicated analysis class (#763)

The subsystem had grown to roughly half of `Util` across #743, #746, #750, #752, #756, and #762, with its own domain records, sentinel, memoization, and interprocedural transfer functions, and only two external calls from `Function.computeNumpyCallsOnParameters`. It now lives in `NumpyParameterFlowAnalysis`, instantiated per function check with the call graph, pointer analysis, and tensor-type index as instance state, which eliminates the eight-parameter threading through `scanForTaintedNumpySinks`. The scan memo is also instance state, so it is shared across the checked function's call-graph nodes instead of being rebuilt per node, a strict caching improvement with unchanged results. The genuinely shared helpers (`resolveStringConstant`, `resolveCalleeFullyQualifiedName`, `isTensorFlowNode`, `pointsToType`) stay in `Util`, widened to package-private.

Behavior-preserving; the numpy-precondition pins cover the move.

Closes #759.


Claude-Session: https://claude.ai/code/session_01M4XLK81HCB6Vp1WQiRPafR

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

291 of 309 new or added lines in 2 files covered. (94.17%)

2135 of 2609 relevant lines covered (81.83%)

0.82 hits per line