ponder-lab / Hybridize-Functions-Refactoring / #2392

81%

main: 83%

Build Type

Pull #750

github

Committed by khatchad

Commit Message

Document the ⊤-permit policy and add ⊤/cross-boundary coverage

Make the precision-favoring treatment of an unprovable shape explicit: numpy
over a ⊤ shape or a descriptor lost across a call boundary is permitted (which
recovers the corpus DenseLayer3d.call idiom, whose input tensor Ariadne types
⊤ - wala/ML#704), and TODO the one-line flip to the sound decline-unless-static
policy for when that lands (#751, blocked on wala/ML#704).

Add two unit reproductions of the policy-divergent path, both permitted under
the shipped policy and inverting when #751 lands:
- testNumpyOnTopShape: numpy over the shape of a tf.constant(np.array(...))
  value (a ⊤ shape, wala/ML#539) - the unit stand-in for the corpus DenseLayer
  scenario that unit tests previously couldn't reach.
- testNumpyOnShapeArgumentToCallee: shape metadata passed as an argument into a
  callee (descriptor lost across the boundary).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

Pull Request Pull Request #750: Make the parameter-flow numpy precondition per-dimension shape-aware

Coverage Stats

178 of 199 new or added lines in 3 files covered. (89.45%)

1 existing line in 1 file now uncovered.

2067 of 2549 relevant lines covered (81.09%)

0.81 hits per line