ponder-lab / Hybridize-Functions-Refactoring / #2497

83%

Build Type

push

github

Committed by web-flow

Commit Message

Bump Ariadne to 0.52.29 (#776)

* Invert testRedefinedFunction: both copies type on Ariadne 0.52.26 (wala/ML#719)

wala/ML#719 fixes the redefined-module-global false negative (#766): each call
binds to the definition in scope with its own argument types, so both live,
tensor-fed definitions of `compute` are typed. Strengthen the count 1 -> 2 per
the pin's TODO.

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

* Permit numpy over an UnresolvedDim shape dimension; bump Ariadne to 0.52.27

Ariadne 0.52.27 carries wala/ML#721 (the `DynamicDim`/`UnresolvedDim` marker
split) and its wala/ML#722 follow-up (the `tf.shape` shape-vector arm, which
keeps evidence-free `tf.shape(x)[i]` axes `Unresolved` rather than guessing).
`numpyOverShapeStaticness` now treats `UnresolvedDim` like ⊤: permitted under
the precision-favoring policy, declined under the sound policy (#751). This
recovers `DenseLayer3d.call`/`3dProj.call` under permit-⊤ while genuinely
runtime-dynamic (`DynamicDim`) shapes still decline.

The permit is optimism, not soundness: for a `tf.shape(x)[i]` extraction whose
axis carries no run-time-`None` evidence, Ariadne classifies `UnresolvedDim`
(wala/ML#722), and permitting it lets numpy run over a dimension a TensorShape
may patch to `None` at run time. The comment names this; the sound policy
(#751) declines it. Consumer side of #772.

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

* Advance the get_loss pred pin to the Unresolved logits form (0.52.27)

On Ariadne 0.52.27, `get_loss`'s `pred` types `{[Unresolved, Unresolved,
Constant,10]}` and a rank-unknown `float32` member. The leading two dims are
`tf.shape(x)[0]/[1]` on a keras-call result whose batch/sequence axes carry no
run-time-`None` evidence, so they classify `UnresolvedDim` per the wala/ML#721
caveat, ... (continued)

Coverage Stats

7 of 7 new or added lines in 1 file covered. (100.0%)

3 existing lines in 1 file now uncovered.

2140 of 2617 relevant lines covered (81.77%)

0.82 hits per line