ponder-lab / Hybridize-Functions-Refactoring / #2807

84%

main: 85%

push

github

Record shape iteration and non-TensorFlow integer sinks as static reads (#884)

Two ordinary consumption surfaces were outside the statically-read-axis
walk's enumerated sets (#882), so a signature was emitted whose first
traced call raises. The shape-target sinks now include tf.image.resize's
size argument and the NumPy buffer constructors (zeros/ones/empty/full),
which require Python integers for their extents exactly as tf.reshape
does; NumPy chains do not root at the TensorFlow module, so the
constructors are recognized by member name on a numpy-module receiver,
with the module test hoisted to Util and shared with the numpy-
application precondition. Iterating the shape vector itself (a for over
x.shape) raises on an unknown rank and tolerates a known one, so it is
now a rank-only surface like len (#809), its products keeping the taint
at lost coverage.

Tuple-unpacking a shape is not an iteration in this IR: it lowers to
constant-index property reads (handleObjectLiteralAssign), so the FPN
_upsample_add shape reaches the verdict through the subscript narrowing
and the new resize sink rather than through the iteration surface, a
correction to the issue's suggested mechanics recorded in the walk's
comments.

The fixture pins three withholding arms (unpack extents into resize over
a wild axis; iteration under differing ranks; a NumPy buffer sized from
a wild axis) against three ranked twins that stay writable. Validated
two-sided: pre-fix the wild arms record nothing and the pins fail; with
the additions the class runs 627/627.

Fixes #882.


Claude-Session: https://claude.ai/code/session_01X8W9PDNUSQFzZRUAYJmRiE

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

25 of 25 new or added lines in 3 files covered. (100.0%)

3192 of 3799 relevant lines covered (84.02%)

0.84 hits per line