ponder-lab / Hybridize-Functions-Refactoring / #2838

85%

push

github

Let a container verdict outrank a parameter's own tensor typing (#890)

* Let a container verdict outrank a parameter's own tensor typing

A parameter whose nesting varies across call sites was typed from the
sites that pass it unwrapped: Ariadne's own evidence made Phase 2 hit,
the container check never ran, and the reduction emitted a single
`TensorSpec` reduced from those sites alone. TensorFlow enforces the
declared nesting, so such a signature admits none of the callers that
pass the sequence, which is the reported harm (#888, observed on
deep_recommenders' CIN.call, whose spec came from the one site passing a
bare array while every conforming site passes a two-tensor tuple).

The container question is now also asked on the Phase 2 hit, gated on a
sequence actually reaching the parameter so an ordinary tensor parameter
pays nothing and keeps its `isTensorContainer()` cache unset, and the
reduction dispatches on the container verdict ahead of the parameter's
own contexts. Either the nested spec is writable or no signature is: for
the mixed form the element extraction already declines (container and
non-container values mixed at the call sites), so the parameter blocks
with `TENSOR_CONTAINER_UNSUPPORTED` and the function hybridizes bare
rather than carrying a specification its callers violate.

The fixture pins both directions on a runnable file: `cin` mixes a tuple
site with a bare-tensor site and withholds its signature, while
`tuples_only` receives tuples alone and keeps its two-element nested
spec. Validated two-sided: pre-fix `cin`'s container verdict is unset
and the single `(2, 3, 5)` spec is emitted; post-fix the verdict is TRUE
and no signature is written. The class runs 632/632.

Addresses the structural half of #888; the negative-test anchoring half
is untouched.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R1emLt6bvQa993huMzTaKu

* Ask the container question for any... (continued)

16 of 17 new or added lines in 2 files covered. (94.12%)

3259 of 3859 relevant lines covered (84.45%)

0.84 hits per line