ponder-lab / Hybridize-Functions-Refactoring / #3149

86%

main: 88%

push

github

Block an unmodeled container even when it carries flat evidence (#977)

* Block an unmodeled container even when it carries flat evidence (#976)

A parameter classified as a tensor container whose element structure is not
modeled reached `TENSOR_CONTAINER_UNSUPPORTED` only when it had no flat tensor
evidence. With flat evidence it fell through to the flat reduction and
received a single `TensorSpec`, which a caller passing a list cannot bind to.

The flat evidence of such a parameter is its elements' types flattened, or a
bare tensor from another call site mixed with the containers. Either way no
single `TensorSpec` admits the callers, so the unmodeled container now blocks
whatever flat evidence it carries. This is what
`Parameter.extractContainerElements` already documents for a `null` element
list.

The fixture reduces NLPGNN's `adjacency_lists`, passed as a bare edge tensor
by one caller and as a list of edge tensors by another.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Read the container verdict from conforming call sites only (#976)

The block added for unmodeled containers read `isTensorContainer()`, which
covers every call-graph node, expected-failure contexts included, while the
element structure is extracted from the conforming nodes alone. A parameter
whose conforming callers pass a bare tensor, while a call the tests declare
must fail passes a list, was therefore blocked as an unmodeled container and
lost the correct flat specification of its conforming evidence. That reports
the form unsupported on the strength of a rejected call, which is what the
conforming-node filter of #892 exists to prevent.

`Parameter` now records whether a container reaches the parameter from a
conforming node, beside the element structure, and the block requires it.
When no node is excluded the existing verdict is reused, so the common case
costs nothing further.

Also corre... (continued)

12 of 12 new or added lines in 2 files covered. (100.0%)

3678 of 4287 relevant lines covered (85.79%)

0.86 hits per line