ponder-lab / Hybridize-Functions-Refactoring / #3135

86%

main: 89%

Build Type

push

github

Committed by web-flow

Commit Message

Record what a parameter's tensor verdict rests on (#972)

`classifyAsTensor` writes `FALSE` for two situations that call for opposite
work. With the enclosing function in the call graph, the tensor-type and
container phases run and neither classifies the parameter as tensor-like; under
contract-compliant generators an empty tensor-type result is the analysis's
bottom, meaning not a tensor, so that is a determination. With the function
absent from the call graph, both phases are skipped and the verdict takes its
default having concluded nothing at all.

The emitted rows are byte-identical in both cases, so no consumer can separate
them and no check over the emitted families could detect the difference. That
matters because a signature covers every parameter or none: one such parameter
suppresses a whole function, and a determination is outside the mechanism while
a default may be a precision limit. Reading the second as the first writes off a
function on the strength of a question never asked.

Records the basis beside the verdict. The distinguishing fact is already
present where the verdict is written, so this carries something the method
knows rather than computing anything new, and no control flow changes: every
addition is an assignment before an existing return.

The basis names what was known rather than what was concluded. A conclusion ages
separately from the row it disposed of and then reads as current, while a basis
stays true about the run that wrote it. `CLASSIFICATION_DID_NOT_RUN` names the
phases not running rather than a negative result, so it cannot be mistaken for
weak evidence of a non-tensor.

The pair is pinned in one fixture. `scale`'s `factor` is a determined
non-tensor and `never_called`'s `z` reaches the same verdict with the phases
never having run, since an uncalled function is absent from the call graph.
Same verdict, opposite reasons, indistinguishable before this change.

Fixes #971

Coverage Stats

20 of 21 new or added lines in 2 files covered. (95.24%)

1 existing line in 1 file now uncovered.

3670 of 4279 relevant lines covered (85.77%)

0.86 hits per line