ponder-lab / Hybridize-Functions-Refactoring / #2096

80%

main: 79%

Build Type

push

github

Committed by web-flow

Commit Message

Abandon the input signature for a mixed sparse/dense parameter (#652)

A parameter that is sparse at some call sites and dense at others now reduces
to bottom in `inferSpec`, matching the discipline the dtype axis already uses.
Previously the sparseness consensus failed silently and the reduction fell
through to a dense `TensorSpec` (pinned by #650). That dense spec rejects the
`SparseTensor` the function accepts at the sparse call site — and a
`SparseTensorSpec` would symmetrically reject the dense tensor — so no single
spec is sound. Emitting one changes the program's acceptance, an RQ4 soundness
violation; abandoning the whole signature is the correct conservative outcome
for an irreducibly-polymorphic parameter.

To report the drop precisely rather than mis-attributing it, the per-parameter
drop diagnostic in `inferInputSignature` is now three-way (heterogeneous dtype,
dtype-⊤, mixed sparse/dense), classified in `inferSpec`'s own precedence order
(dtype before sparseness). A new `AbsenceReason.HETEROGENEOUS_SPARSITY` carries
the reason; the `INPUT_SIGNATURE_INFERENCE` INFO names the parameter (#510).

The #650 pinning test is inverted (`testMixedSparseDenseFallsThroughToDense` ->
`testMixedSparseDenseDrops`) to assert the drop.

Call-site-precise recovery — emitting a tight spec when the polymorphism is only
apparent (one branch live at the real call sites) — is a separate enhancement
gated on Ariadne resolving the parameter's concrete type per call site, tracked
separately.

Closes #642.


Claude-Session: https://claude.ai/code/session_01FuhM8SsTRB6BraLt3Ph7ZC

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Coverage Stats

14 of 18 new or added lines in 2 files covered. (77.78%)

1 existing line in 1 file now uncovered.

1621 of 2024 relevant lines covered (80.09%)

0.8 hits per line