ponder-lab / Hybridize-Functions-Refactoring / 26113123470

72%

push

github

Infer input signature for single-context tensor parameter (#480)

* Implement Algorithm 2 scenario 1: single-context input signature.

Adds `InputSignature` (record wrapping `List<TensorType>`) and
`Function.inferInputSignature(TensorTypeAnalysis)`, plus a regression
test covering scenario 1 of the input-signature inference algorithm:
single call site, single dtype, single concrete shape. The trivial
happy path of Algorithm 2.

`inferInputSignature` walks the function's parameters, skips `self`,
calls `Parameter.getTensorTypes` per parameter, and reduces the
multi-context set to a single `TensorType` via the private static
`inferSpec`. For scenario 1, the set is a singleton and `inferSpec`
returns the input unchanged. Parameters whose `getTensorTypes` return
an empty set are excluded from the signature; this is the documented
practical default pending the lattice-distinguishing Ariadne-side
query (`Parameter.getTensorTypes` Javadoc).

Subsequent PRs will extend `inferSpec` to multi-context cases
(scenarios 2-7) and address the open spec questions on scenarios 9
and 13.

The new test follows the two-assertion discipline: an Ariadne pin
(`getTensorTypes` returns the expected singleton) plus an algorithm
assertion (`inferInputSignature` returns a signature whose only entry
is the same `TensorType` instance, asserted via `assertSame` since
`TensorType` does not override `equals`).

* Reformat scenario 1 fixture with Black.

Local pre-commit hook missed this on the initial commit (likely because
the file was newly created/untracked at the time the hook ran); CI's
`black --check` caught it.

* Address Copilot review on scenario 1.

- `Function.inferSpec`: tighten the singleton-context branch to require
  both a concrete dtype (`getDType()` not null and not `DType.UNKNOWN`)
  and a concrete shape (`getDims()` non-null with every `Dimension`
  being a `NumericDim`). Without this guard, a singleton with an
  unknown/non-concrete dtype or non-numeric dim would ... (continued)

60 of 65 new or added lines in 3 files covered. (92.31%)

1192 of 1651 relevant lines covered (72.2%)

0.72 hits per line