ponder-lab / Hybridize-Functions-Refactoring / #2570

83%

push

github

Emit a per-parameter tensor-type diagnostic CSV (#780) (#797)

* Emit a per-parameter tensor-type diagnostic CSV (#780)

Adds `parameters.csv` to the evaluator: the raw inferred tensor lattice
per parameter at per-dimension granularity, so a wildcard dimension's
cause (a static constant, an unresolved-but-fixed size, or a genuinely
symbolic axis) is visible without a FINE-level provenance run.

Where an input-signature entry collapses every non-constant dimension to
a single wildcard, this preserves the raw `Dimension` class: `Constant,<n>`,
`Symbolic,<name>`, `Unresolved`, `Dynamic`, `Ragged`, a shape-`TOP` row for
unknown rank, and a `scalar` row for rank zero, with the dtype and whether
it is top. The classification reads the raw `Dimension` subtypes directly
rather than through `inferSpec`, which lossily normalizes them. It is a
testable static helper on `Parameter` with synthesized-input unit coverage,
mirroring `InferSpecTest`; the evaluator addition is a thin printer that
leaves the analysis untouched.

Origin and allocator-source columns are omitted: for a parameter both are
invariant (`PARAMETER` and a parameter-local `LocalPointerKey`), so they
would carry no diagnostic signal.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8tXuyy3YWhM7BzvmCxKAn

* Split the #780 diagnostic into parameters.csv and parameter_dimensions.csv

The evaluator emits two files whose names describe their row grain.
`parameters.csv` has one row per parameter, carrying `is tensor`,
`is container`, and the rendered `tensor types` and `container element
types` (rendered for convenience at the parameter grain, since
reconstructing a type across rows or files is not). `parameter_dimensions.csv`
keeps the normalized one-row-per-dimension lattice, because a parameter has
zero-to-many dimensions and a dimension is its own grain.

The rendered type preserves each dimension's raw class (e.g.
`INT32[Constan... (continued)

50 of 56 new or added lines in 1 file covered. (89.29%)

2393 of 2893 relevant lines covered (82.72%)

0.83 hits per line