gather fixtures structure and extension to match A-D level pattern

[lasch]

A-D levels for test fixtures already exist for e.g. reduce.
As a follow-up to #11 and #116 the same pattern should be established for gather.

The table below shows the current status/gaps:

Level	Existing variants (30 total, 14 kernel fns)	Fit
A — shape/distribution	default, y_offset_zero, full_row, slice_large_row, min_block_cols, slice_at_end, wide_slice (gather_kernel, 7); 1core, large_k (gather_kernel_1core, 2); 2d, 2d_serial, 2d_large_table, 2d_large_table_serial (gather_2d_kernel, 4); 1d (gather_1d_kernel, 1); 3d, 3d_large_k, 3d_group, 3d_group_end, 3d_partial, scatter_3d, scatter_3d_partial, 4d, 4d_boundary, 2d_index_gather, 2d_index_roundtrip, 2d_index_3d_block, 2d_index_3d_block_large (9 kernel fns, 14 variants)	Good fit, already exists. All 28 pin fp32/i32 and sweep rank, shape, grid, and distribution — exactly Level A's job. No new variants needed; this is a reclassification, not a rewrite.
B — compute correctness	(none)	Gap. No variant sweeps dtype at the simplest shape on ktir_cpu alone — the closest thing today (1core) is fp32-only and already carries other Level-A-ish framing (single-program shape, not a dtype probe).
C — layout	spyre_stick, spyre_stick_output_only (gather_kernel_spyre, 2); 4d_spyre_stick_output (part of gather_4d_kernel's 3)	Good fit, already exists. All three carry tl.spyre_tensor_layout annotations, no compiles_to_binary, matching Level C exactly.
D — device	(none)	Gap. grep -n "compiles_to_binary" fixtures/gather/meta.py returns nothing — no gather variant reaches dbo-opt or compiles to a real binary today. A and C are largely already there in substance if not in labeling.

The implementation is probably going to be done in at least 2 stages:

1. update/rename existing tests/fixtures to match the pattern (this is a rather straightforward/cheap step)
2. fill in the gaps with new fixtures (this might be split into further sub-steps depending on required effort)

[fabianlim]

@lasch it will be premature to do this, we should wait until the backend supports gathers. It could be quite soon.

[lasch]

...it will be premature to do this, we should wait until the backend supports gathers. It could be quite soon.

This applies to level D only, doesn't it? All other levels should be able to handle gathers and could be prepared?

[fabianlim]

@lasch

This applies to level D only, doesn't it? All other levels should be able to handle gathers and could be prepared?

yes that is true. Yes maybe go ahead to just message it to levels A-C. One thing that I mighjt be doing in the near future, is to move all device related tests to D. So that means any test with a spyre_tensor_layout. For now you can keep those in C like the other fixtures.

[fabianlim]

The Level C half of this needs #187 settled first.

#186 moved physicalization from the ktir stage to the spyrecode stage. The numerical tier runs the ktir stage and stops, so it now reads a logical artifact — the tts.tensor_layout annotation is carried but inert to ktir_cpu. Numerically, Level C is now Level B.

That bears directly on two rows of the table above:

• C — layout: "Good fit, already exists." spyre_stick, spyre_stick_output_only and 4d_spyre_stick_output do carry tl.spyre_tensor_layout and do lack compiles_to_binary, so they still match the shape of Level C. But under the current definition they are Level B variants that happen to carry an annotation — nothing executes their stick-tiled form. Reclassifying them as C on the old reading would record coverage that is not there.
• D — device: gap. This gets more urgent, not less. With C's numerical half gone, a compiles_to_binary gather variant becomes the only thing that would exercise a gather layout end to end. Per Reorganize the KTDP tree, and move the layout out of the tt dialect #167's coexistence section, gather, matmul and softmax are exactly the three families with layout coverage only through lit fixtures.

Stage 1 as you describe it — "update/rename existing tests/fixtures to match the pattern" — is unaffected for A and B and can proceed. It is the C row whose label depends on a definition that is currently in flux.