[cshintov]

What this changes

The review on #2013 asked for keep at the other two shapes Geometry.check admits. The qwen27 attention_io contract now runs pack, append, host keep and device keep at all four: the 27B (24 over 4), its two-Mac rank (12 over 2), GQA16 on one Mac (32 over 2) and the 9B (16 over 4). Test only; no kernel or engine change.

• Issue: n/a, the follow-up the qwen27: keep's gather and scatter stage the binding's KV heads a row, not 4 #2013 review named.
• Stacked on: main (37c2241), which has qwen27: keep's gather and scatter stage the binding's KV heads a row, not 4 #2013.
• atc overlap --json: atc is not installed here; no open pull request touches zig/tests/contract/metal_q27_attn.zig.

Commit:

• contract: qwen27 attention_io at the other two shapes Geometry admits, GQA16's 32 over 2 and the 9B's 16 over 4

Checks on this head

• Head commit: 6b7ae97
• Lean checker, python3 tools/zig/lean_check.py: 424 problems on the base, 424 on this head, 0 new.
• Tests run: zig build test-qwen27, test-qwen27-keep-host, test-qwen27-plans-host and test-kernel-contract pass; zig fmt --check build.zig zig passes.
• CI: pending.
• Not run, and why: no model runs; the change adds contract cases only.

Receipt

• Environment: TensorFold 37c2241 plus this commit, Zig 0.17.0, macOS 26.5, Apple M4 Pro (applegpu_g16, no tensor units). No checkpoint.
• Paths and platforms the change reaches: the kernel contract only.
• Exactness, decode speed, prompt speed: n/a, no engine change.
• Licence of any new weights or drafter: n/a.

Kernel contract, only when a GPU kernel is new or changed

No kernel changes; the existing qwen27 attention_io kernels gain two registered shapes each.

• tf-kernel-contract --only qwen27 on this head: SUMMARY 16 kernels: 16 pass gates 1-4, 0 fail, 0 known findings; 2 controls: 2 red as planted, 0 missed; 0 limit problems.
• With main's pre-qwen27: keep's gather and scatter stage the binding's KV heads a row, not 4 #2013 attention_io.metal put back, the GQA16 keep and device keep fail ("staging keys row 1 head 0 is not stream 0's row 66" at 3x32x64) and the 9B's pass, as they should at 4 KV heads.
• Chips I could not run: M5 (tensor units).

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @cshintov, this is the follow-up we asked for on #2013, and the check with the old attention_io.metal put back is exactly the evidence it needs: GQA16's keep fails where it should and the 9B's passes. The new shapes reach only the attention_io registration loop, so the contract goes from 8 to 16 cases with nothing else touched. We'll run the qwen27 contract on an M5 Ultra, the chip you couldn't, and land it after that.

[ashhart]

Thank you @cshintov. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.