[BobClawblaw]

What this changes

Registers Nemotron's pocket kernels (nemotron_ops.cu, the learned neurons the living-weights learner adds beside the
shared expert) with the kernel contract: relu^2 up into act rows and down into the shared expert's slot, block-row and
tiled, down to fp32 and bf16, against float64. Neurons past the count are NaN so a read of one shows. Two controls:
the up's relu by fmaxf (the bug the previous PR fixes) and the down launched one neuron short.
tf-cuda-test pocket-contract runs the set alone. Stacked on #1628.

Receipt

• Environment: one GB10 (sm_121), nvcc 13.0 and 13.2.
• SUMMARY pocket: 6 kernels, 6 pass gates 1-4, 0 fail, 0 known findings; 2 controls: 2 red as planted, 0 missed
• Gate-5 records for sm_80/86/89/120/121 (pocket_records.zig); nvcc 13.2 gives tf_pocket_in 39 registers on
  sm_120/121, recorded as its override.
• Full contract on the stack's top: SUMMARY 452 kernels: 452 pass gates 1-4, 0 fail, 6 known findings; 325 controls: 325 red as planted, 0 missed; 321 limit problems (the stack on main d33dcfc; none of the limit problems is in this stack's launches). The stack then rebased onto 3f0a715 without conflicts; the three sets, zig build test and the Transcribe clips were re-run there.

Overlap

atc overlap lists #1445 and #1446 as conflicting, in the shared registration lists (registry.zig, cuda_kernels.zig, kernels.zig, kernel-contract.md, tests/cuda/main.zig); whichever lands second rebases.

[ashhart]

Thank you @BobClawblaw. This is in TensorFold 1.2.0, released today: https://github.com/ashhart/TensorFold/releases/tag/v1.2.0

It landed with your commits under your own name, so you're listed as a TensorFold contributor, and it's credited to you in the release notes and the changelog.