[gilp456]

What this changes

Adds separate CUDA repetition sampler entry points and registered numerical contracts while preserving the legacy identity sampler source path. This layer has no production engine caller; GLM wiring is supplied by the final stack PR.

• Issue: Refs GLM CUDA: honor full-prefix repetition_penalty from generation defaults and API requests #3020.
• Stacked on: server: add request-only repetition penalty contract #3153 at bedf1b166. This PR contributes bedf1b166..8ee0d8d23.
• atc overlap --json: selected inspection of GLM-5.3-Flash CUDA: shared rounds draft MTP for every stream at once #2856, GLM on CUDA: kept states, carries and prompt twins under the latent split (TF_GLM_KV_SPLIT, #1431 part 4; stacked on #2937) #2938 and parent server: /v1/models lists each model's input_types; a video it can't take is a 400, not a 500 (#3004) #3019 is complete. GLM on CUDA: kept states, carries and prompt twins under the latent split (TF_GLM_KV_SPLIT, #1431 part 4; stacked on #2937) #2938 shares one file and merges cleanly; server: /v1/models lists each model's input_types; a video it can't take is a 400, not a 500 (#3004) #3019 shares 19 files and merges cleanly; GLM-5.3-Flash CUDA: shared rounds draft MTP for every stream at once #2856 shares none. This is selected coverage.

Stack order: #3153, this kernel/contract layer, then #3155. All three drafts target lead/models-input-types; this PR's review range is the single commit above.

Checks on this head

• Head commit: 8ee0d8d23.
• Lean checker: 345 findings on the target base and head, zero new findings.
• Tests run:
  • zig build test-kernel-contract --summary all: 34/34 steps, 240/240 tests passed.
  • zig build test test-kernel-contract test-golden fuzz-cortheon --summary all: exit 0 with zero failures and 222 platform skips.
  • zig fmt --check build.zig zig, git diff --check, and tools/pr_check.py: pass; 49 changed files cumulatively and zero new lean findings.
• CI: pending on this draft head.
• Not run, and why: CUDA compilation/device execution, compiler-derived resource records, legacy-kernel SASS hashes and the CUDA SUMMARY require an authorized CUDA toolchain/device.
• Plain zig build: 26/29 steps succeeded; tf-qwen35-cuda hits seven Windows std.c/fabric-network errors. The exact target base reproduces the same seven errors and 26/29 result; this stack does not change that code.
• Cluster activity: zero SSH, deployment, service, Docker or model-process commands.

Receipt

• Main against branch: n/a for served inference; this layer has no production engine caller.
• Boxes it reaches: CUDA sampler source, repetition-only entry points, generated reference fixture, registered contract runner and launch inventory.
• New switch? Default, accepted values, profile row: n/a.
• Environment: Windows host, Zig 0.17.0; host contract inventory above; CUDA compiler/device unavailable.
• Paths and platforms the change reaches, which I ran and which I could not: host registration and fixture checks ran; CUDA SM80/86/89/120/121 execution remains pending.
• Exactness: source-pinned float32 fixture and host contract wiring pass; CUDA numerical execution remains pending.
• Main against this branch, same machine and checkpoint: n/a until the final GLM layer supplies a production caller.
• Decode speed before and after: n/a for this uncalled layer; final-stack served measurements remain pending.
• Prompt speed before and after: n/a for this uncalled layer.
• Licence of any new weights or drafter, or "n/a": n/a; no weights or drafter added.

Kernel contract

• Kernels registered with the contract, and the kernel each one replaces: repetition-only greedy, sampled-token and FP32-probability operations are registered; the legacy identity entry points remain the baseline path.
• A planted control for each new kernel, and each one fails as planted: ten controls are registered; CUDA red execution remains pending.
• Each backend's SUMMARY line and exit status: pending CUDA execution; no device SUMMARY is claimed.
• Chips I could not run: CUDA SM80/86/89/120/121; Metal and HIP are outside this CUDA layer.
• Resource and SASS evidence: launch geometry is registered with resource usage explicitly pending; legacy-kernel cuobjdump -sass hashes remain pending.

Checklist

• The title and description match the head commit, receipt and stack.
• The independent fixture and registry checks pass without a CUDA device.
• Comments and docstrings are one line. Measurements remain outside source.
• Publication content contains no personal data, machine names, internal hosts, local paths or AI attribution lines.
• The production identity path retains its existing ABI and source arithmetic.
• No precision policy changes are introduced.
• CUDA planted-control execution and complete SUMMARY.
• Compiler resource records and identity-kernel SASS hashes.
• Final-stack model and performance gates.

[gilp456]

/atc depends 3153