[gilp456]

What this changes

Wires the request contract and repetition-only kernels into GLM CUDA with request-owned full-prefix histories. Each row includes the committed prompt/reply and accepted verification ancestry; prefix extensions append only their new suffix, while shortening or replacement rebuilds membership from authoritative IDs.

• Issue: Fixes GLM CUDA: honor full-prefix repetition_penalty from generation defaults and API requests #3020 when this cumulative stack and its remaining gates land.
• Stacked on: cuda sampler: add separate repetition kernels and contracts #3154 at 8ee0d8d23, which follows server: add request-only repetition penalty contract #3153. This PR contributes 8ee0d8d23..f2a3f894b; 9886d4387 adds only this PR's changelog link.
• atc overlap --json: selected inspection is complete. Parent 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 and GLM on CUDA: kept states, carries and prompt twins under the latent split (TF_GLM_KV_SPLIT, #1431 part 4; stacked on #2937) #2938 each have one conflict in zig/src/families/glm/cuda_follow.zig; final landing-order coordination is required. This is selected coverage.

Stack order: #3153 (request/API), #3154 (kernels/contracts), then this GLM engine layer. All three drafts target lead/models-input-types and identify their own review ranges.

Checks on this head

• Head commit: 9886d4387; implementation commit f2a3f894b.
• Lean checker: 345 findings on the target base and head, zero new findings.
• Tests run:
  • zig build test-glm-repetition-host compile-glm-native-open --summary all: 7/7 steps, 23/23 tests passed.
  • zig build test test-kernel-contract test-golden fuzz-cortheon --summary all: 679/679 steps; 8,699 passed, 222 skipped, zero failed.
  • zig fmt --check build.zig zig, git diff --check, and tools/pr_check.py: pass; 65 changed files cumulatively and zero new lean findings.
• Parent-layer checks: API receipt is in server: add request-only repetition penalty contract #3153; kernel receipt is in cuda sampler: add separate repetition kernels and contracts #3154.
• CI: pending on this draft head.
• Not run, and why: current-head CUDA/SASS/resource qualification and TP2 whole-model exactness/performance require an authorized CUDA/model window.
• Plain zig build: the exact target base and earlier cumulative heads each reach 26/29 steps, then tf-qwen35-cuda hits the same seven Windows std.c/fabric-network errors in unchanged code.
• Cluster activity: zero SSH, deployment, service, Docker or model-process commands.

Receipt

• Main against branch: pending current-head TP2 model comparison.
• Boxes it reaches: GLM CUDA greedy, sampled, grammar redraw, speculative verification, request history, follower transport, refusal/retry and lane lifecycle.
• New switch? Default, accepted values, profile row: server: add request-only repetition penalty contract #3153 adds a request-only field; omitted and null use identity 1; no checkpoint or server-wide default.
• Environment: Windows host, Zig 0.17.0; host GLM and native-open checks above; CUDA compiler/device/checkpoint unavailable.
• Paths and platforms the change reaches, which I ran and which I could not: GLM host state/transport and native-open compilation ran; CUDA SM80/86/89/120/121 and model execution remain pending.
• Exactness: host tests cover full-prefix membership, suffix-only extension, replacement/shortening, accepted ancestry, rejection/reset and follower recovery. Current-head drafted/plain, resumed/fresh, concurrent/solo and grammar model equality remain pending.
• Main against this branch, same machine and checkpoint: pending with an admitted TP2 GLM setup.
• Decode speed before and after: pending at one stream and the widest admitted counts, greedy and sampled.
• Prompt speed before and after: pending cold cache-off measurements at admitted lengths.
• Licence of any new weights or drafter, or "n/a": n/a; no weights or drafter added.

Checklist

• The title and description identify the cumulative stack and this PR's own range.
• History regressions cover suffix-only extension and authoritative rebuilds.
• 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.
• Every token remains inside the existing lane rounds; active repetition uses request-owned state.
• FP32 repetition arithmetic preserves the agreed sign rule.
• Current-head CUDA contract, resource and identity-SASS gates.
• Drafted/plain, resumed/fresh and concurrent/solo TP2 equality.
• Main-against-branch decode and cold-prefill comparison.

[gilp456]

/atc depends 3154