[gilp456]

What this changes

Adds the request-layer repetition_penalty contract while engine capability remains opt-in. Chat, completions, Messages and Responses accept a positive finite float32 factor; omission and null use identity factor 1; a backend without the capability refuses an effective nonidentity factor before generation. If generation_config.json declares an active factor, startup tells the operator to pass it explicitly.

• Issue: Refs GLM CUDA: honor full-prefix repetition_penalty from generation defaults and API requests #3020.
• Stacked on: server: /v1/models lists each model's input_types; a video it can't take is a 400, not a 500 (#3004) #3019 / lead/models-input-types at 0029d0e822f73363cc9e0f774a3e52837bb7a4de. This PR contributes 0029d0e8..bedf1b166.
• 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: this request/API layer, #3154 (kernels/contracts), then #3155 (GLM engine wiring). All three drafts target lead/models-input-types and identify their own commit ranges.

Checks on this head

• Head commit: bedf1b166.
• Lean checker: 345 findings on the base and head, zero new findings.
• Tests run:
  • zig build test-repetition-api-host test-golden --summary all: 8/8 steps; 29/29 focused unit tests; 55/55 fixture/security goldens; 319/331 server goldens with 12 known differences and zero unexpected differences.
  • zig build test test-kernel-contract test-golden fuzz-cortheon --summary all: 677/677 steps; 8,679 passed, 222 skipped, zero failed.
  • zig fmt --check build.zig zig, git diff --check, and tools/pr_check.py: pass; 31 changed files and zero new lean findings.
• CI: pending on this draft head.
• Not run, and why: CUDA, device, model and performance gates belong to the later kernel and GLM layers.
• Plain zig build: 26/29 steps succeeded; tf-qwen35-cuda hits seven Windows std.c/fabric-network errors. The exact parent 0029d0e8 reproduces the same seven errors and 26/29 result; this PR does not change that code.
• Cluster activity: zero SSH, deployment, service, Docker or model-process commands.

Receipt

• Main against branch: n/a for active repetition processing; this layer supplies request admission and refuses active factors until an engine advertises support.
• Boxes it reaches: chat, completions, Messages and Responses parsing, validation, capability reporting and engine request planning.
• New switch? Default, accepted values, profile row: request field only; positive finite float32, with omitted or null equal to 1; no profile, checkpoint or server-wide default.
• Environment: Windows host, Zig 0.17.0; host and golden checks above; GPU and checkpoint n/a for this layer.
• Paths and platforms the change reaches, which I ran and which I could not: native server request paths ran on the host; CUDA device and model paths remain in the later stack.
• Exactness: identity/default propagation and effective-float32 capability refusal are covered by host tests; model token equality remains a final-stack gate.
• Main against this branch, same machine and checkpoint: n/a for this API-only layer.
• Decode speed before and after: n/a; this layer has no active engine implementation.
• Prompt speed before and after: n/a; this layer has no prompt-processing implementation.
• Licence of any new weights or drafter, or "n/a": n/a; no weights or drafter added.

Checklist

• The title and description match the head commit, receipt and stack.
• New startup-note and effective-float32 tests failed before their fixes and pass on this head.
• 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.
• Existing lane rounds, draft settings and precision remain unchanged.
• Final-stack drafted/plain, resumed/fresh and concurrent/solo model gates.
• Final-stack decode and cold-prefill comparison.

[gilp456]

/atc depends 3019