[Feature] Add /v1/decisions for typed decisions on candidate scoring

[rwang5203]

Motivation

Semantic decisions ask a model a fixed question about some input and need one answer plus a probability for every option, without parsing generated text. #40826 added the per-item candidate scoring that gives these probabilities, but each client still has to render the chat prompt, find answer labels that are single tokens, and keep thinking templates out of the way, and mistakes there change decisions without any error.

Changes

• POST /v1/decisions answers typed questions (choice, score, and yes_no) about an input. The server renders each question with the served chat template and thinking off, checks that its answer labels (A to Z, 0 to 9, or yes and no) are single tokens at the answer position, and scores all questions in one score_prompts call without generating. Each answer returns the probability of every option and label_mass, plus choice for a choice question or score for a score question.
• The server owns and versions the prompt wording, and return_prompt_token_ids lets any answer be replayed through /v1/score.
• Templates that would put the answer inside a reasoning block, and setups the route cannot serve faithfully, such as --enable-mis, --dllm-algorithm, a LoRA adapter named in model, or labels that are not single tokens, return an actionable 400.
• Three fixes found while validating. Scoring batched with plain generation no longer stops the scheduler, a grammar that fails to compile with logprob fields returns its 400 instead of stopping the server, and streaming chat returns each token's own top_logprobs inside a multi-token chunk.
• A Decision models page documents the route, and the Qwen3.8-27B and Qwen3.5 model pages link to it.

Validation

On main, Qwen3.8-27B and Qwen3.5-35B-A3B in BF16, one H200 each, with the documented launch command.

• Decisions are bitwise equal to /v1/score replays of the returned ids, cold cache against cold cache, for 23 tested questions, and all 144 SemIf rows return well formed answers on both models (138 and 130 correct, a sanity check only).
• 56 concurrent mixed requests across decisions, /v1/score, and /generate return 200 on both models, on both model pages' deploy commands, and on the 27B under NEXTN. Main stops the scheduler on the same load without decisions.
• Grammar failures with logprob fields return 400 on both models, and on the 35B streamed top_logprobs match the non-streamed ones for every streamed token, where main repeats the first token's list.
• Every tested refusal returns 400 naming the question, field, or unsupported feature, followed by a healthy request. The Decision models examples run on both models, and each model page example on its own model.
• On CPU, a refusal survey of 537 local chat tokenizer directories, copies counted separately, accepts both validated models and 437 in total. Unit tests cover the contract and the three fixes, whose tests fail on main. Pre-commit passes.

Limits

• Nothing is calibrated, and label_mass for yes_no reads low because it counts lowercase yes and no only. Thresholds need labeled data from the workload.
• Reasoning that neither the chat template nor the reasoning parser marks is not detected, and some templates of models that do not reason are refused.
• On these hybrid attention models, answer probabilities moved by up to about 0.07 and label_mass by up to about 0.14 between cold and prefix-cached requests and across batch compositions, with the chosen option unchanged in our checks.
• Leave any temperature in --preferred-sampling-params unset and --enable-mixed-chunk off for decisions, since the first can scale the probabilities and the second can reuse recurrent state never written for the prefix on these hybrid models.
• The server does not check where the chat template puts an answer, and 8 accepted local checkpoints score one token early, so inspect prompt_token_ids before relying on a new model.

---

CI States

Latest PR Test (Base): ❌ Run #36054102942
Latest PR Test (Extra): ❌ Run #36054102642
Latest PR Test (AMD ROCm 10): ⏳ Run #36054102650

[rwang5203]

Closing, as this change landed in main via #41208.