[yatesdr]

Addresses #26 (the scoring primitive; the decision routes follow in separate PRs).

Problem

Decision and distillation clients need the logprob of specific tokens at each generated position: label tokens for classification, or a teacher's top set for KLD. Today Flare returns only the natural top 20, so a label outside it has no value at all.

Change

logprob_token_ids: [id, ...] on Chat, Completions and Responses, matching vLLM's field:

• With generation logprobs enabled, the listed IDs replace the natural alternatives at every generated position, in caller order. The sampled token is appended when absent.
• At most 128 IDs. They must be integers within the vocabulary; duplicates are dropped and null or empty means omitted. Messages rejects the field.
• Values are the raw FP32 logit minus the existing log-sum-exp, before sampling transforms. They are never renormalized.
• The existing finishing kernel gathers the requested IDs with its existing normalization. There is no new launch, full-vocabulary copy or graph family.
• Readback copies only the requested width. Requests without IDs copy exactly what they do today.
• Natural top-k, prompt scores and echo behavior are unchanged.
• Older packages keep their natural limits and reject nonempty IDs before admission, asking for a re-export.

Tests

Helios 893, integration 31, protocol 47 and Flare 134 pass on CPU, plus the Python component tests. An earlier revision passed CUDA, TP2/TP4 replay and 148 live API checks on Qwen3.8, with requested values within 9.0e-7 of an FP64 reference. DeepSeek V4.1 and this revision's GPU runs are still to do.