#1627 [Feature]: logit_bias on /v1/chat/completions (accepted today, but has no effect)

Descripción

**What I would like**
`logit_bias` on `/v1/chat/completions`, at least the "never pick this token" case: `false` in llama.cpp's form `[[id, false]]`, or `-100` in OpenAI's form `{"id": -100}`.

**Why**
I use Qwen3.8-Flash-Next IQ3_XXS in English and Czech. Now and then it switches to Chinese or Cyrillic in the middle of a sentence. With llama.cpp my app prevents this by sending every token that holds a letter outside Latin and Greek as `false`: 103,215 of the 248,320 tokens (control tokens are never in the list). With Strata the list has no effect, so these letters still get through.

**What happens today** (Strata 0.1.41, Windows 11, this model)
The field is accepted (HTTP 200) but ignored. Token 109266 is `你好`:

```
curl http://127.0.0.1:8090/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "qwen3.8-flash-next-iq3_xxs",
  "messages": [{"role": "user", "content": "Write the Chinese word for hello. Only the word."}],
  "max_tokens": 10, "temperature": 0, "reasoning_effort": "none",
  "logit_bias": [[109266, false]]
}'
```

The answer is `你好` with and without `logit_bias`, in both forms. The full 103,215-token list is accepted too, and `你好` and `Привет` still come out. I found no `logit_bias` in `serve/` or in docs/DETAILS.md.

**Possibly useful**
- The list does not change during an answer, so unlike the grammar mask in #1328 it might not need single-token windows.
- Because the list is long, a way to give it once would work for me too, for example a file of token ids named on the server's command line and applied to every request.

**Workaround today**
None that holds. My app can only cut the answer when a banned letter appears and ask again from the start.

I searched the issues and pull requests for logit_bias, logit, ban, script, Chinese and Cyrillic and found no earlier request.