[Other] serve: logprobs on /v1/completions (stacks #3332)

docs/DETAILS.md

@@ -406,6 +406,7 @@ The server listens on `http://127.0.0.1:8080` (`--port` in setup, or edit the ru
 | API | Endpoint |
 | --- | --- |
 | OpenAI Chat Completions (stream and non-stream, tools) | `POST /v1/chat/completions` |
+| OpenAI text completions (stream and non-stream, raw prompt) | `POST /v1/completions` |
 | Anthropic Messages (stream and non-stream, tools) | `POST /v1/messages` |
 | OpenAI Responses (stream and non-stream, tools; stateless, [below](#the-responses-api-and-codex-cli)) | `POST /v1/responses` |
 | Model list and health | `GET /v1/models`, `GET /models`, `GET /health` |
@@ -549,6 +550,15 @@ applies at the next start. Network, key, MCP and program keys are not editable t
   `--parallel`, only a `max_tokens` 1 request gets them: it is the one a batch slot decodes through the single-token
   path, which writes an LP line. A longer reply there is a 400, since the slot's `BT` lines carry none -- entries for
   a prefix only, which a client cannot tell from the model's own answer.
+- **logprobs on `/v1/completions` (vLLM's shape).** `"logprobs": N` (0..20, or `true` with `top_logprobs`) adds
+  `choices[0].logprobs` with `tokens`, `token_logprobs`, `top_logprobs` (one `{token: logprob}` per token) and
+  `text_offset`. Every generated token is listed, the stop token too, so a one-token read whose most likely token is
+  the stop still has its distribution. `"allowed_token_ids"` lists those ids with their full-vocabulary logprobs, not
+  renormalised (unlike chat), so a caller can tell how much of the mass its labels hold;
+  `"return_tokens_as_token_ids": true` keys every token as `token_id:<id>`. The prompt may also be a list of token
+  ids. This is what a decision client reads from `vllm serve`: a typed choice from one next-token distribution
+  (AnyJev's `StrataBackend`). The rules are the chat ones (no cap of its own on `max_tokens`; with `--parallel`, only
+  `max_tokens` 1), and one of its own: logprobs are not streamed on this route, `stream: true` with them is a 400.
 - `STRATA_SPEC_PROB=1` with the default `--spec-min-p` gate: the gate tests the token the draft head drew from its distribution
   q. A draft that fails the gate still counts as a proposal. The verifier keeps q's mass on the tokens that pass the gate
   (not renormalised; call it mu), accepts an offered token D with probability min(1, p(D)/mu(D)), and draws the replacement

serve/frontend.py

@@ -65,6 +65,18 @@ def render(self, messages: list[dict], tools: list[dict] | None = None, add_gene
         return self.template.render(messages=messages, tools=tools, add_generation_prompt=add_generation_prompt,
                                     **kwargs)
 
+    def gen_prompt_tail(self, enable_thinking: bool) -> str:
+        """What add_generation_prompt adds to this template: the generation prompt's tail.
+
+        The raw /v1/completions prompt is passed through untouched; this is what a chat render would have
+        appended when the client asks thinking off. The probe message keeps the render the template needs;
+        both renders get the same kwargs, so the difference is the generation prompt alone.
+        """
+        probe = [{"role": "user", "content": "x"}]
+        base = self.render(probe, add_generation_prompt=False, enable_thinking=enable_thinking)
+        full = self.render(probe, add_generation_prompt=True, enable_thinking=enable_thinking)
+        return full[len(base):]
+
     def _detect_caps(self) -> dict[str, bool]:
         """llama.cpp's capability names, checked at load time against this template and Strata's tool-call format.
         These are rendering hints, not a guarantee that the model will follow a request."""