/metrics: expose the speculative draft counts (offered / accepted) for llm-visuals #457

Hi, I run Strata 0.1.34 with Qwen3.8-Flash-Next IQ3_S at 262K on my RTX 5090 behind llama-swap (119-135 tok/s decode, ~5,000 tok/s prefill, 0.8 s to resume a 71K conversation), thank you for it.

I monitor it live with llm-visuals, a terminal dashboard that polls Strata's /metrics every second (its Strata adapter is on the branch feature/strata). Throughput, context, layers and requests work; the speculative decoding panel stays empty: "Strata does not report draft acceptance".

The server already has the MTP draft counts: serve/server.py reads them from the engine's DONE line (self.last.update(drafts_accepted=..., drafts_offered=...)), returns them in the response timings as draft_n / draft_n_accepted (#83), and the server log prints them ("drafts accepted 601 of 1181").

They are missing from GET /metrics: not in live, not in the requests history, not in totals (checked on 0.1.34). From the response timings I only get the figure once a request has finished; what I would like to see in llm-visuals is the acceptance while the model generates.

Proposal (additive, no change to existing fields):

• in live, while generating: "drafts_offered" and "drafts_accepted" so far for the current request (if the engine can report them before DONE);
• in each requests entry: the same two counts for the finished request (null when the engine did not report them, as draft_n is omitted today);
• in totals: running sums of both.