serve: single-slot MTP speculation in the mux path (KV_SLOTS=1) — #492/#358

[ZacharyZcR]

Closes the single-slot half of #492 and #358: coli serve at KV_SLOTS=1 now keeps MTP speculation instead of disabling it.

Why it's safe at one slot

MTP/n-gram speculation is disabled in the mux (g_draft=0) because the draft/verify batch isn't ragged-safe when slots sit at different positions. But at KV_SLOTS=1 there is no ragged batch — decode is a single contiguous sequence, the exact regime spec_decode already serves in coli chat, coli run, and run_serve. So the gate becomes if(nctx>1) g_draft=0; — multi-slot unchanged, single-slot keeps the engine's resolved draft setting.

How it's wired

• spec_decode gains a logit_out param (the 4 existing callers pass NULL, unchanged behavior) so the mux scheduler can own the continuation logits across calls.
• mux_submit: a single-slot request with drafts active and no per-request grammar takes the new spec path — the prefill logits are stashed on the request, no pending-token cycle.
• Scheduler: the spec slot runs the whole turn through one spec_decode call (same contract as run_serve), streaming tokens via a mux_data emit callback, then mux_done. No chunk-splicing — that would re-enter spec_decode mid-turn and duplicate the boundary token; Ctrl-C still interrupts through g_intr inside spec_decode like every serve path.
• Grammar requests keep the forced-draft path (its acceptance is ~1 where it fires; the two draft sources shouldn't fight over one forward). Multi-slot and the non-spec pending cycle are byte-for-byte unchanged.
• The load-time [MTP] disabled in multiplexed serve banner now fires only for KV_SLOTS>1; single-slot prints [MTP] single-slot serve: speculation active.

Measured (6x RTX 5090, GLM-5.2 int4, frozen usage snapshot)

Correctness gate — the maintainer's bar, "same tokens as coli chat": on a high-confidence prompt (low near-tie margin), single-slot serve spec (DRAFT=3) is token-exact with greedy (DRAFT=0):

prompt "Q: What is 2+2? Answer with one number only. A:"
  DRAFT=3 (spec):   " 4\nWhat is "
  DRAFT=0 (greedy): " 4\nWhat is "   → identical

Determinism: two runs of the spec path on the same prompt produce byte-identical output.

On near-tie divergence: on a low-confidence prompt, spec and greedy do diverge — but so do DRAFT=3 and DRAFT=0 in plain text mode (" What about black and white" vs " What about for printing" on the same prompt). That's the known #163/#510 kernel-family effect: greedy's S=1 forward and speculation's S=1+g verify forward batch the MoE differently, flipping near-ties. It is a property of speculation on this container, present identically in coli chat, not introduced by this PR — and it's sharpest on the per-row int4 container (another reason for the gs64 default in #579). Under SPEC_PIN (default on) and a group-scaled container it's minimized.

Scope

Non-spec paths untouched (verified: KV_SLOTS=1 DRAFT=0 still runs the shared-batch greedy path normally). Speculation on a single-slot server is pure win where verify forwards are bandwidth-bound — the exact ask in #492/#358.