What

Fixes the MTP acceptance regression in #163 by implementing the fix the thread converged on: draft and verify must compute the same function.

Three switches in glm.c were S-dependent, so the draft-side forwards (S=1) and the
verify batch (S≥2) could silently run different kernel families:

1. the int4 IDOT gate S>=g_i4s — asymmetric exactly on ISAs where g_i4s>1
   (x86: S=1 exact / S≥2 IDOT; ARM SDOT and POWER are S-independent, which is why
   the failure signature reversed across ISAs);
2. the S==1-only fused gate+up pair kernel;
3. the Metal GEMM row threshold (S>=16, reachable by verify batches at DRAFT≥15).

SPEC_PIN=1 (default) pins every forward issued while model drafts are live to the
platform's S=1 kernel family, so draft and verify agree by construction — whatever
that family is on the host, per the thread's conclusion (a per-ISA "disable IDOT"
would fix x86 and re-break ARM). Prefill and non-speculative decode keep their
existing gates and their IDOT speed. SPEC_PIN=0 restores the old behavior for A/B.

Also softens the 24-proposal <10% guard (as suggested in the thread): instead of
latching g_draft=0 for the whole session, drafting pauses for 256 tokens and
re-arms on a fresh window — a future rounding regression degrades MTP instead of
silently shutting it down.

Validation

OP's exact config: "Summarize the plot of Romeo and Juliet in three sentences.",
--topp 0.7 --ngen 32, int8 MTP head, DRAFT=3, SEED=1234, single run per cell.

x86 — AMD Ryzen 9 9955HX (Zen 5, AVX-512 VNNI), 92 GB, Linux, --ram 78

The pinned default restores the OP's pre-#68 numbers exactly: 33% (16/48),
2.00 tok/forward — the same cells he reported for main @ 1bdaeee. The greedy
off-distribution opening heals too: old default opens "### Plot Summary",
pinned default opens "Here's a concise summary of the plot of Romeo and Juliet…".
Where both forwards were already exact (IDOT=0), the pin is a proven no-op
(byte-identical text, identical accept counts).

ARM — Apple M4 Pro (NEON/SDOT), 48 GB, macOS CPU-only build, --ram 38

On ARM the pin resolves to int4=idot int8=idot — today's healthy default family —
so nothing collapses and nothing regresses. The small greedy gain is principled, not
noise-mining: under the pin the rare S=1 main-layer forwards (iterations where no
draft is proposed) go through the same IDOT family as the verify batches instead of
the S==1-only fused-pair kernel, removing the last S-dependent switch.

Two cross-checks worth calling out:

• The exact path is ISA-portable: the IDOT=0 cells are the same counts on both
  machines (11/60 greedy, 12/57 sampled at SEED=1234) — Zen 5 AVX-512 and M4 Pro
  NEON agree run-for-run at exact kernels, extending @ZacharyZcR's faithfulness
  trace across ISAs.
• My earlier ARM IDOT=0 collapse (0/24) does not reproduce on current dev:
  four runs today (one seeded + three unseeded) give 21%, 27%, 27%, 24% — 0/24 is
  far outside that spread. That datapoint was n=1 on a build from before glm.c: 4-accumulator NEON SDOT for int8/int4 expert dots (2.4x/core on Apple Silicon) #264's
  4-accumulator SDOT rework merged; I'm not claiming causality, just updating the
  record: on today's dev both ARM families are healthy and S-independent.

(Note: the x86 rig had an idle ollama daemon resident during runs — no model loaded,
acceptance unaffected, tok/s at most slightly conservative.)

make test: 62/62 pass on both hosts. [SPEC_PIN] banner reports the pinned family
at first speculation, e.g. on Zen 5: int4=exact int8=idot; on M4 Pro: int4=idot int8=idot.

What this does NOT claim

• CUDA path untouched: its dispatch condition is S-independent, hence already
  draft/verify-consistent.
• Acceptance percentages under --topp remain non-portable across engine versions
  (discontinuous expert pruning) — the tables compare arms within one build only.
• Speed: this is a correctness fix, not a speed win. On x86 the pinned default runs
  verify batches on exact int4 kernels (slower per forward) but needs ~45% fewer
  forwards (16 vs 29 per 32 tokens); measured end-to-end decode is neutral on this
  host (greedy 0.66→0.71 tok/s, sampled 0.67→0.65). What the fix buys is MTP that
  stays alive and correct — with IDOT prefill untouched.

Closes #163 (with the guard-softening suggestion from the thread included).