Function mtp_max_seqs 

pub fn mtp_max_seqs() -> usize

SSOT for the multi-sequence MTP cap (ATLAS_MTP_MAX_SEQS; default 32 with the K-vs-batch ladder, 4 under ATLAS_NO_MTP_K_LADDER). Value-parsed, not presence-checked. Lives beside the ladder (moved from speculative.rs, originally scheduler/mod.rs) because the two are one policy: the model-side single-sequence MTP structures (catchup ring, refeed labels, carry slot) gate on the same value the scheduler gates dispatch on.

The cap IS the adaptive per-concurrency policy: the scheduler gates dispatch on active.len() <= mtp_max_seqs(). Per-step K comes from mtp_ladder_drafts (task #35): 4:3,8:3,16:1,32:1 — 3 drafts through n=8 (matrix 2026-07-28: C=8 95.84 at 8:3 vs 93.30 at 8:2 on the same binary, and 73.5 MTP-off), then 1 draft through n=16 (wave 19 2026-07-31: C=16 181.9 at 16:1 vs 172.70 at 16:2 on one binary, disjoint ranges — the wave-11 grid’s 16:2 lead does not survive the drop in p1 from 0.859 to ~0.72), then 1 draft through n=32 (2026-07-30, the native-bs32 rung — R = 64 verify rows). ATLAS_NO_MTP_K_LADDER (presence) restores fixed K=4 + cap 4 — the dafd990d adaptive policy. Set ATLAS_MTP_MAX_SEQS=1 to restore single-sequence-only.