pub fn ssm_tail_ckpt_enabled() -> boolExpand description
OPT-IN switch for the tail checkpoint (ATLAS_SSM_TAIL_CKPT=1).
Default OFF. The 3-traj A/B (2026-07-10, 174 samples/arm) showed it is
perf-NEUTRAL: it removes the SSM replay on ~89% of warm turns (mean 254 -> 25
tokens), but the prefill-chunk split needed to land a snapshot on
ssm_tail_boundary costs a median 868 ms extra forward pass for a median of 8
trailing tokens, which cancels the ~1374 ms of replay it saves. It becomes a
clear win only once the SSM state can be captured MID-CHUNK (in the GDN prefill
kernel) instead of via an extra pass. Until then it stays off by default and
ungated for accuracy.