SIGN IN SIGN UP

fix(dots3): exempt positional full-latent slots from SSM companion snapshots

The dots3 deep-context wall is named at last. The referrer-walk census
traced the ~20GB active growth to the SSM companion cache: every stored
checkpoint deep-copies all non-KV slots, and dots3's 13 full-attention
Dots3LatentCache slots look "SSM-like" (a .cache list, not trimmable
KV) but hold O(ctx) positional latent streams. Each boundary snapshot
cloned 13 full-length latents (105 arrays total with k_pe/idx), so the
companion grew QUADRATICALLY with context (+13 latents per boundary at
1024-token spacing; ~8.6GB retained by 36k; the 98->118GB sawtooth the
admission valve then declined against at 60k).

Full positional latents need no companion state: the block lane
reconstructs them positionally and the adoption lane re-types them on
every hit today. Exempt them symmetrically at all three capture sites
(inline vmlx#109, media clean, deferred) and both splice sites (paged
fetch, clean-store base — exempt slots keep the block-reconstructed
content). Sliding (windowed) latents and true recurrent GDN state keep
their companion coverage unchanged. The companion key gains a
schema-v2 salt so pre-exemption entries can never splice misaligned
layer order into the new mapping.
J
Jinho Jang committed
1c282ae236496d8f73904b4570243ab6eb77fc59
Parent: dce7bf8