feat(prefill): add bit-packed KakeyaLattice snapshots

[FluffyAIcode]

Summary

• add an explicit KAKEYA_LATTICE_D4 prefill compression codec to the proto/API
• quantize heterogeneous Gemma-4 K/V tensors with D4 Q=38, bit-pack codes, and reconstruct standard portable KPKV snapshots on import
• frame and checksum packed payloads without weakening existing zlib/raw validation
• pin KakeyaLattice 1.6.1 git commit because bit-packed support is not yet published to PyPI
• require a distinct kakeya-prefill-v3-kl-d4-q38 compatibility format and configure head/worker deployment consistently

Real snapshot evidence

On the live 145-token Gemma snapshot from allens:

• raw KPKV: 33,197,920 bytes
• current zlib wire: 26,267,657 bytes
• KakeyaLattice framed wire: 12,950,269 bytes
• raw-to-wire ratio: 2.56x
• encode: 1.71 s; decode: 0.41 s
• relative K/V MSE: 5.98e-4

This improves storage and warm-hit transfer. It does not reduce allens model prefill compute time.

Tests

• platform-neutral suite — 817 passed, 1 skipped
• TypeScript SDK — 54 passed, 100% coverage; typecheck passed
• real KakeyaLattice 1.6.1 heterogeneous 256/512-head snapshot round trip — passed
• proto lint, format and generated stub drift — passed

Made with Cursor