K2.A.1 evidence: KakeyaLattice KL on/off A/B on vast H200 (1k/4k/16k)

NIAH validation A/B (KL OFF baseline vs KL ON D4 Q=38), gemma-3-1b-it,
SDPA, N=20/arm, default §11.12 ladder.

  ctx   tokens | v04 recall OFF | v04 recall ON | Δrecall | thr OFF | thr ON
  ~1k    1428  |     1.000      |     1.000     |  +0.0pp |  9.92   |  7.72
  ~4k    5598  |     0.350      |     0.300     |  -5.0pp |  4.89   |  4.36
  ~16k  21475  |     0.600      |     0.600     |  +0.0pp |  0.95   |  0.93

Gate (b) [binding]: KL preserves v0.4 recall — 0pp at 1k and 16k; the
-5pp at 4k is single-sample granularity at N=20 (7/20 vs 6/20), not a
real regression. KL ON throughput < KL OFF as expected: K2.A.1 is the
stateless round-trip (compress+decompress per step, no caching); gate (c)
throughput is K2.A.2's target.

NOTE: this evidence was produced with the CUDA bf16 dtype fix for
_round_trip_resident_through_compressor (cast round-tripped K/V back to
the resident dtype/device). That fix is on the K2.A.1 branch but is NOT
yet on main (PR #83 merged before it) — main's KL-ON path will crash on
CUDA bf16 with index_copy_ dtype mismatch until the fix lands.

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: cursoragent