shape_probe on the SHIPPED build (I8_STREAM_K unset, no I8_ONLY_* narrowing),
branch feat/gguf-grid-streamk-parity off feat/gguf-grid-alu-reduction@8ed8b1419,
RTX 4090 sm89, 128 SMs. blocks/SM is the occupancy probe's own figure; a
stream-K column of 0 means the arm carries NO stream-K instantiation, which is
what `_sk_blocks_per_sm` turns into the data-parallel path.

cb=26 is the LUT collapse (bits=4), cb=29 the grid gather (bits=2).

  s247 bits=4 cb=26: dp blocks/SM=3  stream-K blocks/SM=0
  s247 bits=2 cb=29: dp blocks/SM=2  stream-K blocks/SM=0
  s248 bits=4 cb=26: dp blocks/SM=2  stream-K blocks/SM=0
  s248 bits=2 cb=29: dp blocks/SM=2  stream-K blocks/SM=0
  s249 bits=4 cb=26: dp blocks/SM=4  stream-K blocks/SM=4
  s249 bits=2 cb=29: dp blocks/SM=5  stream-K blocks/SM=4
  s250 bits=4 cb=26: dp blocks/SM=4  stream-K blocks/SM=3
  s250 bits=2 cb=29: dp blocks/SM=4  stream-K blocks/SM=4
  s251 bits=4 cb=26: dp blocks/SM=1  stream-K blocks/SM=0
  s251 bits=2 cb=29: dp blocks/SM=1  stream-K blocks/SM=0

s249 and s250 are the 64-wide tiles (TILE_N=64); 247, 248 and 251 are 128-wide.
`SK_COMPILED` in exl3_i8_launch_wg.cuh keys on the TILE -- TN == 64, or the one
256x128, or the 112x128 with WG == 0 -- and not on BI or CBI, which is why the
grid codebook already carries the decomposition at s249 and s250 with nothing
built for it. The LUT arm reads 4 blocks/SM at s249 and the grid arm reads 4
as well: the two are raceable at equal form on this binary.
