THE ARITHMETIC, WRITTEN DOWN BEFORE THE BOOTS, so the live numbers are compared
against a prediction rather than explained by one.

Pack: Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf. Its five IQ2_XXS tensors
(`pack-iq2xxs-cells.txt`):

    3 x 5120x17408  (gate/up)   267,386,880 weights
    1 x 5120x12288  (q_proj)     62,914,560
    1 x 5120x10240  (qkv)        52,428,800
                                ------------
                                382,730,240 weights

The FIELD plane is the only plane that moves. `d`, the sub-scale plane and the
min planes are byte-identical under both spellings -- `grid_spec`/`collapse_spec`
replace `bits`, `grid` and `tile_bytes` and nothing else -- so they cancel out of
the difference exactly.

    field plane, LUT collapse (4.0 bpw) : 191,365,120 B = 182.50 MiB
    field plane, grid tile    (2.0 bpw) :  95,682,560 B =  91.25 MiB
    PREDICTED SAVING                    :  95,682,560 B =  91.25 MiB = 0.0891 GiB

WHAT THIS IS AND IS NOT A PREDICTION OF. It is the `model.weights` pool and
nothing else. It is not a prediction of the whole card, which also carries the
KV pool, the capture pool, the activation floor and the allocator's own
granularity -- any of which can move by more than 91 MiB for reasons that have
nothing to do with this change, and none of which the spelling touches.

AND IT IS SMALL, BECAUSE THE PACK MAKES IT SMALL. Five of this pack's 504
linears are IQ2_XXS: 382.7 Mw of 24,775.2 Mw of linear weight, **1.54%**. The
saving is half of the field plane of 1.54% of the model. A pack quantised mostly
to IQ2_XXS would see ~25x this; this one does not, and the speed win (three of
the five tensors sit at `gate/up`, the cell the grid arm wins 0.757/0.784) is
the larger part of what the promotion buys here.
