WHY THIS PACK CANNOT BE BOOTED FROM THIS LINEAGE, and what the fix is.

Taken from a REAL boot on box 10.2.0.105, sole tenant, of
Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf. Both arms of the intended A/B refuse at the
same place and for the same reason, which is what says the refusal is the
lineage's and not the promotion's.

-- the engine's own served-type census, from the boot log ------------------

  409 tensor(s) in planes across 10 (ggml type, min form) cell(s):
    IQ1_M/none x1,  IQ2_S/none x16,  IQ2_XS/none x9,  IQ2_XXS/none x5,
    IQ3_S/none x144, IQ3_XXS/none x78, IQ4_XS/none x96,
    Q2_K/f16acc x13, Q4_K/f16acc x39, Q6_K/none x8

  GGUFUnservedError: this build cannot decode ggml type Q2_K (min form
  'f16acc'), which 13 tensor(s) in this pack carry --
  model.language_model.layers.10.linear_attn.in_proj_z.weight is one.
  no ladder entry contracts a per-16 weight group with min_form='f16acc',
  signed_subscale=False, has_subscale...

IQ2_XXS/none x5 is this promotion's five tensors, confirmed by the engine's own
census rather than by the header read.

-- the two lineages are disjoint ------------------------------------------

  feat/gguf-grid-alu-reduction (this branch's base)
      HAS   the grid gather, CB_GGUF_IQ2XXS_GRID (29), the grid tile writer
      LACKS the per-16 weight group that carries a min -> cannot serve Q2_K

  fix/gguf-q2k-per16-fused-min (9 commits, in flight)
      HAS   the per-16 min family -> serves Q2_K/f16acc
      LACKS CB_GGUF_IQ2XXS_GRID entirely (grep: 0 hits in codebook_ids.py)

Neither can boot this pack WITH the grid path. Merging them is what would, and
merging them does not work:

-- THE COLLISION -----------------------------------------------------------

`git merge origin/fix/gguf-q2k-per16-fused-min` leaves five conflicts, and one
of them is not textual. Both lineages allocated the SAME shape ids to DIFFERENT
kernel families:

  exl3_i8_launch_wg.cuh, `wg_tile_k`:

    <<<<<<< this branch
        //: Q3_K's per-16 signed family at six bits, tile for tile.
        case 260 ... case 270: return 256;
    =======
        //: THE PER-16 MIN-BEARING FAMILY (Q2_K).
        case 260 ... case 276: return 256;
    >>>>>>> fix/gguf-q2k-per16-fused-min

s260-s270 mean "Q3_K at a six-bit sub-scale" on one side and "Q2_K per-16 with
a min" on the other. Every table keyed on a shape id -- `WEIGHT_D_WIDTH`,
`WEIGHT_SUB_SCALE_BITS`, `WEIGHT_MIN_FUSED_SHAPES`, `GROUP_SCALE_UNRACED_SHAPES`
(`frozenset(range(265, 271))` here), the raced rows -- therefore says two
different things about the same numbers. Picking a side resolves the text and
leaves the tables meaning one family while the launcher compiles another, which
is the plausible-weights failure this whole programme is built to refuse. It is
NOT resolved here, on purpose.

-- WHAT THE FIX IS, for whoever owns both branches -------------------------

Renumber THIS lineage's s260-s270, not the Q2_K family's. They are the six-bit
sub-scale spelling of Q3_K, which is WITHHELD and unserved: the type table says
`sub_bits=8` for Q3_K because the six-bit spelling raced 1.01-1.18x SLOWER
(`docs/receipts/gguf-grid-gather-int8-arm-race-2026-09-22.md` §3), and
`GROUP_SCALE_UNRACED_SHAPES` holds s265-s270 out of every bind. Moving a family
no boot can select costs nothing; moving Q2_K's would move a family this pack
has 13 tensors of.

-- AND THE DENSE-CARRY ESCAPE IS CLOSED ON PURPOSE -------------------------

`build_ablation.Ablation` can carry a type dequantized, which would have let
both arms skip Q2_K identically and kept the A/B's delta honest. It refuses to
join the cache key by design -- "an ablated repack is not this file's repack and
must not be reachable by the content-addressed key every boot of it resolves" --
so a boot cannot be handed one. That guard is correct and was not worked around.
