`arm.build_declines` on a card -- the question a BOOT asks before it binds, not
a table read. Box 10.2.0.105 (`arbi-whitebox-105`), RTX 4090 sm89, sole tenant,
image registry.arbi.work/arbi-serve:test-ef6b061, extension built from the
branch's own csrc.

branch feat/gguf-grid-serve-iq2xxs (the promotion):

  int8 kernel: loaded

  IQ1_M    cb=28 bits=4 tile=nibble       SERVES
  IQ2_S    cb=26 bits=4 tile=nibble       SERVES
  IQ2_XS   cb=26 bits=4 tile=nibble       SERVES
  IQ2_XXS  cb=29 bits=2 tile=grid:iq2xxs  SERVES      <-- the promotion
  IQ3_S    cb=21 bits=4 tile=nibble       SERVES
  IQ3_XXS  cb=27 bits=4 tile=nibble       SERVES
  IQ4_XS   cb=20 bits=4 tile=nibble       SERVES
  Q2_K     cb=22 bits=2 tile=bitstream    SERVES
  Q4_K     cb=22 bits=4 tile=bitstream    DECLINES (see below)
  Q6_K     cb=23 bits=6 tile=bitstream    SERVES

  IQ2_XXS gate/up  5120x17408: GEMV tag=16 rows=[1]
  IQ2_XXS qkv      5120x10240: GEMV tag=16 rows=[1]
  IQ2_XXS q_proj   5120x12288: GEMV tag=16 rows=[1]
  IQ2_XXS down     17408x5120: GEMV tag=16 rows=[1]
  IQ2_XXS k/v      5120x1024:  GEMV tag=16 rows=[1]

Q4_K'S DECLINE IS THE PROBE'S QUESTION AND NOT A REGRESSION. The probe asks
every type with `min_form="none"`, and Q4_K is min-BEARING: no ladder entry
contracts a per-32 weight group with no min, which is the correct refusal for
the question asked. A boot asks Q4_K with its own measured `min_form`. The same
probe run against the nibble control answers identically for Q4_K, which is what
makes this a property of the probe rather than of the promotion.
