perf(GDN-MOE-BF16-OUT): the GDN recurrence output and the z gate are bf16 on MoE too, as both references keep them

[localai-bot]

The GDN recurrence output dcore and the z gate were held in f32 on every MoE
checkpoint while both vLLM and SGLang keep them at the bf16 model dtype. That is
the dtype-polarity case the protocol names: a token gate cannot see it, because
the tokens match and the goldens pass while the path moves twice the bytes on the
largest per-layer GDN activation, twice over -- written by the recurrence, read
by the gated RMSNorm.

GdnOutDType() now takes no model-shape argument, so bf16 is the default on both
arms and VT_GDN_OUT_BF16=0 is the rollback. The redundant dense_model term is
gone from the packed-decode eligibility, which neither reference engine gates on
model shape either.

Removing that term does NOT make packed decode reachable on MoE, and the change
says so rather than implying otherwise. has_packed_ba still gates it, and
in_proj_ba has exactly one writer, the dense loader. That is #1169, filed and
owed, not fixed here.

Two review rounds, both adversarial and both productive. The first found the
advertised lever simultaneously unpinned and self-reddening: severing
GdnOutDType() from its env parser left every suite green, and running the
documented VT_GDN_OUT_BF16=0 rollback red the focused gate, so the rollback
looked like a regression. The repair moved the resolver into detail::, gave the
test its own independent env reader so a drift in the product parser reds instead
of silently agreeing, and registered a second ctest arm under the opposite
environment. The second round found the first repair had exchanged one inaccurate
comment for another, and a sixth surviving copy of the model-shape claim this row
falsifies.

Both rounds are mutation-proven. Severing the resolver reds the f32 arm; forcing
f32 reds the default arm; inverting the parser reds both arms plus the
independence assertion; deleting the production call site reds both arms on
reachability. The two ctest registrations run the same binary and return opposite
verdicts, which is what proves the environment property actually arrives.

Every GPU gate is OWED, not passed, and the row lands at GATING rather than
DONE. There is no nvcc and no GPU on the build host, so nothing here is
measured: the 27B 235/235 dense inertness control, the 35B correctness gate, the
VT_GDN_OUT_BF16 0-versus-1 A/B per leaf, and the nsys memory-format
confirmation are all recorded PENDING on a GPU host. docs/BENCHMARKS.md says
"No number, none claimed". The one control that would prove the dense arm
byte-identical, test_qwen27_paged_engine, was run and SKIPPED with
assertions: 0 -- a skip wearing a pass -- and no result rests on it.

Refs #1168, #1169, #1170, #521.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:Opus-5 [Claude Code]