feat(ops): admit the Q6 fused gate_up shape and evaluate it unfused

[bingchengcc]

What

mlp/gate and mlp/up are 43% of the text weights, and the registered qwen3_8_27b recipe gives
them Q4 - 4.25 bits per weight, 11.5% relative weight error. Q6 reaches 2.6% error at 6.25 bits per
weight. No artifact could use that mixture: linear_swiglu has no Q6 variant, and the generic Q6
GEMM registered only the two vocabulary shapes and the Vision stem.

Three commits, which can be reviewed independently:

• feat(ops) - registers n=34816, k=5120 as a per-shape Q6 file following the existing
  convention, and evaluates a Q6 gate/up pair through the FFN's existing materialized
  decomposition instead of the fused op.
• feat(convert) - the qwen3_8_27b_q6 recipe, the guide section, and a characterization test.
• perf(ops) - tunes the new shape's dispatch over the hot interval and adds the retained
  performance report.

Why the unfused route

linear_swiglu has fp8/nvfp4/q4/q8 but no q6. The FFN already contains an equivalent
decomposition for MTP, so a Q6 gate/up reuses it rather than adding a fifth fused variant. That is
not a compromise: forcing the fused Q4 route through the same decomposition changes the corpus
perplexity below from 1.524538 to only 1.524618.

Tuning

The shape was first registered with the ladder copied from the vocabulary shape, which suits a
geometry with seven times the row-blocks. Measured through production dispatch at every valid T
from 1 to 128:

T	vocabulary ladder	tuned	change
56	330.528 us	228.096 us	-31.0%
64	330.880 us	234.624 us	-29.1%
80	332.544 us	253.792 us	-23.7%
96	334.496 us	274.496 us	-17.9%

docs/maintainer/examples/q6-linear.md is the retained report: the curve at every point, the
priority points and both bulk anchors with the bench's logical byte and FLOP accounting, and one
limitation the tuning did not address. T=8..48 runs at 659..701 GB/s against 1320 GB/s at T=1, so
that interval is bound by the k128 MMA route's weight read rather than by its column capacity.

Evidence

Build: cmake --preset release, then -DBUILD_TESTING=ON; clean, CUDA 13.4 / sm_120a.

Op qualification: tests/ops/linear/test_q6_a16.cpp covers the new shape against the suite's FP32
oracle at the selector's ladder boundaries (1, 4, 5, 6, 7, 8, 9, 16, 17, 24, 25, 32, 33, 48, 49,
50, 128, 129). ninfer_linear_q6_a16_test passes, and was re-run after every ladder change.

Converter: pytest tests/convert - 33 passed, 5 of them new in test_official_recipes.py.

End to end, ninfer-perplexity over a 47,917-token corpus with --context 4096 --stride 2048 --kv-dtype int8:

recipe	MLP gate/up	perplexity
qwen3_8_27b_nvfp4	NVFP4	1.530857
qwen3_8_27b	Q4	1.524538
qwen3_8_27b_q6	Q6	1.520765

The artifact is 19.62 GiB and serves 131,072 tokens of int8 KV with 2.63 GiB still free on a single
RTX 5090. Chat, arithmetic and tool-call smoke checks pass.

Scope notes

• The routing predicate is a pure extension: mtp == true and every non-Q6 format execute the same
  statements as before, so existing artifacts are unaffected.
• Only gate/up moves. mlp/down and the output projections were also measured at Q6 and rejected:
  linear_add's materialized decomposition rounds the projection to BF16 before adding the
  residual, which costs +0.0021 perplexity, more than the weight gain returns. The vocabulary
  endpoints stay Q8, which at 8.5 bits per weight already outranks Q6.
• The tuning measurement ran with an idle resident NInfer server on the same GPU; the bench flushes
  256 MiB of L2 before every sample.

Not included

An unrelated local change (ModelOpt NVFP4 import) is deliberately not part of this branch.