qwen38: the int8 trunk carries one scale per 64 weights (grouped fmt 1 on the GPU), perplexity-neutral

[kreuzzelg]

Follow-up to #1424 (merged): one commit on dev. Also takes the doc note from the #1424 review (greedy-identical is a measurement, not a guarantee; the two reference knobs).

Why

#1424 quantizes the dense trunk to int8 per row. On wikitext-2 (8 x 512, 2048 scored tokens) that costs +1.9 %
perplexity (1.880 against 1.845 BF16). Probing the rows showed the error is spread over the trunk, not one
matrix class: the expert tier is neutral, lm_head contributes nothing, keeping the routers in BF16 recovers 0.3
points, and the GPU's summation order is 0.2 of the 1.9 (1.877 on the CPU reference path). One scale per 64
weights along the input brings it to 1.852 (+0.4 %), the same eight chunk NLLs on the GPU and on the CPU
reference path, greedy text identical to BF16 over 100 tokens.

What changes

• Backend: fmt 1 uploaded with a group size takes the grouped int4 branch of quant_matmul (scales
  [O][ceil(I/gs)], group subtotal times scale); the per-element i / gs becomes a shift when gs is a power of
  two (75 -> 69 ms resident-mm per token on the 553 trunk matrices; per-row was 63). The upload records gs for
  fmt 1 as it does for fmt 4. No new symbols, the DLL loader table is unchanged.
• Tier: qt_dense_init(q, sc, I, O, device, gs); gs > 0 uploads through upload_g and the budget counts the
  scale table; qt_dense_matmul passes the handle's gs. The backend's cached-tensor check refuses a grouped
  tensor asked with gs 0 -- the first hardware run of this branch failed all 553 matrices exactly that way, so
  the fake backend now mirrors the check and the dense test has the case.
• Engine: Q38_TRUNK_GS (default 64, 0 = per row) drives the quantizer, the CPU reference path
  (Q38_TRUNK_CPU_INT8), the selftest, the offered bytes and the placement sum.
• Planner: the trunk estimate carries the scale table (+1/16).
• Docs: knob, trunk paragraph, the perplexity table.

Measured (RTX 3070, cap 224, the 315+100-token prompt of #1424)

	per row (#1424)	gs 64
trunk in VRAM	3.96 GiB	4.21 GiB
resident experts / VRAM share of routes	352 / 22.3 %	312 / 20.6 %
resident-mm / lm-head, ms per token	63 / 5	69 / 7
decode tok/s	1.91	1.88
greedy text, 100 tokens, against BF16	identical	identical
selftest, 553 matrices, GPU against CPU int8 of the same rows and scales	1e-7	1.5e-7
perplexity wikitext-2	1.880 (+1.9 %)	1.852 (+0.4 %)

Tests: the ten tests/test_qwen36_tier_* (dense with the grouped case), test_qwen38_tier_engine (trunk grouped
on the fake tier, cosine 0.99999), Python suites; CUDA builds of qwen36 and qwen38.

Thanks to Claude Code for the help along the way.

https://claude.ai/code/session_01X2HPETYXNwT4RmQkdMMyPS