quantize: default to trellis, and to the variant that has a fused kernel

[cnygaard]

--codebook defaulted to e8_shell while the README recommended trellis and the checkpoint table marked a trellis build as "fastest GLQ decode". Someone following the docs and running glq-quantize got a lattice checkpoint.

Trellis (QTIP TCQ) has been the recommended path since v0.7 and wins where the bits are scarcest — SmolLM3-3B at 2 bpw is PPL 11.94 vs 13.79.

Two defaults, not one

GLQ_TRELLIS_VARIANT defaulted to hyb, but the fused CUDA kernels consume a 3INST checkpoint — a hyb one falls back to the pure-torch decode, "correct but materially slower". Defaulting to trellis alone would have shipped checkpoints with no fast path, which is worse than the lattice default it replaced. Quality is a wash: 3INST-4 matches HYB-4 on SQNR and PPL.

Consequences handled, not discovered later

scripts/batch_quantization.py omitted --codebook when it equalled e8_shell, relying on the CLI default matching its own dataclass default	now always explicit — otherwise a job declaring e8_shell silently becomes trellis, and any fractional-bpw job fails outright
The resume run-key hashed "" for the variant	now hashes the real one. This invalidates existing trellis resume stores, which is correct: they hold hyb work that must never be resumed into a 3inst run
The mixed-precision refusal didn't say what to use instead	now names --codebook e8_shell. --bpw 3.5 used to work with no flag; without the pointer the new default makes it a dead end
README claimed "fractional like 2.5 also works"	corrected — uniform integers only by default

build_parser() is split out of main() so CLI defaults can be asserted without running a quantization; there was no test surface for them before.

Also in this PR

The Blackwell/FlashInfer caveat next to the vllm serve examples: on sm_120 (RTX 5090, RTX PRO 6000) with no CUDA Toolkit, vLLM's sampler ships no prebuilt kernel, compiles one at startup, and takes the engine down before the first token — GLQ's own kernels load fine. glq-chat detects and falls back; vllm serve and LLM(...) do not, and the README now says so.

Verification

• 6 new tests pinning the defaults, the still-reachable lattice codebooks, that the default bpw is one trellis accepts, the variant default, the env override, and that the refusal names e8_shell.
• Full local suite: 1085 passed, 882 skipped, 0 failed.
• Lint: the files touched are clean; the repo's pre-existing 70 findings are unchanged in count.

Note for whoever cuts the next release

This changes what glq-quantize produces for anyone who does not pass --codebook. That is the intent, but it is a behaviour change worth a release-note line rather than a footnote.