Glyd v0.32.1: Gemma 4 writes faster, glyd fetch, clearer refusals

Glyd v0.32.1 is out.

Gemma 4 writes faster. The engine’s attention for a step, and the matrix products of a step and of a check of drafts, now read closer to the card’s memory speed. One user, RTX 4080 SUPER, Gemma 4 12B at wren, without the draft model:

	v0.32.0	v0.32.1
Chat	81.2 tokens a second	88.9
After a 4K-token prompt	74.7	82.1
After a 30K-token prompt	59.5	64.0

The output is the same as before, with and without drafting.

Fixes from our own install test of v0.32.0:

• glyd fetch MODEL downloads a model without running it. The pip-installed glyd sent it to the compression program, and glyd fetch glyd/gemma-4-12B-wren asked for a repo that does not exist.
• glyd verify glyd/gemma-4-12B-penguin downloads the repo first when it is not on the computer yet.
• glyd run of a Gemma 4 or lossy glyd/ repo on a GPU without room for it says so and stops, instead of handing the repo to vLLM, which cannot read it.

pip install -U glyd   # or: brew upgrade glyd
glyd fetch glyd/gemma-4-12B-wren
glyd verify glyd/gemma-4-12B-wren

Full details are in the release notes and the changelog; what is not done yet is on the known gaps.