docs(fold): v0.34.4's record reads done on all three hosts; ADR 0045's knob line names the MLX hosts

[glennneuber]

This fixes three of the five follow-ups from #402's re-review (comment): items 1, 4 and 5, the ones in text the CUDA host keeps. Docs only, two files.

• The fold record's preamble.
  • It said "In progress on the CUDA host — please do not start a parallel 0.34.4 fold", under "Status (2026-09-25)".
  • It now says the fold was merged in fold: upstream v0.34.4 — llama.cpp b11081, MLX 59d600b5, XGrammar 0.2.7 #375, tagged, and deployed on all three hosts on 2026-09-28.
  • The heading now reads "Status (2026-09-28, deployed on all three hosts)", as 0.34.2's record does. Nothing in the repo links to the old heading's anchor.
• Status rows 5 and 6 get the sentence break they were missing before the Metal text.
• Open item 5 still read gfx1151's gate 6 think-on as work "into 2026-09-26". The record's gfx1151 gate 6 section says it finished on 2026-09-27, on all five GGUF models. The item now says gates 4–6 are done on gfx1151 and on Metal.
• ADR 0045 said the knob stays at 0 "on every host". gfx1151 serves no MLX, so the line now names the two hosts that do: CUDA and Metal.

Items 2 and 3, the golden models and the ./mlx/... tests, belong to the Metal host and are in #404.

Merge order with #404. #404 changes status row 4, and this PR changes rows 5 and 6, the next two lines. A trial merge (git merge-tree) reports a conflict in the fold record, which git reports whichever PR merges second. To resolve it, keep both sides. If #404 merges first, I'll rebase this PR.

Checks:

• check_source_paths.py --changed-since origin/main is clean.
• The name scan is clean.
• The in-page anchor #gates-46-on-gfx1151-2026-09-25 resolves; row 6 already uses it.

ai-server/mlx-cuda

🤖 Generated with Claude Code