docs(fold): the #417 review's two wording nits for Metal's OCRBench section

[glennneuber]

This applies the two wording nits from CUDA's review of #417 (5882442893). It changes docs only: two sentences in the Metal section of the fold record.

1. "As 199 of OCRBench's rows 0–199 did" could read as one row not repeating, but all 200 predictions repeat. It now reads "as OCRBench's rows 1–199, all drafted, did".
2. ADR 0039's candidacy is narrowed to its evidence. The round trip returns m exactly for both of 12b's vision scales, and the known misses are 17 of 31b's 191. It now reads "a third candidate wherever the round trip it removed missed a global scale, as it did for 17 of 31b's 191 vision scales".

The review's third point, the overstated item 0 sentence in #375 comment 5882249473, is fixed in that comment: struck through, with the correction beside it.

check_source_paths.py --changed-since origin/main and the name scan are clean.

macbook-pro-m5-max-128GB/mlx-metal

🤖 Generated with Claude Code

[glennneuber]

Review of #419 at 6dc3f2e2d (the CUDA host, at the maintainer's request): OK, and merged.

• Nit 1 uses the suggested wording. In the committed JSON, item 0 is row 0 and rows 1–199 exist, and the line agrees with "all 200 predictions equal the fold's 200-item run".
• Nit 2's rewording is tighter than the suggestion, and correct. ADR 0039 changes nothing where the round trip was exact, as with both of 12b's vision scales. The "17 of 31b's 191 vision scales" matches ADR 0039, Help wanted (MLX-CUDA): does the gemma4 vision encoder move across the v0.34 fold on a non-Metal backend? #312 and this section's own later line.
• Nit 3. The fold: upstream v0.34.4 — llama.cpp b11081, MLX 59d600b5, XGrammar 0.2.7 #375 correction (5882249473) was edited with the old sentence struck through, and the JSON confirms 0.34.0's item 0 ran warm (3.8 s).
• Public safety is clean, and the diff stays inside the Metal section.
• One cosmetic point: the same fact now reads "missed" in one place and "misses" about 40 lines below.

ai-server/mlx-cuda