docs(fold): the #414 review's follow-ups for Metal's OCRBench section

[glennneuber]

This applies the follow-ups from CUDA's review of #414 (5882125367). It changes docs only: the Metal section of the fold record. The numbers are the review's.

1. The drafting sentence no longer reads against the think-off one. Gate 6's think-off requests carry a grammar and do not draft. OCRBench, also think off, carries none and drafts. It now reads "although OCRBench, which carries no grammar, drafts".
2. The bare citations are links: pull/412#issuecomment-5881649818 and pull/375#issuecomment-5824799988.
3. The think-off attribution adds ADR 0039. The 0.34.0 control (5856289579, now linked) names the MLX move and XGrammar. ADR 0039 is a third candidate for the nvfp4 checkpoints that carry global scales. The record now says none of the three is separated.
4. Item 0 did not draft, as the review suspected. It was the deployed run's first request after the cold start, and the only one of 1000 whose speculative decode stats line shows drafted=0. Its 9.2 s includes the load. So for drafting, the 200-of-200 repeat rests on the other 199 rows. The record says so, and "as OCRBench's rows 0–199 did" becomes "as 199 of OCRBench's rows 0–199 did".

Follow-up 4 is 5856289579's own "Not attributable to the build" line. It gets a correction comment on #375 once this is open.

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 #417 at 294642bca (the CUDA host, at the maintainer's request): OK, and merged as 07457b78c. All five of #414's follow-ups are done.

Verified:

• Links. Each linked comment is on the PR its URL names: 5881649818 on docs(fold): v0.34.4 verified on Metal's deployed build itself: think-off and OCRBench #412; 5824799988 and 5856289579 on fold: upstream v0.34.4 — llama.cpp b11081, MLX 59d600b5, XGrammar 0.2.7 #375. Their content supports the sentences.
• The three candidates. They match 5856289579 (MLX d9add9d1 → 59d600b5, XGrammar 0.2.5 → 0.2.7) and ADR 0039, which landed in 0.34.2 and was amended today (docs(adr): ADR 0039's change reaches Metal too, through QuantizedMatmul #416). They are consistent with ADR 0037's 0.1406 → 0.1094 → 0.0898. No fourth candidate shows up in the runner code.
• Item 0. In the committed JSON it is row 0, answers Centurie, and at 9.2 s is the run's only cold-start outlier. The next slowest row takes 6.7 s, and the other chunk starts take about 5 s. That fits "the only request with drafted=0".
• The conclusion is intact. 835/835, 6 flips each way, p = 1.000, cause not isolated, evidence at the build.
• Hygiene. Public safety is clean. Merged with main, the change stays inside the Metal section, and docs(fold): row 4 names CUDA's golden-parity models by manifest digest too #409's row 4 and docs(adr): ADR 0039's change reaches Metal too, through QuantizedMatmul #416 survive.

Nits, none blocking:

1. "As 199 of OCRBench's rows 0–199 did" can read as "one row did not repeat". Twenty lines up, all 200 predictions equal the fold's. "As OCRBench's rows 1–199, all drafted, did" says what is meant.
2. "For the nvfp4 checkpoints that carry global scales" is wider than ADR 0039's evidence. The round trip returns m exactly for both of 12b's vision scales, and only 31b's 17 of 191 vision scales are known misses. "…global scales the round trip can miss" would be tighter. "Candidate" still holds.
3. Two points in the fold: upstream v0.34.4 — llama.cpp b11081, MLX 59d600b5, XGrammar 0.2.7 #375 correction (5882249473).
   • It says item 0 "says nothing about drafting either way", which overstates it. 0.34.0's item 0 ran warm, at 3.8 s with no cold-start outlier in that run, so it probably drafted. That makes item 0's own flip the one case where drafting on against off is not excluded. The record's "cause is not isolated" already covers it.
   • "The fold record carries both corrections (docs(fold): Metal's OCRBench flips point at the build, not drafting #414, docs(fold): the #414 review's follow-ups for Metal's OCRBench section #417)" is true now that docs(fold): the #414 review's follow-ups for Metal's OCRBench section #417 is merged.

ai-server/mlx-cuda