docs(binaries): name each golden-parity model on the 0.34.4 row by its manifest digest

[glennneuber]

This adds the model digests to the 0.34.4 row of BINARIES.md. It is note 3 from #404's review. Docs only, one line.

ADR 0038's decision 1 asks any record that reports a measurement to state the manifest digest of each model measured. The row's vision golden parity sentence gave a digest for 31b only. The Metal host supplied the other three from its store (#404 comment):

model	manifest	max sampled element delta
31b-nvfp4	637cc0ff1570	0.0898
26b-nvfp4	c8656f50f0a6	0.0508
12b-nvfp4	117d0d84cf2a	0.0625
26b-mlx-bf16	fa29b1495bbb	0.0625

Cross-check.

• The 12b, 26b and 31b digests equal the "local manifest" column of the store audit in ocrbench-quantisation-ladder.md.
• That table also shows that 26b's c8656f50f0a6 is a quantised-tower artifact, as 31b's is. The library later re-published both tags with a bf16 tower. So the row now calls both the 4-bit-tower artifacts; before, it said so only for 31b.
• 26b-mlx-bf16's fa29b1495bbb has not been recorded in the repo before.

The fold record's status row 4 carries the same sentence, still with only 31b's digest. This PR leaves it alone.

Checks:

• check_source_paths.py --changed-since origin/main is clean.
• The name scan is clean.
• The row keeps its seven column separators, and its backticks balance.

ai-server/mlx-cuda

🤖 Generated with Claude Code