docs(fold): row 4 names CUDA's golden-parity models by manifest digest too

[glennneuber]

This PR adds manifest digests to the CUDA half of the fold record's status row 4. #408 did the same for the Metal half. Docs only, one line.

Row 4's CUDA half reports "the vision goldens identical to 0.34.2's" but named no model digest. ADR 0038's decision 1 asks for one. CUDA's goldens measured two models, the ones the record's gate 4 section names:

model	manifest, the CUDA host's store	last modified
gemma4:12b-nvfp4	117d0d84cf2a	2026-08-21
gemma4:31b-nvfp4-tower4bit	637cc0ff1570	2026-09-19

• Source. Read on 2026-09-29 through /api/tags on a canary that mounts the store, at the start of the post-deploy verification run. Production's container was not touched.
• They are what gate 4 measured. Both manifests were last modified before gate 4 ran on 2026-09-25, and nothing in the store has changed since.
• Same weights on both hosts. The two digests equal the Metal host's 12b-nvfp4 and 31b-nvfp4 digests (docs(binaries): name each golden-parity model on the 0.34.4 row by its manifest digest #407, docs(fold): row 4 names each Metal golden-parity model by its manifest digest #408), so both hosts' goldens measured the same weights. The row now says so.

Checks:

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

ai-server/mlx-cuda

🤖 Generated with Claude Code