Complete the external-tier Table 1 re-measurement after epic #324 #357

Finish re-measuring the external-tier rows of the paper's Table 1 under the epic #324 code. These are the single-model log-likelihood gap and component correlation on ds002718 sub-002.

Status

The run started on the CUDA workstation at 09b8247 (the epic head after Phase 12, which has the final fitting code):

• command: benchmarks/reproduce_table1.py --tier external --threads 16 --device cuda --native-version v0.3.3;
• protocol: unchanged, 5 seeds at 2000 iterations, Newton off, block_size=512.

Interim per-seed results (pamica on CUDA float64 against the pinned v0.3.3 binary):

Before the epic, Table 1 reported an LL within ~0.0005 of -3.6993 and a component correlation of 0.998.

Remaining

• Complete seeds 204-205, and record the report.json and log under .context/issue-351/raw/.
• Update Table 1's two external rows in paper.md, the single-model bullets in docs/guides/validation.md, README's single-model figure, AGENTS.md, and the external wall-clock in benchmarks/README_dimsweep.md.
• Drop the "measured before epic Epic: make MLX a first-class, end-to-end backend #324" labels that Re-measure parity and sweep docs for epic #324 #351 put on those figures.

Follow-up to #351 (epic #324).