docs([issue-2791]): complete the 4B-vs-9B LoRA fidelity run — verdict: stay as-is

[atomantic]

Summary

• Resumes and completes the empirical fidelity run Re-tune LoRA-training defaults (4B base model + steps) after a fidelity run #2791 required before changing LoRA-training defaults: two matched runs (4B vs 9B-8bit base model, 1200 steps each, identical rank/lr/resolution/seed) trained to completion and compared at steps 0/300/600/900/1200.
• Verdict: stay as-is. Neither proposed change (defaulting single-character LoRAs to 4B, or cutting steps from 1200 to ~600) is supported by the data. Surprisingly, 4B did not lose to 9B on identity in this run — but that rests on a single seed/dataset and needs replication before it should move a default every install inherits.
• Separately documents an unplanned finding: both arms peaked at step 900 and regressed by step 1200, pointing users at the existing checkpoint picker rather than assuming the final step is best.
• A local review pass caught and fixed two factual errors carried over from the original (paused) draft: a caption-leak claim that doesn't hold up against the real validation gate, and a throughput figure recomputed from checkpoint timestamps (corrected ~9.4 s/step → ~6.2 s/step for 4B, ~1.6× → ~2.5× vs 9B).

Closes #2791

Test plan

• Both training runs (9f6fce7e… 4B, 151f3631… 9B) ran to completion via the LoRA training API and are recorded with their run IDs/artifact paths in the doc.
• All ten preview-grid sample images read and visually compared directly.
• Caption-leak and throughput claims re-verified against live code (analyzeCaptionInvariants) and on-disk checkpoint timestamps, not just restated.
• Docs-only change — no code paths affected; TRAINING_DEFAULTS and the base-model picker in server/services/loraTraining/runtimes.js are unchanged.
• Local Code Review gate run; a sub-agent adversarial review found 3 issues (all fixed and re-verified clean).