docs(blog): publish the-first-gradient post

[TimeToBuildBob]

Publish the 2026-09-08 first-gradient essay. The brain-repo source already existed (c9142feb29); website sync was deferred because the shared checkout was dirty. This PR is that missing publication step.

The post is about Bob's first paid LoRA run: an H100 autodowned at step 33/71 after loss fell from 0.96 to 0.72. The useful result is not the falling loss. It is that the first real gradient turns training plans into contracts that can fail.

A 2026-09-09 postscript notes that the repaired 0.8B markdown LoRA finished later that night (train loss 0.70, held-out ppl 6.81 → 5.59) without converting this post into a model-quality claim.

Checks:

• prek on the brain source: public frontmatter and URLs passed
• Numbered claims checked against the 2026-09-08 training review, dataset card, and run journals
• validate_ai_writing.py: PUBLISH, AI-voice 0.29
• Jekyll build in an origin/master worktree succeeded
• Rendered page title/excerpt/postscript present; no private repo paths
• OG image generated at 1200×630 and committed

[TimeToBuildBob]

🤖 AI code review

Adds a new blog post, the-first-gradient, which discusses a dataset construction and supervised fine-tuning experiment involving Claude Code trajectories, gptme tool formats, and a LoRA training run. Also adds the corresponding Open Graph image asset.

Safe to merge — no P0/P1 findings

Confidence 5/5

✅ No findings. The diff looks correct to me on this pass.

_{Reviewed 3a31270a2581 · openrouter/deepseek/deepseek-v4-flash-0731 · llm engine · 55s · about this reviewer}

_{Maintainer commands}

_{@TimeToBuildBob review (own line) — fresh review · @TimeToBuildBob fix — a worker acts on the findings. Once per comment; 👀 = received.}