Add design/screenshot probe type (render + visual judge)

[kwadwoadu]

What

Adds a first-class scoring: screenshot probe type so modelfit can benchmark models on UI/design work. The candidate model returns a self-contained HTML artifact, modelfit renders it headlessly to a PNG, and a multimodal judge grades the rendered image against the rubric. Text probes are unchanged (fully backward compatible).

Flow

design probe -> model returns HTML -> render (Playwright CLI / Chromium) -> result.png
                                                                                |
                verdict <- multimodal judge (sees the image) <-----------------+

Changes

• common.sh: scoring_for_probe, b64_file, render_html (MODELFIT_RENDER_CMD override -> Playwright CLI -> Chromium/Chrome, falling through on failure; MODELFIT_RENDER_VIEWPORT)
• run.sh: screenshot probes write result.html, render to result.png, record rendered:true; render failure -> render_error
• judge.sh: multimodal judge (anthropic image / openai image_url) via jq --rawfile; body streamed over stdin (--data @-) to avoid the 128KB single-arg limit on base64 images; text-only fallback when no PNG
• bin/render.sh + modelfit render dispatch
• probes/example-design.md, prompt/README/limitations docs
• tests/design.test.sh (mock renderer + provider, zero browser/network); mock tests/curl reads --data @-; selftest covers the new paths

Verification

• ./bin/selftest.sh PASS, shellcheck CLEAN
• End-to-end proven live: a model produced pricing-card HTML -> rendered -> multimodal judge graded the pixels (pass, quality 5)