Harden Braintrust tracing and online scoring

[samsel]

Summary

• codify Braintrust online scoring setup with shared scorer slugs and a repeatable setup command
• align Braintrust project naming across production tracing, evals, and scorer pushes
• convert xAI image generation into a child Braintrust span with image artifacts and metadata instead of a flat manual log

Verification

• pnpm exec tsc --noEmit
• pnpm test:unit
• pnpm check:boundaries
• pnpm braintrust:setup-online-scoring:dry-run -- --json
• pnpm exec tsx --import ./evals/server-only-shim.mjs -e "import('./braintrust/online-scorers.ts').then(() => console.log('online scorers loaded'))"
• pnpm build