Evidence — beui.dev interaction reference for the frontend skill

What was tested

1. Manifest partition gate (the machine consumer of designOriginals): bun test packages/shared-skills/frontend-thirdparty-manifest.test.ts run BEFORE the reference file existed, immediately after adding interaction-skill.md to designOriginals in scripts/frontend-refs-manifest.mjs.
2. Builtin-artifact byte-equivalence gate: packages/skills-loader-core/.../shared-skill-extraction.test.ts run after editing the shared SKILL.md and before mirroring the loader-core artifact and TS description.
3. Full battery after implementation: shared-skills package tests (manifest, materialize, provenance, depersonalization, upstreams), skills-loader-core builtin-skills tests, omo-opencode builtin-skills tests, and omo-opencode/src/shared-skills-package.test.ts (packaging + every-skill-parses pin).
4. beui.dev consultation recipe: live curl against every endpoint the new reference documents (/llms.txt, /r, /r/{slug}, /r/{slug}/raw).

What was observed

• RED: 2 failing tests for exactly the right reasons — interaction-skill.md not tracked, .gitignore missing the unignore line (red-manifest-test.txt).
• RED 2: byte-equivalence failure between shared SKILL.md and the loader-core artifact after the shared edit, before mirroring.
• GREEN: full battery passes (green-test-battery.txt); depersonalization gate passes; description stays 787 chars (under the 1024 Codex cap, enforced by the extraction test).
• Endpoints return real payloads: registry index JSON, component detail JSON with files/dependencies, raw TypeScript source (beui-endpoint-verification.txt).

Why it is enough

The change is skill prose plus its machine-consumed registration (manifest, gitignore, ATTRIBUTION, TS description string). Every machine consumer is exercised by its own committed test: the DMCA partition, the materialization set, the byte-equivalence pins across harness copies, packaging inclusion, and frontmatter parse/length budget. Per .omo/rules/test-discipline.md PROMPT TESTS and the shared-skills AGENTS.md note, the prose content itself ships on review, not on wording pins. The consultation recipe is verified against the live service the reference documents, mirroring the lazyweb.md precedent.

What was omitted

• No live opencode/codex session drive: the change adds a reference file and routing prose; skill loading mechanics are unchanged and are covered by the loader/packaging pins above. No new tool, hook, or config surface is introduced.
• beui.dev raw source bodies are truncated in the endpoint capture to avoid vendoring third-party source into the repo.