studio: work inside opencontroller: its roles, its errors, the studio's menu

[patel-lyzr]

What this does

Makes the studio work fully inside opencontroller, which now shows it under Fine-tuning (NeuralgoLyzr/opencontroller#41). Opened on its own, the studio is unchanged.

Embedded, the studio now:

• Hides what the host's role would refuse (allows() in lib/embed.ts):

  • viewers read;
  • operators add datasets and models, train, cancel, take the weights and use the playground;
  • administrators mint machine tokens.

  The host's proxy is what decides; the studio only hides.

• Shows the host's refusals:

  • api.ts reads RFC 9457 errors (detail, title) beside its own error, so refusals are no longer blank;
  • a 403 is shown as a toast in the host, including where a page dropped the error;
  • the artifact download no longer calls alert(), which a framed page can't use.

• Sends its menu to the host over oc-embed/1:

  • menu carries its sections, the island's actions the role allows (Hugging Face token, Clean VRAM), and Docs/GitHub;
  • the host shows the studio's own menu instead of a copy;
  • an action chosen there comes back as action and runs in the studio, where its dialog is.

• Shares its island rows: the Hugging Face token and VRAM rows are now hooks, shared by the island and the host's menu.

Also in this PR:

• the menu now leads with the Playground (its own commit);
• the README and CLAUDE.md explain how to connect the studio to opencontroller;
• the shipped UI is rebuilt.

Testing

• tsc -b passes, and lint shows the same problems as main, with none new.
• Locally, framed in opencontroller (headless Chrome):
  • every section works from the host's menu;
  • as a viewer: change controls hidden, the playground composer disabled with a note, Train and Machines show why;
  • the host's Clean VRAM freed memory here, and its Hugging Face token opened the studio's dialog;
  • no page errors.

🤖 Generated with Claude Code