v0.3.1

Added

• Docs: connecting Claude on the web, and a ChatGPT section. Which clients can carry a bearer
  key and which cannot, and why: a config file can set a header, a browser connector dialog
  cannot. ChatGPT gets its own section covering the Plugins rename, the developer-mode
  prerequisite and where the toggle lives per plan, and why no configuration works directly —
  OpenAI's documentation states ChatGPT cannot present custom API keys, and expects OAuth 2.1
  against the MCP authorization spec. Both blocked clients get something to do rather than a
  shrug: the desktop app for Claude, a header-injecting proxy for ChatGPT, with a Cloudflare
  Worker to copy and the trade stated — the proxy's URL becomes the credential.

Changed

• The per-key tool scope is a tree. It was a three-column grid of blocks whose columns ran
  to different heights, so a tool's modes wrapped into a shape unrelated to the tool above it.
  Each tool is now one full-width row with a disclosure triangle, a picked/total count, and
  its modes as collapsible children under a guide line. A tool with some but not all of its
  modes shows an indeterminate checkbox; unticking the last mode turns the tool off and
  ticking one turns it back on. The header offers All, None and Expand all. Counts
  are rendered by the server, so the scope is readable with JavaScript off.

Fixed

• The loading skeleton now covers every page. site.js looked for .doc, then <main>,
  then document.body, so only the docs pages showed it correctly: the dashboard got the
  markup with no CSS to position or dim it, and the landing, login and error pages prepended
  skeleton lines above the nav. Every template now renders inside one #page wrapper, which
  the script targets and the stylesheet positions and dims. The docs pages still dim only the
  article, so the sidebar stays readable.
• The site footer linked to the repository, which is private. It links to the tool reference
  instead.