[kashishhora]

Summary

Docs follow-up to Google's new WebMCP tooling (webmcp-extension and webmcp-skills).

• README: a short "Testing in Chrome" note. Google's extension reads document.modelContext from an isolated-world content script, so it only sees native registrations and never the polyfill. With chrome://flags/#enable-webmcp-testing on, the provider uses the native API and everything shows up.
• README / api.md: a handler that returns isError: true counts as a success in the hook (lastResult + onSuccess). Only a thrown error sets state.error and calls onError. This wasn't written down anywhere.
• api.md: cross-origin iframes need allow="tools", because the spec gates WebMCP behind the tools Permissions Policy feature, which defaults to 'self'.
• add-tool skill: tips adapted from Google's guidance. Covers keeping results small, waiting for UI updates before returning, throwing actionable errors, what + when descriptions, start_* vs act-now naming, and accepting natural input like "next Friday".

Docs only, no code changes. This doesn't touch the annotation lines that #74 edits.

Test Plan

• pnpm lint
• Checked the isError/onSuccess behavior against runHandler in useMcpTool.ts