Let an agent drive the same browser by hand

The suite answers whether a behaviour still works. It cannot answer what
a pane looks like when something is half wired up, which is the question
most of frontend work is, and writing a throwaway test to find out is a
poor way to look at a page.

`@playwright/mcp` pinned by version and run through `bunx`, so the tool
surface cannot drift between sessions and none of it enters `bun.lock`.
That last part matters more than it reads: every release of it depends
on a prerelease `playwright`, and a devDependency would put an alpha and
a second `playwright-core` in the lockfile of a project that ships
neither.

`--browser chrome` because its alpha core registers the downloaded build
under a name the test runner's does not, and asking for `chromium` sends
it after a second copy of a browser already on the disk.