What it does
Every page asks a capability through the one executor MCP and the HTTP routes use, so the Cockpit cannot answer a question differently from the API. The sidebar's catalogues are the registry's modules, the index's kinds and the graph derivations; a capability's page carries its schemas, its projections, its provenance and a form generated from its input schema that calls its real route. A capability, a kind or a graph added to the backend has a page, a navigation entry and a search entry the first time it exists, with no line of the Cockpit written for it.
The browser layer is an enhancement in the strict sense: a command palette, the runner, a graph drawing and two optional views, under a strict content-security policy with no inline script and no remote origin. With JavaScript off every page still shows everything it knows.
What it does not do
It holds no model, no catalogue and no verdict of its own. It writes nothing of its own
either: what writes is a declared command, which the runner can reach like any other
capability, and a command whose declared effect is repository_mutation carries that warning
on its Run card before it is sent (docs/COCKPIT.md, "What writes is declared").
It is not the published website; that is a static projection of the same registry with no
server behind it.