"Who can pull the plug?" — new page, and recoverability as a named dimension

• New top-level page (plug.html, first in the nav). It opens on the question a board actually asks out loud — if this agent starts doing something no one intended, who stops it, how fast, and what does stopping it cost? — and answers it with a plug profile: who, blast radius, speed, side effects, recoverability.
• A plug profile is the mandate read backwards, and the page says so explicitly. The mandate states what an agent is authorised to reach; the plug profile states what it costs to take that reach away. Same graph, same acceptance record, opposite direction — one artifact, not two.
• Recoverability is the new idea. Our doctrine has been "every risk is already accepted; the only question is how long". Recoverability adds the second half — and how much can we undo if we're wrong — which reorders the queue: a high-exposure agent whose actions replay cleanly can often be accepted for longer than a small one that moves money out of the building.
• Also on the page: the capability-union insight (authorisation is whatever the agent can already do), the read-only architecture with the never-in-the-request-path boundary, a category-level comparison, a break-even framing for the business case, and an FAQ.
• Adapted from the exploratory MVP, not pasted. Ported into the site's own design system (no external font dependency), rebuilt on the standard header/menu/switcher, navigation moved onto the host protocol, and the interactive breaker rebuilt with DOM construction, keyboard support and reduced-motion handling. Claims were cut to what we can defend: no modelled ROI multiple, no invented before/after metrics, no named-competitor deficiency claims, no placeholder logos, no dead links. Pricing is not restated — the page links to the canonical Pricing page.
• A smoke group covers the page: menu parity, the breaker lighting all five dimensions and resetting, and a regression guard asserting the retired claims never come back with it.