A keystroke took a second or more to reach another peer, and in the app up to
twenty seconds to reach the host's own window, which waited on its folder's
watch. Now:

- the host tells its own sections what guests change at once, and tells guests
  when its own open section publishes, rather than waiting on the watch;
- a guest's background settles a report for 20 ms, not a second, since its host
  sends one a commit;
- the host sends what each commit changed (Delta: the header, patches and
  appended bytes), and a guest holding the section takes it on without reading
  the section again; its own commits apply to what it holds too.

Measured with examples/live_latency (20 keys, keystroke stored in one guest's
replica to the change in another's, and host's file to a guest), median:

                         before    after
  loopback LAN           1076 ms   57 ms   (host to guest 1026 -> 34 ms)
  relay on loopback      1071 ms   57 ms   (host to guest 1027 -> 33 ms)
  relay.snowbound...         -    210 ms   (host to guest 137 ms)

Protocol 2 opens with version 2 and answers a peer of another version before
failing, so each end names the other's version: joining says which side must
update Snowbound.

Assisted-by: claude-opus-5.5