test(wake): cover the seeded cursor and the per-session split

The three existing drain tests encoded the old contract — no cursor file meant
"the whole inbound history is new" — and failed once the first run started
seeding to the tip. Seeding is the point: without it every new session would
replay the entire history on its first drain.

Each test now seeds first and then asserts on what arrives after. Adds the two
cases the PR actually changes behaviour for: the first run stays silent and
records the tip, and one message wakes two sessions with different session keys
(with a shared cursor the second one saw nothing).