test(e2e): add live PROBE dump every 5s during turn 4 hang to inspect mock state

Previous diagnostic dumps requests AFTER timeout. This adds an interval
timer that, every 2 seconds while a send() is in flight, dumps the last
3 mock requests if we're already past 50 requests for this turn.

This will let us see HOW the message array structure evolves during
the 600s CI hang — specifically whether:
  - Pre-compaction history gets continuously added back into the wire
  - The synthetic compaction marker is being inserted/re-removed in a loop
  - Each iteration's assistant content keeps growing into the same msg
  - The mock-reset state varies turn-by-turn

CI also gets 'msgs=N lastTwo=...' so we can correlate message count
growth vs request count. Tests pass locally (turns 1-23 in 4-5s).