This issue was found as a part of #22388, but it's a separate problem and should be tracked in a separate ticket.

During reconnect, shortly after synchronization is started, the teacher starts experiencing high virtual pipeline backpressure. If reconnect takes long, the teacher eventually starts switching between ACTIVE and CHECKING and may even fall behind the network and need to reconnect (as a learner).

This happens because the state used for teaching is not released till the very end of reconnect. It prevents virtual map copies from merging / flushing to disk and results in high virtual node cache size and backpressure. This is a regression from #21464. That fix was correct, but it revealed a problem that had existed for years: unreleased copies could be flushed to disk. With 21464 that problem was addressed, but it resulted in a different problem with reconnects.