Rewinding after an underrun should be done like this:
pa_sink_input_request_rewind(s->sink_input, (size_t) (s->sink_input->thread_info.underrun_for == (uint64_t) -1 ? 0 : s->sink_input->thread_info.underrun_for), FALSE, TRUE, FALSE);
(The example is from protocol-native.c.)
That ensures that the received data won't overwrite valid data in case the underrun was so short that the sink still has data left from time before the underrun started.
protocol-esound.c currently does this:
pa_sink_input_request_rewind(c->sink_input, 0, FALSE, TRUE, FALSE);
Giving zero as the rewind amount means that a full rewind will be done, regardless of whether valid data might get overwritten. If valid data gets overwritten, there will be a skip in the audio.
protocol-simple.c has the same bug.