A clean-up that was performed about the same time with
MDEV-14154 used the wrong source for a microsecond timer. The trx_t::start_time_micro is supposed to be initialized from my_hrtime() or my_hrtime_coarse(), not from the monotonic microsecond_interval_timer().
The impact of this is unclear, but it should be minimal. Both internal and recovered transactions ought to have priority over user transactions. On my Debian GNU/Linux system on AMD64, both timestamps look similar. The my_hrtime_coarse() should involve less overhead than microsecond_interval_timer().