queue, scanner, signals and watchdog each had their whole pass inside a while
whose last statement was an unconditional break, so every pass fell out of the
loop and restartDaemon() re-executed console.php: a fresh bootstrap, settings
read and database connect per pass, per node — four times a second in the
signals daemon. The plan's 24 h RSS and queries/s acceptance cannot be measured
on a process that never lives a second.
Every break that means something (a code change, nginx stopped, MAIN's
database gone) stays, and restartDaemon() still re-execs then, which is how a
deploy is picked up. Two things fell out of it: the signals daemon paced itself
inside the branch that read `signals`, so a node whose query failed spun as
fast as MariaDB would answer; and the watchdog's CPU sample was a delta over
its own 2 s sleep every pass, because a fresh process never had a previous
read of /proc/stat.
The numeric cluster settings render as number inputs with the bounds the save
clamps to, instead of text fields that no digits-only script covered.
