From https://bugzilla.redhat.com/show_bug.cgi?id=708537#c3:

------&<---------
'init 5' translates into 'systemctl isolate graphical.target'.
Isolate does more than people expect:

• Services that are not obviously pulled in by the new target will be killed.
  This includes services that are usually loaded on-demand, including
  the stdout-syslog-bridge, which the processes connected to it won't like.
  The getty units are also affected by this now that they're autospawned.
• Oneshot units that have completed successfully before, but don't have
  RemainAfterExit=yes, will be re-run again. This can affect the plymouth
  problem.

So the isolate action is brutal and not really what's expected from runlevel switching. We should not fix isolate, because it is useful in itself, but there should be a new action, say systemctl approach $foo.target. systemd would remember the last approached target and when asked to approach another one, it would compare the set C of units pulled by the current target with the set N of units pulled by the new target. It would stop only units in C-N, it would start units in N-C.
It is obvious that switching from runlevel 5 to 5 would then be a NOP.
------>&---------

Something like this would be useful, not only in the context of switching runlevels, but generally in the context of switching or updating targets. OTOH, maybe the complexity is not worth the trouble.