Consider not installing completions for nonexistent commands

[scop]

We might consider stopping here if the command is not found. That would make things simpler for completions, they would not have to take into account being invoked for nonexistent things. Then again, it would break a flow where a completion is installed for a nonexistent command, then the command is made available. But I don't know if this is something we should support.
(The existence check should be more than type -P if we want to go this way in order to not break ability to install completions e.g. for aliases.)

Originally posted by @scop in #924 (comment)