>     Inherently, changing `real-last-command' is not "nice".
 >
 > I wasn't thinking of changing it; rather, not setting it at all for
 > those kinds of events.

I meant "changing the semantics of `real-last-command'".  We'd change
the semantics by _not_ setting it.

 > That probably works in practice, because the only other package
 > that tests `real-last-command' is table.el, and I think it could
 > just as well use `last-command'.

Yes.  But if we were to suggest not testing `real-last-command' in Lisp
code (aside from repeat.el) we probably shouldn't mention that variable
in the manual in the first place.

 > But you are right that this isn't entirely clean.
 > What could be a cleaner method?
 >
 > Perhaps to make a new variable and set it at the C level, almost like
 > `real-last-command' but excluding certain commands, or certain events.

We could call it `last-repeatable-command' and have repeat.el use that
instead of `real-last-command'.  I attached a patch for keyboard.c.

 >     Disregarding my scruples, how do I tell whether `real-this-command' was
 >     bound to an input event at the time of executing
 >
 > 	      current_kboard->Vreal_last_command = real_this_command;
 >
 > My idea was to test last_command_char.

We maybe should say somewhere that it's a bad idea to change
`last-command-char' (and probably `real-last-command' and
`last-repeatable-command' as well) while executing a command.