Hello,
below is code from raw_getbyte() / zle_main.c:

                case ZTM_FUNC:
                    // MY DEBUG
                    _F = fopen("/tmp/recursive.txt", "a+");
                    fprintf( _F, "queueing_enabled MARK[D] (%d)\n",
queueing_enabled );
                    fclose(_F);

                    while (firstnode(timedfns)) {
                        Timedfn tfdat = (Timedfn)getdata(firstnode(timedfns));
                        /*
                         * It's possible a previous function took
                         * a long time to run (though it can't
                         * call zle recursively), so recalculate
                         * the time on each iteration.
                         */
                        time_t now = time(NULL);
                        if (tfdat->when > now)
                            break;
                        tfdat->func();
                    }
                    // MY DEBUG
                    _F = fopen("/tmp/recursive.txt", "a+");
                    fprintf( _F, "queueing_enabled MARK[C] (%d)\n",
queueing_enabled );
                    fclose(_F);

When not in .recursive-edit, log messages look like:

zlecore() - queueing_enabled (1)
getkeycmd() - queueing_enabled (1)
getkeymapcmd() - queueing_enabled (1)
getkeybuf() - queueing_enabled (1)
getbyte() - queueing_enabled (1)
raw_getbyte() - queueing_enabled (0)
...
...
queueing_enabled MARK[2] (0)
queueing_enabled MARK[D] (0)
queueing_enabled MARK[C] (0)
queueing_enabled MARK[B] (0)
...
...

Timeout is reached every second (I do sched +1 and reschedule), the
logs are produced at that rate. However, when I press Ctrl-C to invoke
a .recursive-edit widget, then:

recursiveedit() - queueing_enabled (1)
zlecore() - queueing_enabled (1)
getkeycmd() - queueing_enabled (1)
getkeymapcmd() - queueing_enabled (1)
getkeybuf() - queueing_enabled (1)
getbyte() - queueing_enabled (1)
raw_getbyte() - queueing_enabled (0)
...
...
queueing_enabled MARK[2] (0)
queueing_enabled MARK[D] (0)
queueing_enabled MARK[C] (1)
queueing_enabled MARK[B] (1)
...
...
queueing_enabled MARK[2] (1)
queueing_enabled MARK[D] (1)
queueing_enabled MARK[C] (2)
queueing_enabled MARK[B] (2)
...
...
queueing_enabled MARK[2] (2)
queueing_enabled MARK[D] (2)
queueing_enabled MARK[C] (3)
queueing_enabled MARK[B] (3)
...
...

This causes Ctrl-C signal to be queued when in .recursive-edit, what
results in need of multiple Ctrl-C presses (errflag is set in middle
of sequence of checks of it's value and raw_getbytes() executes in
weird way). Also, errflag is set at random place after select() in
raw_getbyte(), and as Bart says, this prevents scheduled function to
execute at all, thus chain of rescheduling breaks.

For completeness I attach 10 context lines diff with the debug
messages, but the crucial two are pasted above. To see the efect only
this has to be ran:

wid() { zle .recursive-edit; }
zle -N wid
bindkey '^T' wid
fun() { sched +1 fun; }
fun
^T

PS. Because of problems with ML (maybe they're over?) I send also to
Bart and Peter.

Best regards,
Sebastian Gniazdowski