Investigated this further. The handler isn't exactly late. It is
called before errflag checks, but exits here without setting errflag:

    /* Are we queueing signals now?      */
    if (queueing_enabled) {

The log looks now like this:

-- tmoutp != ZTM_NONE / raw_getbyte() zle_main.c
## zhandler(2) signals.c (debug message in first line of zhandler)
## zhandler(2) EXIT 1 signals.c (debug message in if(queueing_enabled))
-- NOT doing break selret[-1/4/4] / zle_main.c: errflag: 0, retflag:
0, breaks: 0, exit_pending: 0
-- Trying again selret[-1], !errtry[0] / zle_main.c

-- tmoutp != ZTM_NONE / raw_getbyte() zle_main.c
-- (**) NOT doing break selret[0/4/4] / zle_main.c: errflag: 0,
retflag: 0, breaks: 0, exit_pending: 0
--  Passed !errtry(errtry:1) selret[0] / zle_main.c
## zhandler(2) signals.c
## set errflag to 2 (ERRFLAG_INT:2) / signals.c
-- tmoutp->tp <- ZTM_NONE / calc_timeout( do_keytmout: 0 ),
keytimeout: 40 / zle_main.c
== CALC_TIMEOUT() one more chance (timedfns exp100ths: 100)
-- CALC_TIMEOUT() tfnode TRUE no break NOT calling >> DIFF=1 <<
-- CALC_TIMEOUT() 0 != 1 || 100 <= 100 ZTM_FUNC(2)
^^^ LOOP-CALLED calc_timeout: tmout.tp == 2 / zle_main.c

Interesting is (**) line, it still has errno == 4 (last number in
[%d/%d/%d]), apparently errno=0 is needed before select, doing so
clears the value. What's with the queueing and early exit from
zhandler()? It brings disorder to .recursiveedit.

Best regards,
Sebastian Gniazdowski