Better fix for the two-word splitting problem.
It's now clearer to me how IRCONT is meant to be used: the BKBUF buffer
is BSIZE+5 words long, so it's OK to go a word or too over BSIZE. So
you're meant to set IRCONT before you start adding words, and clear it
*before* you add the last one.
So no need for my extra function; just use IRCONT as intended.
The code in CNSTA wrote its last word with IRCONT still set. I think
this is harmless because it was always the first word in a block anyway,
but fixed anyway.