regex->trigram has wrong behaviour when trigrams are question-marked about codesearch HOT 5 CLOSED

A bit of investigation has narrowed my focus to concat(x,y)'s handling when x 
or y has canEmpty=true and a prefix/suffix list rather than a list of exact 
matches.  The hunt continues...

I'm stumped.  I think this bug is beyond my current familiarity with the 
codebase.  I'm certain the issue is with concat(), but after many hours of 
staring at this I really have no idea what the fix should look like.

So, it looks like this particular issue is actually in alternate().  The code 
assumes that either x and y are both exact or both have prefix/suffix lists.  
This obviously isn't the case.

The above bug is caused by the fact that emptyString() has an exact match but 
no prefix/suffix lists, so when being combined we were unioning with a nil 
list.  This didn't crash, it just didn't have the expected behaviour.

This also fixes the case where, when searching for '(123|4567)q' , we would get 
"67q" as the only required trigram, instead of the correct ("23q"|"67q").

After writing that last sentence, I realized that my "corrected" trigrams were 
still missing some.  My implementation of alternate ignored the exact matches.

We now call addExact() before combining the matches.  This allows the (this 
time for sure ;) correct trigram set: 
    ("123")|("456" "567")  ("23q"|"67q")
to be generated.

This issue was closed by revision 56b76ffbf8bb.