I've certainly never tested it with conditionals.
I think it's the conditional indeed, because it works smoothly when I re-write it to not use a conditional:
%
^
\s*
( # <1>
# Single tags like <foo/>
<
\s*
[a-zA-Z:]+
(?:
\s*[a-zA-Z:]*
\s* = \s*
(?:'[^']*'|"[^"]*")
)*
\s*
/\s*
>
|
# Tags in pairs like <foo>content</foo>
<
\s*
([a-zA-Z:]+) # <2/>
(?:
\s*[a-zA-Z:]*
\s* = \s*
(?:'[^']*'|"[^"]*")
)*
\s*
>
(?:[^<>]* | (?1))*
<\s*/\s*\2\s*>
) # </1>
\s*
$
%x
<foo><bar></bar></foo> # Match
<foo><bar></foo></bar> # No match
<foo><bar/></foo> # Match
<foo><bar></foo> # No match
<foo bar=baz/> # No match
<foo bar="baz"> # No match
<foo bar="baz"/> # Match
< fooo / > # Match
<foo/>foo # No match
foo<foo/> # No match
<foo>foo</foo> # Match
<foo><bar/>foo</foo> # Match
#<a><b><c></c></b></a> # No match (WRONG!!)
Now, there's still the three-level-deep problem...
U28geW91IGNhbiBhbGwgcm90MTMgY
W5kIHBhY2soKS4gQnV0IGRvIHlvdS
ByZWNvZ25pc2UgQmFzZTY0IHdoZW4
geW91IHNlZSBpdD8gIC0tIEp1ZXJk