`doctest.OutputChecker` with `doctest.NORMALIZE_WHITESPACE|doctest.ELLIPSIS` flag cannot match ` ... ` with ` `

[user202729]

Bug report

Bug description:

import doctest
optflag = doctest.NORMALIZE_WHITESPACE|doctest.ELLIPSIS
OC = doctest.OutputChecker()

want = """
AB ... CD
"""
got = """
AB  CD
"""
assert not OC.check_output(want, got, optflag)

This is rather unexpected, since substituting the empty string for ... results in a match.

CPython versions tested on:

3.11

Operating systems tested on:

No response

[tim-one]

I agree it can seem surprising, but not sure what can be done about it. Whitespace normalization is done first, which effectively changes got to "AB CD". That is, the original two spaces have been replaced by one. Which indeed can't match the " ... " in want, which still wants at least two spaces. Whitespace normalization has no effect on want, since it's already normalized in that sense.

[user202729]

we could make ... remember that there's a space before it, and allow it to take another space. Which is equivalent to special-casing <space><ellipsis><space> if both flags are true.

Note that want gets normalized too. The following returns True

want = """
AB    ...    CD
"""
got = """
AB x CD
"""
OC.check_output(want, got, optflag)

[tim-one]

Yes, want and got both get normalized. While the combination of options surprised you, how it works is quite straightforward and predictable (runs of whitespace are collapsed to single blanks, then an ellipsis match is done). Perhaps not so much if special cases are piled on (not just "a space", but any flavor of whitespace). As is, the code for ellipsis matching knows nothing about whitespace, and the code for whitespace normalization knows nothing about ellipsis matching. There's value in simplicity and orthogonality too.

[picnixz]

I also wanted to make the output matching more malleable for multiline outputs but I am not satisfied (I have a PR open for that). I think we should acknowledge the limited features of doctest and avoid expanding them into something too generic (and especially if it is a too niche case).

In this case, it would be better to use doctest for a startswith()/endswith() check. I suggest simply improving the docs

[tim-one]

Actually, I've been uncomfortable with almost every change made to make doctest matching "more flexible". WYSIWYG was at the heart of its original design goals, akin to that Black was designed to enforce One True Style.

In this specific case, I think whitespace normalization could be changed to mutate want itself, if ellipsis matching is also in play, by repeatedly substituting matches to regexp

r"\s+(\.\.\.)+\s+"

with " ..." until no more matches are found.

I could live with that.