Add translation tests for argparse

[savannahostrowski]

While #12711 was merged, it seems like there's a separate issue worth tracking around adding translation tests for argparse (see #12711 (comment))

Linked PRs

• gh-124295: Add translation tests for argparse #124803
• [3.13] gh-124295: Add translation tests for argparse (GH-124803) #126046
• gh-124295: Skip translation tests when pygettext is missing #126051
• [3.12] gh-124295: Add translation tests for argparse (GH-124803) (GH-126046) #126054

[tomasr8]

Continuing the discussion here since the other issue was closed.

Quoting Serhiy:

I thought that we should run the parser with different options to produce different error messages with installed special translator that transforms the original text in simply recognizable way. But there may be better ways.

I was originally thinking the same, but to me that seems like a lot of effort compared to the benefit we'd get. Instead, I think that using a message extractor (e.g. pygettext) we can assert both the presence of a specific message and the line number it appears at. This should give us pretty much the same confidence as directly checking the argparse output but with an order of magnitude less work. What do you think?

[serhiy-storchaka]

and the line number it appears at

Would not this make the tests extremely fragile?

[tomasr8]

Would not this make the tests extremely fragile?

Does it matter that much if you don't need to fix the tests manually? You'd have a snapshot of the PO file and a simple command to regenerate it so all you'd need to do is double-check the diff

[serhiy-storchaka]

You would need to regenerate the snapshot after every change in argparse (and it was changed a lot in last few days), and the diff would be so large that it would be hard to notice a real regression in the noise.

But this is a good first step. Before #27667 was merged, Tools/i18n/pygettext.py produced a warning for _() used with non-literals. They should be treated as errors in the test. Also, line numbers should be ignored in comparison and only used to report errors. We should just collect a set of translated strings, compare it with the expected result and report about missing or additional strings.

[tomasr8]

True, if the module is changing a lot, so would the snapshots. I'll try to come up with a PR based on your suggestions :)