toml-test datetime cases use tomlkit's own parse_rfc3339 as the expected value, so they cannot detect datetime parsing bugs

[ArnauFerma]

While running a mutation-based check of the test suite at 4b38bec, one structural gap stood out and seemed worth reporting. Nothing here is a bug in tomlkit's code — every mutation below was a deliberate break, and the code as shipped is correct.

What was observed

tests/test_toml_tests.py builds the expected value for every datetime-typed case with tomlkit._utils.parse_rfc3339:

stypes = {
    ...
    "datetime": parse_rfc3339,
    "datetime-local": parse_rfc3339,
    "date-local": parse_rfc3339,
    "time-local": parse_rfc3339,
}

That is the function under test, so for those four types the parser and the expectation drift together. Breaking the negative-offset sign in parse_rfc3339 (if sign == "-": offset = -offset → never negating) produced:

	red
toml-test compliance cases (680)	0
hand-written tests (378)	3 (test_parse_rfc3339_datetime ×2, test_datetimes_behave_like_datetimes)

By contrast, dropping the integer sign in the parser reddened 6 compliance cases and untranslated string escapes 7 — for int and str the corpus is an independent oracle. For datetimes it is not, and the entire guard is three hand-written tests.

Related, smaller: no input anywhere in tests/ (hand-written or corpus) carries a non-zero minute offset, so +05:30 being parsed as +05:00 (dropping minute_offset from the timedelta) survives the whole suite.

Possible fix

Build datetime expectations independently — datetime.fromisoformat accepts RFC 3339 on Python 3.11+ and the project already requires ≥3.9 with a compat shim, or a tiny hand-rolled regex as the toml-test JSON format is fixed — and add one +05:30-style case to test_parse_rfc3339_datetime.

Also noticed

tests/util.py::elementary_test asserts isinstance(v.unwrap(), unwrapped_type) only, so the nine test_*_unwrap tests detect a wrong type but not a wrong value: Bool.unwrap returning not bool(self) passes test_true_unwrap and test_false_unwrap. Passing the expected value to the helper would close that.

Full write-up with the mutation spec and per-test results, reproducible against 4b38bec: https://github.com/ArnauFerma/falsifiable-tests/blob/main/case-studies/tomlkit.md

Happy to send a PR for either if you'd take one.