Make the fail-under summary agree with the exit code

[OnourImpram]

The defect

--cov-fail-under is decided twice, with two different comparisons.

Where	Predicate
plugin.py pytest_sessionfinish	should_fail_under(self.cov_total, cov_fail_under, cov_precision) — coverage.py's own, compares the total rounded to [report] precision
plugin.py pytest_terminal_summary	failed = self.cov_total < self.options.cov_fail_under — the raw total

When the raw total is below the threshold but rounds up to meet it, the two disagree. The run exits 0 and prints a red FAIL banner at the same time.

Reproduced on pytest 9.1.1 / pytest-cov 7.1.0 / coverage 7.15.2, with the suite's own SCRIPT fixture (88.888…% covered) at the default precision of 0:

$ pytest --cov=... --cov-fail-under=88.9
FAIL Required test coverage of 88.9% not reached. Total coverage: 88.89%
============================= 10 passed in 0.20s ==============================
$ echo $?
0

round(88.888…, 0) is 89, so should_fail_under says pass and the exit code is 0 — correctly, given the configured precision. Only the banner disagrees.

The change

One line: the summary uses the predicate the exit-code path already uses.

The exit code does not change. Only the banner text and its colour move, so no build that passes today starts failing.

On the direction of the fix

The inconsistency is not in dispute, but the two open issues want it resolved in opposite directions, so this is a maintainer call and I want to be explicit about which way I went.

• "Fail under check" is failing because of ignoring the precession config #638 — "'Fail under check' is failing because of ignoring the precession config" — is exactly this: a FAIL shown when precision says the threshold was met. This PR fixes it.
• Why doesn't pytest-cov fail with a non-zero exit code when code coverage threshold isn't met, when running on multiple directories? #728 — the reporter writes that they "rightfully see the message FAIL …" and that the problem is the exit code consulting the rounded value. That would be the mirror fix: change pytest_sessionfinish to use the raw comparison instead.

I chose to align the banner rather than the exit code because:

1. should_fail_under is coverage.py's canonical predicate, and pytest_sessionfinish already delegates to it deliberately — the semantic choice has been made in this codebase.
2. [report] precision / --cov-precision exists to declare the granularity that matters; at precision 0, 88.888…% is 89%.
3. It is the non-breaking direction. Switching the exit code to the raw comparison would start failing CI runs that pass today.

If you prefer #728's reading, the change belongs on the other line and I am happy to redo it that way.

Not changed

The banner still hardcodes Total coverage: {actual:.2f}% rather than using cov_precision, so at precision 3 it can still print a differently-rounded number than the table above it. Fixing that changes an expected string in test_cov_min_float_value_not_reached, so I have left it alone rather than fold an unrelated behaviour change into this PR. Happy to follow up.

Tests

Two regression tests alongside the existing test_cov_min_float_value group — one at default precision, one mirroring #638 with precision = 1. Both fail on master and pass with the change.

Full suite on Windows / Python 3.13: 205 passed, 6 skipped (all pre-existing platform skips). ruff check and ruff format --diff clean.

The remaining fail-under tests are unaffected: their thresholds (88.88 @ p0, 88.89 @ p3, 100 @ p0, 50 @ p0) do not sit in the divergent band, so both predicates return the same boolean.

[ionelmc]

Why does this pr show a huge removal/add change on plugin.py? Is that a display bug in github or what's going on?

[OnourImpram]

Line endings, not a rewrite. The file had been committed with CRLF, so all 477 lines of plugin.py registered as changed. Restored to LF in f3121a1; the pull request now shows +43 −1.

The actual change in plugin.py is seven lines. pytest_sessionfinish decides the exit code with should_fail_under(cov_total, cov_fail_under, cov_precision), while the summary banner compared cov_total < cov_fail_under directly. At a precision boundary the two disagree and the run prints that required coverage was not reached while exiting 0, or the reverse. The banner now calls the same predicate, so it cannot contradict the exit status. The remaining diff is one regression test and a changelog entry.

Thanks for looking at it.