[RonnyPfannschmidt]

Requested by Ronny · project thread

🤖 Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Closes #15201. Supersedes #15204.

Before: CallInfo stores start, stop and duration as three independent floats, and its three construction sites fill them differently. from_call uses timing.Instant. subtests.test() calls time.time()/time.perf_counter() directly, so a test that mocks time inside a subtest gets negative or bogus durations (the #185 class of bug). unittest subTest passes literal start=stop=duration=0.

After: CallInfo stores one private timing.Duration, and start/stop/duration are properties derived from it. All internal construction goes through CallInfo._from_duration(result, excinfo, duration, when). The pattern for any new site is to take a timing.Instant() when the phase begins and pass instant.elapsed(). Where there is nothing to measure, the site passes the named timing.UNMEASURED placeholder (all zeros). unittest subtests are the one such site: unittest reports them after the fact, without timing. Their reported values are unchanged from main; the zeros, inherited from pytest-subtests in #13738, are now just named.

Backward compatibility: external code still uses the float API, so it is kept.

• The float constructor CallInfo(result, excinfo, start, stop, duration, when, _ispytest=True) is used by pytest-subtests (still widely installed, plus Debian/deepin packaging), pytest-asyncio-concurrent and NVIDIA's isvtest. It now builds the Duration from the floats.
• Attribute assignment (call.start = ..., call.stop = ..., call.duration = ...) is used by pytest-asyncio-cooperative and pytest-velo. Setters update the stored Duration independently, so each field keeps its old meaning.
• timing.Instant fields become init-able (still defaulting to "now") so a Duration can represent externally supplied floats.

Visible change: dataclasses.fields(CallInfo) now lists _duration instead of start/stop/duration, and == compares the stored Duration.

Tests: a subtest regression test for mocked time (fails on main), a unittest subtest test asserting timing.UNMEASURED, and test_runner.py coverage for _from_duration and for the float constructor and setters.

[PersuasivePost]

I reported #15201 and tested this branch (claude/callinfo-duration-8dkzw2) against my original reproducer, which patches time.time and time.perf_counter to lambda: 0.0 inside the test and inside a subtest, and sleeps 0.3s.

main @ 420cf37: test_with_subtest reported −1194.217s
this PR: test_with_subtest reports 0.606s (test_normal: 0.301s)

The negative duration is gone. Thanks for the quick fix.