regression in coverage 7.10.3: false negatives with `patch=subprocess` (`# pragma: no branch` is reported as not covered)

[neutrinoceros]

Describe the bug

context: I'm seeing what I believe are false negatives with coverage>7.10.2
xref: neutrinoceros/runtime-introspect#37

To Reproduce
The exact python version shouldn't matter (as long as it's 3.13 or newer)

git clone https://github.com/neutrinoceros/runtime-introspect@e309e37864437548623971a2d4cdc0d122b810b5
cd runtime-introspect
git checkout e309e37864437548623971a2d4cdc0d122b810b5 # latest commit at the time of writing
python -m venv
# ... source env ...
pip install -e . --group covcheck
coverage run -m pytest
coverage combine
coverage report

if you see tests/test_features.py:147 being reported as not covered, you have successfully reproduced this. The previous line starts with an if and is marked with a # pragma: no branch comment.
I bisected this regression to 697d4bb (coverage 7.10.3), and it appears to still be present in the development version.

Expected behavior
I expect line 147 to be omitted from reports. Versions 7.10.0 to 7.10.2 behave as I expect.

Additional context

[neutrinoceros]

I bisect a seemingly different regression seen in the same context to the exact same commit, where a couple lines that are definitely covered are flagged as not covered. I wonder if it's worth a second bug report, considering it might just be two symptoms of a single defect ?

[neutrinoceros]

additional context: I found that if I removed the # pragma: no branch line, that I'll reproduce with its immediate context here for convenience:

@pytest.mark.parametrize("settings", [...])
def test_foo(settings):
	env_dict = {}
	if (COVERAGE_PROCESS_START := os.getenv("COVERAGE_PROCESS_START")) is not None:  # pragma: no branch
	    env_dict["COVERAGE_PROCESS_START"] = COVERAGE_PROCESS_START
	
	xoptions = [f"-X{opt}" for opt in settings.xoptions]
	cp = subprocess.run(
	    [sys.executable, *xoptions, str(script_file.resolve())],
	    env=env_dict,
	    check=True,
	    capture_output=True,
	)

then the other false negatives I refered to in my previous comment also start to show even with coverage 7.10.2

[neutrinoceros]

if I replace it with a (type-incorrect) unconditional setting instead:

env_dict["COVERAGE_PROCESS_START"] = os.getenv("COVERAGE_PROCESS_START")

this does seem work with coverage 7.10.2, but not the dev version, where I get a predictable TypeError in subprocess.run when the env var is None

[neutrinoceros]

supposedly that's precisely because of the crucial change that landed in 7.10.3, where options not originated in a configuration file wasn't correctly passed down to subprocesses.

[neutrinoceros]

okay got it. The following patch unblocked my upgrade:
https://github.com/neutrinoceros/runtime-introspect/pull/39/files

So there is no actual regression. Instead, I think patch=subprocess' documentation could be improved on the following points:

• I didn't understand that COVERAGE_START_PROCESS would need to be set in the child process, even if not originally set in the parent process (which with what the current phrasing the new environment must copy over the COVERAGE_PROCESS_START environment variable. sounds like to me)
• it wasn't clear to me that subprocess.run was going to need it. Maybe that's implied by If an exec*e or spawn*e function is used, ..., but that wasn't obvious to me.

I'm going to open a PR to propose edits to the docs, but feel free to close this PR regardless !