Notebook kernel coverage lost under pytest + nbval + pytest-cov (combine prints "Skipping duplicate data"). Since 7.13.1 #2370

Describe the bug
When the notebook starts a child Python process that measures nothing (e.g. subprocess.run([sys.executable, "-c", "pass"])), coverage recorded for the full notebook kernel is missing from the final report on coveragepy 7.13.1 and later. 7.13.0 reports it correctly.

To Reproduce
with mylib.py:

def work():
    x = 1
    y = 2
    return x + y

.coveragerc:

[run]
source = mylib
parallel = true
patch = subprocess

and mre.ipynb with two cells:

import mylib
print(mylib.work())

import subprocess, sys
subprocess.run([sys.executable, '-c', 'pass'], check=True)

pytest mre.ipynb --nbval-lax --cov=mylib --cov-config=.coveragerc --cov-report=json:cov.json

Then count the executed lines of mylib.py in cov.json. The expected value is 4.

Without cell 2, every version reports 4.

Alternatively see attached .py scripts
mre.py
mre builder with optional patches

When running with combine patched to keep=True (python mre_notebook.py -p keep_combine) I get two data files with the same hash in their names:

.coverage.<host>.pid<P>.<rand>.<host>.pid<C>.<rand>.HNQIp837rWOh    0 executed lines   (the child process)
.coverage.<host>.pid<P>.<rand>.HNQIp837rWOh                         4 executed lines   (the notebook's measurements)

With the optional longer script mre_notebook.py --debug-combine
(it also needs plugins/keep_combine.py), coverage's own debug output (COVERAGE_DEBUG=combine) and a re-run of
coverage combine --debug=combine on a copy of the kept files show:

Combined data file <dir>/.coverage.<host>.pid<P>.<rand>.<host>.pid<C>.<rand>.HNQIp837rWOh
Skipping duplicate data <dir>/.coverage.<host>.pid<P>.<rand>.HNQIp837rWOh
Combined 1 file, skipped 1

Versions:
coverage 7.13.0 (passing), 7.13.1 and 7.17.2 (failing); CPython 3.13; MRE run on macOS (an original, larger notebook run also failed on Linux CI); nbval 0.11.0; pytest-cov 7.1.0; pytest 9.2.0.