fix: stop counting a mutant whose subprocess never ran a test as killed

[tiagoabsantos]

Fixes pestphp/pest#1937.

--mutate can report a 100% score for mutants whose subprocess never ran a single test. MutationTest::hasFinished() reads any unsuccessful exit as a kill, so a subprocess that dies before reaching the test suite is indistinguishable from a legitimately killed mutant. Two inherited arguments make that happen.

ParaTest-only options are forwarded to the mutant (all platforms)

A mutant runs one PHPUnit process, never ParaTest, but the original arguments are passed through as they were. --parallel is already dropped by ParallelOption::remove(); --processes and --passthru-php are not, and PHPUnit rejects them:

INFO  Unknown option "--processes". Most similar options are --covers, --colors, --group, --no-progress, --no-results.

Exit code 2 in a fraction of a second, counted as killed. This is the Linux CI case in the issue, and it applies to every ParaTest-only option, so MutantArguments strips them all rather than only the two that were reported. It also takes over the coverage stripping that was inline, so the rules for what a mutant may not receive live in one place.

The options are removed from the mutant's arguments only, not from the configuration, so the original run keeps --processes and --passthru-php for its own parallel execution — dropping --passthru-php there would take pcov with it and leave the run without coverage.

--processes 2 spends two arguments, so the value is dropped along with the option; a stray 2 would otherwise be read as a test path.

The Pest script is used as the executable (Windows)

new Process([...$filteredArguments, ...]) puts $originalArguments[0], the Pest script, in the executable position. That works wherever the shebang is honoured. On Windows cmd.exe reads vendor/bin/pest as the command vendor:

'vendor' is not recognized as an internal or external command, operable program or batch file.

Exit code 1, counted as killed. The mutant now runs through PHP_BINARY, the same way the original run reached the same script.

Verification

Against tests/.tests/Untested, whose snapshot expects 3 untested out of 4, on Windows 11 with PHP 8.4.25:

	before	after
serial	Mutations: 4 tested, Score: 100.00%, 0.17s	Mutations: 3 untested, 1 tested, Score: 25.00%, 2.36s
--parallel --processes=2 --passthru-php=...	Mutations: 4 tested, Score: 100.00%	Mutations: 3 untested, 1 tested, Score: 25.00%

Both new tests in tests/Unit/MutationTestTest.php fail on the current 5.x and pass here. They run a real subprocess against a stand-in for the Pest script that records its own $argv, so they cover the executable and the forwarded arguments the way a mutant actually sees them, on Linux as well as Windows.

pint --test, phpstan, and pest --type-coverage --min=100 are clean. The 7 failures left in pest on my Windows machine are all present on 5.x before this branch: they are the path and line-ending issues already covered by #36 and #44.

Not included

The issue also suggests not counting a run in which no test was executed as tested, as a backstop for any other environment problem in the subprocess. That needs a signal beyond the exit code — exit code 2 is also what a legitimately killed mutant produces when it makes a test error — so it is a separate change, and #42 is already exploring one shape of it. This PR fixes the two causes that are unambiguous.