[Caching] Add failing reproducer for partially written cache reads

[calebdw]

Targets partial-cache so it lands inside #8426 rather than alongside it. Test only — no production change.

Addresses the one thing blocking that PR:

@TomasVotruba: We'll need a failing reproducer in our CI, so we avoid changing this back and forth.

Fails on main's copy(), passes on this branch's rename(). No threads, no timing, no sleep.

The test opens a read handle on a cache file, saves over the same key, and then reads from the handle it opened before the save — the way a parallel worker holds the configuration_hash file open while require-ing it during container boot. copy() truncates the destination and streams into the very inode that reader is holding, so the reader's file changes underneath it. rename() publishes a new inode and leaves the reader on the complete file it opened.

On main:

1) Rector\Tests\Caching\ValueObject\Storage\FileCacheStorageTest::testSaveLeavesAConcurrentReaderOnACompleteFile
Failed asserting that two strings are identical.
--- Expected
+++ Actual
 return \Rector\Caching\ValueObject\CacheItem::__set_state(array(
    'variableKey' => 'TEST',
-   'data' => 'first',
+   'data' => 'second',
 ));

What it does not prove

It is not a reproduction of the crash. It pins the precondition — a writer mutating a file under a live reader — not a torn read. With a payload this small copy() finishes before the read, so you get complete-but-wrong content rather than the syntax error, unexpected string content / Child process error failure seen in the wild. Making an actual torn read deterministic needs the interleaving, which I don't think is achievable in a test.

What it does give you is a guard that fails the moment copy() comes back, which I read as the point of the request.

Checks

• vendor/bin/phpunit tests/Caching — 14/14 green on this branch, 1 failure on main
• composer check-cs — clean

Context and the Linux reproduction that led here: #8426 (comment)