Fix WindowsFS onClose race condition

[szybia]

• If i create an InputStream on a target file, and then close it, if i concurrently (before onClose finishes) mv tmp target, i shouldn't get an AssertionError
• Happens because with underlying unix FS, getKey will return current inode for the path, if concurrent move happens, the key within openFiles changes, and now it'll fail to remove it from openFiles
• Spotted in ES test failure, where we atomically replace a file
• Reproduced with (tested with my change, no failure):

    public void testAtomicMoveRaceDuringClose() throws Exception {
        Path rawDir = createTempDir();
        WindowsFS windowsFs = new WindowsFS(rawDir.getFileSystem());
        Path dir = windowsFs.wrapPath(rawDir);
        Path target = dir.resolve("target");
        Files.write(target, new byte[] { 42 });
        int writerCount = 16;
        int readerCount = 16;
        Thread[] threads = new Thread[writerCount + readerCount];
        AtomicReference<AssertionError> raceHit = new AtomicReference<>();
        for (int i = 0; i < writerCount; i++) {
            final int id = i;
            threads[i] = new Thread(() -> {
                int counter = 0;
                while (Thread.currentThread().isInterrupted() == false) {
                    try {
                        Path tmp = dir.resolve("tmp-" + id + "-" + counter++);
                        Files.write(tmp, new byte[] { 42 });
                        Files.move(tmp, target, StandardCopyOption.ATOMIC_MOVE);
                    } catch (Exception ignored) {}
                }
            });
        }
        for (int i = 0; i < readerCount; i++) {
            threads[writerCount + i] = new Thread(() -> {
                while (Thread.currentThread().isInterrupted() == false) {
                    try (InputStream in = Files.newInputStream(target)) {
                        in.readAllBytes();
                    } catch (Exception ignored) {}
                }
            });
            threads[writerCount + i].setUncaughtExceptionHandler((t, e) -> { if (e instanceof AssertionError ae) raceHit.set(ae); });
        }
        for (Thread t : threads)
            t.start();
        Thread.sleep(120_000);
        for (Thread t : threads)
            t.interrupt();
        for (Thread t : threads)
            t.join(5_000);
        final var e = raceHit.get();
        if (e != null) {
            logger.info("Race reproduced: ", e);
        }
    }

[2026-04-10T20:43:31,537][INFO ][o.e.x.s.c.c.StatelessHeartbeatStoreTests][testAtomicMoveRaceDuringClose] Race reproduced: 
java.lang.AssertionError: null
	at org.apache.lucene.tests.mockfile.WindowsFS.onClose(WindowsFS.java:80) ~[lucene-test-framework-10.4.0.jar:10.4.0 9983b7ce7fdd04f4d357688fb85c14277c15ea8d - 2026-02-20 18:05:31]
	at org.apache.lucene.tests.mockfile.HandleTrackingFS$1.close(HandleTrackingFS.java:103) ~[lucene-test-framework-10.4.0.jar:10.4.0 9983b7ce7fdd04f4d357688fb85c14277c15ea8d - 2026-02-20 18:05:31]```

[rmuir]

nice find, thank you

[szybia]

@rmuir thanks for the review!

1. should i add a changelog? was scanning it and my best guess is under Other but not sure...
2. i'll also fix the styling...

[rmuir]

@rmuir thanks for the review!

1. should i add a changelog? was scanning it and my best guess is under Other but not sure...
2. i'll also fix the styling...

@szybia Changelog is always a plus I think. "Other" or "Tests" or any other best guess is fine. Thanks!

[szybia]

i added changelog under 10.5.0, i'm assuming could go into a 10.4.1 patch but there's no changelog entry for it, i could add it...

pushed changes as a conversation starter

[rmuir]

10.5 is good, i'll backport it. I don't know if there will be a 10.4.1, but if someone wants to make one we could then cherry-pick it back to that, too.

[szybia]

thank you!

i don't have write access so leaving up to you to merge 🚀

[dweiss]

I'll merge and backport.