Skip flaky TransactionTest2.transactionsWithInterrupts on Windows

[vharseko]

Problem

TransactionTest2.transactionsWithInterrupts is flaky on Windows. The test repeatedly Thread.interrupt()s transaction worker threads while they perform file I/O, then asserts the money-transfer integrity invariant — the sum of all account balances must stay unchanged (0):

final int endingBalance = balance(accountEx);
assertEquals("Starting and ending balance don't agree", startingBalance, endingBalance);

On Windows, interrupting a thread that is blocked in NIO channel I/O closes the underlying FileChannel (ClosedByInterruptException), so Persistit's interrupt handling — and therefore this integrity invariant — can behave differently than on Linux/macOS, producing an intermittent balance mismatch:

java.lang.AssertionError: Starting and ending balance don't agree expected:<0> but was:<-11902>
    at com.persistit.TransactionTest2.transactionsWithInterrupts(TransactionTest2.java:228)

Observed on build-maven (windows-latest, 11). It does not reproduce on Linux or macOS (verified locally on macOS: 1 plain run + 6 runs under full CPU saturation, all green, balance agrees).

Fix

Guard the test with Assume.assumeFalse(isWindows) so it is skipped on Windows and still exercised on Linux and macOS, rather than weakening the integrity assertion (which would hide real regressions on the platforms where the invariant is stable). A comment documents the underlying JDK NIO interrupt-closes-channel behavior.

This is a targeted, honest stabilization: the money-transfer atomicity check keeps full strength everywhere it runs reliably; only the platform where the JVM's interrupt semantics make the stress test non-deterministic is excluded.

Verification

• macOS (real os.name): test runs (not skipped), balance agrees, passes.
• Simulated Windows (-Dos.name=Windows11): Skipped: 1, bails out in ~1.2 s instead of the 10 s stress loop — the skip branch works.
• Full TransactionTest2 class: 5 tests, 0 failures.

[vharseko]

Closing without merging: the mechanism this skip was papering over is addressed by #269 — VolumeStorageV2.writePage accepted short writes, silently tearing pages in the volume file during copyback (analysis in #268). Keeping transactionsWithInterrupts running on Windows preserves the only CI signal that can validate that fix on the affected platform. With #269's hardening, any residual failure would now surface as a loud InvalidPageStructureException instead of a silent balance mismatch — if Windows runs still fail after #269 lands, that is a signal to dig further in #268, not to skip the test.