TransactionTest2.transactionsConcurrentWithPersistitClose residual flake: commit racing Persistit.close() throws raw IllegalStateException instead of PersistitClosedException

[vharseko]

Observed failure

master push run 30078532797, matrix cell build-maven (ubuntu-latest, 21) — 8 of 9 cells green, single failure in persistit/core:

[ERROR] com.persistit.TransactionTest2.transactionsConcurrentWithPersistitClose -- Time elapsed: 1.194 s <<< FAILURE!
java.lang.AssertionError: All threads should have exited correctly expected:<0> but was:<1>

The worker thread's stack trace printed alongside the assertion:

java.lang.IllegalStateException: JOURNAL_FLUSHER is not running
	at com.persistit.JournalManager.waitForDurability(JournalManager.java:1860)
	at com.persistit.Transaction.commit(Transaction.java:897)
	at com.persistit.TransactionTest2.transfer(TransactionTest2.java:380)
	at com.persistit.TransactionTest2.runIt(TransactionTest2.java:313)
	at com.persistit.TransactionTest2$3.run(TransactionTest2.java:272)

Analysis

This is a residual failure mode of the flake previously stabilized by #255, with a different mechanism that the #255 fix (deadline-bounded join before the assert) cannot help with:

1. The test's main thread calls _persistit.close(); JournalManager.close() nulls _flusher (JournalManager.java:1327).
2. A worker thread is concurrently inside Transaction.commit() and reaches JournalManager.waitForDurability(), which finds _flusher == null and throws a raw IllegalStateException("JOURNAL_FLUSHER is not running") (JournalManager.java:1860).
3. IllegalStateException is not a PersistitException, so in TransactionTest2.runIt it bypasses the expected-shutdown catch clauses (PersistitClosedException, PersistitInterruptedException, interrupted-I/O PersistitIOException) and lands in the generic catch (Throwable) branch, which records a failure and does not decrement _strandedThreads.
4. The assertion then fails with expected:<0> but was:<1> regardless of how long the join waits — the worker has already exited, but the counter can no longer reach zero.

Root cause

A transaction that races with Persistit.close() surfaces the shutdown as a raw IllegalStateException instead of the documented PersistitClosedException. Callers (including this test) that treat PersistitClosedException as the expected outcome of a concurrent close cannot distinguish this from a genuine internal error.

Candidate fixes

• Product side (preferred): JournalManager.waitForDurability() should throw PersistitClosedException when the flusher has been stopped by close(), so a concurrent commit() unwinds through the same path as every other closed-state operation.
• Test side (fallback): treat IllegalStateException("JOURNAL_FLUSHER is not running") during shutdown as an expected exit in TransactionTest2.runIt.