Fix RandomAccessReader leak in OnDiskGraphIndexCompactor

[eolivelli]

Summary

OnDiskGraphIndexCompactor.compactLevels() leaks a RandomAccessReader
for every source, on every executor worker thread, for every compaction.

compactLevels() runs its batch tasks on executor (which may be
caller-owned). Each worker thread lazily creates its own Scratch via a
ThreadLocal, and every Scratch holds one GraphSearcher — hence one
RandomAccessReader (via GraphIndex.View) — per source. At the end,
only the calling thread's Scratch is closed:

Scratch s = threadLocalScratch.get();
s.close();
threadLocalScratch.remove();

Every Scratch created on an executor worker thread is never closed,
so its GraphSearchers — and the RandomAccessReaders behind their
Views — leak for the lifetime of the executor (and accumulate across
compaction cycles when the executor is shared).

With an in-memory/mmap reader this is a benign FD/handle leak. With a
remote-backed reader (HerdDB's RemoteRandomAccessReader) each leaked
reader pins an off-heap Netty block buffer; over many compaction cycles
this exhausts the memory budget and stalls ingestion — see
eolivelli/herddb#590,
whose SEVERE: LEAK: ByteBuf.release() was not called traces all
originate in this code path.

A second, smaller leak: resolveEntryNodeSource() called
sources.get(s).getView() and never closed the returned View.

Changes

• OnDiskGraphIndexCompactor.compactLevels(): register every Scratch
  created on any thread in a ConcurrentLinkedQueue and close them all
  in a finally block. On normal completion a close failure is
  surfaced; while a primary exception is propagating, close failures are
  logged so they cannot mask the original cause.
• OnDiskGraphIndexCompactor.resolveEntryNodeSource(): wrap the
  getView() call in try-with-resources.
• New testCompactionClosesAllReaders: wraps each source's
  ReaderSupplier so every vended RandomAccessReader records whether
  it was closed, runs a compaction on a caller-owned multi-threaded
  ForkJoinPool, and asserts no reader is left open once compact()
  returns. The test fails (5 still open) on main and passes with
  the fix.

Test plan

• testCompactionClosesAllReaders passes with the fix, fails without it
• Full TestOnDiskGraphIndexCompactor suite (9 tests) passes

🤖 Generated with Claude Code