Fixes #561.

What

DocStore.lock() used fcntl.flock() for its cross-process mutex, but fcntl doesn't exist on Windows — the ImportError fallback silently yielded with no lock at all. Concurrent submit_document() calls on Windows could both pass the _unique_doc_name() check and save under the same filename, producing duplicate document names.

The Windows branch now takes a real cross-process lock via msvcrt.locking on the same .lock file, and an in-process threading.Lock keyed by storage directory additionally serializes this process's threads — including separate DocStore instances pointed at the same directory. The name-uniqueness check and the save already ran inside with self._store.lock() in submit_document(), so with a working lock the invariant holds. POSIX behavior (fcntl.flock) is unchanged.

Changes

• pageindex/local_store.py: cross-platform lock as above (not re-entrant, documented in the docstring).
• tests/test_local_store.py: regression tests simulating the Windows branch (hidden fcntl, fake msvcrt whose locking() truly serializes):
  • test_lock_windows_branch_serializes_threads: 8 threads x 25 iterations across two DocStore instances on one directory; asserts no lost increments and that the OS-level lock was actually acquired/released each time.
  • test_concurrent_same_name_submits_store_unique_names: the issue's repro — 8 concurrent same-name submit_document() calls receive sample.pdf, sample_1.pdf, ... sample_7.pdf.

Verified

• Both new tests fail on main (duplicates observed, e.g. Document "sample.pdf" was stored as "sample_2.pdf") and pass with the fix.
• Full suite: 712 passed; the 76 remaining failures are pre-existing environment issues (no running event loop / network-dependent LLM credential tests), identical counts on unmodified main.
• The issue also suggests Windows CI coverage: I have a windows-latest job running tests/test_local_store.py ready, but my token lacks the workflow scope needed to push workflow files — happy to add it on request, or it can land as a follow-up.