Make the store lock and the hook timeout work on Windows

[sebplace]

agent-memory does not start on Windows. core/locking.py imports fcntl at module level, so every entry point — mem, mem-hook, mem-mcp — dies with ModuleNotFoundError: No module named 'fcntl' before it can do anything.

This PR makes the two POSIX-only spots portable. Both are narrow.

core/locking.py

store_lock used fcntl.flock for the pipeline-level advisory lock. It now switches on os.name:

• Windows: msvcrt.locking, taking one byte at position 0 with LK_NBLCK and releasing it with LK_UNLCK.
• Everything else: the same fcntl.flock calls as before.

The contention branch now catches OSError rather than BlockingIOError. That is deliberate: msvcrt.locking raises a plain OSError when the region is held, and BlockingIOError is a subclass of OSError, so the POSIX path keeps behaving exactly as it did.

adapters/hook_entry.py

_arm and _disarm used signal.SIGALRM and signal.setitimer, neither of which exists on Windows. Both are now guarded with hasattr, so on Windows the hook runs without a wall-clock guard instead of failing to start.

That felt like the right trade-off given the module's own docstring: a hook that breaks its host is worse than a hook that never fires. Happy to swap it for a thread-based timeout if you would rather keep the guard on every platform.

No behaviour change on POSIX

The same fcntl calls run on the same paths. The Windows branch is only reachable when os.name == "nt".

Verification

Windows 11 ARM64, CPython 3.12.

• mem init, mem record, mem recall complete a full round trip, and BM25 recall returns the seeded entries with sensible ranking.
• mem-mcp answers an initialize plus tools/list handshake over stdio with its five tools.
• Test suite: 377 passed.

The 11 remaining failures are unrelated to this change and were present in the same form before it:

• 9 are the memcore harness attempting to exec a Linux binary (OSError: [WinError 193]).
• 2 are assertions expecting POSIX separators (fact/rogue.md) where the library returns os.sep paths (fact\rogue.md).

That second group points at a separate portability wrinkle: relative paths in IndexReport are built with the platform separator rather than normalised to /. It is arguably inconsistent with the portable-store promise, but it is a distinct concern and I left it out of this PR rather than widening the diff. Glad to open a follow-up if you would like it fixed.