…819)

PR #818's Windows CI cell surfaced six failures that POSIX CI hadn't
caught. Two distinct root causes:

1. Windows ``LockFileEx(LOCK_EX)`` blocks reads from *other* handles,
   not just other lock attempts (POSIX ``flock`` is advisory and lets
   reads through). So ``pid_file.read_text()`` from a fresh handle
   raises ``PermissionError`` (ERROR_LOCK_VIOLATION) while the test
   or ``main()`` still holds the lock.

   - ``test_contended_server_start_preserves_pid_file_content`` now
     reads through the lock-owning ``holder`` (``holder.seek(0); read()``)
     instead of opening a separate handle. Cross-platform safe — POSIX
     was happy either way.
   - ``test_server_main_acquires_portalocker_pid_lock`` cannot reach
     ``main()``'s closure-scoped ``_lock_fp`` to read through it. The
     pid-content equality assertion is now POSIX-only; the
     cross-platform load-bearing assertions (``pid_file.exists()``,
     ``probe_pid_file().alive is True``, post-cleanup ``not exists``)
     still run on every OS. ``probe_pid_file`` is already designed to
     treat the Windows lock-blocks-reads case as alive (catches
     ``OSError`` at open), so the alive-probe is the right check there.

2. ``TestServerAliveRefuses`` got its class-level Windows skip removed
   in #818 because ``_hold_pid_lock`` now uses portalocker. The lock
   mechanics work cross-platform, but four cases assert POSIX-specific
   user-facing text or POSIX-specific ``--force`` behavior:

   - ``test_refuses_when_server_alive_at_legacy_path``,
     ``test_refuses_when_server_alive_at_runtime_path``: assert
     ``"Server still running"`` and ``str(os.getpid())`` in output;
     Windows substitutes a Sysinternals/handle.exe hint that omits
     the recorded pid.
   - ``test_refuses_with_unknown_pid_branch_when_pid_file_empty``:
     asserts ``"lsof"``; Windows uses Resource Monitor / handle.exe.
   - ``test_force_overrides_liveness``: asserts ``--force`` succeeds;
     Windows refuses by design (#730 — ``unlink-while-open`` raises
     WinError 32).

   These four get per-test ``skipif(sys.platform == "win32")`` back
   with a reason that points at #819. The other tests in the class
   continue to run on Windows. Widening the assertions to accept
   the Windows wording is a separate UX-side follow-up.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>