Windows: restore daemon boot shims + Task Scheduler lifecycle (v2.3.0 still can't start on Windows)

[danielhertz1999-bit]

Summary

On Windows, the daemon cannot start at all on current main (v2.3.0): python -m iai_mcp.daemon dies during import/boot on a chain of POSIX-only calls, and iai-mcp daemon install/start have no Windows path. This PR restores the cross-platform shims (which existed in the v1 Windows port but were dropped in the v2.0.0 rewrite and never re-added) and guards the remaining POSIX-only sites, so the daemon boots and is manageable on Windows again.

All changes are no-ops on macOS/Linux — they are guarded by sys.platform/os.name/hasattr/getattr checks or route through a shim that delegates to the original stdlib call on POSIX.

What was broken on Windows

Starting the daemon crashed in sequence on:

1. daemon/__init__.py — top-level import resource (POSIX-only → ModuleNotFoundError)
2. daemon/__init__.py — signal.SIGHUP referenced in an eagerly-evaluated tuple (AttributeError before the surrounding try/except runs)
3. socket_server.py — inspect.signature(asyncio.start_unix_server) evaluated before the existing IS_WINDOWS branch that already avoids it (AttributeError)
4. capture_queue.py, lifecycle.py, lifecycle_event_log.py, lillibrain/io.py — unconditional import fcntl (the _filelock.py shim they used was deleted in the v2.0.0 rewrite 1a333da)
5. doctor/_lifecycle_checks.py::check_c_lock_healthy — local import fcntl
6. cli/_daemon.py — daemon install/uninstall/start call os.getuid() before any platform branch and have no Windows branch (install/uninstall fall into Linux-only systemctl logic; start prints "Unsupported OS")
7. cli/_crypto.py::cmd_crypto_status — unguarded os.geteuid() (AttributeError)
8. capture.py, migrate/_to_lilli_verify.py — os.fchmod not guarded (AttributeError; sibling sites in crypto.py/memory_bank.py already guard it)

Changes

Restore the file-locking shim (commit 1)

• Re-add src/iai_mcp/_filelock.py — a cross-platform flock-compatible shim (LockFileEx/UnlockFileEx on Windows, delegates to stdlib fcntl on POSIX), imported as from iai_mcp import _filelock as fcntl so call sites are unchanged.
• Repoint capture_queue.py, lifecycle.py, lifecycle_event_log.py, lillibrain/io.py, doctor/_lifecycle_checks.py at the shim.
• daemon/__init__.py: make import resource local + gated on sys.platform == "win32" inside _raise_fd_limit() (fd-limit raising is a POSIX-only concept — early-returns on Windows); guard signal.SIGHUP with getattr(signal, "SIGHUP", None).
• socket_server.py: move the inspect.signature(asyncio.start_unix_server) call below the existing IS_WINDOWS branch.

Task Scheduler daemon management + remaining guards (commit 2)

• cli/__init__.py: add _is_windows() and SCHTASKS_TASK_NAME.
• cli/_daemon.py: add _find_pythonw() + _render_schtasks_xml(), and real Windows branches to cmd_daemon_install/uninstall/start (schtasks /Create+/Run, /End+/Delete, /Run). Guard the macOS/Linux os.getuid() calls with hasattr(os, "getuid"). (daemon stop already had its taskkill /F /T branch.)
• cli/_crypto.py: gate os.geteuid() behind os.name != "nt" (mirrors the read-path guard in crypto.py); report mode_secure/uid_matches_process as null on Windows, where st_mode/st_uid are synthetic and access is governed by ACLs.
• capture.py, migrate/_to_lilli_verify.py: guard os.fchmod with hasattr(os, "fchmod"), matching the existing crypto.py/memory_bank.py sites.

Verification (Windows 11, Python 3.12)

• python -m iai_mcp.daemon boots to WAKE and reports version: 2.2.2 (verified on the v2.2.2 base these commits were authored against; rebased cleanly onto v2.3.0 here — all 13 files byte-compile on the new base).
• iai-mcp daemon status → healthy; iai-mcp doctor passes except the pre-existing .daemon.sock-absent check (Windows uses the TCP+auth-token transport, not a Unix socket).
• iai-mcp crypto status, iai-mcp daemon install --dry-run, iai-mcp daemon start all run cleanly (each previously crashed).
• Targeted CLI test suite (test_cli_daemon.py + test_cli_crypto.py) goes 26 → 15 failures on a Windows host (fixed 11, zero regressions vs. the pre-patch baseline). The 15 remaining are pre-existing POSIX-specific test assertions running on Windows (e.g. asserting mode 0o600 where Windows reports 0o666, mocked launchctl paths) — a separate test-suite-port concern, not addressed here.
• A whole-tree sweep of six POSIX-only bug classes (uid family, fchmod/chown, fcntl/resource imports, POSIX signals + os.kill, AF_UNIX sockets, fork/setsid/rlimit) found no further unguarded sites after these changes.

Notes

• No behavior change on macOS/Linux — every change is platform-guarded or shim-delegated.
• Happy to split into two PRs (boot shims / daemon-lifecycle) if you'd prefer to review them separately.

🤖 Generated with Claude Code