feat(platform): declare Windows + macOS support (classifiers, CI matrix, README)

[HumanBean17]

Update (commit 2) — Windows CI surfaced two real product bugs

The first Windows run failed (40 failed + 11 errors). Root-caused and fixed (not skipped):

1. _fdlimit.py — import resource at module scope. The resource module is Unix-only, so the import raised ModuleNotFoundError on Windows. Because both cli.py and server.py import raise_fd_limit, the entire CLI and MCP server failed to start on Windows. Guarded the import (try/except → None); the function already no-ops when RLIMIT_NOFILE is absent. test_fd_limit skipped on win32 (it asserts Unix rlimit behaviour).
2. build_ast_graph.py — db.close() never called. All 8 cleanup paths in incremental_rebuild + write_ladybug closed the Connection but leaked the Database handle. Windows uses mandatory file locking, so the leaked handle holds the .lbug lock and the next open fails with "Error 33: another process has locked a portion of the file." This broke real init/increment on Windows, not just tests. Added db.close() after every conn.close(); same leak fixed in 4 tests that opened their own handle.

macOS leg passes; verified locally that the 53 affected tests pass on macOS (serial) — no regression. Pushing to re-run Windows.

---

Scope

Make Windows a first-class, officially-declared supported platform (alongside macOS). A pre-PR audit confirmed the tool already runs on Windows — every native dependency ships win_amd64 wheels and the Python path is portable — but nothing declared or verified it. This PR closes the packaging / CI / docs gap and fixes the two real Windows bugs the new CI leg exposed.

What Changed

• pyproject.toml — Operating System :: POSIX :: Linux, :: MacOS :: MacOS X, :: Microsoft :: Windows classifiers.
• .github/workflows/test.yml — 3-OS matrix (ubuntu/macos/windows-latest). Ubuntu is the required gate; macOS + Windows are continue-on-error + fail-fast: false until green across several merges.
• README.md — portable smoke-test paths (/tmp/... → tmp/..., gitignored); Linux/macOS/Windows line in Install; Git Bash/WSL note.
• java_codebase_rag/_fdlimit.py — guarded import resource (was crashing Windows on import).
• build_ast_graph.py — db.close() after every conn.close() (was leaking the Database handle → Windows Error 33).
• 4 test files — paired db.close() with conn.close(); test_fd_limit skipped on win32.

Semantics / Non-Goals

• Product behaviour unchanged on Unix (closing db after conn is correct; verified).
• macOS/Windows CI legs remain continue-on-error until observed green — promote by dropping them from the continue-on-error expression.
• Did not change ladybug_queries.py (the MCP server's long-lived read-only singleton — one persistent handle, no conflict in normal operation).

Validation

• Local (macOS): pytest tests/test_incremental_graph.py tests/test_cross_service_resolution_flag.py tests/test_feign_not_exposer.py tests/test_fd_limit.py -q → 53 passed ✅
• py_compile on all 6 changed files ✅; _fdlimit.raise_fd_limit() still works on macOS ✅; 8 db.close() sites confirmed in build_ast_graph.py ✅
• CI: Ubuntu ✅, macOS ✅ on the first run. Windows re-running after the fixes above — that result is the open question this update addresses.

Manual evidence

Native wheels audited on PyPI: cocoindex 1.0.14 (win_amd64), ladybug 0.17.1 (pinned; cp310–cp314 win_amd64+win_arm64), plus established lancedb/pyarrow/torch/tree-sitter wheels.

Out of Scope Confirmed

• Hard-blocking macOS/Windows gates (deferred until green).
• cmd/PowerShell README variants beyond the Git Bash/WSL note.
• /tmp/ cleanup in docs/* operator docs.

Definition of Done

• OS classifiers added.
• CI matrix runs on 3 OSes; Ubuntu remains the required gate.
• README paths portable; Windows note added.
• resource import guarded — CLI/MCP server starts on Windows.
• db.close() lifecycle fixed — init/increment work on Windows.
• Windows CI leg observed green on this re-run → then promote to a hard gate.

🤖 Generated with Claude Code