Problem

OpenClaw on Windows is 2-3x slower than on Mac. CPU profiling (120s full flow) reveals lstat system calls as the dominant bottleneck:

• 81s non-idle time, of which 47.8s (59%) is spent in lstat
• safeRealpathSync / isPathInsideWithRealpath called ~48,000 times → ~976,000 lstat calls
• Most callers do not pass a cache parameter, causing the same path prefixes to be repeatedly resolved with identical results
• Windows single lstat: 49µs vs Mac 1.1µs (45x slower per call) — multiplied by ~1M calls = 47s

Root Cause

Three re-export modules (path-safety.ts, path-guards.ts, boundary-path.ts) forward safeRealpathSync and isPathInsideWithRealpath from @openclaw/fs-safe/path without any cross-call caching. Each invocation starts fresh, even for identical path prefixes that were just resolved moments ago.

Proposed Fix

Inject a global Map<string, string> cache at the re-export layer:

1. path-safety.ts: Wrap safeRealpathSync with a dual-layer cache (global + caller-local)
2. path-guards.ts: Import from path-safety.js (instead of directly from @openclaw/fs-safe/path) so all 25 importing files benefit from the global cache
3. boundary-path.ts: Add input→output result cache for resolvePathViaExistingAncestorSync

Rewrite isPathInsideWithRealpath to call the cached safeRealpathSync from its own module.

Benchmark Results (120s CPU Profile, before → after)

Implementation

• 1 commit, 3 files changed
• Branch: fix/windows-lstat-global-cache