docs(claude): Lesson 26 — jail-at-entry doesn't jail the recursive walk (security review caught what my review missed)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE.md

@@ -447,3 +447,5 @@ Things I discovered about myself building a sister platform across multiple sess
 24. **`onModuleInit` runs before async `onModuleInit` of dependencies completes — register NATS/external subscriptions in `onApplicationBootstrap`.** `NatsService.onModuleInit` connects to NATS (async); `NatsBridgeService`/`HostAgentConsumerService` registered their subscriptions in their own `onModuleInit`, which fired while the connection was still null — so every `subscribe()` hit the `[OFFLINE]` no-op path and the WS bridge was dead-on-boot in *every* production build, silently. Nest guarantees `onApplicationBootstrap` runs only after all module init (including the awaited connect) finishes. Anything that depends on another provider's async startup belongs in bootstrap, not init. The tell: a subscription that "should be there" but the handler never fires and there's no error — trace the *startup ordering*, not the handler.
 
 25. **Fixing a dead code path detonates the live code behind it — budget for the second bug.** The moment Lesson 24's fix made the NATS→WS bridge actually deliver events, the API crashed on the first forwarded heartbeat: `WebSocket.OPEN` was `undefined` at runtime because `esModuleInterop` is off, so `import WebSocket from 'ws'` compiled to `ws_1.default` (undefined). That crash had sat behind the dead bridge since the gateway was written — never hit because no event ever reached it. When you resurrect a path that was silently no-op, everything downstream of it is effectively *untested code running for the first time in production*. Verify the whole chain end-to-end (I watched the DB row appear, then flip offline), don't stop at "the subscription fires now." This is Lesson 10 with a fuse on it. Import-runtime gotcha worth remembering: when `esModuleInterop` is off, prefer instance constants (`client.OPEN`) over class statics (`WebSocket.OPEN`) for `ws`.
+
+26. **A jail check at the entry point does not jail the recursive walk behind it — and my own "line-by-line" review missed it; the automated security review didn't.** The file manager's `jail()` correctly canonicalized and prefix-checked the top-level path, and I traced every escape vector through it and signed off. But `copy_recursive` then walked the directory tree with `fs::metadata` (which *follows* symlinks). A symlink planted inside the jail pointing at `/etc`, then a `copy` of its parent, would dereference it and pull external content *into* the jail to be read — a jail escape the entry check never sees, because the escape is reintroduced by a descendant during traversal. Fix: `symlink_metadata` (lstat) everywhere you recurse, and refuse/never-follow symlinks across the boundary. The transferable rule: **validate at the boundary AND at every step that re-derives a path** (recursion, `read_dir`, glob, archive extraction). And the humbling part — I was confident after reviewing the jail function; the security-review pass caught the HIGH I'd waved through. Trust adversarial verification over your own once-over on security-critical code, especially path/traversal logic.