test(sdk): pin the invariants a reviewer cannot see

[cport1]

Companion to WebDecoy/app's backend guardrails. Same reasoning: every finding in the 0.12/0.13 batch was two places answering one question, and the leftmost-X-Forwarded-For bug survived in two adapters after the same class was fixed elsewhere — because each call site read reasonably on its own and nothing connected them.

Four invariants

• The client IP is resolved only in client-ip.ts
• Decisions are built only through the Decision class, and never spread (a spread silently strips isDenied() / deniedBy() — the trap protect() originally fell into)
• Every starvable rule has a NOT_RUN path, so "checked and fine" stays distinguishable from "never checked"
• No node: import reaches a package that ships to Workers

Writing them found two things, which is the point

A false positive in my own rule. The first version flagged detection/detectors/headers.ts, which merely lists those header names for suspicious-shape detection. That is not IP resolution. The rule now matches actual access — headers['x-forwarded-for'] or .get(...) — rather than a mention.

A real hole in my own fix. It then flagged the Next.js adapter, correctly: the x-real-ip and x-vercel-forwarded-for fallbacks I left inline during the trusted-proxy work were header reads outside the resolver — exactly the drift the invariant exists to stop, reintroduced by the person who wrote the invariant.

They now live inside resolveClientIp, gated on a declared proxy: under trustProxy: false we read no forwarding header at all, because X-Real-IP is as forgeable as the rest. Four new tests cover the relocated behaviour, including that a real forwarding chain still wins over either.

Verification

386 core tests, 20/20 turbo tasks, all three edge entry points still compatible.