docs(security): record the rule-8 guardrail tests as closed (§21)

[ilpanich]

Docs-only. Closes the §15.3.1 item that had been carried through §16.3a, §18.6, §19.6 and §20.5 — rule 8 now has regression coverage in all eleven SDKs.

Merged: go#29 · rust#43 · kotlin#16 · java#37 · c#15 · cplusplus#13 · csharp#33 · swift#15

A correction to the count

Every prior section said nine SDKs outstanding. It was eight. PHP already had Sec085GuardCredentialSubstitutionTest — it is where the pattern was first written, as §15.3.1 itself says — and TypeScript and Python landed in the §16 round. The "nine" came from subtracting only those two from eleven, and was repeated forward without being re-derived.

It was not only a guardrail

The Java Spring filter rejected correctly (401, chain not invoked) but never cleared the ambient SecurityContext, so a rejected caller left behind whatever identity was already on the thread — from an earlier filter, or a pooled container thread. That's the servlet analogue of the SEC-085 substitution.

§21.2 records why §15.1's hand-verification missed it, which is the part worth keeping: that pass asked whether each guard rejects, and every guard does. This is the different question of what a guard leaves behind after rejecting — a property nobody had stated, so hand-verification could not have caught it. Writing the test did.

What the section adds

• §21.3 — which SDKs carry the structural shape SEC-085 exploited (Kotlin, C and C# hand the guard a stateful client rather than a bare verifier), and how the structurally-safe ones assert the absence of a second credential instead.
• §21.4 — what makes the tests non-vacuous: each asserts a substitution was genuinely available before asserting it wasn't taken. Plus the falsification results per SDK.
• §21.5 — verification honesty: dotnet and swift were unavailable locally. C# failed CI on exactly that risk (FakeAxiamServerHandler is private and nested, so unreachable); Swift passed first try on 5.9 and 5.10.
• §21.6 — still carried, unchanged: the SEC-086 timing channel, the gRPC m2m/user parity gap, residuals 4 and 5.

§20.5's stale bullet is replaced with a pointer to §21.

No code changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SkTHvZQMV47t3UwkEtmB1D

---

Generated by Claude Code