I Scanned Dozens of Real Backends into OpenAPI. Here's Why the Naive Approaches Fail.
Hand-writing OpenAPI drifts, regex mistakes cache.get() for a route, and asking an LLM to summarize a repo hallucinates endpoints. Here is the deterministic way to scan real codebases into OpenAPI 3.2: prove the contract, flag every gap, and let AI fill only what it cannot invent.