Architecture Tests Can Pass While Leaving Real Bugs Undetected

A software engineer reviewing a .NET codebase discovered that an architecture test designed to catch unsafe synchronous disposal of dependency-injection scopes was passing despite missing a background job with the same vulnerability. The test's detector logic was sound and included self-tests, but its file discovery was scoped to only part of the codebase, leaving a legitimate source tree unscanned. The incident highlights that architecture guards carry three distinct contracts: the detector's ability to recognize a violation, the discovery process that finds all relevant files, and the boundary definition that determines where a behavior is legally permitted. A passing test count can appear healthy while entire ownership areas remain invisible if scan roots are chosen by convenience rather than by actual ownership rules. The author now recommends combining minimum-count assertions with known sentinel files from each expected source root and periodically auditing the full repository against the guarded file set.