Testing

A SwiftUI screen suite recorded with swift-snapshot-testing at precision: 0.99 stayed green after the row's location text was deliberately wired to the wrong value. On a full device frame, 1% of the pixels is about 30,000, and the wrong text changed about 900. Measuring the mutation and setting precision to 0.9999 turned four of the seven cases red.

Agent progress files can preserve unverified assumptions. Rebuild a small factual baseline, then check inherited claims against their original evidence.

A model may leak encoded prompt fragments yet corrupt the final decode. Capture verbatim tokens, preserve case and order, and transform them deterministically.

Dollar-prefixed action inputs can reach React Flight decoding. Reproduce decoding locally, compare error digests, and avoid mistaking a parser signal for RCE.

Passing a request line to snprintf as its format string turns logging into memory disclosure. Use a constant format and keep secrets out of process memory.

Three times a coding agent read a result from a check that never ran: zsh's `log` builtin (`(eval):log:1: too many arguments`) hid `/usr/bin/log`, `cargo test --exact` matched no test and passed, and an empty variable from "argument list too long" made 400 of 400 runs fail.

An app-hosted iOS test target passed in the console and failed in fastlane. The host app's own login screen made a network request at launch, and its response reached an unimplemented swift-dependencies test value after the tests had started. The failure belonged to no test and was visible only in the result bundle.

A Qt desktop app filtered log lines with msgType < minimumType. Qt numbers its message types in the order they were added, not by severity, so at the default INFO level the log was full of info lines and held no warning or error at all.

A Qt/SQLite history table encrypts its path columns and matches rows by re-encrypting the paths. The update had three independent faults, each one hiding the next. A store-update-load test found them, and a negative test kept the fix from dropping the match rule.

A teardown injected during login defeated a one-time wait. An exclusion gate fixed the reproduction; the production trigger remained unknown.