Eighteen of my guards were down and every log said polling normally.

A health check that asked whether each process had crashed reported zero dead guards for weeks. Eighteen of them were failing on every poll with the same Python TLS handshake timeout, and each wrote one log line identical to an ordinary retry. The per-log failure counts, the loop shape that makes the failure invisible, the one-function fix that covered all eighteen, and the nine-case test whose two most important cases catch a naive fallback that would turn a legitimate 402 into a duplicate side effect.