How I verified prompt-cache reads in a generation-repair workflow

The cache was receiving writes without delivering reuse I found the caching problem in Launcherry's production usage records: repeated generation calls wrote prompt tokens to cache but read nothing ba