What does 90% code coverage tell you?

There are several kinds and they claim different things. Line coverage is the easiest to reach and the weakest. Branch coverage is stronger: it demands both the if and the implicit else, and the forgotten else is precisely where bugs live. Condition coverage goes further, requiring each sub-condition of a compound expression to take both values. Knowing the difference shows you understand the metric rather than reading the number.

A mandated CI threshold has a predictable side effect: people write tests for whatever is easiest to cover — getters, DTOs, generated code — because that's the cheapest way to move the number. Meanwhile the hard-to-cover parts (error handling, unusual paths) are exactly what most needs testing, and they're what gets left. So a mandated threshold can make a suite worse while the number improves.

What I care about more than coverage is mutation testing: a tool makes small edits to your code (flip a comparison, change a constant) and checks whether any test goes red. If the code is broken and the suite stays green, that test verifies nothing. It's much slower so it usually runs on a schedule rather than blocking PRs, but it answers the question coverage cannot.