Every user interview gives me a different feature request. How do I turn messy feedback into a real prioritisation decision?

Raw feature requests are the worst input to a roadmap because users describe solutions, not problems. As a starting point: for every request, dig one layer down to the underlying problem and how often it bites them, then cluster by problem, not by feature, and you will usually find three real problems hiding behind twenty requests. Build for the problem the most users hit most painfully, and let the specific solution be your call, not theirs.