My semi-controversial take is that the best way to use agents for Rust is to literally turn every possible lint and warning on in clippy and require that they're treated as blocking for the agent to fix before ever saying things are done. While a few of them might be overzealous and feel tedious to work around manually, there's no need to care about it being tedious for an agent!
I was pretty explicitly talking about LLM-generated code. At the end of the day, you're either reviewing the code it outputs or you're not. If you are, I'd argue that you're going to be getting a bunch of ugly things in your code that you're going to want to clean up anyhow, and the tradeoff of having everything enabled is well worth it compared to what an LLM will generate with more relaxed lints. If you're not, I don't think you particularly need to care about "ugly" code.
Yeah, but the point about boxing large variants still stands. An unboxed large variant is not done for aesthetics/readability; if it's done, it's either for micro-perf (e.g. cache-coherence), or potentially safety / atomicity.
(And it can even be about semantics/compatibility, if you're doing it across an ABI boundary.)