How much should I trust a 5x codegen win over reflection for populating POJOs?

That result is useful as a signal, but it describes the particular benchmark, not a general rule that generated code is five times faster. The details matter: how the benchmark is warmed up, whether setup and class generation are measured, how many fields are populated, and whether the reflective path caches lookup information.

Yes, I would include a VarHandle version. Repeatedly looking up fields or methods inside the timed operation can make a reflection baseline look worse than a realistic implementation. Resolve accessors once, then compare the cost of the actual writes. Still, generated direct assignments may remain faster, so it is worth measuring rather than assuming.

Also separate one-time costs from steady-state throughput. Generating and loading a class has a cost, while reflection can be ready immediately. If the generated mapper is reused many times, its startup cost may be negligible. If it is used only for a few objects, the overall tradeoff could look quite different.

Check that both approaches do the same work. Type conversion, missing keys, null handling, and access checks can easily change the numbers. I would test a few object sizes and input shapes, and report operations per second alongside the benchmark setup. A single ratio can hide a lot.