thanks!

BTW, for pure evaluation patches, testing with Use NNUE=false is correct, and there is no need to do non-regression tests with Use NNUE=true.

@vondele I understand, but imagine with me this imaginary scenario:
Imagine an evaluation patch that makes stockfish gain let's say 3 ELO at middle game, and losing 1 ELO at endgame
This patch will pass the NNUE=false tests, but when you merge it to the main stockfish, this will harm it, since main stockfish uses the classical evaulation only in endgame (or when material is low) right?
That is why I like to make sure that my patch is good enough to pass the NNUE=false, and I still like to make sure that it doesn't lose anything in the NNUE=true.
What is your thought on this?

I understand the scenario you sketch. However, I want to keep development of classical eval as straightforward as possible, and keep it optimized for classical evaluation.

Furthermore, the impact of classical on use NNUE is small, via hybrid (now mostly for imbalanced psq score, not just endgame or midgame), and most patches to eval will be essentially inconsequential. Yet, our testing will still reject 20% of 0 Elo patches, so we would quite commonly reject a good Elo gainer, just because the impact on use NNUE is 0.

I understand and agree with your vision