BLAKE3 is a nice addition, but in circuits the tree mode is the part I’d worry about, not the compression function itself. The chunk fanout and finalization rules are cheap in software, yet they can turn into a bunch of extra bookkeeping constraints if the DSL doesn’t have a clean way to express the Merkle-ish structure. I’d also be curious whether zkGolf is targeting the full XOF behavior or just the fixed 32-byte hash output, because that choice changes the gadget a lot.