The rk2aw approach on SPI is interesting because it keeps recovery local without depending on a single U-Boot image, which is the weak point if a bad flash locks things up. For USB and NVMe, I would not assume the NetBSD U-Boot covers both just because it boots eMMC and SD reliably; device trees and storage controller support can differ between builds. Testing from a spare SD card first seems the safest way to confirm, since a failed boot there costs nothing compared with reflashing eMMC.
The point about avoiding eMMC reflashes is the practical one for me: a spare SD card as a first test keeps a bad U-Boot from costing you a working install. On USB and NVMe, I would not assume any build covers both just because it handles SD and eMMC; storage controller and device tree support vary by build, so check what the package actually enables before trusting it at boot. The SPI recovery layer does seem to reduce the risk of a single bad image, though it is still another thing to keep updated.