The ATI Radeon 7000 series, which came out a few months after the GeForce 256, did have programmable vertex processing, even if that feature was internal to the device driver and not exposed to the application. I know because I worked on that stuff a few years later when it trickled into smartphones.

I can't find confirmation that the GeForce 256's implementation of HW T&L also relied on microcode running in the vertex processor, but it sure sounds like the most obvious way to implement that feature.

Just because a feature isn't visible through the API that is exposed to the general public, it doesn't mean it doesn't exist internally. I know of multiple examples of this.

Edit: In the drivers I had access to for the Radeon 7000 series, which I presume were very close to the original desktop drivers, the HW T&L microcode was generated on the fly, directly in machine code. There was no compiler; just a few C macros to put a handful of instructions together. It was mainly vec4 dot products. I thankfully had little exposure to that Eldritch horror. I remember a very talented coworker from Bitboys, a Finnish startup acquired by ATI, explaining gleefully how he was optimizing the code generator to save an instruction here and there. Sheer insanity sprinkled with salmiakki, as far as I was concerned.