Yes, he always has some interesting ideas, and fortunately he sometimes realizes one wasn't so good. I remember him saying something like "The map isn't the land," and although I don't remember the details it seems like he was saying there was something deficient about human-readable code. I think that was one he went back on.
I've mentioned here before that I did a test to see how much of the space of both stacks I used on the 6502, and came up with the most stack-intensive ap I had written and then had IRQ interrupts serviced in high-level Forth plus NMIs serviced in assembly language, and found that I used less than 20% of both stacks. 20% of 256 bytes is about 51 bytes. If it were all two-byte cells (it wasn't), that would be 25 cells, which is more than his 18. I wonder if he believes in interrupts. His number does seem to conflict with his ideas of keeping definitions super short and of factoring out things that are only used once, adding to the nest/unnest overhead, not to mention trying to give all these factors names that are any more meaningful than the code itself, and the additional memory penalty incurred by the extra headers. On an entirely different level is multitasking where a separate stack area is given to each task so a task doesn't have to completely finish its stack usage before relinquishing control, like a zero-stack-effect word.
I think he was also against the CASE structure. I cannot agree with him on that.
Nested definitions? Nested macros in assembly language are certainly useful, like if you want one macro to be able to invoke another so you don't have to spell out all the code that there's already another macro for. But isn't that similar to what Forth does all the time? A colon definition uses other colon definitions which use others, until you get down to the primitives, ie, code definitions which are always written in assembly.
I haven't used local variables so far in Forth, but I remain interested. The internals appear simple enough, but I haven't done it to know what challenges I'd run into. I did write about local variables in assembly language in the 6502 stacks treatise at
http://wilsonminesco.com/stacks/loc_vars.html .
Forth is great for us obstinate free-thinkers who want to do things our own way though. We don't have to agree. It lets us make our own tools to get the job done.