Good rundown, especially the biased-reference-counting explanation, that part usually gets glossed over. One thing I'd push on: the use case you're recommending it for, ML preprocessing alongside a Node API, is also the one most exposed to the C-extension caveat you flag near the end. Plain-Python CPU work via ThreadPoolExecutor genuinely benefits today, but the moment that pipeline leans on numpy or pandas internally, which most ML preprocessing does, the actual speedup depends on whether those specific wheels ship free-threaded builds, not on whether 3.14t itself is stable. Worth naming which of the common data-science libs already ship free-threaded wheels as of 3.14 versus which still fall back to re-enabling the GIL at import (it's just a warning, easy to miss in a busy log), since that's the real gate on whether the recommendation holds for that exact workload.