I'm having trouble setting up a rigorous benchmark (Windows doesn't want to install greenlet for pyperformance), but just running a couple of individual files, I got this:

Mean +- std dev: [nbody_main] 208 ms +- 2 ms -> [nbody_specialized] 180 ms +- 2 ms: 1.16x faster
Benchmark hidden because not significant (1): pidigits
Mean +- std dev: [chaos_main] 127 ms +- 2 ms -> [chaos_specialized] 120 ms +- 1 ms: 1.06x faster
Mean +- std dev: [spectral_norm_main] 190 ms +- 10 ms -> [spectral_norm_specialized] 175 ms +- 1 ms: 1.09x faster
Mean +- std dev: [raytrace_main] 588 ms +- 48 ms -> [raytrace_specialized] 540 ms +- 4 ms: 1.09x faster

Hopefully those are accurate.

(Windows doesn't want to install greenlet for pyperformance)

I had the *exact* same issues, I eventually found a workaround for it after many hours spent guessing.

Initially, setuptools complained that I needed MSVC++ 14.0 or later (even after I had the latest one installed). I found that for some strange reason, *only* 14.0 worked, 14.2x etc. don't. After installing MSVC 14.0, there was then some strange complaint about missing some .exe/.dll. Searching that entire error message led me to a result on StackOverflow advising copying said files from the Windows SDK in Visual Studio over to the MSVC 14.0 folder. This finally allowed greenlet to compile. I've since lost the exact SO links, but I hope this leads you somewhere.

Anyways, I don't recommend benchmarking on Windows for stable results (trust me, I've tried ;-). pyperf system tune doesn't work on Windows. This leads to very inconsistent results unless you manually disable turbo boost, set core affinities, etc. Also, PGO for _PyEvalFrameDefaultEx might be broken on Windows on the main branch (see bpo-45116).

Eventually I gave up and just used Linux for stable benchmarking. pyperformance compile_all also works properly there, which allows you to automate your benchmarks https://pyperformance.readthedocs.io/usage.html#compile-python-to-run-benchmarks

PS. are your numbers with PGO and LTO? If so, they're spectacular!

Hm the above was not PGO. I tried again with PGO and it is not so good:

Mean +- std dev: [nbody_main_pgo] 177 ms +- 4 ms -> [nbody_specialized_pgo] 190 ms +- 2 ms: 1.07x slower
Mean +- std dev: [pidigits_main_pgo] 208 ms +- 1 ms -> [pidigits_specialized_pgo] 210 ms +- 2 ms: 1.01x slower
Mean +- std dev: [chaos_main_pgo] 106 ms +- 1 ms -> [chaos_specialized_pgo] 110 ms +- 1 ms: 1.04x slower
Mean +- std dev: [spectral_norm_main_pgo] 169 ms +- 7 ms -> [spectral_norm_specialized_pgo] 167 ms +- 1 ms: 1.02x faster
Benchmark hidden because not significant (1): raytrace

If some misses are caused by mixed int/float operands, it might be worth investigating whether these occur in loops.

Most JIT compilers perform some sort of loop peeling to counter this form of type instability.

E.g.
x = 0
for ...
x += some_float()

x is an int for the first iteration, and a float for the others.

By unpeeling the first iteration, we get type stability in the loop

x = 0
#first iteration
x += some_float()
for ... #Remaining iterations
    x += some_float()  # x is always a float here.

New changeset 3b3d30e by Dennis Sweeney in branch 'main':
bpo-45367: Specialize BINARY_MULTIPLY (GH-28727)
3b3d30e