Indeed, this reveals comparisons of float results using =.
I've also been able to reproduce locally (may require bumping count as nan is produced relatively rarely with Gen.float).
The most interesting part is that the result is sometimes printed as -nan:

### OCaml runtime: debug mode ###   
### set OCAMLRUNPARAM=v=0 to silence this message
random seed: 534962093
generated error  fail  pass / total     time test name
[✗]   982     0     1   981 / 10000     0.0s STM Float Array test sequential
[✓]     1     0     1     0 / 10000    11.0s STM Float Array test parallel

--- Failure --------------------------------------------------------------------

Test STM Float Array test sequential failed (7 shrink steps):

   Set (6, nan)
   To_seq


+++ Messages ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Messages for test STM Float Array test sequential:

  Results incompatible with model

   Set (6, nan) : Ok (())
   To_seq : [1.; 1.; 1.; 1.; 1.; 1.; -nan; 1.; 1.; 1.; 1.; 1.; 1.; 1.; 1.; 1.]

Looking more into this, it seems the cmd printer is using Util.Pp.float which uses fprintf fmt "%F"
whereas the result printer is using the Float combinator printer relying on QCheck.Print.float which uses string_of_float. We should probably use the same printer in both arguments and results to avoid false alarms.

CI summary: 16 workflows failed all due to nan float comparisons using =:
32bit 5.3
32bit 5.4
Bytecode 5.4
FP 5.4
Linux 5.3
Linux 5.3 debug
Linux trunk debug <--- CI log contains 223k lines of updating memprof. and is then cut off since ocaml/ocaml#14057
Linux-ARM64 5.3
Linux-ARM64 5.4
macOS-ARM64 trunk
MinGW bytecode 5.3
MinGW bytecode trunk
MSVC 5.4
MSVC bytecode 5.4 (still going)
musl 5.3
musl 5.4

Out of 50 workflows 16 failed all with a genuine error

CI summary: all 50 workflows passed - merging

CI summary for merge to main: 51 workflows passed