[ocaml5-issue] Assertion failure on live words in GC test

[jmid]

The new Gc test in #469 triggered an assertion failure in Linux debug 5.3.0+trunk:
https://github.com/ocaml-multicore/multicoretests/actions/runs/10581716289/job/29319700079?pr=469

random seed: 369983017
generated error fail pass / total     time test name

[ ]    0    0    0    0 / 1000     0.0s STM Gc test sequential
[ ]    0    0    0    0 / 1000     0.0s STM Gc test sequential (generating)
[ ]  728    0    0  728 / 1000    61.9s STM Gc test sequential
[✓] 1000    0    0 1000 / 1000   104.3s STM Gc test sequential

[00] file runtime/shared_heap.c; line 1392 ### Assertion failed: local->stats.pool_live_words == pool_stats.live
File "src/gc/dune", line 4, characters 7-16:
4 |  (name stm_tests)
           ^^^^^^^^^
(cd _build/default/src/gc && ./stm_tests.exe --verbose)
Command got signal ILL.
[ ]    0    0    0    0 / 1000     0.0s STM Gc test parallel

[jmid]

This one is easily reproducible.
I've commented out the sequential test and can trigger the assertion failure on every run, regardless of the seed:

$ OCAMLRUNPARAM="s=4096,v=0,V=1" dune exec src/gc/stm_tests.exe --profile=debug-runtime -- -v
random seed: 494408706
generated error fail pass / total     time test name
[ ]   41    0    0   41 / 1000     7.1s STM Gc test parallel[00] file runtime/shared_heap.c; line 1392 ### Assertion failed: local->stats.pool_live_words == pool_stats.live
Illegal instruction (core dumped)
$ OCAMLRUNPARAM="s=4096,v=0,V=1" dune exec src/gc/stm_tests.exe --profile=debug-runtime -- -v
random seed: 97187661
generated error fail pass / total     time test name
[ ]   16    0    0   16 / 1000     2.6s STM Gc test parallel[02] file runtime/shared_heap.c; line 1392 ### Assertion failed: local->stats.pool_live_words == pool_stats.live
Illegal instruction (core dumped)
$ OCAMLRUNPARAM="s=4096,v=0,V=1" dune exec src/gc/stm_tests.exe --profile=debug-runtime -- -v
random seed: 330070200
generated error fail pass / total     time test name
[ ]    5    0    0    5 / 1000     1.1s STM Gc test parallel[00] file runtime/shared_heap.c; line 1392 ### Assertion failed: local->stats.pool_live_words == pool_stats.live
Illegal instruction (core dumped)

[jmid]

This is known upstream as a sporadic failure ocaml/ocaml#13090.
I've shared a standalone reproducer on that issue.

[jmid]

I've arrived at a simpler reproducer and shared details here on the upstream issue: ocaml/ocaml#13090 (comment)

[jmid]

This has now been fixed upstream on trunk in ocaml/ocaml#13774 but may continue to trigger on 5.3.