Verification 1021 not called by the same chainings (mlang vs compir)

[ddeclerck]

Verification 1021 is called by the following chainings when compiled with mlang :

verif_saisie_cohe_corrective
verif_saisie_cohe_primitive

However, it is called by the following chainings with compir :

verif_calcul_primitive
verif_calcul_corrective

This is weird because verif 1021 only refers to input variables (context and family), so it should indeed end up in verif_saisie_xxx.
It's even weirder because verif 1019 just above also uses only input variables (context and family) and correctly ends up in verif_saisie_xxx both with mlang and compir.

The only difference I see is that rule 1021 uses a for "loop", although I can't see why this would make any difference.

[Keryan-dev]

We could either look the code of compir up, or pester people around to see what the expected result should be (and inevitably report the bug, or cringe at the weird choice made). I choose the latter.

[ddeclerck]

Having looked for an explanation in compir's source code, I too believe we should resort to the second option :D

[ddeclerck]

Investigating compir's source code a bit more, I found that this behavior is a consequence of a "fix" for some verifications that were supposedly not called under "for" constructs (semexp.c, expr_prepare_verifs_indiv). Removing the "fix" in compir did not result in missing verifications though, and the verifications were called in the right chaining. So, we're just going to ignore this weird behavior, and keep this issue here for reference.