fix: report const generic args of the wrong type

[hongjr03]

This pr fixes #22953 (supersedes #23045). The panic from the issue was already fixed upstream by removing borrowck in #23043 .

The root cause is that although the ConstArgHasType constraint is registered when indexing an array, the resulting SelectionError::ConstArgHasWrongType was dropped implicitly in solver_errors::from_fulfillment_error, so no diagnostic was produced.

[ChayimFriedman2]

More diagnostics is always a good thing and I would not oppose it, but... are you sure this really fixes the root cause? We have no logic in MIR lowering to skip lowering if there are diagnostics (we should have though, but it's tricky to decide what errors should prevent lowering).

Also about this diagnostic specifically, rustc does emit a diagnostic here but I'm not personally sure we should: of course we should emit a diagnostic, but when the the mismatch actually occurs (i.e. when you're using a i64 as an array length or similar) and not when the constant is used (rustc emits at both).

View changes since this review

[hongjr03]

More diagnostics is always a good thing and I would not oppose it, but... are you sure this really fixes the root cause? We have no logic in MIR lowering to skip lowering if there are diagnostics (we should have though, but it's tricky to decide what errors should prevent lowering).

You're right. The change now just plugs ConstArgHasWrongType into the same type-mismatch path that lower_to_mir_with_store already uses to bail out — emit_trait_errors now reports it and inference marks it in nodes_with_type_mismatches. So no new policy needed, it just treats const-arg mismatches as another type mismatch.

Also about this diagnostic specifically, rustc does emit a diagnostic here but I'm not personally sure we should: of course we should emit a diagnostic, but when the mismatch actually occurs (i.e. when you're using an i64 as an array length or similar) and not when the constant is used (rustc emits at both).

I dropped the ConstArgHasWrongType diagnostic path and instead emit from type lowering, at the point where the constant is lowered against its expected type. So we now get the actual "mismatch occurrence" site.

Thanks for the review!