@chrysn What would be the idomatic way to do this change backward-compatible? The struct member hdr on the C side will no longer be present in the PR, at least when non-UDP transports are (also) used.

I'll check, but the riot-wrappers shouldn't expose any of the fields anyway, so they can just switch over to the accessors in a compatible way.

If there is transition, I think we can just do this as-is already, at latest when coap_pkt_set_code is available.

Sadly, the helper coap_pkt_set_code() is not yet in RIOT (a getter for the code was already present, though). So just switching to the setter will not build with RIOT until the PR in question is merged.

I could split out a PR to add the setter and backport that to the latest release, hoping on getting a release maintainer approval for a feature backport. But that would still bump the RIOT version requirements of the rust-riot-wrappers.

With the plan laid out in RIOT-OS/RIOT#20900 (comment), it should be fair to have this PR's change conditional: if any of the new modules are in (say, #[cfg(any(riot_module_gcoap_tcp, riot_module_gcoap_websockets))), the new version is used; otherwise, the old is used. If that's creating too much churn, it might be an option to create a pseudomodule gcoap_nonudp that the new ones depend on, but I don't anticipate that there'll be an inrush of transports right away.

We don't have a dependency on the cfg-if crate in the wrappers, so the old implementation would need to be kept in sync with a #[cfg(not(any(riot_module_gcoap_tcp, …)))] statement; I think that's fine, but if you think it is ugly, cfg-if could be added. (But all this will be out first thing after the next RIOT release anyway).

The proposed path did not work out: Neither c2rust nor bindgen support the unnamed-union trick. (In hindsight, that could have been clear: There are no unnamed unions in Rust).

The next best thing is probably introducing a marker that decides based on the presence of the new accessor whether or not to do the new thing (as Rust's rust-lang/rust#64797 is not there yet).

As that kind of markers is not around, I'm taking the liberty to add that.

To be fair: The unnamed union trick is also implementation defined behaviour in C. The assumption that two differently typed pointers to the data address space with the same memory as target hare bitwise identical is not backed by the C spec.

So this trick is exactly the kind of bad black magic that made a lot of people switch to rust.