feat(buffa): add MessageFullName trait

[yordis]

• Event-sourced systems identify each event on the wire by its fully-qualified type name (e.g. user.UserCreated). To use a buffa-generated message directly as a domain event. Skipping the usual hand-written mapping layer between "domain type" and "storage type". Callers need that name as a compile-time constant on the type itself. Today it isn't there: the closest surface, ExtensionSet::PROTO_FQN, is bundled with the extension machinery and only emitted when a message opts into unknown_fields=true, which adds a hidden __buffa_unknown_fields field to the struct and forces every struct-literal construction in domain code to trail ..Default::default(). Reaching for buffa::ExtensionSet just to read the FQN therefore drags in storage and ergonomic costs the caller doesn't want.
• Following the same shape as protocolbuffers/protobuf#27111, this PR introduces buffa::MessageFullName as a standalone trait with a single const FULL_NAME: &'static str item. Kept out of Message precisely so the FQN concern stays orthogonal and adding it doesn't break any hand-written impl Message in downstream code. Generic call sites import the trait (use buffa::MessageFullName;) and read T::FULL_NAME at compile time with no runtime indirection, so event stores, type registries, and Any type-URL construction can dispatch on it instead of maintaining a hand-written match arm per message.
• ExtensionSet::PROTO_FQN is intentionally left in place to keep this change additive; the two consts are emitted from the same proto_fqn source in codegen so they cannot drift, and that invariant is documented at both sites.
• The name MessageFullName / FULL_NAME mirrors the convention used across the protobuf ecosystem (and matches protocolbuffers/protobuf#27111 to keep the two PRs aligned). Not attached to it. Happy to rename if a different spelling reads better here.

Related: TrogonStack/trogonai#108.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[yordis]

To be extra clear, intentionally opt-out from unknown_fields=true since that causes extra fields that I wish to avoid to codegen, having the trait to a very specific things helps with the Interface segregation principle, which is effectively what is going on for me

[yordis]

@iainmcgin could help me on this one? right now I decided to manually maintain the string since the extra extension stuff add a lot of noise in the PR, I am hoping this would be the direction to avoid the interface segregation anti-pattern. Most likely, breaking changing the extension set I think

[iainmcgin]

Hi yes I've been tinkering with this one a bit, thanks for the contribution. It's a very busy week but I'll try and get something together in the next few days for this - likely as part of a 0.6.0 release.

[yordis]

Aye, I am around to do the work, just let me know and I jump in

[iainmcgin]

[claude code]

Thanks for the PR — the motivation (compile-time FQN access for type-erased dispatch without descriptor machinery) is a real gap and the trait shape is the right tool.

I pushed two commits to the fork (maintainerCanModify is on, hope that's OK):

1. ef0d8bc — merge of main. The view-Serialize (view: impl Serialize for generated view types (#83) #106), empty-ancillary-file cleanup (fix(buffa-codegen): omit empty ancillary content files #107), and feature-gated codegen + buffa-descriptor regen (feat(buffa-codegen): gate generated json/views/text impls on crate features #117/feat(buffa-descriptor): regen with views/json/text/arbitrary behind crate features #118) all touch the generated code and the codegen modules, so the rebase was non-trivial.

2. b0fa38a — API reshape, since the trait surface gets semver-frozen the moment it ships:

   • MessageFullName → MessageName — matches prost::Name, the ecosystem precedent.
   • Add PACKAGE, NAME, TYPE_URL alongside FULL_NAME — prost::Name ships PACKAGE + NAME + runtime full_name()/type_url(); buffa already computes all four at codegen time, so they ship as &'static str literals (zero runtime cost). PACKAGE and NAME matter independently because FULL_NAME can't be split unambiguously (foo.Bar.Baz could be package foo.Bar + message Baz, or package foo + nested Bar.Baz); codegen knows which. Adding required consts to a published trait is a breaking change, so the full set lands now.
   • Drop the Message supertrait, implement for view types too — the use case you describe (event registries, dispatch by FQN) wants T::FULL_NAME to work on a zero-copy FooView<'a> without to_owned_message(), and a name-keyed registry shouldn't have to prove it can encode just to register a type.

   The trait's TYPE_URL is the same string as the inherent Foo::TYPE_URL const codegen has emitted since 0.4.0 — documented and tested. docs/migration-from-prost.md now has the prost::Name → MessageName row.

If any of this doesn't fit your use case, happy to discuss — the buffa::MessageFullName form is preserved nowhere, so it's a clean rename from your branch's perspective.