Feature: configurable string field type (e.g. `smol_str::SmolStr`), mirroring use_bytes_type

[uqio]

buffa-build 0.6 has use_bytes_type() / use_bytes_type_in(&[paths]) to map proto bytes → bytes::Bytes instead of Vec<u8>. There is no symmetric control for proto string: generated code always uses String.

Request: a parallel knob — e.g. use_string_type() / use_string_type_in(&[paths]) (or string_type(StringRepr)) — mapping proto string to a configurable immutable string type such as smol_str::SmolStr.

Motivation: We migrated a large hand-written proto3 surface (~138 messages) to buffa. The in-memory model is read-mostly and never mutates strings in place, so String's growable buffer is unnecessary; SmolStr gives inline storage for short strings (identifiers, codes, language tags) and O(1) clone for long ones, with no wire-format change. use_bytes_type already covers our bytes fields; string is the only remaining gap.

Implementation note: the codegen already parameterizes the field type for bytes via bytes_fields; a symmetric string_fields + chosen type appears to mirror that path. The selected type would need the same trait surface buffa relies on (Clone/PartialEq/Debug/Default, plus the serde and arbitrary impls already feature-gated like today's json/arbitrary). SmolStr satisfies these. No wire-encoding impact (UTF-8 bytes unchanged).

Happy to send a PR if the direction is acceptable.

[iainmcgin]

Thanks for the detailed write-up and the offer — the direction makes sense and we'd like to land it. The motivation (read-mostly surface, no in-place mutation, cheap clone of long strings) is exactly what an Arc-backed inline string buys you, so this is a good fit.

We're going to pick up the implementation directly, because it touches a few codegen seams that need to be generalized carefully rather than copied from the bytes path:

• merge_string currently reuses a &mut String's buffer in place, so the decode path needs a new generic decode_string_to::<S>() constructor instead.
• json_helpers::proto_string::deserialize is hardcoded to String today; it has to become generic over S: From<String> (the bytes module was already made generic over From<Vec<u8>> for the same reason).
• clear() becomes S::default(), since the immutable candidates have no clear() (same move the Bytes path made).

Rather than SmolStr-only, we'll make it a pluggable knob — string_type(StringRepr) over a small curated set: SmolStr (default), EcoString, and CompactString. They all satisfy the trait surface buffa needs (Clone + PartialEq + Default + Debug + AsRef<str> + From<&str> + From<String>), plus serde; SmolStr and CompactString ship native Arbitrary, and we'll add a small shim for EcoString. buffa will take these as optional, feature-gated deps and re-export them (mirroring how bytes and serde_json work), so generated code keeps referencing ::buffa::* and downstream crates don't have to declare the deps.

One note for anyone watching: SmolStr gives O(1) clone of long strings via Arc<str>; CompactString, while it's the most-downloaded option, copies on clone and is mutable — a different tradeoff, which is why we're exposing both rather than picking one.

We'll tag you on the PR for review since you have the real-world workload. Tracking this now.