compilers/protobuf: revive the Protocol Buffers compiler from #21 against the current contract

[fuad-daoud]

What exists

A complete Protocol Buffers / gRPC compiler was drafted in #21 (branch feat/protobuf-compiler, head 2345b07, +10,170 lines across 67 files). It lowers proto2, proto3 and the 2023 edition through github.com/bufbuild/protocompile: messages and enums into the ID-keyed registry, field numbers onto Property.WireID, oneof as a WireTagged union behind a Flatten wrapper, repeated/map as hoisted containers, the well-known types, extensions and reserved ranges — with a conformance corpus under testdata/conformance/protobuf and goldens under testdata/golden/protobuf. The PR body is the design record; read it before starting over.

It was opened 2026-07-21, last pushed 2026-07-23, and has not been reviewed. README milestone 5 places the Protobuf compiler after AsyncAPI and GraphQL, so this is the tracking issue for that row, not a request to land the draft now.

Why the draft cannot land as it stands

Measured against main at e1e92a0 (166 commits ahead of the draft's base):

• Contract. compilers.Compiler now requires Formats, Detect and Compile (compilers/compilers.go, added by feat(engine)!: let compilers own format detection and options #341). The draft implements Formats and Compile only, so it does not build.
• IR version. Every golden the draft ships writes "irVersion": "0.1.0"; main is at 0.4.0 after three breaking revisions (feat(ir)!: add Payload.Required and Parameter.Provenance #437–feat(ir)!: bump IRVersion to 0.4.0 for the stack's shape changes #442, feat(ir)!: land the IR 0.4.0 stack #464). The lowering targets struct shapes that no longer exist — ErrorCase is a response, exclusive bounds are carried separately, Payload.Required and Parameter.Provenance exist, and so on.
• Package layout. The draft predates the micro-compiler restructure (docs/micro-compiler-design.md) and is one flat package. internal/archtest's rules map, the 20-method cap on compiler types, and the per-package allowlists all postdate it; the arch_test.go conflict is that map having been rewritten underneath it.
• Shared grammar. canonicalWords and the identifier grammar were extracted into compilers/compile by compilers/compile: own the identifier grammar shared by every compiler #162/compilers/compile: own the canonical naming grammar #163/refactor: move the naming and ID grammars into the framework #183 specifically because the draft's copy had diverged (compilers: canonical naming grammar diverges across the three compilers #161). The draft still carries its own.
• go.sum conflicts, and CI was green only against the July main.

What reviving it needs

1. Rebase is not the shape of the work — start a fresh branch from main and port the lowering, using the draft as the reference for the proto→IR mapping decisions.
2. Implement Detect (a .proto source is recognisable from syntax = / edition = at the top; bounded sniff, same discipline as compilers/openapi/detect.go).
3. Lay it out as compilers/protobuf + compilers/protobuf/internal/*, each internal package with its own archtest rules entry; use compilers/compile for naming and IDs rather than a local grammar.
4. Regenerate the goldens at the current IRVersion and re-derive the conformance matrix witnesses (docs/ir-spec-matrix.md row keys) for the rows Protobuf can express.
5. Run the full oracle sweep: go run ./cmd/morphic-harness testdata/conformance/protobuf.

Decision needed

Whether #21 stays open as a draft or is closed with this issue as its pointer. Either way the branch should not be deleted — the design work in it is the asset.

Refs #21.