[kixelated]

Stacked on kixelated/rust2#71 (the moq pin bump), for quest/m1/ffi-export.md's "Several crates" bullet.

Problem

"Foreign type" is the census's largest refusal on moq-net. moq-net re-exports moq-pattern's Pattern, Patterns, InvalidPattern, Segment, and Specificity, and moq-clock's Instant, and names them in its API: origin::Producer::scope and allowed, auth::Grant's fields, stats::Config::exclude, server::Handshake::with_deadline, frame::Frame::arrived. The exporter refused every re-export of another crate's item, and so every signature naming one. moq-ffi uses scope and Grant today.

Approach

• A re-exported item of another crate is walked like a local one, at its re-export path. rustc's queries (ADT shape, inherent impls, visibility, docs) work across crates, and the glue already reaches moq-net from outside, so the glue needed no change.
• A signature naming another crate's type maps when the API re-exports it, and stays a foreign type refusal otherwise.
• A re-exported module of another crate (pub use kio;, pub use bytes;) stays refused, now as re-export of another crate's module: its API is that crate's own export.
• Tests: a dep fixture crate whose record and object the api fixture re-exports and names. The C transcript drives them and must match the native one. The reject fixture refuses pub use dep;. moq/moq.c now scopes an origin with moq-pattern's patterns, so root and absolute show a real root.

moq-net census at the pinned revision:

	Before	After
Types exported	136	142
Methods and fields exported	532	575
Constants exported	5	7
foreign type refusals	24	18
exclude entries	97	88

The new excludes are the moq-pattern and moq-clock methods that do not map (Pattern::new's generic, Patterns::iter, Instant::from_std, ...).

Decisions needed

1. Whose names re-exported types take in C. moq-pattern's types are now moq_pattern/moq_patterns, and moq-clock's Instant is moq_time_instant, all through moq-net's export.
   • A: export at the re-export path (this PR, recommended). Unblocks moq-net's scoping and auth API now. It also holds later: once several crates export into one library, a re-export of an item that another export owns becomes a name for that type, and items no export owns keep exporting here.
   • B: each crate is its own export, as the quest bullet sketched: moq-pattern and moq-clock get their own ffi.toml and census, and moq-net's re-exports name their types. This needs the several-crates machinery first (IR shared between rustc invocations, one glue crate), which is an L on its own. Until then, these refusals stay.
   • C: A only for types a signature needs, leaving unused re-exports refused. This saves some header lines, but a re-export is public API, so it adds an inclusion rule for no consumer.

Impact

• rust2-ffi: another crate's item that the API re-exports is now exported. The re-export of another crate's item refusal is now re-export of another crate's module, for modules only. Documented in ffi/README.md.

Checks

• just check, just test: pass.
• just ffi moq and just ffi census at the pinned moq: pass.

Alternatives

• Treat a re-export like a type alias, exporting nothing. Every signature naming the type would still be refused.

Follow-ups

• A re-export of a type the rules already map, such as pub use bytes::Bytes, would also export it as an object. No moq crate does this; if one does, its methods fail loud in the census.
• The quest bullet now says what remains: exporting moq-tokio and the media crates into one library.

(written by Opus 5.5)

🤖 Generated with Claude Code

https://claude.ai/code/session_0115AT2ue6x9cT5e5J9SZtCn

[kixelated]

Closing: the maintainer chose option B. moq-pattern and moq-clock become their own exports, and moq-net's signatures name their types across crates, instead of exporting them at moq-net's re-export path. That work continues under quest/m1/ffi-export.md's "Several crates" bullet.

(written by Opus 5.5)