[robobun]

Behaviour change: none

Problem

• bun_runtime exports nothing, but 271 of its items are bare pub. unreachable_pub reports none, and dead_code skips what they reach.
• rustc counts a pub type as exported when an impl of a foreign trait for a foreign type names it. strum::IntoStaticStr emits one: impl From<T> for &'static str.
• The largest chain starts at the unused body::Tag (src/runtime/webcore/Body.rs:588) and hides 172 items.

Fix

• Add #![deny(unnameable_types)] to src/runtime/lib.rs. It reports each such type: 97 on main, 0 here.
• Narrow every bare pub item to pub(crate). Delete body::Tag and the EnumTag derive on body::Value.
• Delete the dead code this exposes, or gate it with #[cfg] when one platform uses it. The Notes name each item.
• Verified: bun run rust:check-all, 12 of 12 targets. The Notes list the other checks and the tests.

Background

• unreachable_pub reports a pub item no other crate can reach. dead_code skips an item that rustc counts as exported.
• Nominal visibility caps that count: a pub(crate) type is never exported.
• Weighed: fix only the body::Tag chain (55 of 97 types stay exported), or a source-lint test (an extra CI pass).

Downsides

• A bare pub type that an impl exports is now a hard error with no automatic fix.
• 27 new #[cfg] lines gate 9 enum variants and 2 enums to one platform.
• Not covered: pub written inside another crate's macro, and the other workspace crates.

Notes

Order. #42819 (Remove libuv on Windows) deletes Writable::Pipe, SourceHandle::ShellWritable and MultiPartUpload.global_this. This PR must land after it. It stays a draft until then. After the rebase, the #[cfg(windows)] gates for those items go away. Related: #36895 constructs Writable::Pipe on POSIX again, #42266 is a dead-code sweep in the same files.

Deleted (nothing reads them on any target).

• webcore::body::Tag, and Value::tag() with the derive that emits it (0 callers).
• streams::IntoArray.value and its 11 initializers in ByteStream.rs, FileReader.rs, ByteBlobLoader.rs.
• MultiPartUpload.global_this and its 2 initializers in s3/client.rs.
• The payload of shell::subproc::Writable::Fd (now a unit variant, like Readable::Fd).
• shell::subproc::WritableInitError. Writable::init returns Writable.
• The payload of AsyncState::Done (shell/states/Async.rs).
• Error::eol, Error::fmt, Error::InvalidCharacter and their name() arms. Only cpus_impl_linux built them, and From<crate::Error> for OsError dropped the variant. The helpers return OsError::Any directly, and that From impl is deleted with them.

Gated (one platform uses them).

• #[cfg(windows)]: Writable::Pipe with Writable::on_close, the stdin wiring block, the finalize arm and the FileSink import. SourceHandle::ShellWritable with its 4 match arms. Error::FailedToInitPipe, Error::FailedToBindPipe. cron::TaskXmlError.
• #[cfg(target_os = "macos")]: Error::FailedToCreateCoreFoudationSourceLoop, Error::FailedToSpawnFSEventsThread, cron::CalendarError.
• #[cfg(not(windows))]: Error::WatchFailed, CreatePtyError::NotSupported, InitError::NotSupported and their arms.

Narrowed by hand (no rustc lint reports these). 34 pub tokens inside bitflags!, opaque_ffi!, jsc_host_abi!, jsc_promise_handler!, multi_array_columns! and impl_field_parent! invocations, 2 object_pool! aliases, and 4 items in the shell_builtins! macro body. The #[no_mangle] host functions among them keep their symbols: 11 host functions in this crate are already private.

Executed code that changes, and why the result is the same.

• ByteStream::on_data: two let v = self.value(); become self.pending_value.with_mut(|pv| pv.clear_without_deallocation()). value() did that clear as a side effect, and nothing read v after it was stored.
• POSIX Writable::init: Writable::Fd(result.unwrap()) becomes Writable::Fd. The unwrap() cannot fail: Stdio::Fd(fd) maps to PosixStdio::Pipe(fd), and spawn_process always sets the spawned stdin for it.
• Windows Writable::init: when start_with_current_pipe() fails, the same panic!("unexpected error while creating stdin") now runs in init. Before, init returned the one-variant error and its only caller panicked with that message.

Not changed.

• The EnumTag derive still emits impl From<&Data> for Tag. It has 0 callers in the workspace. Once this crate is capped it hides nothing, so it is a separate cleanup.
• The other 12 #[enum_tag(existing = ..)] sites. Each has real .tag() callers.
• strum::IntoStaticStr enums in the other crates. Many are real exports. The attribute is not in [workspace.lints] for that reason.

Measurements.

• lint matrix: 0 new diagnostics over 12 triples x dev/release, --all-targets, clippy, and the bun_asan + bun_debug + socket_fault_injection and bun_codegen_embed cfg sets (0 triples skipped).
• unnameable_types reports in bun_runtime (linux x64 / windows x64 / darwin arm64): 97 / 98 / 98 on main, 0 / 0 / 0 here.
• bare pub item declarations in src/runtime: 271 lines + 2 object_pool! aliases on main, 0 + 0 here.
• guard probe: a new #[derive(strum::IntoStaticStr)] pub enum gives 1 unnameable_types error and 0 others. As pub(crate) it gives 0 and 1 dead_code.
• per-platform gates: +27 #[cfg] lines under 3 predicates (15 windows, 7 not(windows), 5 target_os = "macos").
• regenerated codegen: 0 files differ.
• mordant: 0 (lint, file) pairs above the baseline.
• release binary (linux x64, llvm-size): file 80,873,032 -> 80,868,936 bytes (-4,096), .text 58,187,957 -> 58,185,141 bytes (-2,816), every other section equal. Debug binary: body::Tag symbols 2 -> 0 (535 B -> 0 B).
• dynamic exports: 0 added / 0 removed (692). bun_runtime rlib global symbols: 12,417 -> 12,292. C-ABI (unmangled) names in the rlib: 2,513 -> 2,513, same list.
• type sizes (-Zprint-type-sizes): IntoArray 16 -> 8 bytes, MultiPartUpload 344 -> 336 bytes, streams::Result 48 -> 48 bytes.
• stream pull functions (release symbol sizes): ByteStream::on_data 3,420 -> 3,272 B, FileReader::on_read_chunk 2,264 -> 2,127 B. CPU time of a fixed stream-read script, 15 interleaved runs: -0.75% median against an A/A spread of +0.71%.
• rustc CPU time for the bun_runtime unit with and without the attribute, 8 paired non-incremental runs: +1.4% median, inside the same-variant spread of 5.0%.
• Not available in this environment: perf, valgrind and bloaty. So there are no instruction counts, and sizes come from llvm-size and llvm-nm.

Tests (debug + ASAN build, linux x64). test/js/web/streams/streams.test.js, test/js/web/fetch/body-stream.test.ts, body.test.ts, fetch.test.ts, test/js/bun/spawn/spawn.test.ts, test/js/bun/shell/bunshell.test.ts, file-io.test.ts, test/js/bun/s3/, test/js/bun/terminal/, test/js/bun/cron/, test/js/node/os/os.test.js, test/js/workerd/html-rewriter.test.js, test/js/bun/http/serve.test.ts, test/internal/source-lints/. Each failure also fails with a build of the merge-base on the same machine (5 s timeouts under load, root user, no IPv6).

[robobun]

Status: draft, waiting for #42819 to land first.

How to reproduce on main:

• Add #![deny(unnameable_types)] after line 5 of src/runtime/lib.rs and run cargo check -p bun_runtime. It reports 97 pub types that rustc counts as exported (linux x64).
• Or delete body::Tag and the two derive lines on body::Value in src/runtime/webcore/Body.rs and run the same check. It reports 172 unreachable_pub and 4 dead_code errors.