feat(transport): add io_uring frame reader for the monoio data plane

[beinan]

First step of #285. Purely additive — adds transport::uring alongside the existing Tokio frame reader, with no caller changes, so it can land and be verified in isolation.

Why a second reader instead of an abstraction

Tokio's AsyncRead borrows a buffer for the duration of a read. io_uring is completion-based: the kernel owns the buffer while the operation is in flight, so monoio's AsyncReadRent moves it in and hands it back with the result:

// tokio
stream.read_exact(&mut buf).await?;

// monoio
let (res, buf) = stream.read_exact(buf).await;

A borrowed-buffer signature cannot be made sound over a completion ring, so the reader is rewritten rather than made generic over both. The wire format is identical — same FrameHeader, same payload bytes.

The limits are the point

read_frame's DoS protections (#111) are what make this more than a thin wrapper, and they are reimplemented verbatim with tests that fail if they regress:

protection	test
per-type cap checked before allocating	rejects_oversized_control_payload_before_allocating — sends a header claiming MAX_CONTROL_PAYLOAD_LEN + 1 and no payload; if the reader allocated first it would block instead of returning immediately
Ping carries no payload	rejects_ping_with_payload
timeout on a stalled peer	times_out_a_stalled_payload — header then silence; asserts it returns well inside the deadline
clean EOF is Eof, not an error	clean_eof_at_frame_boundary
truncated payload is an error, not a short frame	truncated_payload_is_an_error — would otherwise desync the protocol
stream left at the next frame boundary	reads_consecutive_frames

Plus the happy paths: a normal frame, a zero-length payload (must not allocate or block on a second read), and the write helper returning its buffer for reuse.

Ten tests, all on a real io_uring runtime over loopback TCP. CI can execute them — the capability probe in #281 confirmed PROBE_RESULT=SUPPORTED on ubuntu-latest, which is why this needs no feature gate.

Dependency note

monoio is added with the sync feature, which provides the spawn_blocking the zero-copy sendfile path will need to stay off the ring. Worth knowing for the next PR: unlike tokio's, it requires an explicitly attached thread pool and panics at runtime without one (execute blocking task without thread pool attached).

Verification

• cargo test --workspace --all-features --locked — 28 suites pass, including the 10 new ones
• cargo clippy --workspace --all-targets --all-features -- -D warnings — clean
• cargo fmt --all --check — clean
• cargo doc --workspace --no-deps --all-features — clean

Next in #285

Thread-per-core accept loop behind a config flag, then porting handle_conn/handle_put/handle_delete to use this reader.

Refs #285, #273

🤖 Generated with Claude Code