[284km]

Addressed UDP currently has no native FFI, so mere-ruby cannot implement unconnected UDPSocket#send or obtain a datagram's sender. In its upstream DNS test, the sender raises NotImplementedError and another thread waits until the 600-second file limit.

Add sock_sendto and sock_recvfrom to the shared C/LLVM native runtime. Destinations are numeric IPv4/IPv6 addresses; payloads use the byte arena, flags pass to the kernel, and failures preserve thread-local errno. Receive returns the syscall count together with the numeric sender address, including for empty datagrams. Both helpers validate arena bounds; callers must clamp a MSG_TRUNC count before reading arena bytes. Sending preserves the existing descriptor, binding, and peer, including the host's refusal when Darwin rejects an addressed send on a connected socket. A zero-capacity receive without sender metadata returns EAGAIN; queued packets, including empty ones, still retain their sender and do not write to payload storage.

Validation:

• dune runtest: 2,946 passed, 0 failed.
• New bounded loopback probe passes on C and LLVM on macOS, and on C in an Ubuntu container. It covers IPv4/IPv6, binary/empty packets, multiple destinations, sender identity, MSG_PEEK, truncation, nonblocking errors, connected sockets, closed descriptors, invalid arguments, arena bounds, and zero-capacity receives with and without queued packets. The no-packet regression failed on the previous implementation and passes after the fix.
• Existing sockaddr gate: 116/116 passed.
• CI assignment/failure controls, actionlint, and git diff --check pass. The new gate runs in the runtime group, increasing the explicit inventory to 188 checks.

The interpreter, Wasm, and RISC-V do not gain these capabilities. The runtime CI gate also checks C and LLVM on Linux.

[284km]

Reviewed 76023a6f against base f350170f. Found one Minor correctness issue in zero-capacity receives; see the inline comment.

The shared native-runtime implementation keeps C and LLVM behavior aligned. The subtraction-based arena bounds checks avoid overflowing an offset-plus-length calculation, and the loopback fixture checks source identity and preservation of the socket's binding and peer.

Validation on an isolated snapshot of the PR head, on macOS:

• dune runtest: 2946 passed, 0 failed.
• datagram_check.sh: C and LLVM passed; sockaddr_check.sh: 116/116 passed.
• Additional C/LLVM probes passed at -O0 and with UBSan at -O2: oversized descriptors, offsets, lengths and flags, empty sends, host-verified MSG_TRUNC counts, and thread-local errno isolation.
• The additional non-threaded probes also passed ASan/UBSan. ASan with the threaded probe was inconclusive because the sanitizer runtime failed to deallocate memory.
• CI configuration self-tests retained all 188 gates, including the new runtime gate, and passed the routing, result and setup negative controls.

The zero-capacity failure was reproduced on both C and LLVM. Linux execution was not repeated locally.

[284km]

[Minor] Handle zero-capacity receives before reporting a datagram

cap == 0 passes validation, but on macOS a zero-capacity recvfrom can return n == 0 with sl == 0 even when no packet has been sent. The cleared sockaddr then becomes "other 0 ", and this helper reports "0 other 0 " with fd_last_errno() == 0. Callers therefore see a successful receive with invalid sender metadata rather than a timeout/error.

Reproduced on both C and LLVM with the existing extern declarations:

let fd = socket 2 2 0;
let _ = sock_bind fd "127.0.0.1" 0;
let _ = tcp_set_timeout fd 100;
let p = mem_alloc 1;
print (sock_recvfrom fd p 0 0)

No sender is needed. Changing the capacity to 1 correctly returns "" with EAGAIN after the timeout.

Please either reject zero receive capacity with EINVAL and document the positive-capacity requirement, or handle zero-capacity receives so that success requires an actual packet and valid sender metadata. Add a no-packet regression and retain coverage for genuine empty datagrams received with positive capacity.

[284km]

Fixed in 7b101ac. sock_recvfrom now maps a zero-capacity result with no sender (sl == 0) to an empty result with EAGAIN, instead of constructing 0 other 0 . Zero capacity remains supported when the kernel actually supplies a packet and its sender.

Added IPv4/IPv6 regressions for no queued packet on timeout-configured and nonblocking sockets, plus queued empty and nonempty packets received at zero capacity. The latter checks verify sender identity, unchanged payload storage, and removal from the receive queue. Existing positive-capacity empty-datagram coverage remains in place.

The no-packet regression fails on the old implementation. The expanded gate passes on C and LLVM on macOS and C on Linux; the existing sockaddr gate passes all 116 checks.

[284km]

Good regression coverage: sending to two destinations through the same bound socket and checking its binding and connected peer helps catch implementations that reconnect or replace the descriptor while adding addressed sends.

[284km]

Re-reviewed 7b101ac7, including the changes since 76023a6f. The previous Minor finding is addressed. No new actionable findings.

The missing-sender check prevents a zero-capacity receive from reporting a nonexistent packet. Queued empty and nonempty datagrams still return their actual sender, and the documentation now states the behavior explicitly. Polling before the queued-packet assertions avoids arrival races.

Validation on an isolated snapshot of the updated head, on macOS:

• The original reproducer now returns "" with EAGAIN on both C and LLVM.
• The updated datagram gate passed on C and LLVM for IPv4 and IPv6. Separate runs passed at -O0 and at -O2 with ASan/UBSan.
• Removing the new guard from generated C, and separately from the LLVM companion, made the new no-packet regression fail with the expected assertion. This confirms that the added coverage detects the original bug.
• Existing socket-address checks: 116/116 passed.
• CI configuration self-tests retained all 188 gates and passed their negative controls; git diff --check passed.

GitHub CI still has queued/running jobs. This follow-up used focused runtime validation; the full unit suite and Linux execution were not repeated locally.

[284km]

Verified: this resolves the previous zero-capacity finding on both native backends. Checking for missing sender metadata preserves successful zero-capacity receives of queued datagrams, while the original no-packet reproducer now returns an empty result with EAGAIN. The new regression also fails when this guard is removed.

[284km]

Follow-up review of 7b101ac7: no remaining actionable findings. The previous zero-capacity receive issue remains resolved. The code is unchanged since my last review, so the earlier local C/LLVM, sanitizer and negative-control results still apply.

I have now checked the completed CI results for this exact head:

• All 17 GitHub checks passed, including the four macOS/Ubuntu build configurations, both distribution compatibility jobs, and all nine gate groups.
• The Ubuntu 24.04 runtime job actually executed the new datagram gate: its log reports both datagrams: c passed and datagrams: llvm passed. The existing socket-address check also reports 116/116 passed.
• The compiler gate job reports 2946 passed, 0 failed for the full unit suite.

The shared implementation and regression coverage now have successful native-backend validation on both macOS and Linux. CI verification is complete; no additional code changes are requested.

[284km]

Confirmed in the completed Ubuntu 24.04 runtime-job log: this step executed successfully and reports both C and LLVM passing. Together with the earlier macOS checks, this validates the shared runtime and the zero-capacity regression on both supported host platforms.

[284km]

Rebased onto 01ed96805060232c47a839c11f62e9c464de603e; the new head is 991f299b20f18639713685fe3297e087ddc5b1d7. This retains the v0.1.711 compiler changes now required by mere-ruby main. git range-diff confirms both UDP commits are unchanged by the rebase.

Validation after rebasing: dune runtest reports 2,946 passed and 0 failed; the bounded datagram gate passes on C and LLVM on macOS, and the existing sockaddr gate passes 116/116. The downstream mere-ruby UDP regression also matches Ruby 4.0.9 on macOS and Linux with the hash index ON and OFF. CI is rerunning on the new head.