ABC-012: socket layer: NETInfo returns garbage, dual-stack bind broken

Now: OPEN
Sev: HIGH

NET/TCP/UDP/CURL-integration batch (2026-07-07 review). NETInfo reports OK without ever copying the resolved host into the output buffer; the aNETAddress macro memcpys past a stack pad on long hosts; the bind walk aborts on the first EADDRNOTAVAIL instead of trying the next address. CURL's event-loop glue can stall transfers. Most of this is unreachable by tests today because WITH_INET is never on in CI (ABC-017).

Input

Context

NET.h:40-54 NETInfo — getnameinfo writes stack locals host/ service; Bu8rewind(text, range) only moves boundary pointers, no memcpy into text ever happens → OK with stale/zero payload. Uncalled in-tree (hence unnoticed); exported API.
NET.h:29-38 aNETAddress — Bu8rewind return ignored; the two memcpys run unconditionally → stack OOB write for host ≥ 254 bytes, misplaced port on smaller failures. Hand-assembled pad vs typed feeds.
NET.c:12 — a_pad(u8, host, 64) vs NETmaxhost 222 / DNS max 253: legal hostnames > 63 chars make Resolve/Connect/Listen fail SNOROOM.
TCP.c:50-56 + UDP.c:19-25 — first bind() failure returns TCPFAIL instead of trying the next addrinfo (dual-stack v6-first breaks v4); dead if (rc == 0) break; betrays the mangled loop.
TCP.c:9 — TCPConnect(..., b8 nonblocking) param silently ignored (INDEX.md advertises it); POL callers get a blocking connect.
TCP.h:12-18 TCPAccept — accept's addrlen discarded, no Bu8rewind(addr, ...) (cf. UDP.h:25) → NETraw(addr) stays empty; no EINTR retry; EAGAIN == hard TCPFAIL. UDP.h:20-36: same EINTR/EAGAIN collapse + no MSG_TRUNC check (silent datagram truncation).
TCP.c:64 TCPListen — no SO_REUSEADDR: EADDRINUSE for the whole TIME_WAIT window on restart.
CURL.c:26-38 — write/header callbacks ignore u8bReserve/u8bFeed failures and return bytes → on OOM, truncated body delivered with HTTP 200. Return 0/CURL_WRITEFUNC_ERROR instead.
CURL.c:44-59 curl_pol_cb — returns the PRE-action events mask, clobbering interest changes curl made during socket_action → stalls on IN→OUT transitions. CURL.c:77 POLTrackEvents return dropped.
CURL.c:8-9,83 — process-global curl_multi/curl_running/timer vs POL's thread_local queue: cross-thread use silently corrupts.
POL.c:34 — pollerbAllocate return unchecked in POLInit; POL.c:301 int pollms overflows for multi-week deadlines → negative timeout → poll blocks forever.

Goals

Repro where feasible (loopback tests behind WITH_INET; fault-inject the CURL callbacks). Fix per item; EINTR/EAGAIN policy consistent with FILE.c's retry convention.

Constraints

Requires ABC-017's WITH_INET CI lane to keep any of this tested.
Lifecycle items in CURL.c belong to MEM-010 — don't collide; this ticket touches the callbacks/event glue only.

WIP

Design decisions

EAGAIN is a NEW soft code NETAGAIN (ok64-generated) on TCPAccept/UDPDrain; EINTR retries in place; UDP MSG_TRUNC → NETNOSPACE.
New POLEvents(fd, *events) getter (POL.h/.c, INDEX.md) so curl_pol_cb returns the POST-socket_action mask.
No repro seam for CURL cb OOM / POL multi-week deadline — defensive fixes guarded by the existing CURL/POL suites.

TODOs

NETInfo copies via checked Bu8rewind (NETtestInfo); aNETAddress checked, oversize → empty (NETtestAddressOversize)
host pad 254 (NETtestResolveLongHost); bind walk continues (TCP/UDPtestBindWalk, real ::1→127.0.0.1); SO_REUSEADDR (TCPtestReuseAddr); nonblocking honored (TCPtestNonblock)
TCPAccept addr rewind + NETAGAIN (TCPtestAcceptAgain); UDP EAGAIN + MSG_TRUNC (UDPtestAgain/Trunc)
CURL cb failures → CURL_WRITEFUNC_ERROR; pol post-action mask; thread_local mismatch documented (2-line comment)
POLInit alloc checked; poll timeout clamped 60s

Blockers and bummers

Untestable in CI until WITH_INET exists there (ABC-017); verified locally with -DWITH_INET=ON (cache var, CMake untouched).
WITH_ASAN unusable on the Alpine host (libasan link error, pre-existing) — suite ran without ASAN.

Outcome

LANDED by gritzko as ?ce0c1d9b "ABC-012: assorted fixes to the networking code" — but the post MISSES test/NET.c (was unk, never put): the commit carries the NETtest CMake target with no source, so WITH_INET failed to configure until ABC-017 guarded the block with if(EXISTS). RECOVERY: jab put test/NET.c + post from work/ABC-012 (file still sits there), then drop ABC-017's guard note. Suite was 53/53 x4 with WITH_INET=ON pre-landing.