fix(net): enforce namespace ARP reply ownership

[fslongjin]

Summary

• Authorize Ethernet ARP replies against the ingress network namespace's existing FIB, independently of AnyIP IP admission.
• Preserve default weak-host replies across local interfaces, reject foreign targets and martian sources, and support zero-source DAD without learning the zero address.
• Prepare an ARP route read view in both direct and NAPI polling without changing the selected IP output backend or local-input polling mode.
• Pin the dependency hook from fix(arp): separate reply authorization from AnyIP admission smoltcp#39 at a2c9cc0da88f41baab5ecc603c33af68384d4cea and add eight packet-level dunitest regressions.

Root cause

AnyIP veth interfaces answered ARP requests for the bridge gateway with a container MAC. Packet capture showed competing gateway replies and subsequent published-port response packets forwarded to the wrong container. The bridge forwarded the incorrect destination MAC correctly; resetting caches or disabling AnyIP would not fix address ownership.

The optional dependency hook separates response permission from neighbor learning. DragonOS reuses its pre-acquired namespace route view, with no new per-packet lock, device scan, or ownership cache.

Validation

• Before the fix: a published-port HTTP request timed out after eight seconds with zero bytes; six of seven original ARP regression cases failed on the baseline guest.
• Linux reference: all eight final ARP tests passed.
• DragonOS, built against the fixed git revision: ARP 8/8, Netfilter 107/107, bridge/veth 9/9, forwarding MTU 3/3, and conntrack netlink 9/9 passed.
• Two independent snapshot cold boots with default Docker bridge/nft/OverlayFS and a second restored container: published-port HTTP first requests succeeded in 94 ms and 5.4 ms, followed by twelve successful requests each. Capture showed only the actual bridge gateway MAC answering gateway ARP.
• Container removal/recreation and IP reuse: first HTTP response succeeded in 6 ms. Daemon normal shutdown completed.
• make fmt, make kernel, diff checks, and three-role adversarial review passed. The reviewed martian-source boundary was fixed and covered by a Linux/guest test.
• Dependency validation: 759 unit tests, seven doctests, noalloc IPv4 and alloc IPv6-only checks, and smoltcp PR CI passed.

Scope and remaining boundary

This fixes namespace ARP reply ownership, not every optional proxy-ARP/NUD policy. An independent pre-existing HTTP body repetition issue remains: raw TCP reads exceed Content-Length, while curl stops at the declared length. The suspected sendfile offset-update defect needs a separate syscall-level fix; HTTP 200 here validates timeout recovery, not full raw-stream correctness.

Dependency merge update

smoltcp #39 has merged into dragonos/v0.12.0. The pin now uses its merged commit. Its Git tree is identical to the previously validated dependency commit; no protocol behavior or other dependency version changes in this update.

[fslongjin]

@codex review

[fslongjin]

@codex review