[Japabu]

What

The T14 session that #1138 staged ran Linux, ToyOS at 4f52c30, then Linux, and every run exited 0. It decides nothing: ToyOS's osbench_lan found the peer, then refused the whole group with no netstack link holds 10.0.2.15.

The cause is on main, not in [tcp]:

• std's TcpStream::socket_addr (sdk/std/sys/net/connection.rs) returns a hard-coded 10.0.2.15 with the stream's local port. libc's getsockname does the same.
• Since netstack runs every link a driver brings: ranked, with failover and return (P5) #1122, osbench's sys::link_of picks a connection's link by matching that address against netstack's leases.
• The T14's lease comes from the bench's router, so no link matches. QEMU's guest test passes only because slirp leases exactly that address.

Change (issues only)

• Ramp issue: records the void session, and the per-connection spread from its two Linux runs.
  • At 100 ms the ten connections read 5.13–6.63 MB, a standard deviation of 478 KB. That is 16 times the 30 KB miss.
  • The two runs agree under the exit's exchangeability check: two-sided rank test p = 0.69 at 100 ms and 0.42 at 300 ms.
  • The issue stays open; the next session waits on the address fix.
• getsockname-answers-an-address-nobody-asked-netstack-for: changes from finding to defect, with this evidence. It names std's copy of the literal and the fix: netstack's connect and accept answers carry the local address, and the SDK, std and libc return it.

Checks

Issue text only. No IP shape outside RFC 1918.

Evidence: https://github.com/ToyOSOrg/evidence/tree/35f09268e09714a8547496bdd766bcd3ca998325/r1

Risk

• Until the address fix lands, osbench_lan's whole group refuses on every network that doesn't lease 10.0.2.15, so no LAN row is measurable on metal.

🤖 Generated with Claude Code