fix(udp): treat only n < len(b) as a short write (#1029)

[MaurUppi]

What

Minimal correctness fix for the UDP data-path retry storm: UdpEndpoint.WriteTo now treats only n < len(b) as a short write, instead of n != len(b).

A proxy WriteTo reports the on-wire byte count, which includes the outbound protocol's encapsulation overhead (shadowsocks AES-128-GCM + IPv4 = +39 bytes), so n can legitimately exceed len(b) on a fully successful write. UDP datagrams are atomic; only n < len(b) is a genuine short write. The old check retired healthy endpoints and exhausted the retry limit, surfacing as Touch max retry limit. warnings.

Evidence

• Pre-fix: instrumented WriteTo errors showed 10655/10655 retry-exhaustion drops were exactly n = len(b)+39 over-writes (delivered on the wire), zero genuine under-writes.
• Post-fix: 22.3h production window logged zero Touch max retry limit. events.

Scope

• Correctness fix only — no instrumentation. Branched from upstream dae/main and cherry-picked clean (2 files, gofmt-clean).
• Includes two regression tests: +39 over-write stays alive; genuine partial write still retires + io.ErrShortWrite.
• This is the same change submitted upstream as fix(udp): treat only n < len(b) as a short write (#1029) daeuniverse/dae#1030; merging here lands it in our fork build ahead of upstream.
• The diagnostic toolkit that found this (health-check gauges, UDP retry diagnostics, snapshot/correlate scripts) stays on the archived branch fix/healthcheck-retry-evidence-1026-1029, intentionally excluded here.

Replaces #36 (closed).