Multipath: path-health aware scheduling and control knobs

[aswanthk777]

A path can blackhole while remaining schedulable: with sparse traffic
the congestion window never fills and smoothed_rtt stays frozen at its
last-known-good value, so RTT-based schedulers keep picking a dead path
forever (measured: 10s link blackout = 100% loss with live alternatives
available). A path can also be 'zombie': up and delivering, but parking
packets in oversized carrier queues for seconds (measured on a real
cellular link: srtt up to 37s at only 7% loss) — worse than dead for
interactive traffic, and invisible to PTO-based health because ACKs do
eventually arrive.

• recovery: expose consecutive_pto_count() (resets on any ACK — a
  self-reviving health signal)
• path: Path::unhealthy() (>= 2 consecutive PTOs) and unhealthy_with()
  adding an optional smoothed-RTT ceiling; PathMap::max_srtt holds the
  ceiling
• schedulers (minrtt, round-robin): two-pass selection — prefer healthy
  paths, fall back to blackhole suspects only when no healthy path can
  send, so a fully-degraded path set still transmits
• connection: set_multipath_algorithm() switches the scheduler live,
  set_scheduler_max_rtt() sets/clears the RTT ceiling, path_health()
  returns per-path (addrs, srtt, consecutive PTOs, active, unhealthy)
  for external control planes

With the keepalive PING API an application can probe all paths
periodically; a dead path accumulates PTOs and is avoided within ~1s,
and the first ACK after link recovery restores it automatically.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LqZK49rKWmBNsihzCrsbgD