feat: add application-level idle keepalive to detect zombie connections

[javi11]

Summary

• Adds KeepaliveInterval and KeepaliveCommand fields to Provider for optional application-level idle keepalive
• When enabled, sends a lightweight NNTP command (DATE, HELP, or CAPABILITIES) periodically on idle connections to detect zombie TCP connections before a real request discovers them
• On probe failure the connection is closed immediately; runConnSlot reconnects automatically on the next request

Motivation

Network paths (NAT tables, firewalls) can silently drop idle TCP connections while the client-side socket still appears open. Without keepalive, zombie connections are only discovered when a real request times out — causing seconds of latency on the next request. This feature detects dead connections proactively during idle periods.

Implementation notes

The probe is injected as a synthetic *Request through the normal write pipeline so readerLoop maintains FIFO ordering with any other in-flight pipelined requests. The inflightSem slot acquired before the blocking select is owned by the keepalive request; readerLoop releases it at the normal point (after delivering the response), exactly as for real requests.

Configuration

Provider{
    KeepaliveInterval: 45 * time.Second,  // 0 = disabled (default)
    KeepaliveCommand:  "DATE",            // optional; defaults to "DATE"
                                          // use "HELP" or "CAPABILITIES"
                                          // for providers that don't support DATE
}

When SkipPing: true and KeepaliveCommand is empty, keepalive is disabled automatically (no safe probe command is known for that provider).

Test plan

• TestKeepalive_KeepsConnectionAlive — server responds correctly to DATE; verifies probe fires and connection remains usable for real requests
• TestKeepalive_DeadConnection — server drops connection on DATE; verifies Run() returns and slot can reconnect
• make check (lint + race detector) — all pass

🤖 Generated with Claude Code