fix: cleanup stale discovery nodes correctly

[LycanW]

Summary

• add a focused regression test for stale-node cleanup behavior
• fix ServiceRegistry::cleanup time-unit mismatch (seconds vs milliseconds)
• keep change minimal: convert now and timeout to milliseconds before retain filter

Reproduction Evidence (before fix)

A standalone Rust repro (dependency-free) that mirrors current cleanup logic shows stale nodes are incorrectly retained:

node_present_after_cleanup=true
thread 'main' panicked at <anon>:17:5:
BUG reproduced: stale node was not removed

Repro scenario:

• store node timestamp in milliseconds (now_ms - 20_000)
• call cleanup with timeout_secs = 10
• buggy code computes now in seconds and compares against millisecond timestamps, so stale nodes never expire

Fix

src/discovery.rs:

• now switched from as_secs() to as_millis()
• timeout_secs converted to timeout_ms
• retain condition now compares milliseconds to milliseconds

Why this is high-confidence

• Heartbeat timestamps are written in milliseconds at send time.
• Cleanup previously compared those values against seconds.
• Unit mismatch deterministically prevents expiry.

Validation

• Repro before fix: fails as shown above.
• Post-fix equivalent logic behaves correctly (stale node removed).

Environment Notes

In this cloud environment, crate-level cargo test is currently blocked by toolchain/build issues unrelated to this logic:

1. stable cargo 1.83 cannot parse time = 0.3.47 (edition2024 metadata)
2. nightly build then fails in zmq-sys C++ header resolution (<string>, <vector> not found)

So validation used a dependency-free Rust repro to satisfy /repro and verify behavior.