After 60abeaa flipped the excess-IP selector to "oldest wins,
newest loses" (to protect the original/current connections), the
per-client IP table in `inbound_client_ips.ips` never evicted IPs
that stopped connecting. Their stored timestamp stayed ancient, so
on every subsequent run they counted as the "oldest protected"
slot(s) and whichever IP was actually using the config now was
classified as "new excess" and re-banned via fail2ban.
This is exactly the #4077 scenario: two IPs connect once and get
recorded, the ban lifts after the configured duration, the lone
legitimate IP that reconnects gets banned again, and again, and
again — a permanent 3xipl.log loop with no real abuser anywhere.
Fix: when merging the persisted `old` list with the freshly
observed `new` log lines, drop entries whose last-seen timestamp
is older than `ipStaleAfterSeconds` (30 minutes). A client that's
actually still active refreshes its timestamp any time xray emits
a new `accepted` line for a fresh TCP, so the cutoff is far above
even idle streaming sessions; a client that's genuinely gone falls
out of the table in bounded time and frees its slot.
Extracted the merge into `mergeClientIps` so it can be exercised
by unit tests without spinning up the full DB-backed job.
Tests cover:
- stale old entry is dropped (the #4077 regression)
- fresh old entries are still carried forward (access-log rotation
is still backed by the persisted table)
- newer timestamp wins when the same IP appears in both lists
- a clock-skewed old `new` entry can't resurrect a stale IP
- a zero cutoff never over-evicts
Closes#4077
