Reap datalogger connections that are abandoned without a FIN

[FezVrasta]

grottserver never closes an idle socket. Its select loop sets only SO_REUSEADDR on the listener and nothing on accepted connections, so a session the datalogger abandons without a FIN stays ESTABLISHED for the lifetime of the process and its send_queuereg entry is never reclaimed.

That would be survivable if dataloggers opened connections freely, but they allocate source ports from a small fixed pool — a ShineLan-X was observed cycling through exactly seven (1029–1035). Every orphan permanently consumes one. Once all seven 4-tuples are held open, the datalogger cannot open a connection at all: it stops reconnecting, stops sending, and the feed goes completely silent.

Evidence

Observed on a ShineLan-X in offline/grottserver mode. After ~15 hours of total silence — no data records, not even a connection attempt — the datalogger reconnected within 15 seconds of a grottserver restart, on three separate occasions:

restart	first reconnect
14:39:55Z	14:40:13Z
15:01:36Z	15:01:47Z
16:00:27Z	16:00:42Z

/proc/net/tcp showed the orphaned sockets idle — no queued bytes, no timer, no retransmits:

port   state    tx_queue  rx_queue  timer   when(s)  retrans
1030   ESTAB    0         0         none    0.0      0
1033   ESTAB    0         0         none    0.0      0
1029   ESTAB    0         0         none    0.0      0

Not blocked, not stalled — simply never closed. A restart drops every server-side socket and frees the whole pool at once, which is why a restart appears to fix it and why the silence returns once the pool is spent again.

Fix

Enable SO_KEEPALIVE on accepted connections so the kernel reaps them without operator intervention. With 120 s idle and 3 probes at 30 s, a vanished peer is detected in about 3.5 minutes. The probes are also traffic, so a stateful firewall stops expiring a session that is legitimately quiet between records.

The TCP_* options are Linux-specific and applied through getattr guards, so platforms lacking them still get SO_KEEPALIVE at the system default interval. A failure to set any option is logged and swallowed rather than refusing the connection.

After the change, on the same hardware: Close connection entries appear for the first time, and the pool holds steady at 2 of 7 rather than draining to zero.

port=1031  ESTAB  timer=KEEPALIVE  when=78.9s
port=1030  ESTAB  timer=KEEPALIVE  when=7.2s
held: 2 of 7-port pool

Tests

Five new tests in tests/test_grottserver_keepalive.py cover the option being set, the configured probe timings, the guard for platforms lacking the TCP_* knobs, and an AST check that the accept path actually calls the helper.

No regressions: clean master gives 294 passed, this branch gives 299 passed, with the same pre-existing failures in test_addon_runtime_config.py and test_release_validation.py (the latter requires GitHub API credentials) on both.

Notes

Independent of #12 — the leak is in master regardless of offline mode, so this can merge in either order. The add-on version is deliberately not bumped and the changelog entry is filed under Unreleased, matching the existing release process.