Greptile r2 on #133283:
- #1: probe_store built one byte string the size of jobs.json; a large store made cron status,
doctor and the idle recovery check allocate that much. It now writes the same total in 64 KiB
chunks.
- #2: the host-wide store gauges read every degraded record, so a profile that left this
gateway while its store was down kept it at writable=0 until a save that would never come.
register_ticked_homes already discards a departed home's worker pool; it now also drops that
home's store state (store_health.forget_homes). The export test pins it.
b403581f36