A week of infrastructure triage

Not every week in a homelab is about shipping something new. This one was
about three different services quietly getting more stable, none of it
glamorous, all of it necessary.

It started with a GPU passthrough issue that turned out to be tangled up
with a disk I/O bottleneck — fixing the passthrough meant first rightsizing
the host so it wasn't fighting itself for resources. That kind of
compounding problem always takes longer to diagnose than to fix.

From there: a local-first LLM router went up in front of Ollama, so
requests can be routed sensibly instead of hardcoded to one model or
endpoint. Then a mail-security scanning platform got deployed, which came
with its own set of ingestion quirks to work through — the kind of thing
where the install is the easy 20% and getting the pipeline actually parsing
input correctly is the other 80%.

Last was a memory-pressure fix on a VM that needed more headroom than it
had — zram swap, documented so future-me doesn't have to rediscover why it
was needed. Nothing here is a feature. It's just the week where the
plumbing gets better, and three separate things are quieter for it
afterward.