Refresh Instance Catalog

Re-read this registration's tools from upstream, now.

TOS-4521: this used to call resolver.invalidate() and nothing else, which clears ONE pod's in-process cache. The next read then falls straight back to the durable index, so the endpoint reported success while changing nothing an agent could see - and awen-api runs several pods, so it did not even reliably clear the cache for the caller's own next request.

The only path that actually refetched upstream was the tool-catalog-index sweep, on a 6-hour cron, so a deliberate change to a connector's advertised tools stayed invisible for up to 6 hours. That is not cosmetic: thinqOS ranks MCP tools by embedding similarity over name + description, so a description decides whether a tool is reachable at all (COM-729).

Runs the same refresh the sweep runs, on this registration only. Kept as a sync endpoint so the live upstream fetch happens on FastAPI's threadpool rather than the event loop, matching how the scheduler entrypoint drives the same coroutine.