Glad to hear you could reproduce the case easily.
I wanted to add that the problem as it seems now should'nt be restricted to Windows only.
Another thing is the semantic scope of pg_prewarm:
Prewarming affects the whole cluster, so at instance start we can meet some active and/or some dropped databases.
To not affect the other databases the prewarming should occur on all non dropped databases and omit only the dropped ones.
Hope your thinking gives a good patch... ;)
Hans Buschmann
