I'll agree with the long running daemon option if you are allowed to run it. Once I wrote a scheduler daemon under similar circumstances, but unfortunately it would die after a day or two for no reason. After much hair pulling, it turns out that the web host was periodically killing off "unauthorized" long running processes. It seems that running your own daemon violated their terms of service.

My solution was to write an externally triggered scheduler as a cgi program, and a separate triggering daemon running on a computer that I controlled. The daemon would send an http request to the scheduler once an hour. On receiving the request the scheduler would check a file for queued jobs, run them and exit.