Thanks Matt, Ed, Ruben, Nicklas, Patrick, jdow, Chis et.al. for all the replies, you can stop sending them now, unless you get far below one. At least I now know for sure that I'm stumped :-)
For the record: I have no bayes_ignore anywhere, and I don't believe I have missed something in my setups. Look at the following test-sequence, and note that the same testspam sometimes hits BAYES_99 and sometimes hits no bayes. Prompt is timestamp in GMT+1, so as you can see, there are no timeouts. There is no config "bayes_be_stocastic" is there? ;-P 14:48:28)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:32)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:36)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:40)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:43)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES (14:48:47)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:50)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:54)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:48:57)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:49:01)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:49:05)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES (14:49:09)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES X-Spam-Score: tests=ADDRESS_IN_SUBJECT=0.533,BAYES_99=3.5,BIZ_TLD=2.013, 3.5 BAYES_99 BODY: Bayesian spam probability is 99 to 100% (14:49:13)$ spamc -d <host> -p 7830 < spam-1740.txt | grep BAYES (14:49:18)$