Rule Development
At the moment we do this through Bugzilla; a rule that looks good is taken by the dev team and checked into SVN in the developer's sandbox, whereby a distributed group of contributors will test it against their corpus in the nightly and "preflight" mass-check runs. The resulting collated accuracy data can be viewed at [http://buildbot.spamassassin.org/ruleqa/ this site].
This is documented better in RuleSandboxes.
MassesOverview describes a number of tools that are useful in evaluting the quality of prospective rules.
CustomRulesets illustrates rules available from third parties and has links to their mass-check results.