<<< A majority of that effort is rewriting tests to make win9x happy, not rewriting behavior to fix win9x applications.
Understood
<<< Few people (if any) want to intentionally break win9x applications, but spending a large amount of developer effort to maintain the tests there isn't really the best investment, when it could instead be spent fixing real bugs.
Regressions are a normal part of large scale development. Regressions are real bugs. Unless you are becoming what you imitate:-)
<<< I could drop the Win9x tests and concentrate on Windows 2000 and higher. Would this be a good course of action?
Drop seems harsh, prioritize seems reasonable. If you are 199% confident nothing you do would affect Win9x (famous developer last words) Or AJ says put the Win9x users in the lifeboat then it's a good course for you And a not so good one for Win9x users.