Currently, the framework is lagging unit testing. Add some unit test.

Some senders use noreply emails addresses and it's easy to distinguish them from something useful based on the reply-to header.

Example: replyto:[email protected]

+* Criteria:
+    query: replyto: "[email protected]"
+  Actions:
+    archive
+    never mark as important
+    apply label: spam

http://linux-quirks.blogspot.com/2011/07/how-to-create-gmail-filter-based-on.

The label distribution in the description for continuous labels should match pandas describe.