Update:
I have decided to reduce my overthinking on this problem. I had been swept away by the documentation and had turned on the Headers in pine and saw a ton of stuff I wanted to segregate into incriminating and irrelevant with precision.
I discarded that ideal in favor of plain view, exporting the message to a file and deleting the outermost layers of forwarding to leave behind only the original From: To: Subject: etc section. Then I fed each of those files to the SpamAssassin learning module via sa-learn --spam --file filename(s).
Now I have to go through and balance all those spams with a similar quantity of Hams (the logical complement to spam: wanted, valid email) so the
Bayesian filters have the
right tools to work with.