I've been leaning towards local processing on an mbox folder... that would also let me build up multiple Bayesian filters to try categorizing my other mail - why does almost everyone only auto-categorize spam? Google kinda broke out of that with their "social" and "important" semi-categories, but it'd be pretty easy to build in local clients, yet I've never seen it.
For bayesian junk filtering to work, you train with messages marked either as junk or nonjunk. I suppose you could train based on positive tag, and assume the message has the opposite signal if it didn't get that tag. My historic email is only classified for (non)junk. If I would start more classifications, I would have to ignore the existing messages.
Fyi, I'm working on https://github.com/mjl-/mox, which also includes a bayesian filter. This sounds implementable.
All reasonably identifiable by word choices, and often at extremely different urgency / priority levels.
But I would personally also have an "interesting" category for newsletters and whatnot that are more likely to get a full read (newsblur has a weak version of this for my RSS, it's nice). And an "urgent" category that notifies more visibly than others. "Recruiters" is also periodically useful - a bunch of cold emails filled with buzzwords, should be identifiable a mile away.
But underneath it all it's mostly that it confuses me that there's so little ability to experiment here. Shared always-updating spam filters makes sense, but there's a lot of similar things people could try unique to their needs that are just totally unsupported. I don't want better global categorization (gmail), I want personalization. Email is weirdly non-personal in the vast majority of systems, and I suspect it's part of why people burn out on it and switch to a never-ending churn of apps that are a better fit for one piece or another.