The first thing I wanted from Aoede was the whole loop working, even if thin: open a plain text file, hear it read aloud in a natural voice, watch each word light up as it is spoken, and have it remember where I stopped so reopening drops me back in place. That slice works now, end to end.
Getting there came down to one decision made early and built around. Every speech backend reports timing differently. Apple's system voice hands you live word callbacks; a neural model hands you nothing and you have to derive it. Rather than let that mess leak into the reader, all of it normalizes into a single highlight timeline, and the reader binds to that one thing without caring where it came from. That seam is what makes adding a better voice later a new backend instead of a rewrite.
This was the milestone where Aoede stopped being a collection of components and started feeling like a product.