You may already know this, but the voice engines implementations have a lot of room for performance improvement, and you could probably squeeze out even more voices, routing patterns, secondary features, etc if you tighten them up.
They're written clearly but kind of like textbook/reference implementations rather than the more aggressively fine-tuned versions you'd see in proprietary commercial synths.
You can probably coach Claude through improving them, but you generally need to do that explicitly.
Lowest hanging fruit for your design would be to have it organize control flow (and apply effective math tricks) to minimize branching and to organize data references with cpu cache more in mind.
On a more ambitious scale, working over blocks instead of samples, and using vector instructions where practical would go a long way. The latency impact of working on small blocks (i.e ~8/16/32 samples) instead of individual samples is irrelevant but the performance opportunities that open up for many algorithms is huge.