The "engineers find it annoying" reason gets dismissed too quickly — morale and velocity compound over time, and a codebase that constantly frustrates your best people has real business costs that don't show up cleanly in any metric. That said, the bar for a full rewrite should still be high.

The real test I'd add: can you still experiment quickly on it? If the system is stable but every A/B test takes three weeks to instrument, that's not "leave it alone" territory — that's slow death.

The "leave it alone" framing assumes stability is neutral, but systems that don't get touched also don't get understood. When something finally does need to change, nobody knows how it works anymore — and that's when rewrites become genuinely expensive.