One thing that helped was resetting prosody at the language boundary instead of trying to blend it. We inserted a very short neutral prosody anchor like a micro pause with flattened pitch before the switch, then reintroduced the target language's rhythm. Before that, listeners described the switch as "rushed" or "accent bleeding through." After the change, comments shifted to "clean switch" and "more natural," and MOS at switch points moved from roughly okay but noticeable to mostly unremarkable, which is exactly what we wanted.