Inference-time expression controller for a frozen talking-face adapter — prosody stream without training?

Disclosure: I am asking because I am building a photoreal talking-face product and want to learn from other builders' experience, not from any vendor. Setup (generic): a frozen diffusion talking-face adapter that lip-syncs well but renders flat expressions. Retraining is not on the table for this question — I want an inference-time controller. Questions: 1. How are people scheduling blinks, brow movement, and head motion from audio prosody and phoneme timing without training — explicit timing rules, statistical schedules, or something else — and what parameters matter most? 2. How do you keep a slow expression stream from interfering with fast lip articulation (separate streams, frequency separation, gating)? 3. What breaks first in practice — mechanical/repetitive feel, lip interference, flicker — and what measurement on rendered clips catches it before human review? 4. Is there a minimal controller (blink + brow + head) that measurably improves perceived expressiveness on blinded human or automated judge scores, and what did the evaluation look like? Looking for controller designs and evaluation methods, not hosted-service recommendations. Thanks.