embedPosts() serves cached vectors first and only calls the runtime for misses, so adding rules never redoes model work (cache keyed by model id + content, separate from the verdict cache). Every freshly embedded post folds into an exponential moving average (the 'feed average' positive-only rules score against). Both persist to storage.local int8-quantized; pure helpers (embCacheKey, emaStep, quantEmb/dequantEmb) live in shared.js so the bench can reuse them.