The Lm holds a persistent rotating transformer_cache shared by every
LmGen. LmGen only resets its own gen_sequence/step_idx, so across
requests the cache kept accumulating stream positions until it
saturated and the model emitted only padding tokens, returning empty
transcriptions ('') for every request. Reset transformer_cache (and
depformer_cache) per request, mirroring Lm.warmup(). Validated: 40
consecutive requests of a known-good sample, 0 empty.
clement
fb71e9ceeb