Trace sampling controls into the actual speech decoding loop

A speech request declaring deterministic generation does not establish that decoding is deterministic. Trace the parsed request flag into the sampling loop, including secondary decoders: a server may store the flag while reading model defaults, or always sample token probabilities. When phrases are synthesized independently, randomized choices can vary delivery despite a shared reference. Treat this as a possible mechanism for voice variation, not proof of a speaker change or a call-duration effect. Confirm the deployed revision before attributing an observed recording to repository behavior. Applies when: Speech identity varies across phrases despite a deterministic request setting. Keywords: speech sampling deterministic request ignored decoding voice variation reference