TxVoiceProcessor runs two egress resamplers to produce identical L/R output

Follow-up from PR #4875 (48 kHz TX voice refactor), finding 6 of @jensenpat's review. Deferred there because collapsing the pair would narrow a public contract; recorded here for that discussion.

The observation

TxVoiceProcessor runs two independent r8brain instances for the 48 → 24 kHz egress conversion, one per channel:

std::unique_ptr<Resampler> m_outputLeftResampler;
std::unique_ptr<Resampler> m_outputRightResampler;

On the production mic path the two inputs are byte-identical by construction: processCapturedInt16() collapses capture to a canonical mono signal and duplicates it to L/R, and nothing in the strip decorrelates the channels. ClientReverb does not — its 23-sample stereo spread is allocation headroom, not active decorrelation, and #4875 corrected the header comment that implied otherwise.

So the most expensive DSP in the block runs twice to produce the same numbers. CPU and memory only — it cannot reduce latency, since the two run in parallel and contribute one group delay.

Measured context from #4875: the full strip costs under 1% of one core, so this is not urgent.

Why it is not a trivial deletion

processFloat48() is a public entry point that accepts arbitrary interleaved stereo, not necessarily duplicated mono. Collapsing to a single resampler would either narrow that contract or require a runtime "are these identical?" check on the audio thread.

Related: reconcileEgressFrameCounts() and its recovery path exist solely to handle the two instances disagreeing on frame count — a condition that cannot arise with one resampler. Removing the pair would let that machinery go too.

Decision needed

Either narrow processFloat48()'s contract to duplicated-mono and use one resampler, or keep the pair and accept the cost as the price of a genuinely stereo-capable seam.

Start with TxVoiceProcessor::processFloat48() and processCapturedInt16(), then inspect m_outputLeftResampler, m_outputRightResampler, and reconcileEgressFrameCounts(). Resolve whether processFloat48() should guarantee duplicated-mono input or retain arbitrary stereo support; done means documenting and implementing the chosen contract, including the recovery-path consequences.