docs: STT whisper->parakeet evaluation and resident swap

[ai-hpc]

Replaces the deployed STT engine (whisper.cpp ggml-small) with parakeet.cpp tdt_ctc-110m on the Jetson Orin Nano Super, and documents the head-to-head that motivated it. Addresses the STT-accuracy open item in #2.

Comparison (50 clean LibriSpeech clips, identical on-device conditions)

Metric	whisper ggml-small (original)	parakeet tdt_ctc-110m-q8 (current)
WER (Whisper-normalized)	1.02%	0.82%
RTFx (amortized)	19×	116×
Per-call latency (resident server)	~500 ms	~170 ms
Peak RAM Δ	+726 MB	+410 MB
Model on disk	466 MB	170 MB

parakeet is more accurate, ~6× faster compute, ~3× lower per-call latency, and ~300 MB leaner — and it's the small parakeet model. Three live ESP32-LyraT mic captures transcribed correctly.

What's in the PR

• docs/stt-parakeet-evaluation.md — full report: method, comparison, real-capture check, caveats, build notes, optional streaming.
• deploy/parakeet-stt/ — reproducible artifacts:
  • serve-mode.patch — adds a resident serve subcommand to parakeet.cpp (load model once; no per-call reload).
  • parakeet-stt-server.py — HTTP shim with the same POST /inference → {"text": ...} contract as whisper-server, on :8178.
  • genie-parakeet.service — systemd unit (resident, Restart=always, on boot).
  • README.md — build / install / revert.

Integration & reversibility

genie-core is unchanged (it still posts to :8178; parakeet now answers there). Swap = systemctl disable --now genie-whisper && systemctl enable --now genie-parakeet; revert is the inverse.

Caveats

Clean LibriSpeech subset (both score low) — noisy-24 kHz real-capture WER is not yet quantified at scale (only the single live capture). tdt_ctc-110m is the lean model; tdt-0.6b-v2 is more accurate but heavier (swap via the service's PK_MODEL). q8_0 quant, device-specific.