feat: 100 consecutive voice cycles with zero stalls and zero silent d…

[andriypolanski]

Summary

Adds a reproducible M1 soak harness under tests/voice/ that drives (or analyzes) 100 end-to-end voice cycles on Jetson hardware, records per-cycle latency and failure signals, and emits a single pass/fail summary plus a JSON ledger for the
closing PR.

Closes #109
Related: M1 roadmap in README

Changes

• Add tests/voice/100-cycle-soak.sh with run (tee genie-core --voice stderr) and analyze (parse log → JSON ledger + exit code).

Real Behavior Proof

• I have built and run the affected code locally (or noted why I could not).
• I have verified the change end-to-end on Jetson hardware OR explained the equivalent verification path I used.

What I ran

cd /path/to/genie-claw
cargo build -p genie-core --release

# Jetson Orin Nano Super 8 GB — full stack via deploy scripts
export GENIEPOD_CONFIG=/opt/geniepod/config/geniepod.toml
./deploy/scripts/start_all.sh   # or equivalent systemd stack

# 100-cycle soak (interactive wake / push-to-talk)
./tests/voice/100-cycle-soak.sh run --cycles 100

# Optional: re-analyze a saved log
./tests/voice/100-cycle-soak.sh analyze tests/voice/out/soak-<timestamp>.log \
  --expect-cycles 100

Test plan

• shellcheck --severity=warning tests/voice/100-cycle-soak.sh
• cargo build -p genie-core --release
• ./tests/voice/100-cycle-soak.sh analyze on a short sample log (unit-style smoke)
• Jetson (required for M1 close): ./tests/voice/100-cycle-soak.sh run --cycles 100 → PASS, ledger attached
• Confirm 0 silent drops: every non-empty transcript has a GeniePod: line in the same cycle
• Confirm 0 stalls: 100 Total cycle lines, each under max-cycle budget

Acceptance criteria

• 100 / 100 cycles complete without stall (timeout-bounded)
• 0 silent drops (every cycle has either a real reply or a logged error)
• Reproducible from a clean Jetson Orin Nano Super 8 GB on genie-ai-runtime v1
• Ledger artifact attached to this PR

Notes for reviewers

• The harness parses stderr markers already emitted by voice_loop.rs (Total cycle, You said, GeniePod). It does not yet inject canned WAV clips; cycles are operator-driven on hardware (wake word or push-to-talk).
• Silent TTS (Piper produced no audio) is tracing-only today; extend markers in a follow-up if ledger needs audio-out confirmation without a microphone loopback.
• VAD is not in the live capture path yet; soak measures fixed-window record → STT → LLM → TTS as shipped.
• CI runs shellcheck on the script but does not execute a 100-cycle hardware soak.

[ai-hpc]

Closing this as not intended for #109.

The plan for issue #109 is to test the voice loop with a real human on hardware, not by shell command.

Thanks @andriypolanski