LAJKTO NEURAL AUDIO RUNTIME (LNAR)

Backend: local ACE-Step C++ resident worker
Default runtime root on this development PC: L:\ACE-Step-CPP
Optional override: LAJKTO_ACE_CPP_ROOT

Generation profile:
- direct DiT path, no 5Hz song-planning LM
- fixed 4 inference steps for BOTH SFX and ambience
- compute selection: Auto / AI chip-NPU / GPU / CPU
- Auto priority: usable NPU backend -> Vulkan GPU -> CPU
- Windows NPU detection uses DXCore machine-learning hardware enumeration
- SFX: minimum 32 latent frames (~1.28 s internal context), then trim to requested duration
- Ambience: minimum 96 latent frames (~3.84 s context), 4-12 s generated blocks
- generated assets are cached under runtime/lnar/cache
- authored game audio remains the automatic fallback
- generator/machinery hum is ALWAYS the authored facility_hum.wav in both modes
- E interaction + machine activation/cancel sounds are ALWAYS the authored interaction WAVs in both modes
- AI may replace ambience, footsteps, jump/landing, mission-complete and future variable SFX

No VPS or RunPod is used by Protocol X LNAR. The resident ACE worker binds only to 127.0.0.1 and is owned by the game process.
