Add a production PyTorch MRT2 backend for Linux and Windows

[brxs]

Summary

Add a production PyTorch backend for Magenta RealTime 2 on Linux and Windows, consuming Apolinario's implementation as a pinned upstream dependency behind a thin LSDJ adapter. Keep MLX as the macOS backend.

Dependencies

• Refactor LSDJ's cross-platform application foundation #107 for shared storage, installation, executable, process, audio, and CI contracts.
• Spike: validate Apolinario's PyTorch MRT2 port for LSDJ #109 must reach the documented go criteria and select the production service topology.
• Add third-party model licensing, attribution, and download acknowledgement #108 owns user-facing licensing, notices, and download acknowledgement and must be complete before release.

Architecture

• Define a runtime-neutral MRT2 service contract used by the Rust host and frontend.
• Select MLX on supported macOS systems and PyTorch on supported Linux/Windows NVIDIA systems.
• Keep platform/backend selection explicit and diagnosable; do not silently choose an unsupported CPU path.
• Consume a versioned upstream package/tag when available, otherwise an immutable VCS revision. Do not copy or fork the runtime into LSDJ.
• Keep the LSDJ-owned layer limited to API adaptation, lifecycle integration, input/output normalization, and diagnostics.
• Use the worker topology selected by Spike: validate Apolinario's PyTorch MRT2 port for LSDJ #109 and supervise it through Refactor LSDJ's cross-platform application foundation #107, including readiness, heartbeat/watchdog, bounded logs, graceful shutdown, crash detection, and full process-tree cleanup.

Scope

Backend parity

• Implement prompt, negative prompt, guidance, temperature, top-k, seed, steering/conditioning, reset, warm-up, and model selection with the semantics validated in Spike: validate Apolinario's PyTorch MRT2 port for LSDJ #109.
• Preserve the existing chunk/context/overlap contract, sample rate, channel layout, and continuous audio behavior.
• Surface any intentional cross-backend semantic difference in UI/help and diagnostics.

Runtime and model management

• Install a pinned Python/PyTorch/CUDA runtime without requiring system Python, Git, shell tools, or a user-installed CUDA toolkit.
• Use the verified, atomic install/update mechanism from Refactor LSDJ's cross-platform application foundation #107.
• Track runtime and model state separately: unavailable, downloading, verifying, installing, ready, updating, failed, and recoverable previous version.
• Record exact upstream runtime, model, driver, and adapter versions in diagnostics.
• Keep technical provenance tied to immutable revisions; user-facing terms and notices come from Add third-party model licensing, attribution, and download acknowledgement #108.

Performance and scheduling

• Support both LSDJ scheduling modes validated by the spike:
  • 25 frames / approximately 1 second;
  • 5 frames / approximately 200 ms when both decks are armed.
• Preserve the approximately 1.5-second playback-ring safety model.
• Prevent background install, logging, or model-management work from blocking the real-time audio callback.
• Expose generation latency, queue depth, worker health, and engine-reported underruns for diagnostics and smoke tests.

Failure behavior

• Fail clearly for unsupported hardware, insufficient VRAM, incompatible drivers, corrupt assets, and worker startup/crash conditions.
• A failure on one model worker must follow the isolation behavior chosen in the spike and must never orphan GPU processes after app exit.
• Preserve the previous verified runtime/model when an update fails.

Tests and documentation

• Adapter contract tests using deterministic/fake backends.
• Cross-backend parity fixtures for controls and audio shape/timing.
• Runtime install/update/rollback tests on Ubuntu and Windows CI.
• Process readiness, crash, timeout, and cleanup tests.
• Hardware smoke checklist reused by the Linux and Windows release issues.
• Contributor documentation for updating the pinned upstream dependency without creating an LSDJ fork.

Acceptance criteria

• Linux and Windows select the PyTorch backend; supported macOS continues to select MLX.
• The production dependency is pinned to an immutable upstream release/revision and no runtime source is vendored or forked into LSDJ.
• Users do not need system Python, Git, shell tooling, or a CUDA toolkit.
• All required LSDJ controls pass the parity fixtures agreed in Spike: validate Apolinario's PyTorch MRT2 port for LSDJ #109.
• On the proposed minimum NVIDIA hardware, both decks run for 10 minutes in both scheduling modes with zero engine-reported underruns.
• Diagnostics report backend, dependency/model revision, driver/runtime version, latency, queue depth, and worker state.
• Failed installs/updates retain the previous working version, and app exit leaves no worker descendants.
• CI covers shared adapter/install/lifecycle behavior on Ubuntu and Windows; real-hardware results are recorded in the platform release checklists.

Out of scope

• Replacing the macOS MLX backend.
• Maintaining an LSDJ fork of the PyTorch port.
• CPU-only MRT2 support.
• AMD or Intel GPU support.
• Platform installers and hardware certification.