Add a portable Stable Audio 3 TFLite backend

[brxs]

Summary

Add the official Stable Audio 3 TFLite runtime as the portable CPU baseline for Linux and Windows, behind the same application-facing contract as the existing macOS MLX backend.

Dependencies

• Refactor LSDJ's cross-platform application foundation #107 for shared storage, verified installation, child-process, executable, and CI contracts.
• Add third-party model licensing, attribution, and download acknowledgement #108 owns user-facing licensing, notices, provenance display, and download acknowledgement and must be complete before release.

Architecture

• Define a runtime-neutral Stable Audio service contract.
• Keep MLX on supported macOS systems.
• Select the official TFLite implementation on Linux and Windows as the dependable baseline, including on NVIDIA systems.
• Keep backend selection and diagnostics explicit.

Scope

Feature parity

Support every Stable Audio feature currently exposed by LSDJ where the pinned TFLite API permits it:

• text-to-audio music and sound-effects modes;
• audio-to-audio, continuation, and inpainting;
• positive and negative prompts;
• seed, duration, step count, CFG and APG-related controls;
• Small and Medium model selection;
• LoRA selection/application;
• output naming, preview, cancellation, and model/runtime status.

Document any unavoidable backend difference and do not silently ignore a control.

Input/output normalization

• Normalize accepted user audio internally to the sample rate, channel layout, and WAV representation required by the runtime.
• Do not require users to install or invoke ffmpeg.
• Preserve LSDJ's current supported duration range, including up to 380 seconds for the Medium workflow where the upstream model/runtime supports it.
• Validate outputs before exposing them to the library/player.

Runtime and model management

• Package or install a pinned TFLite Python runtime without relying on system Python or shell tools.
• Use Refactor LSDJ's cross-platform application foundation #107's authenticated manifest, checksum, archive-safety, atomic promotion, rollback, and retry behavior.
• Implement explicit state for runtime/model unavailable, download, verification, install, warm-up, ready, update, failure, and fallback.
• Disclose storage before download using measured values for the pinned revisions (currently approximately 2.3 GB for Small and 9.5 GB for Medium; update these when pinned assets are selected).
• Tie technical provenance and cache cleanup to exact model and LoRA revisions.

Real-time protection

• Run generation outside the audio callback and enforce bounded process/thread/CPU/RAM use.
• Measure Stable Audio generation while both MRT2 decks are active.
• Queue, constrain, or pause background generation when necessary to protect live playback.
• Expose progress, cancellation, worker health, memory use, and failure state.

Tests and documentation

• Contract tests shared with the MLX adapter where semantics match.
• Fixture coverage for each mode and control, including audio normalization, LoRA, cancellation, corrupt output, and long-duration validation.
• Install/update/rollback and path tests on Ubuntu and Windows.
• Performance characterization on the platform minimum CPU/RAM configurations while both decks are active.
• User and contributor documentation for backend selection, storage, downloads, cache removal, and diagnostics.

Licensing dependency

All user-facing terms, attribution, gated-access behavior, credential handling, and acknowledgement UX are owned by #108. This issue must provide the revision/provenance metadata that flow consumes.

Acceptance criteria

• Linux and Windows use the official TFLite backend; supported macOS continues to use MLX.
• Every LSDJ Stable Audio mode/control either passes parity tests or has a documented, user-visible backend limitation.
• Input audio is normalized internally with no user-installed media tooling.
• Small and Medium downloads show measured storage requirements before starting and recover safely from interruption or verification failure.
• Stable Audio generation cannot block the real-time audio callback and is tested while both decks are active.
• Supported generation, cancellation, LoRA, and long-duration workflows complete without orphaned workers or corrupt library entries.
• CI covers contracts, paths, installation, rollback, and fixtures on Ubuntu and Windows.
• Add third-party model licensing, attribution, and download acknowledgement #108 is complete before this backend is included in a public Linux or Windows release.

Out of scope

• Replacing the macOS MLX backend.
• TensorRT or custom CUDA acceleration.
• Linux/Windows application installers.
• ARM, AMD, or Intel GPU acceleration.
• Training or maintaining model forks.