Add gated Windows SA3 CUDA foundation

[brxs]

Summary

• add a gated Windows PyTorch/CUDA candidate for Stable Audio 3 Small Music and SFX while keeping TFLite as the release-ready default and fallback
• consume the pinned upstream magenta-community/magenta-realtime-2 runtime as a dependency; no fork or vendored runtime code
• use one candidate Windows GPU environment for MRT2 and SA3, with a cross-process priority/VRAM broker so realtime MRT2 can preempt disposable background SA3 jobs
• isolate SA3 generation in an authenticated one-shot worker with bounded requests, cancellation, crash/OOM containment, provenance checks, and canonical 44.1 kHz stereo PCM16 output
• fail closed at install/runtime boundaries until the gated model inventory and physical Windows NVIDIA qualification are complete
• serialize shared MRT2 worker model switches so the old CUDA generation is reaped before replacement allocation, avoiding transient minimum-VRAM OOMs

Stack note

This draft targets codex/107-integration. The base already contains #110, but not the portable SA3 TFLite work, so this diff intentionally includes #127 (#111) until integration absorbs it. The managed service-manifest launcher is landing separately; this branch's worker contract is ready for its sa3-pytorch-cuda service key but no production CUDA manifest is enabled yet.

Release gates still open

• Unified runtime requalification: the resolvable shared lock uses PyTorch/torchaudio 2.7.1+cu126, Hugging Face Hub 1.7.1, Transformers 5.8.0, and NumPy 2.3.5. MRT2 must be requalified on this exact stack before it can replace the Add a production PyTorch MRT2 backend for Linux and Windows #110 baseline.
• Gated artifact inventory: Stability AI Community/Gemma-gated access is required to record every Small config/nested asset hash. The pin deliberately says releaseReady: false and the installer rejects it. Add third-party model licensing, attribution, and download acknowledgement #108 owns the authenticated compliance/provenance workflow.
• Physical hardware: the provisional CUDA 12.6 driver floor is Windows NVIDIA driver 560.76, but driver/runtime compatibility, measured VRAM reservations, parity, performance, cancellation, OOM recovery, process cleanup, and the 10-minute dual-deck zero-underrun run remain unqualified on real Windows hardware.
• Medium model: Medium remains on TFLite because it needs FlashAttention 2 and there is no qualified official Windows wheel; unofficial wheels are not accepted.
• Selection behavior: Auto remains on TFLite. Explicit GPU requests fail clearly while gated and never silently fall back. A developer-only opt-in exists solely for qualification.

Validation

• focused CUDA policy, worker, broker, MRT2 integration, pin, and SA3 contract suite: 73 passed
• Rust shared stop/reap-before-spawn and launch/model-load recovery regression: passed
• existing nonshared per-deck launch-first restart regression: passed
• Rust clippy for app library/tests with warnings denied: passed
• Python Ruff lint and format checks: passed
• pin audit with --allow-incomplete: passed, correctly reporting the candidate as release-blocked
• broader branch validation: frontend 659 passed and production build passed; Rust workspace 277 passed, 2 ignored, with two sandbox-denied loopback/process tests passing when run with their required local permissions
• Windows cross-check reaches native dependency compilation, then is blocked on this macOS host by the absent Windows SDK (assert.h), so native Windows CI/hardware remains required

Progresses #114.

[brxs]

Superseded by and fully consolidated into #131. The combined branch includes this lane's substantive changes, subsequent integration fixes, and exact-head cross-platform validation. Please continue review in #131.