Magenta RT 2 backend family (sidecar-hosted, append-only)

[ryanontheinside]

DRAFT: Magenta RT 2 as a backend family (sidecar-hosted, append-only)

Merge order

Draft, stacked: opened against ryanontheinside/feat/models/backend. Merge #224 -> #220 -> models/backend before this.
Stack: main <- #224 <- #220 <- models/backend <- this (draft). Sibling leaf: sa3 (draft, same parent).
Heads-up for whoever merges the second leaf: families.py will conflict (both leaves register their family in the same dicts); streaming/session.py resolves clean apart from one comment line.

First non-diffusion model family behind the GeneratorBackend seam, validating that the seam absorbs a generator with a completely different shape: Google's Magenta RealTime 2, running in a WSL sidecar process, append-only (no refinement, no positional source), rolling-window song instead of fixed-length.

What's here (two commits)

a9ff406 mrt2 backend.

• acestep/streaming/mrt2/protocol.py: versioned, language-neutral length-prefixed frame protocol (JSON control + raw PCM audio frames) over TCP. Deliberately generic: this is the template for absorbing ANY future generator in any runtime.
• scripts/mrt2_sidecar.py: the model host. Owns the Magenta RT 2 checkpoint inside its own venv/runtime (WSL), streams PCM frames on credit grants. Model variant is picked at sidecar launch (--model=X).
• acestep/streaming/mrt2/backend.py: MRT2Backend behind the seam. Frontier-driven emission with crossfade splicing into a rolling window; credit accounting; prompt and style-blend forwarded to the sidecar's MusicCoCa conditioning (handle_set_prompt, handle_set_prompt_blend). Capability mask is all-off except generation: every ACE-shaped command (swap, timbre, structure, LoRA, depth) fails loudly via command_failed.
• create_mrt2_session registered in families.SESSION_CREATORS: the family bypasses the ACE create path entirely (no TRT profiles, no model load, no demucs, no conditioning encode). Honest nullable metadata (no bpm/key) per the contract surface.
• tests/unit/test_mrt2_backend.py: protocol framing, frontier/window math, credit flow, prompt forwarding, capability mask (13 tests, no sidecar needed).

dc2dc9c magenta pedal. /magenta web page: a minimal pedal-style UI (prompt, blend, temperature) built on the @demon/client SDK and gated by the served capability mask, demonstrating that a non-ACE family gets a working frontend without touching the shipped app.

Transactional create (carried from #222's semantics)

The sidecar TCP link is a real acquired resource, so a failed session create cannot strand it: MRT2Backend.__init__ closes a self-acquired client if construction fails after connect (injected test clients stay caller-owned), StreamingSession.__init__ closes the backend if init fails after make_backend, and the factory registers the audio engine on an ExitStack with pop_all() ownership transfer, mirroring the ACE create path. StreamingSession.close() closes backend-owned resources first.

Status / why draft

Live-verified end-to-end against the WSL sidecar (append-only frontier contract, prompt steering). Remaining before ready-for-review: sidecar deployment story for pods, and a golden-harness scenario for the family. scripts/mrt2_live_check.py is the manual smoke (needs the sidecar running).

Validation (2026-06-06, restacked tip)

160 unit tests passing at this tip (mrt2 suite + all inherited guards), web typecheck clean.