Add bounded MRT2 render worker protocol

[brxs]

What changed

• add a dedicated authenticated --render-worker mode for PyTorch/CUDA MRT2 clip rendering
• make integer frames authoritative on requests; the native gateway owns the documented half-up seconds-to-frames conversion
• bind every job to a strictly increasing u64 sequence and exact {jobId, sequence, frames, pcmBytes} response oracle
• define strict request, cancel, begin, chunk, end, and sanitized error frames with exact scalar types
• cap all JSON, control, metadata, PCM, and chunk sizes and reject malformed, duplicate, out-of-order, truncated, or oversized input
• keep one warmed MRT2 engine alive for serial jobs while terminating the disposable process on cancellation, EOF, write deadline, or overlapping/out-of-sequence requests
• poll control before accepting a completed render, before BEGIN, between chunks, during bounded writes, and before/during END so cancellation and EOF cannot produce late output
• consume the launch capability before its exact first-frame write, preventing reconnect or retry with the same child token
• add a native-facing response parser that validates exact audio identity, aligned bounded chunks, exact PCM byte total, and SHA-256
• add focused protocol tests for framing, limits, warm reuse, half-frame rounding, huge/deep JSON, exact scalar types, sequence ordering, one-shot authentication, cancellation/EOF at generation and chunk boundaries, and sanitized failures

Why

The portable managed-runtime work needs a bounded contract between the native host and an out-of-process MRT2 clip renderer. Keeping render traffic separate from real-time deck control prevents unbounded PCM or malformed control sequences from entering the deck-worker protocol, while process termination gives cancellation deterministic CUDA cleanup even though upstream inference is not cooperatively cancellable.

This is a stacked Wave A change based on codex/107-managed-runtime-final and contributes to #107.

User impact

There is no standalone UI change in this PR. It provides the reusable, warm PyTorch MRT2 render worker that the native Windows/Linux integration can supervise and consume without regressing the existing deck-worker behavior.

Validation

• ruff format --check backend — 43 files already formatted
• ruff check backend — passed
• pytest backend/tests/test_render_sidecar.py backend/tests/test_sidecar.py -q — 48 passed
• pytest backend/tests -q — 358 passed

[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Superseded by and fully consolidated into #131.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.