[Feat] burn-engine / ONNX interop — embed or deploy inference

[ax1s-x1zz]

Context

docs/ROADMAP.md Track F-4. Burn is pivoting toward an embedded inference + LoRA/QLoRA engine (burn-engine, Burn 0.22 era). Xazz de-risks coupling by introducing a backend trait at the MLOp lowering boundary so Burn stays the first provider but burn-engine / ONNX Runtime are swappable behind the same Typed IR.

Scope

• Backend trait at the MLOp lowering boundary (Burn first provider; burn-engine / ONNX swappable)
• Runner subprocess keeps both modes: embedded engine (in-process) or remote server (mirrors today’s xazz-runner + xazz-server split)

Acceptance

The same .xzz runs inference via embedded burn-engine and via ONNX with identical outputs.

Dependencies

D2 (ONNX export/import) partially; F1–F3 (#70–#72) — guardrails must exist before inference calls are first-class. Timed to burn-engine release.

---

See docs/ROADMAP.md Track F for the full plan and execution order.

[ax1s-x1zz]

Status — 2026-09-07: the guardrails are ready; this issue is the last remaining piece and is blocked on the burn-engine release.

The F-track dependencies are now complete:

• F1 ([Feat] GenAI prompt input gate — static scan of prompt literals #70) prompt input gate ✅
• F2 ([Feat] GenAI LLM output gate — runtime re-scan + audit chaining #71) output gate + audit chaining ✅
• F3 ([Feat] Fine-tuning data sanitization — safe LoRA/QLoRA prep #72) fine-tuning data sanitization ✅
• F5 ([Feat] Model provenance — weights & license metadata guard #74) model provenance registry ✅

This issue (F4) waits on the burn-engine (Burn 0.22) release to wire the actual inference/fine-tuning calls behind the guardrails. When the engine ships, the remaining work is:

1. Backend trait at the MLOp lowering boundary (Burn stays first provider; burn-engine / ONNX swappable)
2. Runner subprocess keeps both modes — embedded engine or remote server
3. The sanitization artifact (F3) feeds the fine-tuning call; the model registry (F5) gates weights; the output gate (F2) audits prompt/response