spec: qwen3-moe-streaming-sse-v1 — follow-up contract to M32d KV cache

[noahgift]

Summary

Registers a new provable contract `qwen3-moe-streaming-sse-v1` for per-token SSE streaming on the qwen3_moe chat-completions path. Natural follow-up to #1832 (M32d KV cache).

Why now

Pre-M32d, streaming SSE was meaningless on qwen3_moe — full-prefill-per-token mode at ~0.5 tok/s meant the client would see no output for ~30 minutes before all 256 tokens arrived at once. Post-M32d at 9.62 tok/s sustained, per-token emits become valuable for chat UX.

Contract gates

• V1_001: chat-completions with `stream=true` emits per-token SSE events (not buffered)
• V1_002: `stream=false` still returns a single JSON response (regression check)
• V1_003: streaming throughput ≥ 2 tok/s median inter-event time

Implementation phases (engineer playbook)

• Phase 1 (~2hr): callback variant of `run_qwen3_moe_generate` accepting `&mut dyn FnMut(u32) -> bool`
• Phase 2 (~4hr): wire into `try_qwen3_moe_backend` — `tokio::task::spawn_blocking` + mpsc channel + axum SSE stream
• Phase 3 (~2hr): cargo integration test (env-gated, `#[ignore]` by default; mirrors `qwen3_moe_serve_dispatch_v1.rs`)

Total ~6-8 hours; operator-actionable once #1832 merges.

NOT in scope

• MoE inference correctness (`qwen3-moe-serve-dispatch-v1` covers it)
• KV cache mechanics (M32d / feat(M32d): KV cache for qwen3_moe inference path — 19× speedup #1832)
• Streaming for dense models (already exists via `OwnedQuantizedModelCachedSync` continuous batching)
• Tool-call streaming deltas (separate contract)

Test plan

• Pure contract YAML; no code touched
• CI: `pv validate` (if wired)

🤖 Generated with Claude Code