feat: Streaming Responses - Progressive output for multi-duck tools

[nesquikm]

🦆 Duck Enhancement Proposal

💡 The Problem

All responses are currently buffered — the client waits until every LLM finishes before seeing anything. For duck_council (all providers), duck_debate (multi-round), and duck_iterate (back-and-forth), this means long waits with zero feedback.

🚀 Proposed Solution

Stream partial results as each duck completes its response, rather than waiting for all ducks to finish.

What to stream:

• compare_ducks / duck_council → emit each provider's response as it arrives
• duck_debate → emit each round as it completes
• duck_iterate → emit each refinement step
• Single-duck tools (ask_duck, chat_with_duck) → stream token-by-token

Technical notes:

• OpenAI SDK already supports streaming (stream: true) — currently hardcoded to false in DuckProvider
• MCP spec supports progress notifications for long-running operations
• The upcoming multi-turn SSE transport SEP will further improve streaming support

📎 Related

• feat: Long-Running Tasks support (MCP 2025 spec) #15 (Long-Running Tasks) — complementary feature
• MCP roadmap: multi-turn SSE transport is an active SEP

[nesquikm]

Partial progress in d9d6802: all multi-provider tools now emit notifications/progress as each duck responds, and duck_iterate/duck_debate use the async Tasks API so clients can poll for results instead of blocking.

Still open:

• Token-by-token streaming (stream: true in OpenAI SDK)
• SSE transport for real-time updates
• Emitting partial results (e.g., each council response as it arrives) rather than just progress counters