fix: route muse-spark to /v1/responses (Zen requires Responses API)

[ConTresillo]

Fixes Muse Spark always showing as flaky / Test ✕ in the dashboard.

Root cause: Upstream Zen moved muse-spark-1.2/1.3-contributor-free to the Responses API. POST /zen/v1/chat/completions now returns 500 {"type":"error","message":"Internal server error"} for those IDs (repro 100%, issues #44659 #44847 #45744), while POST /zen/v1/responses with {"input":"hi"}\ returns 200. Docs at https://opencode.ai/docs/zen now list Spark under /v1/responses. Zen-proxy previously always hit /chat/completions, so Spark was bucketed as flaky and dashboard Test never succeeded even though the model is healthy.

Validated that the local \opencode serve\ path (\C:\DevLearning\opencode-as-cloudllm\ gateway via \POST /session\ -> /session/:id/message) still works for Spark — this is why the same model works in that project but not via zen-proxy's remote Zen path. Direct upstream probing confirms: \muse-spark-*\ -> 500 on chat, 200 on responses;
emotron-3.5\ intermittent timeout; \deepseek-v4-flash-free\ -> 400 model unavailable.

Fix:

• Route /muse-spark/i\ models to \POST {upstream}/responses\ with \chatMessagesToInput()\ translation and back-translation via
  esponsesOutputToText()\ -> OpenAI \chat.completion\ shape (so existing /v1/chat/completions\ clients keep working).
• Ignore tiny \max_tokens\ for Responses models. Spark uses ~900 reasoning tokens for a trivial prompt; forwarding \max_output_tokens:5\ truncates to empty \output_text. Only forward when >=256.
• Apply same routing to auto-sync probing and /api/test, so sync now marks Spark as \working\ instead of \laky.

Verified:

• After fix, \GET /api/status\ working list includes both Spark variants (before: 3 working / 5 flaky).
• \POST /v1/chat/completions\ with \model=muse-spark-1.2-contributor-free\ -> 200 \Hello there, it's lovely to meet you!\ via zen-proxy (previously 500).
• \model=muse-spark-1.3-contributor-free\ -> 200 as well.