Go: hy3 stream returns empty event-stream (non-stream 200 OK)

Model: hy3 via https://opencode.ai/zen/go/v1/responses (Go models list contains hy3/hy3-preview)
Auth: ZEN_API_KEY (sk-Eh5x...)

Repro (non-stream OK):

curl -X POST https://opencode.ai/zen/go/v1/responses \
  -H "Authorization: Bearer $ZEN_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"hy3","reasoning":{"effort":"low"},"input":[{"role":"user","content":[{"type":"input_text","text":"write a haiku about code"}]}]}'

-> 200 JSON with output_text OK, all efforts minimal/low/medium/high/max/xhigh/ultra/none pass

Repro (stream FAIL - empty):

curl -N -v -X POST https://opencode.ai/zen/go/v1/responses \
  -H "Authorization: Bearer $ZEN_API_KEY" -H "Accept: text/event-stream" -H "Content-Type: application/json" \
  -d '{"model":"hy3","reasoning":{"effort":"low"},"stream":true,"input":[{"role":"user","content":[{"type":"input_text","text":"write a haiku about code"}]}]}'

-> HTTP 200 content-type: text/event-stream but 0 events for >3s, only blank lines. No response.created/output_text.delta/response.completed.

Control (same request, other models OK):

• deepseek-v4-flash stream:true -> has response.created/in_progress/delta/completed (native responses)
• mimo-v2.5 stream:true -> has naked delta (needs proxy _complete_sse_frame, but not empty)
• hy3 non-stream is the only one that works, so Wire is not chat-adapted; stream path seems not wired

Impact: Codex is streaming-only, so hy3 via hy3-go hangs as "thinking forever" then dumps truncated implicit reasoning as garbled text. Non-stream alone cannot be used. Proxy cannot fix empty stream (only missing frames, see vision_proxy.py:_complete_sse_frame). We have removed hy3-go from local models.json (13 models, hy3/hy3-preview kept in _SEARCH_FALSE as guard).

Expected: hy3 stream:true should emit standard Responses SSE like deepseek-v4-flash / gpt-5.6-luna.

Env: opencode.ai Go, 2026-08-21 19:59 CST, hy3 context 262144 / max 128000 / text->text / 295B MoE 21B active (OpenRouter tencent/hy3), curl 8.7.1, h2. hy3-preview same behavior, occasionally Model is unavailable.

Thanks for fixing similar deepseek-v4-pro web_search 400 (#42090) and PR #40210 — same Go gateway.

Start by reproducing the empty SSE response with the provided curl request, then inspect the Go gateway path and the referenced vision_proxy.py and models.json handling. Compare hy3 streaming with the working deepseek-v4-flash and mimo-v2.5 cases; done means hy3 and hy3-preview emit response.created, output_text.delta, and response.completed events rather than blank lines.