Release v0.4.3 — fix kling-v3-omni v2v/v2v-edit video param schema

[nosuchperson]

Upstream ReferenceVideoToVideoRequest and VideoToVideoEditRequest schemas (mule-run src/aag/mule-router/docs/api/kling/v3-omni-reference-video-to-video.yaml and v3-omni-video-to-video-edit.yaml) declare video as a single URL string and keep_audio as a boolean. The CLI defined video as an array of {video_url, refer_type, keep_original_sound} objects and exposed first_frame / last_frame parameters that upstream rejects via additionalProperties: false. As a result every invocation of either endpoint failed at mule-router's pydantic layer with error_code 2001 "'video' expected Input should be a valid string".

Align both endpoints with the upstream schema:

• video: array → string (URL)
• new keep_audio: boolean (default false)
• v2v: drop first_frame / last_frame
• v2v: note in description that aspect_ratio must be set explicitly when no first_frame is given. The openapi marks it optional with auto-infer, but generation fails at runtime with 4002 "Aspect ratio must be specified when no first image and not video editing".

Verified end-to-end against api.mulerouter.ai:

• v2v + video=<url> + keep_audio=false + aspect_ratio=16:9 + mode=std + duration=5 → task 56ec9e50… produced result_00.mp4
• v2v-edit + video=<url> + keep_audio=true + mode=std → task 09439f64… produced result_00.mp4
• Old array shape returns HTTP 400 / error_code 2001 immediately, confirming no downstream rescue path

Re-ran the full 34-endpoint regression on mulerouter.ai with the patched CLI binary: pass rate goes from 25/34 (pre-fix) to 33/34 (post-fix), with the remaining failure being an environmental openai/gpt-image-2/generation upstream slowness that also failed on 0.4.2.

Bump @mulerouter/core and mulerouter to 0.4.3 (publish workflow verifies versions match).

Reported by: github-mulerouter-skill harness session 65f136d7.