Direct `models/infer` image requests were defaulting to `image.text2image`, which dropped reference inputs and produced unrelated generations. Route image-edit payloads to `image.img2img`, accept canonical `image` URLs during reference resolution, and preserve reference-aware prompt hints.
Constraint: The test server must continue accepting both canonical `image` and legacy alias fields from callers already in circulation
Rejected: Fix only reference URL parsing | direct overrides would still be misrouted as text2image before reference resolution runs
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Any future image-model route inference must inspect input shape before defaulting to text2image
Tested: `env GOCACHE=/tmp/popiartserver-gocache go test ./internal/server -run 'TestInferRouteKeyForModelRecognizesViduAsVideo|TestInferRouteKeyForModelRecognizesImageEditInputs|TestResolveImageToImageReferenceAcceptsCanonicalImageURL|TestGenerateEditedImageRefsUsesSeedreamImagesGenerationsEndpoint'`; deployed to `101.42.99.35` and verified Seedream gateway requests include `image`
Not-tested: Full manual validation against every non-Seedream image-edit provider
wtgoku
a73b85c45f