[billxc]

Summary

Enable image generation and editing through GitHub Copilot upstreams. Copilot advertises GPT Image 2.5 Flare with Images API endpoints, but Floway previously discarded those capabilities and rejected image calls.

• Derive generation and edit capabilities independently from the upstream model catalog, without hardcoding model IDs.
• Forward image generation and JSON/multipart image editing through the existing authenticated Copilot transport.
• Construct JSON bodies and their media types together; have the image-edit serializer return the encoded body with its headers.
• Keep content-type decisions out of Copilot authentication, and let the transport generate multipart boundaries.
• Update Custom/Azure image-edit callers to consume the serialized request without changing their protocol behavior.
• Preserve upstream responses and existing image usage accounting.

This change is backend-only; the Playground does not yet expose image-generation models.

Validation

• 139 tests passed across nine explicitly selected request-encoding, provider, authentication, model-discovery/merging, and gateway image test files.
• ESLint passed for all 14 changed files.
• TypeScript checks passed for provider, provider-copilot, provider-azure, provider-custom, and gateway.
• Real end-to-end requests were repeated after the cleanup through an isolated local Floway server using an authorized Copilot account:
  • /v1/images/generations: HTTP 200 in approximately 10 seconds.
  • /v1/images/edits with a multipart upload: HTTP 200 in approximately 12 seconds.
  • Both outputs fully decoded as 1024×1024 PNGs; the edit changed blue foreground pixels to red.
  • Both requests and separate text/image token quantities were recorded in the local usage database.
• SSE/partial-image streaming and real JSON-reference editing were not tested.

No local credentials, databases, logs, or generated images are included in this PR.