feat: expand /v1/rerank, /v1/images/generations, /v1/responses beyond OpenAI-only

Summary

Today the gateway restricts /v1/rerank, /v1/images/generations, and /v1/responses to OpenAI providers only (per #168 + #211, parallel to existing /v1/responses §4.6 restriction). This is currently defensible as a fail-fast against operator misconfiguration — a non-OpenAI Model on these routes would otherwise dispatch to an upstream that 404s.

But the restriction is more conservative than the de-facto reference implementation in this category, which treats these endpoints as multi-provider unified surfaces. Concrete real-world offerings the gateway currently rules out:

/v1/rerank — 14+ providers in the reference

• Cohere — the canonical rerank API; most common non-OpenAI rerank request shape
• Jina AI — popular for embedding + rerank pipelines
• Voyage AI — rerank-2 model is a frequent ask
• Azure AI, Together AI, Bedrock, NVIDIA NIM, Hosted vLLM, DeepInfra, Fireworks AI, Watsonx, Infinity, HuggingFace

/v1/images/generations — multi-provider

• Stability AI — SDXL, Stable Diffusion 3
• Black Forest Labs — FLUX.1
• AWS Bedrock — Titan Image, Amazon Nova Canvas, SDXL via Bedrock
• Azure (drop-in), Vertex AI Imagen, Gemini, Recraft, Fal AI, DashScope

/v1/responses — multi-provider

• Azure — drop-in compatible (lowest-friction next step)
• xAI — Grok with Responses API shape
• GitHub Copilot, VolcEngine, Manus, Perplexity, Databricks (GPT-gated), OpenRouter, Hosted vLLM

Why this matters

Each provider has its own native wire format for these endpoints. The reference handles this by shipping per-provider transform modules that convert the OpenAI-shaped request into the provider's native shape, and the response back. The gateway today has the bridge pattern (aisix-provider-{openai,anthropic,gemini,deepseek}) that already does this for chat completions; extending the same shape to these endpoints is a natural fit.

Suggested phases (incremental, ordered by user pain × implementation cost)

Phase 1 — /v1/rerank Cohere

• Most common non-OpenAI rerank request; many RAG pipelines depend on Cohere reranker.
• Cohere's wire shape is well-documented and stable.
• Add transform in aisix-provider-cohere (new crate or fold into an existing provider crate).
• Relax /v1/rerank restriction docs to mention OpenAI + Cohere.

Phase 2 — /v1/images/generations Stability + BFL

• Both expose well-defined APIs distinct from OpenAI's /v1/images/generations.
• Stability's API shape requires a transform; BFL similar.

Phase 3 — /v1/responses Azure

• Drop-in compatible (Azure deployments are bit-for-bit compatible with OpenAI's wire shape modulo authentication).
• Lowest-cost expansion: just allowlist Azure provider on the route.
• Then xAI (Grok), which is API-similar but auth-different.

Each phase ships:

• Provider-specific request/response transforms
• Rust unit tests covering the transform shape
• E2E tests against the live mock for the provider
• Docs §4.x update relaxing the OpenAI-only restriction to the supported provider set

Out of scope

• Streaming for these endpoints — separate concern; some providers (e.g. Bedrock images) don't support it.
• Tool calling on /responses — already supported on OpenAI; cross-provider parity is part of #170's tool-translation work.

Companion

• #168 / #211 — established the current restrictive stance
• #212 — e2e coverage for the current restriction (will need rewriting per provider as the expansion lands)

Surfaced by

CLAUDE.md §6 research-discipline triage on the follow-up issues queue. Compared the gateway's endpoint coverage against the documented-and-tested provider matrix of the canonical multi-provider routing implementation in this space.

Start with CLAUDE.md §6, the current provider restrictions, and the existing aisix-provider bridge pattern. Scope the phases in order, beginning with Cohere, then review #212 for the restriction coverage that will need rewriting. Done means provider-specific transforms, Rust unit tests, provider mock E2E tests, and the relevant docs §4.x updates for each supported phase.