A few things before this can land:

1. Reuse _CohereCompatibleRerankClient instead of re-implementing it

The DashScope compatible endpoint speaks the same wire format as the existing helper at hindsight_api/engine/cross_encoder.py:530 ({model, query, documents, top_n} → {results: [{index, relevance_score}]}). _rerank_compatible duplicates that logic. Compose the helper the way SiliconFlowCrossEncoder (line 768) and ZeroEntropyCrossEncoder (line 726) do — you get the query-grouping and request shape for free.

2. Drop _COMPATIBLE_MODELS

Hardcoding frozenset({"qwen3-rerank"}) means any new variant (qwen3-rerank-plus, future qwen3-reranker-*, etc.) silently routes to the native endpoint and fails with a shape mismatch. Pick one:

• Only support the compatible endpoint and remove the native path entirely (simplest — qwen3-rerank is the headline model here).
• Or expose endpoint selection as an explicit config flag rather than inferring it from the model name.

3. Documentation is missing

• hindsight-docs/docs/developer/configuration.md:556 — add alibaba to the HINDSIGHT_API_RERANKER_PROVIDER value list, plus rows for HINDSIGHT_API_RERANKER_ALIBABA_API_KEY / _MODEL next to the SiliconFlow block (~line 583).
• hindsight-docs/docs/developer/models.mdx:475 — provider table entry and an example env-var block (mirror the SiliconFlow one at line 553).

Also: the class docstring says auth comes from DASHSCOPE_API_KEY or HINDSIGHT_API_RERANKER_ALIBABA_API_KEY, but from_env only reads the latter — either drop the claim or add the fallback (Cohere does it at config.py:1581).