τ³-bench: offer the reranker and qwen search methods through the relay #28

Priority

extreme

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Today

τ³-bench's banking_knowledge offers 11 of the upstream's search methods. Nine are recorded as not offered in upstream-settings.json, for two reasons, both checked against the upstream at v1.0.1:

• The four reranker methods (bm25_reranker, bm25_reranker_grep, openai_embeddings_reranker, openai_embeddings_reranker_grep). The pointwise reranker asks a chat model through the OpenAI client (knowledge/postprocessors/pointwise_llm_reranker.py:59, model gpt-5.2), which reads OPENAI_BASE_URL. In a run that variable is the embeddings route, whose one model is text-embedding-3-large, or it is not set at all.
• The five qwen methods (qwen_embeddings, ..._grep, ..._reranker, ..._reranker_grep, alltools-qwen). Their embedder calls OpenRouter at an address written in the code (knowledge/embedders/openrouter_embedder.py:118, base_url="https://openrouter.ai/api/v1") with OPENROUTER_API_KEY, past the relay, which never hands a driver a key.

Wanted

The nine methods offered, with every call still through the run's relay.

• A relay route that sends a request by what it asks for: an embeddings request as the embedding model, a chat or responses request as the reranker's model. The driver's one OPENAI_BASE_URL then serves both, and a method without embeddings (bm25_reranker) gets the route for its reranker alone.
• For qwen, a route to OpenRouter on the run's relay, and the embedder pointed at it. The address is written in the upstream's code, so this needs either a change upstream (reading it from the environment) or one line of ours setting it in the driver; the second is code beyond glue and is the owner's yes to give.

Acceptance

• A relay route can pin one model for embeddings requests and another for chat requests, each recorded as the model sent.
• The four reranker methods are values of retrieval, each with what it needs said (needs).
• The owner has said how the qwen embedder is pointed at the relay.
• The five qwen methods are values of retrieval, or stay recorded as not offered with the reason.
• A run of each offered method on a few tasks makes every model call through the relay (its record shows the reranker's and the embedder's calls).

Dependencies and related

none