feat: add OrcaRouter as a named backend

[zsanig22-dotcom]

Overview

RLM's whole pitch is swapping llm.completion(prompt, model) for a recursive rlm.completion(prompt, model) that can programmatically examine, decompose, and recursively call itself over near-infinite contexts. When you're chaining sub-LM calls that way, every hop goes through your model provider — and the way you reach those models shouldn't force you to pick a single vendor or a bare custom base_url.

This PR adds orcarouter as a first-class backend, mirroring how the existing openrouter backend is wired in rlm/clients/__init__.py and rlm/clients/openai.py.

What's included

• rlm/clients/__init__.py: new orcarouter branch in get_client() that dispatches to OpenAIClient with base_url defaulting to https://api.orcarouter.ai/v1 — exactly the same pattern as the openrouter and vercel branches.
• rlm/clients/openai.py: auto-loads ORCAROUTER_API_KEY when api_key is omitted and base_url matches the OrcaRouter endpoint, so the key is picked up from the environment like OPENROUTER_API_KEY.
• rlm/core/types.py: "orcarouter" added to the ClientBackend literal.
• Docs: README (Model Providers), docs/api/rlm.md (backend literal + kwargs table), docs/getting-started.md, and the docs site's backends and api pages.
• Tests: dispatch test in tests/test_imports.py asserting the orcarouter backend resolves to OpenAIClient with the correct base URL.

Usage looks like the existing OpenRouter snippet:

rlm = RLM(
    backend="orcarouter",
    backend_kwargs={
        "api_key": os.getenv("ORCAROUTER_API_KEY"),
        "model_name": "orcarouter/auto",  # or vendor/model, e.g. anthropic/claude-opus-4.8
    },
)

Why OrcaRouter

OrcaRouter is an OpenAI-compatible AI gateway built for both models and agents. Like OpenRouter, it exposes a provider/model namespace across many models — but it also combines adaptive routing, automatic failover, zero-markup inference, observability, guardrails, and agent-tool governance behind the same endpoint. Adding orcarouter as a first-class provider means RLM users can use that stack directly, without treating OrcaRouter as an anonymous custom base URL. It also runs gateway-level, zero-trust security for AI agents on the same endpoint — screening every prompt/response and governing every tool call on a default-deny basis, with no application code changes.

Verification

• ruff check . and ruff format --check . pass (0.14.10).
• pytest tests/ --ignore=tests/repl/test_modal_repl.py --ignore=tests/clients/ — 251 passed, including the new test_orcarouter_backend_dispatch.
• Live-tested against https://api.orcarouter.ai/v1 through the new backend path: client.completion("Reply with exactly: ORCAOK") returned ORCAOK (200, token usage tracked).

Discord: discord.gg/YEubt8enRA · X: https://x.com/OrcaRouter

I'm an engineer on the OrcaRouter team.