feat(providers): add "local/" routing prefix for OpenAI-compatible local servers

[np6126]

Summary

A common deployment is running an OpenAI-compatible inference server locally — Ollama, LM Studio, vLLM, llama.cpp — and routing claw-code's OpenAI provider client at it via OPENAI_BASE_URL.

The obvious model id formats actively misroute today:

• qwen/qwen3:14b or qwen3-coder → metadata_for_model maps the qwen family to DashScope (DASHSCOPE_API_KEY, dashscope.aliyuncs.com). A user serving Qwen3 locally on Ollama is silently routed to Alibaba's cloud.
• kimi/kimi-k2.5 → same DashScope routing.
• grok-* → routes to xAI.

The existing openai/ prefix would work as a workaround, but mentally conflates "this is my local Qwen3" with "the OpenAI API itself", and a user who does want the existing OpenRouter behaviour (slug preserved on the wire for openai/...) can't have it both ways.

This PR adds local/ as an explicit routing prefix that says "this is a local OpenAI-compatible server, route as OpenAI client, strip the prefix on the wire". Routing semantics are identical to openai/: OpenAi provider, OPENAI_API_KEY (typically an unused placeholder for local servers), OPENAI_BASE_URL.

Diff

// providers/mod.rs
-if canonical.starts_with("openai/") || canonical.starts_with("gpt-") {
+if canonical.starts_with("openai/")
+    || canonical.starts_with("local/")
+    || canonical.starts_with("gpt-")
+{
     return Some(ProviderMetadata { provider: ProviderKind::OpenAi, ... });
 }

// providers/openai_compat.rs (wire_model_for_base_url)
-if matches!(lowered_prefix.as_str(), "xai" | "grok" | "qwen" | "kimi") {
+if matches!(lowered_prefix.as_str(), "local" | "xai" | "grok" | "qwen" | "kimi") {
     return Cow::Borrowed(&model[pos + 1..]);
 }

Two places only — both pure additions.

No change to strip_routing_prefix — it's #[allow(dead_code)] and only referenced from its own unit tests; adding local there would be cosmetic.

No change to the OpenAI gateway slug preservation logic used by OpenRouter and similar gateways with a non-default OPENAI_BASE_URL — openai/... behaviour stays exactly as upstream.

Example

export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_API_KEY=ollama-placeholder
claw --model local/qwen3:14b "..."

Wire request: POST $OPENAI_BASE_URL/chat/completions with model=qwen3:14b. No misrouting to DashScope.

Test plan

• cargo check --workspace clean
• Existing provider tests: cargo test -p api
• Manual: route local/qwen3:14b against a local Ollama, verify the request goes to the local server with bare model id on the wire

Companion discussion

A separate issue (#3045) covers two related open questions: <think>...</think> block filtering in streamed output, and whether the openai/-slug preservation logic should remain non-configurable once local/ is available. Posted separately because both involve tradeoffs upstream may want to weigh in on before code lands.

[shakoorshkh]

Model id formats like qwen/qwen3:14b silently misroute to DashScope when the user is running a local Ollama server. No error — wrong destination.

Adds a new prefix convention (local/) that users must learn. If OPENAI_BASE_URL is not set, local/ routes to OpenAI with a stripped model id, which will likely 404.

Explicit prefix over inference. Users targeting local servers now carry that intent in the model string rather than relying on routing heuristics to guess correctly.