fix(rx): support dsh reasoning effort controls

[samzong]

What’s changed?

• Publish protocol-scoped model capabilities from the bundled rx provider snapshot.
• Route verified Tokener DSH endpoints through the Responses API and project supported reasoning effort choices into DSH settings.
• Keep runtime discovery limited to availability, prevent capability inheritance on custom endpoints, and restrict Tokener Dev to models verified on that endpoint.

Why

• Tokener-launched DSH used Chat Completions and model entries without effort metadata, so reasoning effort switching was unavailable and tool requests could reject selectable effort values.

Verification

• make check
• Built rx and verified the generated DSH route is openai-responses with GLM 5.2 limited to off, low, and medium.
• Verified a production GLM 5.2 Low request through Responses with tools returns HTTP 200.