Add Opus 5 and Kimi K3 benchmark scores

[qinxuye]

Summary

• add a bundled Claude Opus 5 profile with official model ID, limits, pricing, and source links
• record Opus 5 GPQA Diamond 93.2 from Artificial Analysis and LiveCodeBench 89.0 from Vals
• fill Kimi K3's previously missing LiveCodeBench score with the new Vals result, 87.2
• add registry regression coverage for both models and both capability benchmarks

Why

The production capability axis uses gpqa_diamond and livecodebench. Opus 5 is now available, and Vals has published LiveCodeBench results for both Opus 5 and Kimi K3, so both profiles can provide the complete benchmark pair.

Validation

• XROUTER_LLM_ENTERPRISE_DISABLE_AUTOPATCH=1 PYTHONPATH=src python -m pytest -q — 81 passed, 21 skipped
• PYTHONPATH=src python -m compileall -q src tests
• git diff --check
• repository secret scan

[gemini-code-assist]

Code Review

This pull request adds a new model configuration for claude-opus-5 and updates the existing kimi-k3 model configuration with new source URLs and benchmark scores. It also updates the corresponding unit tests to verify the loaded profiles. The review feedback points out an inconsistency in the source_quality field for claude-opus-5, which is marked as official despite relying on third-party benchmark evaluations, and suggests changing it to third_party to align with other models like kimi-k3.