refactor(agent): centralize resolved configuration

[zpzjzj]

Summary

• build one role-aware ResolvedAgentConfig after YAML, CLI, and credential loading
• resolve legacy slashed CLI model references once without mutating the loaded eval config, then pass the resolved value directly to adapters
• reuse the same resolver for agent_judge while inheriting the runner engine lifecycle, and report the resolved runner identity

Compatibility

• preserves legacy --model provider/name, opaque slashed model IDs, YAML provider/name pairs, provider environment precedence, local-login delegation, and Qoder auto semantics
• lets an explicit CLI model bypass superseded YAML provider/name environment references while continuing to resolve active base URL and model params
• does not add --provider, protocol-aware nested credentials, capability validation, or a public judge-engine schema

Testing

• make verify
• make test
• make test-action
• make e2e (all non-live paths passed; the local live Qoder CLI 0.2.0 run returned a successful response but did not expose token usage, so its existing non-zero token assertion failed)

Part of #196

[JHWang-1997]

Requesting changes for one immutability issue in the new resolved configuration. Kwargs and ModelParams are cloned, but the custom-engine configuration is still shared with the loaded eval config.

[JHWang-1997]

LGTM