fix: wire persisted provider configuration into runtime

[DaBear987]

Wires persisted provider configuration from runtime settings into actual provider execution. Provider kind is authoritative when explicitly configured, with regression coverage for Anthropic, xAI, OpenAI, DashScope, and local OpenAI-compatible routing. Validation completed with cargo fmt --check, runtime/API/CLI test suites, and git diff --check.

[1716775457damn]

Good change — making provider kind authoritative removes the ambiguity of a persisted config silently overriding the routing. One question for clarity: with kind authoritative, what wins against an explicit runtime hint, e.g. a --provider flag or a model prefix on the command line? If the stored kind always wins regardless of the current invocation, a user doing a one-off provider switch could be surprised. Either a short precedence note in the docs or a test covering flag-vs-persisted precedence would make the intent unambiguous.

[1716775457damn]

补充建议形成闭环：优先级建议定为「显式运行期参数（--provider / 命令行 model 前缀）> 持久化配置中的 provider kind > 默认推断路由」。这样持久化配置保证会话间一致，命令行参数保留单次覆盖能力，符合多数 CLI 惯例。若同意此语义，可补一条 flag-vs-persisted 的优先级回归测试；runtime 测试套件这边需要的话我可以协助补。

[1716775457damn]

5adb8be (fix: honor explicit model provider over persisted config) 正好落实了我上次建议的优先级语义——显式运行期参数优先于持久化 provider kind，符合 CLI 惯例，赞。合入前一个小建议：可以合并掉那批临时的 CI patcher / stage / export commits（82eae82～29c904a），用一个干净的功能 commit 承载 5adb8be 的改动，评审与 bisect 都会更清爽。

[1716775457damn]

Re-checked at 5adb8be — the "explicit args > persisted kind > default inference" priority is now correctly applied through the injection path, matching what we discussed; the fallback chain reads clean. Also noticed the temporary PR3299 CI hook has been reverted (530a37f), good cleanup so it won't pollute the final merge. Worth re-running the full CI once more on the cleaned head before merge to confirm nothing regressed.

[1716775457damn]

Reviewed the force-push to f8ad763 — the head stays consistent with the "explicit args > persisted kind > default inference" semantics we settled on, and the temp CI hook is out of the final diff. No new concerns; please run the full CI once on this cleaned head before merge and we can get this in.