fix(settings): persist custom model reasoning efforts correctly

[AtlaxTech]

Summary

Fix custom-model reasoning effort settings so they persist using the adapter's canonical reasoningEfforts configuration field instead of the runtime-only reasoning descriptor. This makes configured reasoning levels available in the conversation model picker for custom providers.

What changed

• Persist per-model reasoning levels through reasoningEfforts.
• Preserve existing wire aliases for configured levels.
• Migrate valid legacy reasoning.efforts entries on edit.
• Remove legacy reasoning in the same model update.
• Remove the per-model default-effort selector, since the default belongs to provider/runtime resolution.
• Add regression coverage for persistence, legacy migration, adapter schema compatibility, and complete custom-model object propagation.

Verification

• Clean npm ci / patch-package postinstall: PASS
• Focused regression: 9/9 PASS
• Typecheck: PASS
• Build: PASS
• git diff --check: PASS
• LAN tests with local-bind permission: 49/49 PASS
• Full suite with local-bind permission: 670/671 PASS; the sole remaining failure is the pre-existing Feishu previous-tag assertion
• Real runtime smoke using a custom OpenAI-compatible provider backed by local Ollama:
  • model: qwen3:8b
  • picker exposed Default / Low / High
  • High selected
  • request completed normally as a full assistant turn
• Legacy lowercase reasoning configuration migrated successfully to canonical reasoningEfforts; legacy reasoning and defaultEffort were removed after save/reload

Related

• fix(settings): expose reasoning effort editor for custom provider models #215 introduced the custom-model reasoning controls this fixes.
• [功能建议] 支持在 provider 级别配置 reasoningEfforts 默认值 #230 is related provider-inheritance work and is outside this PR's scope.