[studio] fix: scope LLM operations to the current form

[ai-yang]

Summary

• rebase the fix onto the latest rocketmq-studio branch and adapt it to the redesigned LLM settings flow from feat: add claude-code and qoder agent engines with streaming prompt enhance #954
• bind connection-test and explicit-save callbacks to the current component lifecycle, form generation, and operation generation
• ignore late success, error, and loading callbacks after provider/engine/form changes, a newer operation, or unmount
• prevent a stale save from clearing a replacement API key or publishing a success notification for an obsolete form
• keep the current no-auto-save connection-test contract and temporarily disable connection tests while an explicit save is in flight

Fixes #738.

This remains separate from the initial configuration/model request ownership fixed by #737.

Root cause

handleTest and handleSave allowed every Promise callback to update the current form and shared loading/result state. A request started for an older form could therefore publish its result after the user changed provider, engine, or fields. An older explicit save could also clear a newly entered API key when its response arrived late.

The fix uses a shared operation generation for result ownership and per-request generations for loading ownership. Form changes invalidate connection tests immediately, while an already-sent save keeps its loading guard until it settles so another mutation cannot be started accidentally.

Compatibility

• no public API, payload, configuration, or persistence-format changes
• preserves provider, engine, enabled: true, empty-key retention, and the no-auto-save test behavior introduced by feat: add claude-code and qoder agent engines with streaming prompt enhance #954
• an explicit save still submits exactly once; only stale UI callbacks are suppressed

Verification

• latest base: 228ad5298dbbeaa769f1c1de242117c6cfbe4216
• unmodified rebased baseline: 6 deterministic ownership regressions failed in the targeted suite
• targeted operation-ownership suite: 20 isolated processes, 8/8 each (160/160)
• clean Node 20.20.2 container: 81 test files / 396 tests passed
• ESLint: 0 errors (4 pre-existing Fast Refresh warnings)
• TypeScript/Vite production build: passed, 7,997 modules transformed
• Prettier check and git diff --check: passed

[RockteMQ-AI]

Summary

Well-implemented fix for stale state updates in LlmSettings. The generation counter pattern (beginOperation/ownsOperation) correctly prevents async callbacks from updating state after the component unmounts or the user switches providers. This fixes a common React race condition where a slow test/save request could overwrite newer state.

LGTM — clean implementation with good test coverage.

---

Automated review by github-manager-bot

[lizhimins]

Closing: this area belongs to Track 3 (AI Native), which already has a complete design in progress (see #1024). Further LLM/AI changes should follow that design discussion instead of ad-hoc PRs. Thanks for the effort! / 关闭说明：该改动属于赛道三（AI Native）范围，赛道三已有完整的设计方案正在推进（见 #1024），LLM/AI 相关改动请跟随该设计讨论，不再单独接收零散 PR。感谢贡献！

[lizhimins]

Reopened — the previous close was applied in error; it does not fall under the Track 3 batch closure. / 重新打开：刚才的关闭属误操作，此 PR 不在赛道 3 批量关闭范围内。