fix(changelog): derive config labels from run data / 修复变更记录：根据运行数据生成配置标签

[cquil11]

Summary

• Derive changelog config labels and chart highlights from the benchmark configs actually produced by the selected workflow run.
• Use raw changelog keys only to scope the affected model/config family and as a fallback for historical runs without benchmark coverage.
• Add regression coverage for Kimi K3 H200, where the raw agentic key omitted MTP while the run content reported spec_method: mtp.

Root cause

The previous formatter shared the legend's display builders but still inferred MTP by searching for a literal mtp segment in the raw changelog key. Config keys are selectors, not authoritative benchmark metadata, so labels could disagree with the plotted points.

Validation

• bun run test:unit — 3,570 tests passed
• bun run typecheck
• bun run lint
• bun run fmt
• E2E_FIXTURES=1 bun run build
• Cypress component suite — 195 tests passed
• The production-mode Cypress integration rerun exposed pre-existing stateful Agentic spec failures after its first passing test; no failures referenced the changed changelog code.

中文说明

变更内容

• 根据所选工作流实际产出的基准测试配置生成变更记录中的配置标签和图表高亮。
• 原始变更记录键仅用于限定受影响的模型与配置范围；对于缺少基准测试覆盖数据的历史运行，继续作为兼容回退。
• 补充 Kimi K3 H200 回归测试：原始 agentic 键未包含 MTP，但运行数据中的 spec_method 为 mtp。

根因

此前的格式化逻辑虽然复用了图例的显示构建器，但仍通过搜索原始变更记录键中的字面量 mtp 来判断 MTP。配置键只是选择器，并非权威的基准测试元数据，因此标签可能与实际绘制的数据点不一致。

验证

• bun run test:unit — 3,570 项测试通过
• bun run typecheck
• bun run lint
• bun run fmt
• E2E_FIXTURES=1 bun run build
• Cypress 组件测试 — 195 项测试通过
• 生产模式 Cypress 集成测试重跑在首项测试通过后暴露了既有的 Agentic 有状态用例失败；失败均未涉及本次修改的变更记录代码。

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 4, 2026 8:52pm