Fix Best per SKU in MP4 replay / 修复 MP4 回放中的“每个 SKU 最佳配置”

[Oseltamivir]

Summary

• Recompute the metric-aware Best per SKU winner after interpolation for every replay frame, so the MP4 preview and export can switch hwKeys as benchmark history advances while keeping one winner per physical SKU.
• Reuse the current parent chart winner's color for every historical winner of the same physical SKU, preventing older hwKeys from falling back to gray while their config labels change.
• Give historical winners for one physical SKU a stable roofline identity and encode a deterministic ~480 ms geometry morph directly into replay fractions. The live preview and exported MP4 now consume the same intermediate line positions instead of relying on a wall-clock D3 transition that the encoder could miss.
• Offer 3×, 4×, and 5× playback only while Best per SKU is enabled; disabling the mode clamps playback back to the regular 2× ceiling. Best-per-SKU MP4 mode keeps its explicitly requested animation active even when the ambient reduced-motion preference is enabled; regular replay still uses the stepped, non-animated fallback.
• Apply the selected Best-per-SKU speed to MP4 duration while retaining 30 FPS, cutting a maximum-length export from 900 frames at 1× to 180 at 5×. Wait for one paint boundary instead of two before rasterizing each frame; regular replay export remains unchanged.
• Build fixed replay axes from the configurations that can actually appear, and honor the current precision, quick-filter, and hardware visibility state.
• Carry visible unofficial-run data into replay, date-gate it at the playhead, apply the same Best per SKU policy, and avoid writing historical winners back into the live parent chart.
• Add unit and Cypress regression coverage for changing historical winners, stable SKU colors, encoded intermediate winner geometry, exact settled geometry, speed gating, and unofficial-run overlay filtering.

Unofficial run overlays

Works for both official runs and ?unofficialrun= overlays. Overlay rows respect active hardware visibility, per-run dismissal through the assembled overlay input, precision filters, quick filters, and the faster Best-per-SKU playhead. Unofficial runs have one observed benchmark date, so they have no cross-date winning-config geometry to morph; they remain date-gated and use their run colors. The overlay path is covered by the unit regression; the official replay winner transition, historical color reuse, deterministic geometry, and speed gating are verified by unit and Cypress tests. Vercel preview is deployed and ready.

Testing

• bun run typecheck
• bun run lint
• bun run fmt
• bun run test:unit — 4,112 tests passed across workspaces
• bun vitest run src/components/inference/replay/__tests__/replayFrameData.test.ts — 31 tests passed; verifies an encoded intermediate MP4 frame differs from the snapped winner and that the line settles exactly on the incoming config
• bunx cypress run --spec cypress/e2e/inference-replay-best-per-sku.cy.ts — 1 test passed; forces reduced motion on and verifies continuous progress within 300 ms, stable path reuse, deterministic geometry updates, 5× selection, and removal of 5× outside Best per SKU
• bun run test:e2e:quick:component — 27/27 component smoke tests passed
• DB-backed smoke integration remains unavailable locally because DATABASE_READONLY_URL is not set

中文说明

• 在每个回放帧完成插值后，按当前指标重新计算“每个 SKU 最佳配置”，使 MP4 预览和导出能够随着基准测试历史推进动态切换 hwKey，同时确保每个物理 SKU 始终保留一个胜出配置。
• 同一物理 SKU 的所有历史胜出配置沿用当前父图表胜出配置的颜色，避免旧 hwKey 在配置标签变化时回退为灰色。
• 为同一物理 SKU 的历史胜出配置复用稳定的 roofline 标识，并将约 480 ms 的确定性几何变形直接编码到回放进度中。实时预览和导出的 MP4 现在使用完全相同的中间曲线位置，不再依赖编码器可能错过的实际时间 D3 过渡。
• 仅在启用“每个 SKU 最佳配置”时提供 3×、4× 和 5× 回放速度；关闭该模式后，播放速度自动降至常规上限 2×。即使系统启用了“减少动态效果”，Best-per-SKU MP4 模式仍会保留用户明确请求的动画；常规回放则继续采用分步且无动画的降级方式。
• MP4 导出时长遵循所选的 Best-per-SKU 速度并保持 30 FPS，使最长回放的帧数从 1× 时的 900 帧降至 5× 时的 180 帧。每帧光栅化前只等待一次绘制边界而非两次；常规回放导出保持不变。
• 固定坐标轴只覆盖回放中实际可能出现的配置，并遵循当前 precision、快速筛选和硬件可见性状态。
• 将可见的 unofficial run 数据纳入回放：根据播放时间点控制出现时机，应用相同的“每个 SKU 最佳配置”规则，同时避免把历史帧的胜出配置写回实时父图表。
• 补充单元测试和 Cypress 回归测试，覆盖历史胜出配置动态变化、SKU 颜色稳定性、编码后的中间曲线几何、最终稳定几何、速度范围限制以及 unofficial run overlay 的筛选逻辑。

Unofficial run overlay 支持

该改动同时支持 official run 和 ?unofficialrun= overlay。Overlay 数据遵循硬件可见性、组装 overlay 输入中的单次 run 关闭状态、precision 筛选、快速筛选和更快的“每个 SKU 最佳配置”播放进度。Unofficial run 只有一个观测日期，因此不存在可进行跨日期变形的胜出配置曲线；它们仍按播放时间点出现并使用对应 run 的颜色。单元测试覆盖 overlay 路径；单元测试和 Cypress 测试共同验证 official replay 的胜出配置切换、历史 SKU 颜色复用、确定性几何变化和速度范围限制。Vercel preview 已部署并就绪。

测试

• bun run typecheck
• bun run lint
• bun run fmt
• bun run test:unit — 所有 workspace 共 4,112 个测试通过
• bun vitest run src/components/inference/replay/__tests__/replayFrameData.test.ts — 31 个测试通过；验证 MP4 中间帧不同于直接跳变后的胜出配置，并验证曲线最终精确落在新配置上
• bunx cypress run --spec cypress/e2e/inference-replay-best-per-sku.cy.ts — 1 个测试通过；强制启用“减少动态效果”，并验证回放在 300 ms 内持续推进、稳定复用 SVG path、更新确定性几何、支持 5× 选择，以及关闭“每个 SKU 最佳配置”后移除 5×
• bun run test:e2e:quick:component — component smoke suite 27/27 通过
• 由于本地未配置 DATABASE_READONLY_URL，依赖数据库的 smoke integration 无法运行

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 21, 2026 8:26am

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.