feat(inference): default to the best line per SKU / feat(inference)：默认显示每个 SKU 的最佳曲线

[Oseltamivir]

Summary

• rank configuration Pareto frontiers within each physical hardware SKU using nine-point normalized AUC over the candidates' shared measured x-domain
• default /inference to the winning configuration while preserving manual legend selection and a shareable Best per SKU toggle
• apply the same selection policy to official data and ?unofficialrun= overlays, including overlay scope changes
• add Simplified Chinese UI copy (每个 SKU 仅显示最佳配置) and focused official/overlay tests

Why

Several frameworks, deployment modes, and speculative-decoding variants can produce many lines for the same chip. Showing one representative winner per SKU makes the initial comparison readable without removing access to the other configurations.

Validation

• bun run typecheck
• bun run lint
• bun run fmt
• focused Vitest: 44 passing
• Cypress component smoke suite: 25 passing, including unofficial overlay rendering
• full unit suite: 3,233 passing; one unrelated existing time-zone-sensitive failure in visit-tracking.test.ts (2026-05-15T22:00:00Z is the next local calendar day in Asia/Singapore)
• Cypress integration suite requires DATABASE_READONLY_URL; the local run was stopped after availability returned 500 because that variable is not configured

Works for both official runs and ?unofficialrun= overlays; verified through the component overlay suite.

---

Note

Medium Risk
Changes default legend selection and URL semantics on /inference, but manual toggles and i_best=0 preserve prior behavior; ranking is localized to chart filtering, not auth or data APIs.

Overview
Adds a Best per SKU mode on the inference dashboard so the legend starts with one winning hwKey per physical chip instead of every framework/deployment variant.

Selection uses a new bestSeriesPerSku helper: group series by SKU, build each candidate’s Pareto frontier for the active chart roofline direction, then pick the winner by normalized AUC over the shared measured x-range (so wider curves don’t win by span alone). Ties break on hwKey.

InferenceContext owns bestPerSku (on by default on the inference tab unless i_best=0), computes bestHwTypes from the current interactivity/e2e graph and y-metric, and uses that set for automatic hardware resets when model/sequence/metric/x-mode changes. Manual legend edits, select-all, or presets turn the mode off. Share URLs gain i_best.

ScatterGraph exposes a legend switch (EN + zh copy) that reapplies best selection for official + overlay preview runs. ChartDisplay applies the same policy when unofficial overlay scope changes so tables/charts stay aligned.

Vitest coverage for the ranking helper; Cypress mocks updated; visit-tracking.test fixes flaky calendar-day handling under non-UTC timezones.

^{Reviewed by Cursor Bugbot for commit 0aaf379. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 10, 2026 1:47pm

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b269baa. Configure here.}

[adibarra]

@claude fix the failing time zone sensitive test

[github-actions]

Claude finished @adibarra's task in 1m 35s —— View job

---

Fix failing time-zone-sensitive test

• Read visit-tracking.test.ts and locate the TZ-sensitive assertion
• Fix the test so it passes regardless of local time zone
• Run the focused unit test to confirm (default and TZ=Asia/Singapore)
• Commit & push

---

Root cause: visit-tracking.ts computes calendar days in local time (getFullYear/getMonth/getDate), but the test fed UTC ISO strings (…Z). In UTC+8, 2026-05-15T22:00:00Z rolls over to 2026-05-16 local, so the "does NOT increment when a new session happens on the same calendar day" case saw a different day and incremented — failing only in non-UTC zones.

Fix: setNow now strips the trailing Z so timestamps are parsed in the local zone, matching the code under test. No production code touched, and every ISO literal in the test bodies is unchanged.

function setNow(iso: string) {
  // …parse as local time to match visit-tracking's local-calendar-day logic…
  vi.setSystemTime(new Date(iso.replace(/Z$/u, '')));
}

Verification — bunx vitest run packages/app/src/lib/visit-tracking.test.ts:

• default TZ → 13 passed
• TZ=Asia/Singapore (the zone that reproduced the failure) → 13 passed

Pushed as 53215a9 to agent/best-line-per-sku.
· branch agent/best-line-per-sku

[adibarra]

@claude also resolve the cursor pr comments

[github-actions]

Claude finished @adibarra's task in 7m 52s —— View job

---

Re-review + resolve Cursor Bugbot comments

All three Bugbot findings are valid. I implemented and verified fixes for each — but this run is the read-only review job, and its token is denied push access (403 … denied to github-actions[bot]). The commit is ready locally (6d90d09, verified below); the full patch is at the bottom so you can apply it directly, or re-trigger an implement-capable job (@claude frontend) to push it.

• Read the changed files & understand the selection policy
• Fix chore: bump the all-minor-patch group with 2 updates #1 (High): Ranking ignores selected precision
• Fix chore: bump anthropics/claude-code-action from 04f907ba2a4c77706a56b4078c682d250716cc0c to cd77b50d2b0808657f8e6774085c8bf54484351c #2 (Medium): Overlay toggle/remove keeps Best per SKU on
• Fix chore: bump the all-minor-patch group with 7 updates #3 (High): Overlay + official miss default best selection on first load
• Add regression test
• Typecheck / lint / fmt / focused tests
• Push — blocked: read-only review-job token (403)

---

What each fix does

#1 — Ranking ignores selected precision (best-series-per-sku.ts, all 4 call sites)
bestSeriesPerSku now takes an optional isVisible(point) predicate and skips points that fail it, so a config can no longer win a SKU from a hidden precision / quick-filtered row and default the chart to a line with no visible series.

• InferenceContext.bestHwTypes → effectivePrecisions.includes(point.precision) (mirrors hwFilteredPoints).
• ChartDisplay.scopedBestSelections (official + overlay) → selectedPrecisions.includes(...) && matchesQuickFilters(...) (mirrors visibleComparisonRows).
• ScatterGraph toggle-on handler (official + overlay) → same predicate.

#2 — Overlay toggle keeps Best per SKU (ScatterGraph.tsx)
handleToggleHwType / handleRemoveHwType now call setBestPerSku(false) on the overlay (unified-selection) branch, matching the official path (toggleHwType/removeHwType already did). A manual overlay selection is no longer silently reseeded by a later scope change. The programmatic toggle-on path still sets bestPerSku (it calls commitUnifiedSelection directly, not the handlers), so it's unaffected.

#3 — Overlay misses default best selection (ChartDisplay.tsx)
scopedActiveOverlayHwTypes now returns scopedBestSelections.overlay whenever bestPerSku is on — not only during a scope change — so first loads no longer leave overlays fully selected while the official side already narrowed. This is safe because of #2: any manual overlay change clears bestPerSku, so forcing best here never overrides an explicit choice. selectedOfficialHwTypes also uses scopedBestSelections.official on the scope-change frame instead of the full officialScope.

Verification

• bun run typecheck → clean
• bun run lint → clean · bun run fmt → clean
• bunx vitest run …/best-series-per-sku.test.ts → 6 passed (added ignores points the chart does not render when picking a winner)

Patch (apply on agent/best-line-per-sku)

git apply this patch — resolves all three findings

diff --git a/packages/app/src/components/inference/InferenceContext.tsx b/packages/app/src/components/inference/InferenceContext.tsx
index 23ddf01..33b5f62 100644
--- a/packages/app/src/components/inference/InferenceContext.tsx
+++ b/packages/app/src/components/inference/InferenceContext.tsx
@@ -913,9 +913,13 @@ export function InferenceProvider({
     ) {
       return hwTypesWithData;
     }
-    const best = bestSeriesPerSku(graph.data, direction);
+    // Rank over the points the chart actually renders (the active precisions),
+    // matching hwFilteredPoints — otherwise a hidden-precision line could win.
+    const best = bestSeriesPerSku(graph.data, direction, (point) =>
+      effectivePrecisions.includes(point.precision),
+    );
     return best.size > 0 ? best : hwTypesWithData;
-  }, [graphs, hwTypesWithData, selectedXAxisMode, selectedYAxisMetric]);
+  }, [graphs, hwTypesWithData, selectedXAxisMode, selectedYAxisMetric, effectivePrecisions]);
 
   const setBestPerSkuAndApply = useCallback(
     (enabled: boolean) => {
diff --git a/packages/app/src/components/inference/ui/ChartDisplay.tsx b/packages/app/src/components/inference/ui/ChartDisplay.tsx
index f80f823..962a19c 100644
--- a/packages/app/src/components/inference/ui/ChartDisplay.tsx
+++ b/packages/app/src/components/inference/ui/ChartDisplay.tsx
@@ -464,8 +464,12 @@ export default function ChartDisplay() {
       return { official: officialScope, overlay: overlayScope };
     }
     const overlay = overlayDataByChartType[wantedType];
-    const officialBest = bestSeriesPerSku(graph.data, direction);
-    const overlayBest = bestSeriesPerSku(overlay?.data ?? [], direction);
+    // Rank only over rendered points so a hidden precision / filtered-out row
+    // cannot win the SKU (mirrors visibleComparisonRows' eligibility).
+    const isVisible = (point: InferenceData) =>
+      selectedPrecisions.includes(point.precision) && matchesQuickFilters(point, quickFilters);
+    const officialBest = bestSeriesPerSku(graph.data, direction, isVisible);
+    const overlayBest = bestSeriesPerSku(overlay?.data ?? [], direction, isVisible);
     return {
       official: officialBest.size > 0 ? officialBest : officialScope,
       overlay: overlayBest.size > 0 ? overlayBest : overlayScope,
@@ -478,6 +482,8 @@ export default function ChartDisplay() {
     overlayScope,
     selectedXAxisMode,
     selectedYAxisMetric,
+    selectedPrecisions,
+    quickFilters,
   ]);
   const overlayRowsScopeKey = `${selectedModel}|${selectedSequence}|${selectedPrecisions.join(
     ',',
@@ -486,7 +492,7 @@ export default function ChartDisplay() {
   const overlayRowsScopeChanged =
     isUnofficialRun && appliedOverlayRowsScopeKey !== overlayRowsScopeKey;
   const selectedOfficialHwTypes = overlayRowsScopeChanged
-    ? officialScope
+    ? scopedBestSelections.official
     : isUnofficialRun
       ? (localOfficialOverride ?? activeHwTypes)
       : activeHwTypes;
@@ -496,8 +502,21 @@ export default function ChartDisplay() {
     const activeScopedOverlayKeys = new Set(
       [...activeOverlayHwTypes].filter((key) => overlayScope.has(key)),
     );
+    // With Best per SKU on, default overlays to the winning line even on the
+    // first load — not only on a scope change. The provider seeds every overlay
+    // key on load; without this the overlay stays fully selected while the
+    // official side already narrowed. A manual overlay toggle clears bestPerSku
+    // (see handleToggleHwType / handleRemoveHwType), so forcing best here never
+    // overrides a user's explicit choice.
+    if (bestPerSku) return scopedBestSelections.overlay;
     return overlayRowsScopeChanged ? scopedBestSelections.overlay : activeScopedOverlayKeys;
-  }, [activeOverlayHwTypes, overlayScope, overlayRowsScopeChanged, scopedBestSelections.overlay]);
+  }, [
+    activeOverlayHwTypes,
+    overlayScope,
+    overlayRowsScopeChanged,
+    scopedBestSelections.overlay,
+    bestPerSku,
+  ]);
   useEffect(() => {
     const merged = new Set(activeOverlayHwTypes);
     overlayScope.forEach((key) => merged.delete(key));
diff --git a/packages/app/src/components/inference/ui/ScatterGraph.tsx b/packages/app/src/components/inference/ui/ScatterGraph.tsx
index 921bda5..69e20b2 100644
--- a/packages/app/src/components/inference/ui/ScatterGraph.tsx
+++ b/packages/app/src/components/inference/ui/ScatterGraph.tsx
@@ -685,10 +685,21 @@ const ScatterGraph = React.memo(
       commitUnifiedSelection,
     ]);
 
-    // When no overlay data, delegate to context's toggleHwType (preserves setActivePresetId)
+    // When no overlay data, delegate to context's toggleHwType (preserves
+    // setActivePresetId and clears bestPerSku). With an overlay loaded the
+    // toggle goes through the unified selection, which never touches bestPerSku,
+    // so clear it here — otherwise a later scope change reseeds the best set and
+    // wipes this manual choice.
     const handleToggleHwType = useCallback(
-      (key: string) => (overlayData ? unifiedToggle(key, false) : toggleHwType(key)),
-      [overlayData, unifiedToggle, toggleHwType],
+      (key: string) => {
+        if (!overlayData) {
+          toggleHwType(key);
+          return;
+        }
+        setBestPerSku(false);
+        unifiedToggle(key, false);
+      },
+      [overlayData, unifiedToggle, toggleHwType, setBestPerSku],
     );
 
     // Legend "X" (remove) — same overlay split as handleToggleHwType. With an
@@ -704,11 +715,12 @@ const ScatterGraph = React.memo(
           removeHwType(key);
           return;
         }
+        setBestPerSku(false);
         const next = new Set(resolvedUnifiedSelection);
         next.delete(key);
         commitUnifiedSelection(next);
       },
-      [overlayData, removeHwType, resolvedUnifiedSelection, commitUnifiedSelection],
+      [overlayData, removeHwType, resolvedUnifiedSelection, commitUnifiedSelection, setBestPerSku],
     );
 
     // --- Theme ---
@@ -3282,8 +3294,15 @@ const ScatterGraph = React.memo(
                           direction === 'lower_left' ||
                           direction === 'lower_right'
                         ) {
-                          const selection = bestSeriesPerSku(data, direction);
-                          for (const key of bestSeriesPerSku(overlayData.data, direction)) {
+                          const isVisible = (point: InferenceData) =>
+                            selectedPrecisions.includes(point.precision) &&
+                            matchesQuickFilters(point, quickFilters);
+                          const selection = bestSeriesPerSku(data, direction, isVisible);
+                          for (const key of bestSeriesPerSku(
+                            overlayData.data,
+                            direction,
+                            isVisible,
+                          )) {
                             selection.add(`overlay:${key}`);
                           }
                           commitUnifiedSelection(selection);
diff --git a/packages/app/src/components/inference/utils/best-series-per-sku.test.ts b/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
index 266b757..305acbb 100644
--- a/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
+++ b/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
@@ -57,6 +57,25 @@ describe('bestSeriesPerSku', () => {
     expect(baseSku(point('GB200-NVL72', 'gb200_dynamo-trt', 1, 1))).toBe('GB200');
   });
 
+  it('ignores points the chart does not render when picking a winner', () => {
+    const withPrecision = (hwKey: string, x: number, y: number, precision: string) =>
+      ({ ...point('B200-8', hwKey, x, y), precision }) as InferenceData;
+    // b200_hidden scores higher but only on a precision the chart hides, so the
+    // visible predicate must exclude it and pick the best visible series.
+    const selected = bestSeriesPerSku(
+      [
+        withPrecision('b200_hidden', 10, 100, 'fp4'),
+        withPrecision('b200_hidden', 20, 90, 'fp4'),
+        withPrecision('b200_visible', 10, 80, 'fp8'),
+        withPrecision('b200_visible', 20, 60, 'fp8'),
+      ],
+      'upper_left',
+      (candidate) => candidate.precision === 'fp8',
+    );
+
+    expect(selected).toEqual(new Set(['b200_visible']));
+  });
+
   it('ranks unofficial-run overlay series with the same SKU policy', () => {
     const overlayPoint = (hwKey: string, x: number, y: number) =>
       ({
diff --git a/packages/app/src/components/inference/utils/best-series-per-sku.ts b/packages/app/src/components/inference/utils/best-series-per-sku.ts
index db863e0..dc34c02 100644
--- a/packages/app/src/components/inference/utils/best-series-per-sku.ts
+++ b/packages/app/src/components/inference/utils/best-series-per-sku.ts
@@ -28,10 +28,19 @@ interface ScoredSeries {
  * so a curve cannot win merely because it spans a wider range. The chart's
  * existing Pareto direction and monotone interpolation are reused to keep the
  * ranking aligned with the line users see. Ties are deterministic by hwKey.
+ *
+ * `isVisible` restricts scoring to the points the chart actually renders (e.g.
+ * the selected precisions and quick filters). Without it, a configuration could
+ * win from hidden-precision points and default to a line with no visible series.
  */
-export function bestSeriesPerSku(points: InferenceData[], direction: Direction): Set<string> {
+export function bestSeriesPerSku(
+  points: InferenceData[],
+  direction: Direction,
+  isVisible: (point: InferenceData) => boolean = () => true,
+): Set<string> {
   const bySku = new Map<string, Map<string, InferenceData[]>>();
   for (const point of points) {
+    if (!isVisible(point)) continue;
     if (!isFrontierEligible(point) || !Number.isFinite(point.y)) continue;
     const sku = baseSku(point);
     const key = String(point.hwKey);

中文说明
Cursor Bugbot 的三个问题均属实，已实现并本地验证修复（提交 6d90d09），但当前是只读的 review 任务，其 token 无推送权限（403），因此无法推送。请直接应用上方补丁，或用 @claude frontend 重新触发可写任务来推送。

• chore: bump the all-minor-patch group with 2 updates #1（高）精度过滤：bestSeriesPerSku 新增可选 isVisible 谓词，仅在图表实际渲染的点（当前精度 + 快速过滤）上排名，避免隐藏精度的配置胜出而导致该 SKU 没有可见曲线。
• chore: bump anthropics/claude-code-action from 04f907ba2a4c77706a56b4078c682d250716cc0c to cd77b50d2b0808657f8e6774085c8bf54484351c #2（中）overlay 手动切换未清除 bestPerSku：overlay 分支的切换/移除现在调用 setBestPerSku(false)，与官方路径一致，后续 scope 变化不再覆盖手动选择。
• chore: bump the all-minor-patch group with 7 updates #3（高）overlay 首次加载未应用最佳选择：当 bestPerSku 开启时，overlay 始终默认最佳曲线（不再仅限 scope 变化）；由于 chore: bump anthropics/claude-code-action from 04f907ba2a4c77706a56b4078c682d250716cc0c to cd77b50d2b0808657f8e6774085c8bf54484351c #2 保证手动操作会清除该开关，此处强制不会覆盖用户选择。
  验证：typecheck / lint / fmt 通过，聚焦测试 6 项通过。
  · branch agent/best-line-per-sku