fix(agentic): default percentile to p90, surface only p90/p99

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: functionstackx

packages/app/cypress/support/mock-data.ts

@@ -189,7 +189,7 @@ export function createMockInferenceContext(
     workflowInfo: null,
     selectedYAxisMetric: 'y_tpPerGpu',
     setSelectedYAxisMetric: namedStub('setSelectedYAxisMetric'),
-    selectedPercentile: 'median',
+    selectedPercentile: 'p90',
     setSelectedPercentile: namedStub('setSelectedPercentile'),
     selectedXAxisMetric: null,
     setSelectedXAxisMetric: namedStub('setSelectedXAxisMetric'),

packages/app/src/components/inference/InferenceContext.tsx

@@ -134,7 +134,7 @@ export function InferenceProvider({
     () => getUrlParam('i_e2e_xmetric') || null,
   );
   // Latency percentile applied to the chart x-axis for agentic scenarios.
-  // Values: 'median' | 'p90' | 'p99' | 'p99.9'. Non-agentic charts ignore.
+  // Values: 'p90' | 'p99'. Non-agentic charts ignore.
   const [selectedPercentile, setSelectedPercentile] = useState<string>(
     () => getUrlParam('i_pctl') || 'p90',
   );

packages/app/src/components/inference/hooks/useChartData.ts

@@ -83,7 +83,7 @@ export function useChartData(
   selectedRunDate?: string,
   enabled = true,
   latestAvailableDate?: string,
-  selectedPercentile = 'median',
+  selectedPercentile = 'p90',
   /** When set, only series for these two registry GPU keys are shown (compare pages). */
   compareGpuPair?: readonly [string, string] | null,
 ) {
@@ -261,12 +261,7 @@ export function useChartData(
             selectedPercentile,
           ) as keyof AggDataEntry;
           if (adjusted !== xAxisField) {
-            const pctlWord =
-              selectedPercentile === 'median'
-                ? 'Median'
-                : selectedPercentile === 'p99.9'
-                  ? 'P99.9'
-                  : selectedPercentile.toUpperCase();
+            const pctlWord = selectedPercentile.toUpperCase();
             xAxisLabel = xAxisLabel.replace(/^(Median|Mean|P90|P99(?:\.9)?)\b/iu, pctlWord);
             xAxisField = adjusted;
           }

packages/app/src/components/ui/chart-selectors.tsx

@@ -315,7 +315,7 @@ export function PercentileSelector({
       <LabelWithTooltip
         htmlFor={id}
         label="Latency Percentile"
-        tooltip="Percentile of the latency distribution used for the chart x-axis. Agentic runs carry median/p90/p99/p99.9 variants; switch percentiles to see tail-latency behavior."
+        tooltip="Percentile of the latency distribution used for the chart x-axis. Switch between p90 and p99 to see tail-latency behavior on agentic runs."
       />
       <Select
         value={value}

packages/app/src/lib/data-mappings.ts

@@ -186,15 +186,17 @@ export const SEQUENCE_OPTIONS = Object.keys(SEQUENCE_CONFIG) as Sequence[];
 /**
  * Percentile of the latency distribution used for the chart x-axis when
  * viewing agentic traces. Agentic rows carry median/p90/p99/p99.9 variants
- * for ttft, ttlt (=e2el), and itl (and intvty derived from itl); p90 and
- * p99 are surfaced in the UI.
+ * for ttft, ttlt (=e2el), and itl (and intvty derived from itl); only p90
+ * and p99 are surfaced in the UI.
  */
 export enum Percentile {
   P90 = 'p90',
+  P99 = 'p99',
 }
 
 const PERCENTILE_CONFIG: Record<Percentile, { label: string }> = {
   [Percentile.P90]: { label: 'p90' },
+  [Percentile.P99]: { label: 'p99' },
 };
 
 export const PERCENTILE_OPTIONS = Object.keys(PERCENTILE_CONFIG) as Percentile[];

packages/app/src/lib/url-state.ts

@@ -67,7 +67,7 @@ export const PARAM_DEFAULTS: Record<UrlStateKey, string> = {
   i_seq: '8k/1k',
   i_prec: 'fp4',
   i_metric: 'y_tpPerGpu',
-  i_pctl: 'median',
+  i_pctl: 'p90',
   i_xmetric: 'p99_ttft',
   i_e2e_xmetric: '',
   i_scale: 'auto',