Fix GrowthBook exposure did mis-attribution across account switches

Exposure events (experiment:viewed / feature:viewed) could be logged with
a did that did not match the unit GrowthBook actually bucketed on, surfacing
as "Multiple Exposures" warnings on did-based experiments.

Two causes in AnalyticsFeaturesContext:

- The variation came from result.key but the did came from ambient React
  session metadata, so they were sourced independently. Stamp the event's
  did with result.hashValue so did and variation come from the same
  evaluation. Only did-bucketed experiments are overridden.
- setAttributes triggers a synchronous re-evaluation that can fire exposure
  events. It ran before the new render's tracking callback was registered,
  so events went through a stale callback closing over the previous
  account's metadata after a key={did} remount. Register the callbacks
  before setAttributes to close that window.

Kept registration synchronous rather than moving to a useEffect:
setAttributes must run before children read gates, and the feature usage
callback has no deferred-replay, so a feature evaluated before registration
would lose its event.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/analytics/index.tsx

@@ -1,5 +1,6 @@
 import {createContext, useContext, useMemo} from 'react'
 import {Platform} from 'react-native'
+import {type Result} from '@growthbook/growthbook-react'
 
 import {Logger} from '#/logger'
 import {
@@ -169,6 +170,39 @@ export function AnalyticsContext({
   return <Context.Provider value={childContext}>{children}</Context.Provider>
 }
 
+/**
+ * GrowthBook attribute name for the did. Must match the key used in
+ * `setAttributes` (`#/analytics/features`) and the assignment unit configured
+ * in the GrowthBook dashboard.
+ */
+const DID_HASH_ATTRIBUTE = 'did'
+
+/**
+ * Builds the session metadata override for an exposure event so the event's
+ * `did` is sourced from the unit GrowthBook actually bucketed on
+ * (`result.hashValue`) rather than from ambient React session metadata. This
+ * keeps the `did` and the variation in sync, since both then come from the
+ * same evaluation. See APP-2461.
+ *
+ * Only did-bucketed experiments are overridden. Experiments bucketed on
+ * another attribute (e.g. `deviceId`) have a `hashValue` that is not a did, so
+ * we leave their session metadata untouched and let the ambient did stand.
+ */
+function sessionMetadataForResult(
+  parentContext: AnalyticsBaseContextType,
+  result: Result<unknown>,
+): MergeableMetadata | undefined {
+  if (result.hashAttribute !== DID_HASH_ATTRIBUTE) return undefined
+  const {session} = parentContext.metadata
+  if (!session) return undefined
+  return {
+    session: {
+      ...session,
+      did: result.hashValue,
+    },
+  }
+}
+
 /**
  * Feature gates provider. Decorates the parent analytics context with
  * feature gate capabilities. Should be mounted within `AnalyticsContext`,
@@ -185,22 +219,46 @@ export function AnalyticsFeaturesContext({
    * Side-effects: we need to synchronously set these during the same render
    * cycle. These calls do not trigger re-renders, they just set properties on
    * the singleton GrowthBook instance.
+   *
+   * Order matters here. We register the tracking callbacks _before_ calling
+   * `setAttributes`, because `setAttributes` triggers a synchronous
+   * re-evaluation that can fire exposure events. Registering first guarantees
+   * those events run through this render's callback (with this render's
+   * metadata) rather than a stale callback left over from a previous render,
+   * e.g. after an account switch remounts this provider via the
+   * `<Fragment key={did} />` breaker in `App.<platform>.tsx`. See APP-2461.
+   *
+   * We deliberately keep these synchronous rather than moving them into a
+   * `useEffect`: `setAttributes` must run before children evaluate gates (or
+   * they bucket on the previous account's attributes), and the feature usage
+   * callback has no deferred-replay, so a feature evaluated before it is
+   * registered would lose its `feature:viewed` event entirely.
    */
-  setAttributes(parentContext.metadata)
   feats.setTrackingCallback((experiment, result) => {
-    parentContext.metric('experiment:viewed', {
-      experimentId: experiment.key,
-      variationId: result.key,
-    })
+    parentContext.metric(
+      'experiment:viewed',
+      {
+        experimentId: experiment.key,
+        variationId: result.key,
+      },
+      sessionMetadataForResult(parentContext, result),
+    )
   })
   feats.setFeatureUsageCallback((feature, result) => {
-    parentContext.metric('feature:viewed', {
-      featureId: feature,
-      featureResultValue: result.value,
-      experimentId: result.experiment?.key,
-      variationId: result.experimentResult?.key,
-    })
+    parentContext.metric(
+      'feature:viewed',
+      {
+        featureId: feature,
+        featureResultValue: result.value,
+        experimentId: result.experiment?.key,
+        variationId: result.experimentResult?.key,
+      },
+      result.experimentResult
+        ? sessionMetadataForResult(parentContext, result.experimentResult)
+        : undefined,
+    )
   })
+  setAttributes(parentContext.metadata)
 
   const childContext = useMemo<AnalyticsContextType>(() => {
     return {