Discrepancy in Daily user engagement between Firebase Analytics dashboard and BigQuery

Question

I'm trying to develop a query against Firebase Analytics data linked to BigQuery to reproduce the "Daily user engagement" graph from the Firebase Analytics dashboard (to include in a Google Data Studio report).

According to Firebase Help documentation, Daily user engagement is defined as "Average daily engagement per user for the date range, including the fluctuation by percentage from the previous date range." So, my attempt is to sum the engagement_time_msec (the additional engagement time (ms) since the last user_engagement event according to https://support.google.com/firebase/answer/7061705?hl=en) for user_engagement events, divided by the count of users (identified by user_dim.app_info.app_instance_id) per day. The query looks like this:

SELECT ((total_engagement_time_msec / 1000) / users) as average_engagement_time_sec, date FROM
(SELECT
  SUM(params.value.int_value) as total_engagement_time_msec,
  COUNT(DISTINCT(user_dim.app_info.app_instance_id)) as users,
  e.date
FROM `com_artermobilize_alertable_IOS.app_events_*`, UNNEST(event_dim) as e, UNNEST(e.params) as params
WHERE e.name = 'user_engagement'
AND params.key = 'engagement_time_msec'
GROUP BY e.date)
ORDER BY date desc

The results are close to what's displayed in the Firebase console graph of Daily user engagement, but the values from my query are consistently a few seconds higher (BigQuery results shown here on the left, Firebase Console graph values on the right).

To note, we're not setting user_dim.user_id and not using IDFA, so my understanding is the correct/only way to count "users" is the user_dim.app_info.app_instance_id, and I imagine the same would be true for the Firebase console.

Can anyone suggest what might be different between how I'm determining the average engagement time from BigQuery, and how that's being determined in the Firebase console graph?

To note, I've seen a similar question posed here, but I don't believe the suggested answer applies for my query since 1) the discrepancies are present over multiple days, 2) I'm already querying for user_engagement events and 3) the event date being used in the query is stated to be based on the registered timezone of your app (according to this).