The Phantom Watchers

BingeBox, a streaming service, has a data-quality poltergeist. Some rows in the watch logs have a null minutes_watched (the player crashed before reporting) and some have a null device (ancient smart TVs that predate the tracking SDK). The analytics team calls them the phantom watchers — and right now the phantoms are silently vanishing from every report, because comparing anything to null in Spark yields null, and a null filter condition drops the row.

Your job: exorcise the nulls without losing the rows.

The watch_logs DataFrame:

Your task — produce a cleaned engagement report:

1. minutes_clean: minutes_watched with nulls replaced by 0.0.
2. device_clean: device with nulls replaced by "unknown".
3. engagement: derived from minutes_clean —
   • "binge" when minutes_clean >= 60
   • "casual" when minutes_clean >= 10 (but under 60)
   • "phantom" otherwise (under 10 — including the resurrected zero-minute rows)
4. Keep only rows where the engagement is "binge" or "phantom" — the two groups the retention team is studying. A phantom row with a null device must survive all the way to the output.
5. Return exactly log_id, show, device_clean, engagement.

Assign the DataFrame to result. Row order doesn't matter.