Hi, I could not find these three cases documented either, so I will separate what the docs say from what I would treat as an assumption.
What is documented (Convert external tables to managed): Databricks keeps the data in the old external location for 14 days to allow rollback. After 14 days, if predictive optimization is enabled, it deletes that old data automatically. If predictive optimization is off, you run VACUUM on the converted managed table after the 14 days (needs DBR 17.3 LTS or above, or serverless). The doc also notes that even with predictive optimization on, the old data may not get deleted at day 14 for things like rarely used tables. Only the data files in the external location are removed. The Delta log and the Unity Catalog reference stay.
On your three questions:
1) VACUUM before day 14. The docs frame the cleanup of the old location as something that happens after the 14 days, and VACUUM is described as the manual route for that. They do not say what a VACUUM before day 14 does to the old location. I would not rely on "it only touches the new managed location". If rollback matters, simply do not run VACUUM (manually or through a job) on that table inside the window, and check whether predictive optimization is enabled on the schema.
2 and 3) delta.deletedFileRetentionDuration. The conversion page does not mention it at all. My reading is that this property controls which unreferenced files VACUUM may remove from the table's own storage, while the 14 day window is a separate fixed retention applied to the old external copy. That would mean a 7 day value does not shorten the rollback window and a 20 day value does not extend it. That is an inference, not a documented guarantee, so please do not treat my reading as confirmed.
Since you asked whether anyone has tested it, the safe way to settle it is a quick experiment on a small throwaway external Delta table: SET MANAGED, then run VACUUM with different retention settings and check whether the files in the old external path are still there (list the path before and after). Doing that on a test table costs minutes, and the result is better evidence than any of our assumptions. If you need it guaranteed for production, I would raise it with your Databricks account team so they can confirm in writing, since the docs leave this interaction unspecified.
Practical rule until then: after SET MANAGED, leave the table alone with respect to VACUUM and predictive optimization for the full 14 days if you want rollback to stay safe.