With a 90-day incremental refresh configured, the 100-day old record would not be picked up because it's outside the refresh range. The existing record from a previous refresh will remain unchanged.

The Detect Data Changes section of the incremental load documentation indirectly lists approaches to this issue when it addresses cardinality and memory consumption:

• Use a power query (M) function to persist the latest value.
• Using XMLA scripts to pre-process records however you like during incremental load (of course, it doesn't tell you it's for Premium Subscription only)
• Offloading the task to a separate ETL process (ADF, Databricks, etc), having it flag the data PBI should consume. This is what I do to centralize data validation and avoid implementing unique solutions for every data consumer.