Hot partition throttled on demand

Your leaderboard service starts returning elevated errors and the SDK logs fill with ProvisionedThroughputExceededException from DynamoDB, even though the table is on on-demand capacity and CloudWatch shows consumed capacity well under the account peak. The ThrottledRequests metric is spiking on one partition. A new feature shipped this morning that writes every score update under a single gameId for the current tournament. Client p99 is climbing and your retry storm is making CloudWatch consumed-capacity look higher than real demand. How do you triage and mitigate?