Directly Linked Users

Stripe Fraud Ring Tasks
This problem is Task 1 of 3 in the Stripe fraud-ring sequence.
Task 1: Directly Linked Users (current)
Task 2: Fraud Ring Size
Task 3: Risky Fraud Ring
Financial crimes are rarely committed in isolation. Bad actors may spread activity across multiple user accounts, but shared identifiers can reveal connections.
Each transaction log entry is formatted as user_id,device_id. Two different users are directly linked if they use the same device. A single user may appear multiple times with different devices.
Given the log and a target user, return all users directly linked to the target user. The target user should not appear in the output. Return the user IDs in lexicographic order.
Input
transactions: transaction rows formatted as user_id,device_id.
targetUser: the user to investigate.
Output
Return the directly linked user IDs in lexicographic order, excluding targetUser.

Function
findDirectlyLinkedUsers(transactions: String[], targetUser: String) → String[]

Examples
Example 1
transactions = ["Alice,D1","Bob,D1","Charlie,D2","David,D3","Eve,D1"]
targetUser = "Alice"
return = ["Bob","Eve"]
Alice used D1. Bob and Eve also used D1.
Example 2
transactions = ["Alice,D1","Alice,D2","Bob,D2","Cara,D3"]
targetUser = "Alice"
return = ["Bob"]
Alice is associated with D1 and D2, and Bob shares D2.

Constraints
0 <= transactions.length <= 10^5
User IDs and device IDs contain no commas.

The trick is two passes. Pass one: scan the log and collect every device the target user touched into a set. Pass two: scan again and, for each row whose device is in that set, add the user to a result set, skipping the target. Sort lexicographically at the end. The common pitfall is returning duplicates. A user who shares two devices with the target, or appears in repeated rows, must show up once, so use a set, not a list. Another trap is forgetting to exclude the target when they appear on multiple devices. Also handle an empty log and a target who isn't in the log, both of which should return an empty array. Split each row on the first comma only, though the constraints say IDs contain no commas. This runs in O(n log n) with the sort, which is fine for 10^5 rows. Build the device-to-users map now, because Tasks 2 and 3 reuse it. If you freeze live, StealthCoder is the hedge that reads the problem and hands you the structure.

If this hits your live OA and you blank, StealthCoder solves it in seconds, invisible to the proctor.