> This says to me a subquery (which always returns a single value) represents an empty rowset as `null` in that return value, ...

A subquery used in a context that expects a single, scalar value needs to return a single, scalar value. Such contexts are when used as an expression or with operators that expect a single, scalar value. When no scalar value exists, it returns `NULL` for precisely the reason you thought: it means "unknown". Think of a subquery in these contexts as being a replacement for a scalar UDF. It will always return a single something (T-SQL does not allow for a return type of `void`). So, what you have is effectively: `select @testB = dbo.Function(6);` assuming that "dbo.Function" is defined as:

    CREATE FUNCTION dbo.Function (@Input INT)
      DECLARE @Temp INT;
      SELECT @Temp = TestTable.[id]
      FROM   (VALUES (5)) TestTable([id])
      WHERE  TestTable.[id] = @Input;

      RETURN @Temp;

More examples:

    SELECT 1 AS [one], (SELECT 2 WHERE 1 = 0) AS [nothing];
    -- 1   NULL


    DECLARE @MoreThanOneRow INT;
    SELECT @MoreThanOneRow = (SELECT [object_id] FROM sys.objects);
    Msg 512, Level 16, State 1, Line XXXXX
    Subquery returned more than 1 value. This is not permitted when the subquery
      follows =, !=, <, <= , >, >= or when the subquery is used as an expression.


    DECLARE @MoreThanOneColumn INT;
    SELECT @MoreThanOneColumn = (SELECT TOP (1) [object_id], [schema_id] FROM sys.objects);
    Msg 116, Level 16, State 1, Line XXXXX
    Only one expression can be specified in the select list when the
      subquery is not introduced with EXISTS.




> ... but attempting to select a single column from an empty rowset does not return `null`. Rather, it kind of says "I have no value, so there is nothing to assign".

I wouldn't think of an empty result set in terms of "I have no value, so there is nothing to assign". Instead, think of it in terms of: an assignment will be performed _for every row returned_, and no rows returned means that there is no action to take. It might be a subtle difference, but the focus really needs to be on "rows = actions" more so than "returned value is something that can be assigned".

This distinction is not so subtle when:  

1. You are executing a T-SQL scalar UDF in a query and passing in a column. Assuming we are _not_ dealing with Scalar UDF inlining (starting in SQL Server 2019), then scalar UDFs are executed per each row, hence the performance issues associated with them and the impetus for the new feature in SQL Server 2019 to inline UDFs (if they meet the requirements).
1. You are setting a value like you are showing for `select @testA = id from ...`. The assignment will happen per each and every row returned. And the final value of the variable will be the last row returned. That can be predictable if you use an `ORDER BY`, or you can live life on the edge and go for unpredictable results by not using an `ORDER BY`. For example:

        DECLARE @Temp INT;
        SELECT @Temp = tab.col
        FROM   (VALUES (1), (3), (4), (2)) tab(col);

        SELECT @Temp AS [NotOrdered];
        -- 2

        SELECT @Temp = tab.col
        FROM   (VALUES (1), (3), (4), (2)) tab(col)
        ORDER BY tab.col;

        SELECT @Temp AS [Ordered];
        -- 4

Another way to look at this is that you are using a result set as a loop:

<!-- language: lang-none -->

    -- pseudo-code
    while (ResultSet.GetNextRow())
        set @testA = ResultSet.GetColumnValue("id");

1. If the first call to `ResultSet.GetNextRow()` returns false (due to no rows in the result set), then the loop is never entered and the assignment never happens.
1. If there are multiple rows in the result set, then the assignment happens for each and every row returned, and in the order that they are returned.

In fact, this is how string concatenation works when done via result set (**please note:** as Martin Smith pointed out in this [Stack Overflow answer][1], string concatenation done in this manner is _not_ guaranteed to produce the expected result, even if most of the time it does; however, that does not take away from this method proving my point here):

    DECLARE @List NVARCHAR(MAX) = N'**';
    SELECT @List = @List + N', ' + QUOTENAME([name]) 
    FROM   sys.schemas
    WHERE  1 = 0
    ORDER BY [name];
    SELECT @List AS [NoChange];
    -- **
    SELECT @List = @List + N', ' + QUOTENAME([name]) 
    FROM   sys.schemas
    ORDER BY [name];
    SELECT TRIM(N'*, ' FROM @List) AS [SQL2017orNewer];
    --SELECT SUBSTRING(@List, 5, 80000) AS [PreSQL2017];

I even made use of this behavior to generate a complex T-SQL script:

[Uppercase All String Columns (and in a Single Query!)][2]

  [1]: https://stackoverflow.com/a/15163136/577765
  [2]: https://sqlquantumleap.com/2019/03/20/uppercase-all-characters-and-in-a-single-query/