How are these null values stored in a NOT NULL column

The solution is dbms_redefinition package. Basically redefition will move, online, to a new table (called interim table), then move all dependent objets like indexes and last exchange the interim table with the original one.

What you have to do is:

• Create the interim table withe correct column definition
• Run redefinition
• Drop the old table

Use dbms_redefinition in the following way:

-- DETERMINE IF THE ORIGINAL TABLE CAN BE REDEFINED ONLINE
BEGIN
DBMS_REDEFINITION.CAN_REDEF_TABLE('SCHEMA','YOURTABLE', DBMS_REDEFINITION.CONS_USE_ROWID);   
END;
/

-- BEGIN THE REDEFINITION
BEGIN
DBMS_REDEFINITION.START_REDEF_TABLE(
UNAME => 'SCHEMA',
ORIG_TABLE => 'YOURTABLE',
INT_TABLE => 'INTERIM_YOURTABLE'
OPTIONS_FLAG => DBMS_REDEFINITION.CONS_USE_ROWID);
END;
/

-- IF THE TABLE HAD DEPENDENCIES (INDEXES ... CONSTRAINTS ... TRIGGERS)
-- THIS WOULD BE THE POINT AT WHICH THEY WOULD HAVE BEEN COPIED
DECLARE 
error_count pls_integer := 0;
BEGIN
DBMS_REDEFINITION.COPY_TABLE_DEPENDENTS('SCHEMA', 'YOURTABLE', 'INTERIM_YOURTABLE', 
dbms_redefinition.cons_orig_params, TRUE, TRUE, TRUE, FALSE, error_count);
DBMS_OUTPUT.PUT_LINE('errors := ' || TO_CHAR(error_count));
END;
/

-- FINISH THE REDEFINITION
exec DBMS_REDEFINITION.FINISH_REDEF_TABLE('SCHEMA','YOURTABLE','INTERIM_YOURTABLE');

If you run redefinition with SYSTEM user there is no problem. If you want to execute it with a less privileged user you have to trick some privileges in order to get it working. Privileges required are:

• Execute privilege to DBMS_REDEFINITION
• Create any table
• Alter any table
• Drop any table
• Lock any table
• Select any table

Tables with the following characteristics cannot be redefined online:

• [9.0.1]Tables with no primary keys
• Tables that have materialized view logs defined on them
• [9i] Tables that are materialized view container tables and AQ tables
• [10g] Tables that are replicated in an n-way master configuration can be redefined, but horizontal subsetting (subset of rows in the table), vertical subsetting (subset of columns in the table), and column transformations are not allowed.
• The overflow table of an IOT table
• Tables with fine-grained access control (row-level security)
• Tables with BFILE columns
• Tables with LONG columns can be redefined online, but those columns must be converted to CLOBS. Also, LONG RAW columns must be converted to BLOBS. Tables with LOB columns are acceptable.
• Tables in the SYS and SYSTEM schema
• Temporary tables

Other restrictions:

• A subset of rows in the table
• Only simple deterministic expressions, sequences, and SYSDATE can be used when mapping the columns in the interim table to those of the original table. For example, subqueries are not allowed.
• If new columns are being added with no column mappings, then they must not be declared NOT NULL until the redefinition is complete.
• There cannot be any referential constraints between the table being redefined and the interim table.
• Table redefinition cannot be done NOLOGGING.
• [10g] For materialized view logs and queue tables, online redefinition is restricted to changes in physical properties.
• You cannot convert a nested table to a VARRAY.

This answer shows an implementation of the techniques mentioned by Michael Green in comments to the question. It uses SQL Server, but the basic idea is applicable to any SQL-compatible database engine that supports dynamic SQL. Support for UNPIVOT and PIVOT is not required.

The broad idea is:

1. Convert rows to key-value pairs
2. Compare with the reference row to identify differences
3. Convert the differences to one row per WERKS using dynamic SQL

It is the use of dynamic SQL that gets around the issue of not knowing which columns to SELECT - they are determined dynamically.

Sample table and data

CREATE TABLE #MARC
(
    c1 integer PRIMARY KEY,
    c2 integer NULL,
    c3 integer NULL,
    c4 integer NULL,
    c5 integer NULL
);

INSERT #MARC
    (c1,c2,c3,c4,c5)
VALUES
    (1, 2, 3, 4, 5), -- The reference row
    (2, 2, 3, 4, 5),
    (3, 2, 3, 4, 5),
    (4, 2, 3, 7, 9), -- Differences: 7 in c4; 9 in c5 
    (5, 6, 3, 4, 8); -- Differences: 6 in c2; 8 in c5

Unpivot and find differences

-- Holds differences found (WERKS keys and column key-value pairs)
CREATE TABLE #Differences
(
    c1 integer NOT NULL,
    name varchar(2) NOT NULL,
    value integer NOT NULL
);

WITH 
    Ref AS
    (
        -- Unpivoted reference row
        SELECT 
            M.c1, 
            CA.name,
            CA.value
        FROM #MARC AS M
        CROSS APPLY 
        (
            VALUES
                ('c2', M.c2),
                ('c3', M.c3),
                ('c4', M.c4),
                ('c5', M.c5)
        ) AS CA (name, value)
        WHERE
            M.c1 = 1
    ), 
    NonRef AS
    (
        -- Unpivoted non-reference rows
        SELECT 
            M.c1, 
            CA.name,
            CA.value
        FROM #MARC AS M
        CROSS APPLY 
        (
            VALUES
                ('c2', M.c2),
                ('c3', M.c3),
                ('c4', M.c4),
                ('c5', M.c5)
        ) AS CA (name, value)
        WHERE
            M.c1 <> 1
    ), 
    Differences AS
    (
        -- Find column value differences
        SELECT 
            NonRef.c1, 
            NonRef.name, 
            NonRef.value
        FROM NonRef
        JOIN Ref
            ON Ref.name = NonRef.name      -- Same column name
            AND Ref.value <> NonRef.value  -- Different value
    )
INSERT #Differences
    (c1, name, value)
SELECT
    D.c1,
    D.name,
    D.value
FROM Differences AS D;

Differences

-- Show differences for illustrative purposes
SELECT
    D.c1,
    D.name,
    D.value
FROM #Differences AS D;

╔════╦══════╦═══════╗
║ c1 ║ name ║ value ║
╠════╬══════╬═══════╣
║  4 ║ c4   ║     7 ║
║  4 ║ c5   ║     9 ║
║  5 ║ c2   ║     6 ║
║  5 ║ c5   ║     8 ║
╚════╩══════╩═══════╝

Dynamic SQL

-- Header
DECLARE @sql varchar(max) =
    'SELECT D.c1' + CHAR(13);

-- Add aggregation
WITH D AS
(
    SELECT DISTINCT D2.name
    FROM #Differences AS D2
)
SELECT @sql +=
    (
        SELECT 
            ', MAX(CASE WHEN D.name = ' + CHAR(39) + D.name + CHAR(39) + 
            ' THEN D.value END) AS ' + QUOTENAME(D.name) + CHAR(13) AS [text()]
        FROM D
        FOR XML PATH (''), TYPE
    ).value('(./text())[1]', 'varchar(max)')

-- Footer    
SET @sql +=
    'FROM #Differences AS D' + CHAR(13) +
    'GROUP BY D.c1;'

The constructed SQL statement in @sql is:

SELECT D.c1
, MAX(CASE WHEN D.name = 'c2' THEN D.value END) AS [c2]
, MAX(CASE WHEN D.name = 'c4' THEN D.value END) AS [c4]
, MAX(CASE WHEN D.name = 'c5' THEN D.value END) AS [c5]
FROM #Differences AS D
GROUP BY D.c1;

That is the manual PIVOT syntax. The final result is obtained by executing the SQL:

EXECUTE (@sql);

Results:

╔════╦══════╦══════╦════╗
║ c1 ║  c2  ║  c4  ║ c5 ║
╠════╬══════╬══════╬════╣
║  4 ║ NULL ║ 7    ║  9 ║
║  5 ║ 6    ║ NULL ║  8 ║
╚════╩══════╩══════╩════╝

This is the minimal set of columns describing the column differences with respect to the reference row, as required.

-- Clean up temporary tables
DROP TABLE 
    #Differences,
    #MARC;

There is some SQL Server-specific syntax above, but the manual UNPIVOT can be achieved in other ways if APPLY or LATERAL is not supported e.g. with a cross join to a suitable identity matrix table. Generating the dynamic SQL uses FOR XML PATH as an efficient way to concatenate the strings, but again this can be done in any language.

Stack Exchange Data Explorer Demo