Oracle Self-Join on multiple possible column match

I have a query requirement from ----. Trying to solve it with CONNECT BY, but can't seem to get the results I need.

Table (simplified):

create table CSS.USER_DESC (
    USER_ID          VARCHAR2(30) not null,
    NEW_USER_ID      VARCHAR2(30),
    GLOBAL_HR_ID     CHAR(8)
)

-- USER_ID       is the primary key
-- NEW_USER_ID   is a self-referencing key
-- GLOBAL_HR_ID  is an ID field from another system

There are two sources of user data (datafeeds)... I have to watch for mistakes in either of them when updating information.

Scenarios:

A user is given a new User ID... The old record is set accordingly and deactivated (typically a rename for contractors who become fulltime)
A user leaves and returns sometime later. HR fails to send us the old user ID so we can connect the accounts.
The system screwed up and didn't set the new User ID on the old record.
The data can be bad in a hundred other ways

I need to know the following are the same user, and I can't rely on name or other fields... they differ among matching records:

ROOTUSER    NUMROOTS   NODELEVEL  ISLEAF    USER_ID    NEW_USER_ID  GLOBAL_HR_ID   USERTYPE      LAST_NAME         FIRST_NAME  
-----------------------------------------------------------------------------------------------------------------------------
EX0T1100    2          1          0         EX0T1100   EX000005                    CONTRACTOR    VON DER HAAVEN    VERONICA    
EX0T1100    2          2          1         EX000005                00126121       EMPLOYEE      HAAVEN, VON DER   VERONICA    
GL110456    1          1          1         GL110456                00126121       EMPLOYEE      VONDERHAAVEN      VERONICA

EXOT1100 and EX000005 are connected properly by the NEW_USER_ID field. The rename occurred before there were global HR IDs, so EX0T1100 doesn't have one. EX000005 was given a new user ID, 'GL110456', and the two are only connected by having the same global HR ID.

Cleaning up the data isn't an option.

The query so far:

select connect_by_root cud.user_id RootUser, 
       count(connect_by_root cud.user_id) over (partition by connect_by_root cud.user_id) NumRoots, 
       level NodeLevel, connect_by_isleaf IsLeaf, --connect_by_iscycle IsCycle,
       cud.user_id, cud.new_user_id, cud.global_hr_id,
       cud.user_type_code UserType, ccud.last_name, cud.first_name
from   css.user_desc cud
where  cud.user_id in ('EX000005','EX0T1100','GL110456')
-- Using this so I don't get sub-users in my list of root users...
-- It complicates the matches with GLOBAL_HR_ID, however
start with cud.user_id not in (select cudsub.new_user_id 
                               from   css.user_desc cudsub 
                               where  cudsub.new_user_id is not null)
connect by nocycle (prior new_user_id = user_id);

I've tried various CONNECT BY clauses, but none of them are quite right:

-- As a multiple CONNECT BY
connect by nocycle (prior global_hr_id = global_hr_id)
connect by nocycle (prior new_user_id  = user_id)

-- As a compound CONNECT BY
connect by nocycle ((prior new_user_id = user_id)
                     or (prior global_hr_id = global_hr_id 
                         and user_id != prior user_Id))

UNIONing two CONNECT BY queries doesn't work... I don't get the leveling.

Here is what I would like to see... I'm okay with a resultset that I have to distinct and use as a subquery. I'm also okay with any of the three user IDs in the ROOTUSER column... I just need to know they're the same users.

ROOTUSER    NUMROOTS   NODELEVEL  ISLEAF    USER_ID    NEW_USER_ID  GLOBAL_HR_ID   USERTYPE      LAST_NAME         FIRST_NAME  
-----------------------------------------------------------------------------------------------------------------------------
EX0T1100    3          1          0         EX0T1100   EX000005                    CONTRACTOR    VON DER HAAVEN    VERONICA    
EX0T1100    3          2          1         EX000005                00126121       EMPLOYEE      HAAVEN, VON DER   VERONICA    
EX0T1100    3         (2 or 3)    1         GL110456                00126121       EMPLOYEE      VONDERHAAVEN      VERONICA

Ideas?

Update

Nicholas, your code looks very much like the right track... at the moment, the lead(user_id) over (partition by global_hr_id) gets false hits when the global_hr_id is null. For example:

USER_ID   NEW_USER_ID   CHAINNEWUSER   GLOBAL_HR_ID   LAST_NAME   FIRST_NAME
FP004468                FP004469                      AARON       TIMOTHY
FP004469                                              FOONG       KOK WAH

I've often wanted to treat nulls as separate records in a partition, but I've never found a way to make ignore nulls work. This did what I wanted:

decode(global_hr_id,null,null,lead(cud.user_id ignore nulls) over (partition by global_hr_id order by user_id)

... but there's got to be a better way. I haven't been able to get the query to finish yet on the full-blown user data (about 40,000 users). Both global_hr_id and new_user_id are indexed.

Update

The query returns after about 750 seconds... long, but manageable. It returns 93k records, because I don't have a good way of filtering level 2 hits out of the root - you have start with global_hr_id is null, but unfortunately, that isn't always the case. I'll have to think some more about how to filter those out.

I've tried adding more complex start with clauses before, but I find that separately, they run < 1 second... together, they take 90 minutes >.<

Thanks again for you help... plodding away at this.

You have provided sample of data for only one user. Would be better to have a little bit more. Anyway, lets look at something like this.

SQL> with user_desc(USER_ID, NEW_USER_ID,  GLOBAL_HR_ID)as(
  2    select 'EX0T1100',  'EX000005',  null          from dual union all
  3    select 'EX000005',   null,       00126121      from dual union all
  4    select 'GL110456',   null,       00126121      from dual
  5  )
  6  select connect_by_root(user_id) rootuser
  7       , count(connect_by_root(user_id)) over(partition by connect_by_root(user_id)) numroot
  8       , level nodlevel
  9       , connect_by_isleaf
 10       , user_id
 11       , new_user_id
 12       , global_hr_id
 13    from (select user_id
 14               , coalesce(new_user_id, usr) new_user_id1
 15               , new_user_id
 16               , global_hr_id
 17            from ( select user_id
 18                        , new_user_id
 19                        , global_hr_id
 20                        , decode(global_hr_id,null,null,lead(user_id) over (partition by global_hr_id order by user_id)) usr
 21                    from user_desc
 22                 )
 23         )
 24  start with global_hr_id is null
 25  connect by prior new_user_id1 = user_id
 26  ;

Result:

ROOTUSER    NUMROOT   NODLEVEL CONNECT_BY_ISLEAF USER_ID  NEW_USER_ID GLOBAL_HR_ID
-------- ---------- ---------- ----------------- -------- ----------- ------------
EX0T1100          3          1                 0 EX0T1100 EX000005    
EX0T1100          3          2                 0 EX000005                   126121
EX0T1100          3          3                 1 GL110456                   126121