SQL Server 2019

By using db<>fiddle, you agree to license everything you submit by Creative Commons CC0.
select *
into t
from (values 
      ('a', 4	),
      ('a', 7	),
      ('a', 12),
      ('a', 12),
      ('a', 13),
      ('b', 1	),
      ('b', 2	),
      ('b', 3	),
      ('b', 4	),
      ('b', 5	)
     ) v(id, value)
10 rows affected

with ts as (
      select distinct t.id, t.value, dense_rank() over (partition by id order by value) as seqnum
      from t
     ),
     cte as (
      select id, value, value as grp_value, 1 as within_seqnum, seqnum
      from ts
      where seqnum = 1
      union all
      select ts.id, ts.value, 
             (case when ts.value >= cte.grp_value + 3 then ts.value else cte.grp_value end),
             (case when ts.value >= cte.grp_value + 3 then 1 else cte.within_seqnum + 1 end),
             ts.seqnum
      from cte join
           ts
           on ts.id = cte.id and ts.seqnum = cte.seqnum + 1
     )
select *
from cte
where within_seqnum = 1
order by id, value


id
value
grp_value
within_seqnum
seqnum


a
4
4
1
1

a
7
7
1
2

a
12
12
1
3

b
1
1
1
1

b
4
4
1
4


      select distinct t.id, t.value, dense_rank() over (partition by id order by value) as seqnum
      from t


id
value
seqnum


a
4
1

a
7
2

a
12
3

a
13
4

b
1
1

b
2
2

b
3
3

b
4
4

b
5
5