Custom aggregate last_value_when

At work I've found the following custom aggregate function very
helpful.  But I would like to extend it and I'm not sure how to do
this.

The function is last_value_when.  When you call last_value_when(value,
cond) over w it gives you the last value in the window where cond was
true.

Here is my implementation.

-----

CREATE OR REPLACE FUNCTION last_value_when_sfunc(state anyelement,
value anyelement, cond BOOL) RETURNS anyelement AS
$$
    SELECT CASE WHEN cond THEN value ELSE state END;
$$
LANGUAGE SQL;

COMMENT ON FUNCTION .last_value_when_sfunc (anyelement, anyelement,
bool) IS 'Helper function for tracking last matching in window';

CREATE AGGREGATE last_value_when(anyelement, bool) (
    SFUNC = expression.last_value_when_sfunc,
    STYPE = anyelement);

COMMENT ON AGGREGATE last_value_when (anyelement, bool) IS 'Aggregate
function for tracking the last value when a condition was true';

-----

Here is an example of its use.

SELECT timestamp,
  , measurement_type
  , value
  , timestamp
  , last_value_when(timestamp, measurement_type = 'foo') over w as
last_foo_timestamp
  , last_value_when(value, measurement_type = 'foo')  over w as last_foo_value
FROM some_table
WINDOW w AS (
    ORDER BY timestamp,
      CASE WHEN measurement_type = 'foo' THEN 0 ELSE 1 END
    ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW
)

This gives me for every row the last time/value for foo, no matter how
many rows back it might have been.

For working with time series I've found this incredibly powerful.  In
fact I think that it is worth including in Postgres!

But here is my problem.  I would like to build equivalents of lead/lag
with the same idea.  That is where I currently do lag(some_column, 2)
I'd like to instead be able to say lag_when(some_column, 2,
some_condition).  So I could pick out not just the last measurement of
foo, but the last 3 measurements of foo.

How would I do that?