This issue seems similar to #94290 but is applicable even when we don't use the leaf txn in the local only flow (as of #94399). Consider the following setup:

CREATE TABLE data (a INT PRIMARY KEY, b INT);
INSERT INTO data SELECT i, i FROM generate_series(0, 9) AS g(i);
ALTER TABLE data SPLIT AT SELECT i FROM generate_series(1, 9) AS g(i);
SELECT * FROM data FOR UPDATE;
SELECT * FROM data FOR UPDATE;

The second SELECT FOR UPDATE takes much longer (about 500ms vs 3ms for the first run). (Note that I think that the delay is about 50ms x number of ranges involved - 1.)

Jira issue: CRDB-22874