Using a jigsaw nightly build properly working with jtreg v4.1, we performed a full run of the client regression tests on Linux (apprx 5000 tests). While some of the tests are supposed to run on other platforms, we still could use the results as a good estimation for the number of tests using internal API.

Now, total number of tests failing because of modular issues is 417. Of them, 382 essentially require some internal API, to the best of my knowledge. The rest is divided in 3 parts:
(1) 10 tests I can fix myself;
(2) 12 tests should be reevaluated, and this work perhaps has priority;
(3) 13 tests should be eventually fixed by dev team.
I'll create 3 subtasks for that.