TableViewMouseInputTest.test_rt_38464_cellSelection has a code comment hinting to a problem with using the test utils:

     // setup: get cell and create a mouse firer on it
     TableCell cell_4_2 = VirtualFlowTestUtils.getCell(table, 4, 2);
     MouseEventFirer firer = new MouseEventFirer(cell_4_2);
     ....
     // while testing cell states
    if (row == 4 && column == 2) {
        // bizarrely cell (4,2), i.e. the bottom-right cell consisting
        // of Michael Brown's email address, doesn't exist.
        continue;
     }

The reason here is that the test relies on the auto-creation of a StageLoader for the table, done both by the test utils and then again by the firer - the latter does so for the isolated cell, implicitly moving it from the table's scenegraph into the new scene.

Fix is to manually add the table to a scene (via a stageloader, f.i.) before (at least) creating a firer.

Probably should take a closer look at similar test patterns and see whether they might effect the test outcome (in either direction). The scenegraph seems stable enough when moved around (as a whole) across multiple scenes, f.i. when repeatedly getting cells from the test utils (didn't find any weirdness, but didn't try too hard ;)

A test demonstrating the mis-/behavior with/out commented code line:

    @Test
    public void testMouseFirerInScene() {
         // get row and cell in that row (doesn't matter which exactly
        TableRow<?> tableRow = (TableRow) VirtualFlowTestUtils.getCell(tableView, 1);
        TableCell<?, ?> tableCell = (TableCell) VirtualFlowTestUtils.getCell(tableView, 1, 0);
        assertSame("sanity: parent of cell is expected row", tableRow, tableCell.getParent());

        // manually install scene to keep table's scenegraph stable
        // stageLoader = new StageLoader(tableView);
        new MouseEventFirer(tableCell).fireMousePressAndRelease(KeyModifier.SHIFT);
        assertSame("parent of cell must be unchanged", tableRow, tableCell.getParent());
    }