Regarding "This is just the situation, which the quantum mechanics doesn't cover.", it was a situation described in quantum formalism - which clearly doesn't fulfill energy conservation.
In statistical physics type of effective theories, QM strongly resembles, there can be conservation of expected energy - that energy is most likely conserved.
But a really fundamental theory, in some scale effectively described by QM, should have real ultimate energy conservation - like any Lagrangian mechanics.
Let's go back to the problematic Stark effect for Lyman-gamma (4->1) ...
I have decided to perform the calculations (pdf file) as described here for n=3.
So we need matrix <n,l,m| z^hat|n,l',m'> for fixed n (assuming degeneracy) and all n^2 possibilities for l and m.
Possible shifts are given by eigenvalues of this matrix (times a*E*e).
For n=3 we get eigenvalules: {-9, -9/2, 0, 9/2, 9} - it fits to Frerichs' results assuming we don't see the 0 line.
For n=4 we get eigenvalues: {-18,-12,-6, 0, 6, 12, 18} - visually it seems to fit Frerichs' results assuming we don't see the {-6, +6} lines.
However, he got (10^8/lam): {102630.5, 102684.2, 102823.6, 102964.4, 103021.7},
after subtracting average value we get {-194.38, -140.68, -1.28, 139.52, 196.82}
The proportions suggest that the external lines should be ~140 * 1.5 = 210, so the observed ones are essentially narrower than predicted.
How to repair this discrepancy?
Maybe someone has some more recent experimental results for Lyman-gamma (4->1)?
update:
I have just found 1984 paper ( http://journals.aps.org/pra/ab…/10.1103/PhysRevA.30.2039 ) which starts with "Recent measurements of the Stark profiles of the hydrogen Lyman-alpha and -beta lines in an arc plasma have revealed a sizeable discrepancy between theoretical and experimental results" ...