Different kind of tablebase usage in top engines?

[Jouni]

[d]8/8/2k5/2r5/8/3RP3/4K3/8 w - - 0 1

Analysis by PlentyChess 7.0.37:
...
1.Kf3 Rf5+ 2.Kg4 Rf8 3.Rd1 Rg8+ 4.Kf5 Re8 5.Re1 Kd5 6.Kf6 Ke4 7.Kg7 Kd3 8.e4
= (0.00) Depth: 82/59 00:00:06 99919kN, tb=1607291

Analysis by Reckless 0.9.0-dev-0dd5b9ac (TB access working):
...
1.Kf3 Rf5+ 2.Kg4 Rf8 3.e4 Re8 4.Rd4 Kc5 5.Rd5+ Kc6
= (0.00) Depth: 50/10 00:00:02 22944kN

Analysis by Stockfish 17.1:
...
1.Kf3 Rf5+ 2.Kg4 Rf8 3.e4 Re8 4.Kf5 Rf8+ 5.Ke5 Re8+ 6.Kf4 Rf8+
= (0.00) Depth: 58/13 00:00:03 10870kN, tb=17

0/17/1.6 million?

[syzygy]

Different ways of using tablebases when the root position is already in the TBs (and is drawn).

[jdart]

Different ways of using tablebases when the root position is already in the TBs (and is drawn).

I believe the Stockfish method is that you filter the moves at the root so that none of them have a worse score than the best tablebase move. But then you disable probing inside the search. So you can get analysis, including multipv analysis if desired, from the engine, but be assured that it won't in the end play a suboptimal move. (Arasan does this, too).