Instead of loudspeaker, it is better idea to try the amp out with lower cost power resistors. Luckily, in my spare parts room, I've found some of those thick 30W resistors with a nice golden colored casing, I've combined them together to 4.7 Ohm and 9.4 Ohm load and powered the amp on. I was still using Rigol DP832A bench supply at +/-30V.
Nothing has burned.
Good, now I could continue analysis. I got max 50Vpp output before clipping started. Meaning that minimum 30 - 50/2 = 5V was needed for each Vce to keep working. I wonder if this can be reduced through increased biasing. EDIT: I've tried it. No, it can not.
Corresponding RMS value was 17.5V. On 4.7Ohm load, this means 65W output power, or 3.7 A rms output current. Having in mind that power supply was feeding 52W on each side, the efficiency is bit above 60%, similar to what was estimated earlier.
Each transistor dissipates cca 20W and this made my heatsink pretty warm really quick. I found some 2.7K/W heatsink in the workshop, but apparently it wasn't good enough. A colleague told me it's 2.7K/W per meter, while I had cca. 10 cm in length. Oh, well, going to find me a better heatsink.
Spectral performance was somewhat worse than expected. I could see some noise on peaks of the sine waveform, and spectrum showed some harmonics. I wonder if this was because of purely resistive load or sign of some design downfall. Further experiments will show if there THD as low as 0.001% is possible.