I've done a bit more Read-noise and Dynamic Range number crunching based on some more files, and at different ISOs. The conclusion is that the 16-bit lossless mode does add DR compared to the 14-bit 'lossless' mode, but it remains a bit below the claimed 15 stops. Also, there is no real difference in readnoise between nominal ISO 50 and ISO 100. Read noise for ISO 3200 and above remains unchanged, so becomes basically ISO-less.
Attached are the summarized 14-bit mode shot results of the 'Roundtower' sequence. The numerical differences between the R/G/B/G2 color planes are marginal, they all show very similar values.
I've used RawDigger Version 1.2.4 Build 437 (which is supposed to read the 16-bit mode correctly) to read the files. The DR of 13.65 stops at nominal ISO 50 in 14-bit mode, change to a DR of 14.39 stops in 16-bit mode (not quite the claimed 15-bits but still very high).
Read noise drops from a standard deviation of 5.03 DN to 3 DN, at ISO 50 (and I assume at ISO 100, but I have no file to verify yet). Base ISO therefore seems to be nominal ISO 100. We'll see when DxOlabs publish their analysis how they rate the real ISO sensitivity.
At ISO 800 the DR changes from 11.66 stops in 14-bit mode, to 12.06 stops in 16-bit mode.
The observed 4-DN combing of the histograms in 14-bit mode, indeed disappears in 16-bit mode, so the electronics do seem to produce 16-bit native signal levels, with a noise floor in the 1.61 LSBs (least significant bits).
Low ISO read-noise is very low, and DR is impressive, assuming these files are representative for a larger population of sensors.
Cheers,
Bart