Basically you can comp certain types of shots quite happily in log with no loss if you are not doing anything which requires a linear computational colour space.
if you are using CGI then you need to bring it and your film plate together into ONE colour space to composite them. Most people prefer to do this in Linear space. This will actually require you to apply a lookup (in the case of shake a logLin) to BOTH the CGI and the scan. Doing this at floating point bit depth is the safest way to ensure no data is clipped.
As you are capturing data which is already compressed and colour managed to a broadcast spec (4:2:2 YUV) none of this should really be an issue as you are alrady locked to a video colour space paradigm !
It seems that lot of the confusion on this thread is because people are unable to find good solid information about film compositing practises and that there are a lot of assumptions going on based on terms tha thave not been correctly defined.
I can see that some folks are on top of all this stuff to. for all those who aren't
Log : is a colour encoding format which is based on Kodak's method of compressing the data on a film negative into 10bits of digital information designed to represent a MUCH larger range of Colour values. This is typified by the logarithmic graph used in the encoding of the values. Colors in the toe and head of this graph contain lots of dat but are not easily viewable on conventional display devices. Hence the use of refernce black and white points !
Lin : a colour space in which colour value of a pixel is assumed to have a linear relationship with its brightness. Tape and digital imagery (CGI) are not linear as they also have a gamma correction embeded by default. Linear footage is obtained by negating BOTH logarithmic and video gamma encoding
Flaoting Point /Bit Depth : Log Lin and video have no baring on bit depth whatsoever. Bit depth is a measure of the ammount of data stored perpixel in each colour channel and can be any of the conventional values 8, 10, 12 (for those of you with discreet systems) 16 or Float (32). but can also be other values.