If you can't tell any difference hot or cold then it's probably not worth the hassle. You need to conduct some listening tests. I perceive it as bad imaging and fuzzy soundstage.
I am only just beginning to get into this but looking at the data sheets it seems that on the 94, the SAA7210 does interpolate up to 1 uncorrectable error, and the SAA7220 up to 8 sequential errors (based on EFAB (pin 4) which can be nailed). Grounding that pin should (I imagine) cause an audible "tick". If not a LED, scope or counter can be connected and the errors observed during warm up (or other tweaks), all with the intent of reducing uncorrectable errors (unreliable samples).
On other Phillips ICs I have seen, the error information is a data stream similar to to serial data, with a number of bits positions in the error word indicating more detailed information, but on the SAA7210 it simply shows that the current sample is "reliable" or not.
It could be that if one has bad tempco parts on say your Demod PLL filter it is less effective when cold at maintaining lock or similar. Also that the various adjustments were made when the transport was at operating temperature, and the transport is sub-optimal at other times (or the supplies are not stable etc).
I think the CIRC does not intruduce any artefacts but the interpolation (error concealment) does, and this is the only way the sound can be effected (as from there I go via coax to an offboard DAC, permenantly powered). After all, the concealment/filtering is there to hide itself. This just makes it more difficult to detect by ear.
Perhaps unreliable samples are much more common than we think, that it takes time for the brain to form the soundstage perception, and that process is spoiled by "bad information".
peter