The text from the product user guide states the following:
Following supervised classification of smoothed NBAR data, a set of
post-processing steps that incorporate prior probability knowledge and
adjust specific classes based on ancillary information are applied to
the classification results (McIver and Friedl, 2002; Friedl et al.,
2002). The final class-conditional probabilities have substantial
levels of inter-annual variability caused by residual noise in input
time series, missing data, and changes within the training database
(Friedl et al., 2010). To reduce interannual variability caused by
classifier instability, we developed an approach based on Hidden
Markov Models that post-process map results for each year, which
dramatically reduces inter-annual variability in the product
(Abercrombie and Friedl, 2016). After stabilization, the
classifications are condensed into the final set of six legends and
associated QA information. Despite improving the stability to the
product, we urge users not to use the product to determine
post-classification land cover change. The amount of uncertainty in
the land cover labels for any one year remains too high to distinguish
real change from changes between classes that are spectrally
indistinguishable at the coarse 500-m MODIS resolution. For more
detailed information about the development and accuracy of the C6
MCD12Q1 product see Sulla-Menashe et al.
(Hightlighting done by me).
Based on the above quotation, it would seem that the "Hidden Markov Models" stabilization referenced in the QA layer can fail for some reason. Pixels with such failures are then labelled with code 8 in the QA, and the resulting value in MCD12Q1 is then the pre-stabilization value. I have been unable to find further details on what can cause such failures in the stabilization.
However, I'd personally interpret the code 8 as an indication of a more uncertain classification, since the interannual noise reduction failed. Although it could be interesting to evaluate the correlation between code 8 and actual land cover changes.