How much stability ADSL2+ needed to lower SNR Margin?

[#1]

Question: How much stability is needed for the ADSL2+ system to return the downstream SNR Margin back to 3dB from 6dB?
Background: A couple of months ago interleaving was turned off at my request. Up to that point the downstream SNR noise margin had been running at 3dB with a speedy 18000+ kbps sync speed. After 10 days continuous connection with interleaving off (still at 3dB/18000+kbps) the downstream noise margin climbed to 6dB, dropping the downstream sync speed to 16000+kbps (this isn’t a problem by the way!).
After a bit (more) tinkering with my setup, I have a nice stable setup which has maintained an unbroken DSL connection of almost 32 days and counting!  For the record, the TG582N router I'm currently using instead of my Linksys is working great!
The error stats from the router from almost 32 days (and 60+Gb of traffic) are:
Error Seconds (Local/Remote): 73,130/ 0
FEC Errors (Up/Down): 0 / 0
CRC Errors (Up/Down): 472 / 510,399
HEC Errors (Up/Down): 333 / 334,009
PN ticket support say 'that the line is targeted to be at 6dB'. Does this mean that it is 'locked' at this SNR Margin or could the system ever decide automatically to bring the downstream SNR margin down to 3dB again? As I say, it is not particularly an issue, just a question born out of curiosity.
Thanks in advance for your words of wisdom.

[#2]

According to one of the powerpoint documents within DLMWorkshop_291009_material.zip, a key figure is the mean time between errors (MTBE) - with an MTBE above 600 seconds needed for the DLM to make changes towards greater bandwidth.
I don't think the 32 days constant sync maintained matters very much to the DLM - I think the DLM just records a max value of 86400 seconds for the mean time between retrains (MTBR) for each day that has no resyncs. I think the DLM collects stats for each day.
If I attempt to calculate your MTBE from those stats, I get:
(number of error free seconds) / (number of error free intervals)
((32*24*60*60)-73130)/(73130+1) = 36.8
This seems very low, it's lower than the red quality threshold of 60 for the DLM to consider increasing the target SNR margin. Either I've calculated it wrong again, or the stat counters were not reset at the start of the 32 days. You may be able to get some error counters for the past 24 hours via telnet. An MTBE of 600 should be about 142 errored seconds per day.
I don't know how many days of red or green line quality are required for the DLM to actually make a change, there are other factors besides the MTBE and MTBR, and the DLM has a long memory, so it will have stored how the line performed when running at 3db, and factor that in, if it performed badly in terms of error rates, it'll be disinclined to try it again.

[#3]

Wow! Thanks for the detailed response. It all now makes sense and telnet-ing into the router suggests that there are too many error seconds per 24 hours to make the line green. My calculations suggest it falls within the 'amber' banding, hence no changes.
Thanks again for the info.