I have done some more tests, now that the line has been stable for a few days. Previous erratic results might have had something to do with Openreach vans being around this area like wasps at a picnic.
If one unplugs one power supply and plugs in another, the modem retrains, normally to a different data rate, making it difficult to compare SNR values. Each retrain gives a new QLN spectrum, but as these are single measurements they may not be repeatable. To get around this, I wired up a DPDT switch with make-before-break action, which allowed me to switch 12V and 0V supply lines without interrupting the operation of the modem.
The results in this post come from a single Netgear modem, with 5 different supply configurations. The supplies were two +12V switchers that came with Netgear and TP-Link modems, a class I supply that I purchased separately, and the same Netgear & TP-Link supplies with an added strap between 0V and mains earth. NOTE: I do NOT recommend the strapping procedure to anyone unfamiliar with the hazards of domestic wiring, and probably not to anyone else either.
@PhilipD: All three supplies were made in China. Not that I have any fondness for such. I do have a pukka British power supply, but it was excused on grounds of age and infirmity. See the attached inspection label.
For the test, I switched supplies every 5 minutes, as plotted in the attached time chart with X axis in minutes. Since a measurement was taken every 2 seconds, each 5 minute section has 150 fairly consistent readings. For both Netgear and TP-Link switching supplies, adding the earth strap improved upstream SNR (labelled "^ SNR") by 1.0 dB, and downstream SNR by 2.0 dB. [Y axis labelling is a division count rather than a dB count]
In the upper graph of the separately attached SNR spectrum, the red and dark green spectra are from the Netgear power supply with and without an earth strap. The lower graph shows the improvement from adding the earth strap, i.e. the difference between the two upper graphs. The spectrum is ADSL2+, with 512 samples.
As far as possible I avoided moving cables during the test, and as cable layout may be a significant factor I do not claim that anyone else will see the same result.