I don't think this is intended (edit: legendary is bugged and giving huge amounts of fury) but I've cleaned up the rotation/priority some and managed to stay in Metamorphosis for over 6m from a pull starting at 200 fury. I didn't let Meta drop at 6m, I just made it all the way back up to 1000 fury and was tired of dpsing the dummy. My play was sloppy but not too bad.
The key is a priority on fury generation: Auto attack, Corruption, Felguard, Hand of Gul'dan, with Wild Imps and Grimoire: Felguard helping to boost all the way back up when they come off cd. Dropping or delaying Shadowflame/Soulfire and using only Demonic Slash when below ~200 fury seems to be enough to stay in Meta since their Damage Per Execution Time/fury is lower than Slash's. Delaying Bane of Doom by a few seconds is a good idea too even though its DPET/fury is huge, it probably isn't worth it to fall out of Meta. Holding back Felstorm would work too.
Since a single Hand of Gul'dan is only 3/4 the damage of our filler it might be best to let it stack back up in times of very high fury.
Spell Fury Cost Execution Time Damage DPET DPET/fury
Bane of Doom 90 1.3 333264 191531 2128
Demonic Slash 40 1.3 19500 11206 280
Soul Fire 110 1.74 47000 27011 246
Shadowflame 100 1.3 34650 19914 199
Felstorm 24 1 7218 4148 172
Hand of Gul'dan
+25
1.3
11450
8304
^Note that all the tests above were in my 402 ilvl gear reforged for Mastery, with about 15% haste, and a Tarecgosa. I'm not sure how much this has to do with it, and I plan to run another test on the premade.
-Soul Fire's fury cost has an added 10, 5 for each auto attack skipped.
-Felstorm has a cost of 60 energy, the same cost as Legion Strike which grants 24 fury. Felstorm stops melee attacks too and single target makes it only a small dps gain. I need to test Voidwalker again, he does similar damage and superior fury generation. Would be fun to use a couple of them on single target fights since I haven't seen mine in a couple years.
This video was made before I came to a lot of these conclusions and doesn't reflect the best methods, but it does show most of it.