[glub]

One last report (I'll give up after this one, promise):
The leak doesn't happen when the convar sv_alltalk is set to 1, no matter which plugins are loaded. Whenever it is set back to 0, the memory leak happens.

I don't know what differs from the way Basecomm and Sm_deadalltalk plugins handle voice, but I thought it'd be good to know that detail.
Also, sm_deadalltalk plugin seems to cause less memory leak than Basecomm, the increase is not as steep with it loaded for some reason...

Another thing: I noticed that Basecomm (and deadalltalk both) tend to break voice hooks randomly after a while, no clear indication as to what causes it. Sometimes they revert back to being fixed for a few minutes (and opposite teams can't hear each other again) and then break again. Really confusing. :(
Changing sm_deadtalk does not fix anything at all.

It's also interesting to note that when using voice_enable 0, then 1 (client side), the console outputs many
"ProcessVoiceData: Voice_AssignChannel failed for client XX". Could be related, or maybe not....

Here is what I have tested to get to this conclusion:
sv_alltalk is active, 2 players broadcast from opposite teams, 20 bots added to increase memory leak pace; Tested while being both alive, and dead (spectating).
-No SM or SM with no addon loaded = no leak (same things, I know)
-SM with only Basecomm = no leak.
-SM with baseplugins & sm_deadalltalk = no leak.
-unloaded sm_deadalltalk and loaded Basecomm = no leak.

-sv_alltalk now set to 0; sm_deadtalk convar reset from 0 to 1: leak!
-sv_alltalk now reset to 1: no more leak!


Honestly, I don't expect a fix for this issue. Neotokyo is now too much of a niche game to justify spending time and effort on figuring the problem out. Also, we will never have its source code anyway, and it must be too difficult to debug the binaries. Just wanted to testify for anyone who might find this useful.
Waiting for Source2 now. :)