Then how does one get higher scrypt hash rate by increasing mem clocks, and why does GPU-z clearly show the allocation of 512 MB when cgminer starts?

Perhaps, at a particular combination of core and mem clocks. The point is that to maximize scrypt hashing speed, one is forced to underclock the core down to 50 - 60% of mem.

At least, how can I tell each cgminer instance to use only X shaders out of 1536 ? Should I specify "shaders" parameter and not specify "thread-concurrency" ?

I can't stop thinking that SHA2 and scrypt need different kinds of resources and that probably a careful merge of scrypt.cl and one of the sha2 kernels would yield interesting results.

SHA2 needs fast core and little ram with dataset probably fitting the Radeons L2 cache, while scrypt needs fast and ample ram with underclocked cores with a particular core/ram speed ratio, being it often 0.57 of RAM speed or even 0.5 on my 7870 XT.

Now, I used bitminter for bitcoins and cgminer for litecoins concurrently, adjusting intensity on the latter up and down until I had the perception that I was gaining something even with applications that are blindly competing for GPU resources. How could one gain with custom kernel doing both sha2 and scrypt simultaneously?

More directly to the point, with an example. Let's say that your 7970 yields 700 Mh/s or 700 Kh/s when dedicated to mine one coin. I wonder if it could do, for example 450 Mh/s AND 450 Kh/s, mining 2 coins simultaneously...