i don't want to copy my code here because my code is scattered in places and there are a lot of dependencies.
but the environment is this: fb size 1920x1080 , tex size 1024x512 , tex has several regions and each has borders to avoid clearing other regions , setvirtualdimensions(1024,512) is called once for the tex before calling setRenderTarget.
i tested removeing rendering of object3d, found that simply resetting the fov (emptyworld.renderscene) can lead to the situation of needing offset but not scaling. which means World.renderScene() changes the situation. test result is consistant.

there is another problem which we know already: the rendered image is flipped in texture. but it can be counteracted from app level. i just think it's strange.

the question is not about the color , but about the border of texture being cleared , when clearAll option is set to false,
while the java doc says otherwise.

anyway, now i use blitting (from a pure color texture) to replace clearing.

Hi Egon,
i thought of a situation which i am not sure is necessary at all, that temporarily switching the render target.
if that happens, a getRenderTarget method can ease the problem of (the engine user) keeping track which texture is the current target.
like this:
b=getRenderTarget();
setRenderTarget(a);
render something;
setRenderTarget(b);

but if the engine doesn't call setRenderTarget, only the app code call it, then the user can use a global variable to keep track of the target texture. then getRenderTarget is unnecessary.
and temporarily switching render target might be a wrong thing to do.

i found some feature of es3.0 that i like to have, such as occlusion query (a better object3d.wasVisible() function)
is it possible to enable es3.0?

hi Egon,
the java doc says getPixels costs a lot of time (for realtime rendering).
is the time related to the size of framebuffer?
if the framebuffer is of size 1x1, does it still cost that much time or is the time much shorter?
i think of using getpixels to do collision detection, is it possible?
the method is like this, place a camera at a bullet, set farclip to the distance that bullet travels in delta time, read the resulting depthbuffer to know whether/where/when the bullet hits something.
if this is possible, is it better or worse comparing to the ray-polygon collision detection in jpct engine?

one of the best screenshots i have so far. most recent 3D model,(i made dozens of models before this one , as practising)

then i 'd better add a limit at the output of fragment shader.
the output vertexcolor of vertex shader should not matter because it's just another varying value passing from vertex shader to fragment shader.