The screenshot shows that for each meta tile only one of the output elements is visible. My wild guess is that the JAI slicing code is changing the origins of each tile in the meta tile to create it without actually creating a new image, and the jpeg encoder is not aware of it, and it's probably using some sort of direct access.
Shall we go back to the old code that used to build separete BufferedImage objects as tiles?

Simbos, it seems that wit native jai around the following exception occurrs:

29 lug 19:14:52 WARN [geoserver.ows] - 
java.lang.NullPointerException
	at com.sun.media.jai.mlib.MediaLibAccessor.findCompatibleTag(MediaLibAccessor.java:537)
	at com.sun.media.jai.mlib.MlibBandSelectOpImage.computeRect(MlibBandSelectOpImage.java:88)
	at javax.media.jai.PointOpImage.computeTile(PointOpImage.java:969)
	at com.sun.media.jai.util.SunTileScheduler.scheduleTile(SunTileScheduler.java:904)
	at javax.media.jai.OpImage.getTile(OpImage.java:1129)
	at javax.media.jai.RenderedOp.getTile(RenderedOp.java:2257)
	at com.sun.media.imageioimpl.plugins.clib.CLibImageWriter.getContiguousData(CLibImageWriter.java:272)
	at com.sun.media.imageioimpl.plugins.clib.CLibImageWriter.getMediaLibImage(CLibImageWriter.java:712)
	at com.sun.media.imageioimpl.plugins.jpeg.CLibJPEGImageWriter.write(CLibJPEGImageWriter.java:240)
	at org.geotools.image.ImageWorker.writeJPEG(ImageWorker.java:2132)
	at org.vfny.geoserver.wms.responses.map.jpeg.JPEGMapProducer.formatImageOutputStream(JPEGMapProducer.java:69)

Sigh?

Mumble, yet this is happening only on GS 1.6, with 1.7 and the same jdk it seems to be fine...

I applied the fix only on gs-1.7.x

I am back and forward porting it... but yeah, I can just forward port it

Fixed on 1.7.x and trunk

noticed that same thing seems to apply to the imagewriter that comes with jdk

committed on trunk

Simone, you've committed some debugging code that woudl write images in your user home directory. Please remove it? (eventually along the other commented out code?)

sure thing

fix is in