Looks like the 128k default block size is pretty well chosen:
it saves ~6% of image size compared to 64k, and subsequent
differences are ~3% per doubling the block size up to 1M
(thanks led@ for carrying out the tests).

So we'll stick with 256k for "normal" xz compression (inodes
uncompressed) and get 512k back for "tight" one (compressed).
The runtime performance issues are to be examined yet when
bootchart or the like is deployed, nothing drastic though.

With "fast" (gzip/lzo) squash compression inodes go unmolested.

For the record, tight live-webkiosk builds as 95M image in 3:40,
and tight live-flightgear.iso builds as 669M image in 6:34.  Nice.

There's no much sense going for 1M block size: e.g. live-webkiosk
would drop to 93M (3:46) but its load time would increase up to
2:07 as compared to 1:48 for -b 524288 and 1:42 for -b 262144 -noI
on a Duron 500/512M system given the very same DVD+RW media.