D071 - D070's rule was wrong, and the corrected one cannot be tested by avoiding it

D070 concluded that a console refuses an image whose inner filesystem is smaller than the cache its header declares, and had selfish pack clamp the declared value down to fit. The evidence was one A/B: padding a small package's inner image past 0xD0000 cleared the error.

That was the wrong reading, and the clamp does nothing. Built with the declared cache set to exactly the inner size and sent to a console, the mount fails identically:

declared cache	inner image	result
851968	720896	Failed to enable GDDR5 cache, EINVAL
720896	720896	the same, at the same line
851968	1769472	mounts past it
851968	11141120	mounts past it

The declared field does not decide it. The inner filesystem itself has to exceed a fixed size, somewhere in (720896, 1769472] - 0xD0000 is the plausible boundary and is not established; it is only the value the header happens to carry. The padding A/B changed two things at once, the inner size and its relation to the declared cache, and D070 attributed the result to the wrong one.

So the clamp is removed. selfish pack warns instead, and the warning says that lowering the declared size was tried and does not help, because that is the obvious next idea for anyone who hits this. Builder::cache_size stays: the override is harmless and a caller may have a reason, but nothing here uses it, and writing a value no real package carries in exchange for nothing was the mistake.

The part that matters more than the field

This was supposed to be the control for whether selfish's indirect signature block causes a console panic. A package needs payload_blocks > 12 to use one, and a payload block is 0x10000:

   no indirect   needs   inner <=  720896
   mounts at all needs   inner >  ~851968

Those cannot both hold. At this block size every package large enough to mount necessarily uses an indirect block, so "does it still panic without one" is not a question hardware can be asked. The control is impossible, not merely unbuilt.

That is not a dead end for the investigation, because the indirect block was already cleared by inspection - examples/indirect_probe shows ours and a real package's are the same structure, and a real package with 944 blocks mounts. What dies here is the experiment, and knowing an experiment cannot exist is worth recording: the next session will otherwise design it again.