D597 - Nothing ever wrote the command, and a second summary read as none
Status: measured Date: 2026-09-08
The storage was never touched, which is different from being zero
The command buffer's storage has read as zero in every run since it
became readable, and every reading of that has assumed the guest wrote
zeros or wrote elsewhere. ORBISTOUN_DIRECT_FILL settles it
in one boot:
what it points at, 0x740009100000: 0xd1d1d1d1d1d1d1d1 …
non-zero in it: +0x0:0xd1d1… +0x8:0xd1d1… … (sixteen words, all the fill byte)
Nothing wrote a byte of it. The fakelib sets a storage pointer, sets a size, increments a command count and sets a flag - and encodes nothing at all. A header that claims one command over storage that was never touched is not a partially built buffer; it is bookkeeping with no payload behind it.
The same run shows the buffer object carrying the fill byte
at +0x28 and +0x38, so parts of it are
uninitialised too.
This is the diagnostic doing what it was built for (D325): a field nobody filled in and a deliberate zero are currently indistinguishable, and here the distinction was the finding.
A second summary that read as nothing
D595 found 0 KiB standing for four hundred and two
bytes. Auditing the report for the same shape found one more:
standing 467528 of 467537 calls answered by an implementation (0% on stubs)
Nine calls land on stubs, and those nine are the
only calls in the run worth looking at. 0% on stubs reads
as nothing is stubbed. The share is now printed beside the
count rather than instead of it.
Both are integer division producing a true number that reads as a different claim, and both were in the block a reader looks at first. Two found; the audit covered every division and percentage in the report and found no third.
The ninth
Box::leak, and why it is not fixed here
orbistoun_fs::open::next_handle hands file handles from
the host heap, so a FILE * is a different address every run
- the leak D584 fixed at eight sites in orbistoun-kernel,
in a ninth nobody looked at because the search was scoped to one
crate.
Fixing it needs a decision rather than an edit.
orbistoun-fs depends on neither
orbistoun-kernel nor orbistoun-mem, so the
shared allocator is out of reach, and the options are to move
blocks down the spine, to add a dependency, or to install
it through a hook as the asynchronous file path's reader is.
Which crate owns guest-visible blocks is a structural
question, and answering it by whichever import is easiest to
add is how a spine stops meaning anything. Recorded, and left for a
deliberate choice.
What this does not establish
Why the encoder writes nothing. It is called, it returns success, and the count is incremented on that success - so the fakelib believes it encoded something. Where it believes it put it is the remaining unknown on this wall.
Nor that the fill byte proves the storage is the right buffer. It proves nothing wrote that region. A guest writing commands somewhere else entirely would produce the same evidence, and that reading is still open.