2026-09-03 - poisoning a scalar out-parameter, and a list of what orbistoun cannot decide

Two changes, both from the consuming side. orbistoun read this project's report and found twenty-four measurements that measured nothing.

The probe was reading back its own initialiser

106-encoder/path-probe reports res after each sceKernelLoadStartModule, and res starts at zero. Every one of the twenty-four came back 0x0 - which is what a platform that never touches the out-parameter produces, and equally what one that writes zero produces.

Now (int)0xC7C7C7C7u, the pattern byte obs_layout_patterns already uses. Untouched is a visible answer. D303 has the argument and the residual.

The general shape is the interesting part: obs_report_written has always made this case for buffers, and nothing made it for scalars. Swept into the backlog rather than fixed one site at a time.

And a backlog entry written from the other side

docs/backlog/022-measurements-orbistoun-is-blocked-on.md. Four entries, each with what would settle it:

• scePthreadSetprio / scePthreadSetaffinity - census-only, no arity. The unblock is a lawful reference confirming the FreeBSD analogues, not a console run.
• What sceKernelLoadStartModule does to a module already placed - needs hardware, and the run that settles it is loading a module this project built and reporting whether anything of its own ran. "Nothing ran" is as useful an answer as a name.
• The width of a vendor error return - every probe reports (uint64_t)(uint32_t)x, so the upper thirty-two bits are the cast. One assembly thunk reporting rax verbatim retires a caveat repeated at every comparison site.
• The out-parameter sweep above.

Written here rather than in orbistoun because the side that owns a console is the side that can close them, and a gap nobody wrote down is one the next capture will not close.

And the decision gate has been failing, checking nothing

Found while reading this repository's own report. obscene-tool decisions built the index for a single-file DECISIONS.md; the log was split into one file per entry long ago and the subcommand was never repointed, so it has been printing no entries found and exiting 1. A gate in verify.sh that checks nothing, for weeks.

Rewritten to check what is actually here - every entry has exactly one row, every row a real file, no number claimed twice - and to refuse an empty directory, because a gate with nothing to check has not passed. D314.

Watched failing before being believed, and both failures were worth having. Hiding one entry gives D303-...md: no row, exit 1; restoring gives consistent, exit 0. The first parser matched the bare string decisions/, which the front matter uses to describe the layout, and ran to the next ) - a "link" six rows long. The unit test written for exactly that caught it.

Numbered 314 rather than 304: worklogs 136-145 cite the ten numbers from 304 upward in their titles with no files behind them, and writing into that range would give those citations a wrong target. doccheck has been reporting the gap all along.

And a second gate looking in only one place

doccheck reported src/porthole/README.md: make elf - no such rule. It is rule 61 of src/porthole/Makefile, and the README names the directory two lines above. The gate resolved every rule against the root Makefile alone, so a document beside its own Makefile could not name its own rules. Now resolved against both. D315.

Same shape a third time: a negative from a gate that only looked in one place is a fact about the gate. Watched failing - a fake rule appended to that README is still caught.

Eight rows in the index had no title at all

Visible the moment the gate was repointed and the index regenerated:

| 🟢 | D302 | [](decisions/D302-the-conformance-matrix-and-its-naming.md) | decided | 2026-09-03 |

The splitter reads a title from # D<NNN> - Title and nothing else. Nine entries had been written # D302: Title, so nine rows rendered as a bare link with no text - an index of unlabelled links, which is the one thing an index must not be. Headings normalised, index regenerated, eight empty titles to zero, and the gate now refuses an untitled row so it cannot come back quietly.

State

Host build clean: exit=0, zero errors or warnings under -Werror. Identity scan clean. Tool builds, 5 gate tests pass, and the gate reports 302 entries consistent.

clippy on tool/ still has 5 pre-existing errors, all in main.rs (redundant_closure, four collapsible_if) and 7 warnings. None are in the rewritten file, and they are the standing question with the user rather than this unit of work.

Nothing committed. The day holds D302, D303, D314, D315, worklogs 153-154, the matrix subcommand, the rewritten decisions gate, the doccheck fix, the MODULE-FORMAT.md note, and backlog 022.