2026-09-10 - A guard that covered one of three, and five copies of one table

REQ-20260910T0015Z-8e44, the documentation audit that was not the CLI vocabulary. Most of it was prose catching up with code, and two items were real.

The guard was a third of what its own docs claimed

./bin/selfish links exists for one reason, stated in docs/BUILDING.md: prove the linker tests ran rather than skipping for a missing ld.lld, because a skip would pass having verified none of the segment layout or tag conventions.

It ran cargo test -p selfish-elf --test links and nothing else. Three test targets link a module against link/module.ld - links, current and module - so two of the three could skip silently under the very verb that exists to stop that.

Now a loop over all three, failing on the first skip and naming which target. Verified by running it here, where ld.lld is absent: it stops at links, says so, exits 1. The guard refusing is the guard working.

The prose was wrong in the other direction too - "two integration tests" where there are three files - and per the convention the fix is to remove the number rather than write today's. Both copies now describe the shape: the integration tests that link.

Five copies of one table, two already drifted

The category table - applicationCategoryType, the DMEM budget and display ownership - existed in writing.md, this repo's GLOSSARY.md, the meta glossary, param.rs and the CLI. The two prose copies now point at the collection's four-axes table and at selfish_title::category, which is what code actually reads.

Kept in place at both sites: the one consequence people hit, which is that a GUI under system-app gets zero direct memory and sceVideoOutOpen fails 0x80290001 - the app starts and cannot draw. A cross-link that drops the thing somebody needed at the point they needed it is not a simplification.

The "oracle, never a source" rule was stated four times; README.md already deferred to CLAUDE.md, and the other two now say which statement is canonical rather than restating it independently.

What was deliberately not touched, and why that mattered

The request opened with the exclusions, which was the right order. Confirmed untouched:

• SCE_*_LEGACY in crates/selfish-elf - 13 references, all intact. These are dynamic tag numbers a loader matches on, and dynamic.rs pairs each with a *_CURRENT sibling, so a half-rename breaks Tags::detect - which this repository's own CLAUDE.md records as having been wrong once already.
• data/*.tsv, pkg-keys.toml, sdk-versions.toml - no changes at all. Quoted terms inside provenance citations; OpenOrbis-PS4-Toolchain is a project name, not vocabulary.
• The generated indexes, which are rebuilt from per-entry files.

git diff --stat crates/ data/ was empty at the end, which is the check rather than the intention.

The shape of it

Three of the audit's findings were prose describing behaviour that never existed - extract --out, positional symbol names for sections, --all as a filter when it is a truncation. Each would send a reader to an error rather than to an out-of-date flag, which is the worse failure: a stale name looks stale, and a confident wrong description does not.