2026-09-02 - (/loop) Bulk port batch 9: the attribute accessors, and three guards that fired

documented   715 needed, 297 missing   ->   715 needed, 267 missing

Fifteen functions, each with its posix_-prefixed twin delegated to it (D475), so thirty gap entries closed.

What went in

The remaining POSIX list is dominated by attribute accessors, and orbistoun already had the machinery: pthread_attr_init allocates an object this crate owns the layout of, and attr_get/attr_set read and write fields at fixed offsets. So the work was mostly declaring new offsets:

• Thread attributes: getguardsize, getinheritsched, getschedpolicy, getscope, setscope.
• Mutex attributes: getpshared/setpshared, getprioceiling/setprioceiling.
• Condition-variable attributes: getclock/setclock, getpshared/setpshared, destroy.
• pthread_equal.

The thread attribute object grew from eight words to sixteen. Safe, and the existing comment says why: "this crate defines the layout, defensible only because nothing else reads it".

pthread_equal answers non-zero for equal. That is the opposite sense to strcmp and to the _np comparison functions, and a caller reading it as a difference gets every answer backwards - which is why it is worth a sentence rather than being left to look obvious.

A note on what these do and do not promise. Storing pshared and prioceiling and reading them back is the whole contract of an attribute object - D272 established exactly that, when a Gettype was reading the guest's stack because Settype wrote nothing. Whether a lock then acts on the value is the lock's business, and scePthreadMutexInit already declares (D474) that it does not read the attribute block. So these are complete as written; the gap is declared where it actually lives.

Three guards fired, and two were my own mistakes

• every_served_name_is_declared caught the whole declaration block missing.
• The kernel's every_implementation_is_also_declared_here_or_says_why_not caught the same for its side. Both exception lists are deliberately named lists, not blanket exemptions, so each entry went in with its argument written beside it.
• pthread_equal turned out to be already declared with a provisional arity of zero. POSIX gives it two thread identifiers; corrected.

Both of my mistakes were the same one: an idempotence guard keyed on a string that already existed for another reason - '"pthread_equal" =>' matched an old declaration, and '"pthread_attr_getguardsize",' matched the registration row rather than the exception list, so the edit silently did nothing and the tests caught it. A guard string has to be unique to the thing being guarded, which is now a note in the loop's tooling list.

State

clippy --tests clean, fmt clean, kernel/posix/libc/fs tests pass, identity scan clean, nothing committed.

Next: 267 left. getsockopt/setsockopt are still deferred and want a per-option table. Then the two structural items, both worth raising rather than just doing: the TitleOwn loader (unbuilt code, no research, unlocks PS5Util's 36k corpus calls) and the obSCEne differential, which is the only check for "implemented but wrong" and has never been run against any of this.