2026-09-07 - The futex and the stack attributes stop being census entries

Two behavioural sections for the sibling emulator's open threading premises, both promoted out of the census into real calls. Documented in docs/decisions/D322 and docs/backlog/024.

What was added

• 031-stackattr (three checks): what a thread attribute set says about a running thread's stack. self-describes runs the sequence three retail titles run - scePthreadSelf, then scePthreadAttrGet on the answer, then the address and size - and records both. address-is-the-base settles which end of the stack the reported address is, by placing the address of one of its own locals against the reported region. fresh-attr-names-no-stack records what a set nothing configured reports.
• 032-syncaddr (four checks): the platform's futex, which nothing has ever called. wait-returns-on-mismatch, wake-releases-a-waiter, compare-width and wake-with-no-waiter. Every wait runs on a worker nobody joins; the main thread only sleeps, reads a state counter and calls the wake.

Five names moved from the census to platform.h, imports.c and the @called-elsewhere block: scePthreadAttrGet, scePthreadAttrGetstackaddr, scePthreadAttrGetstacksize, sceKernelSyncOnAddressWait and sceKernelSyncOnAddressWake. The OBS_SURFACE_SYNC_ON_ADDRESS group held exactly those last two and is gone; the sized variants in ..._ADDRESS2 stay censused. 239 -> 246 checks, 48 -> 50 sections, 39532 -> 39527 symbols.

Every check was made to fail before being believed

make host passes all seven. That is the weaker half of the evidence; each was then broken on purpose:

break	result
the host wake does nothing	wake-releases-a-waiter failed naming the state the worker stopped in; compare-width skipped; the run finished
the host compares 32 bits	compare-width reported 32 and recorded the worker's release
the address is the stack's top	address-is-the-base reported a top
the region contains no frame	address-is-the-base failed, naming the address

The first of those is the one that matters most: it proves a platform whose futex never returns costs one worker thread and not the run, which is the entire safety argument for calling a blocking function at all (D322).

Surprises

• A C toolchain was reachable after all, through Docker. Worklog 158 recorded that no WSL distribution and no native compiler existed here, so the two sections it added were never compiled. The only WSL distribution is docker-desktop, and a gcc:13 container with /oops mounted builds and runs the host suite (make host CC=gcc, since the Makefile pins clang). 016-syncbounds and 019-posixerr have now been compiled and run for the first time as a side effect, and both pass. That is worth knowing before the next session assumes it cannot build.
• Adding two sections broke the build, and the right thing caught it. OBS_SCREEN_MAX was 48 against a registry of 50, and D259's compile-time assertion in registry.c failed the build naming itself. Raised to 56 - what two columns hold by screen.c's own note, not arbitrary headroom.
• The attribute family had no host stubs at all, so 010-kernel/thread-attributes has been skipping since it was written. It now runs and passes. A skip on the host build is not necessarily inert: it can be a check nobody has ever validated.
• The guards gate caught a real violation, exactly the D058 shape: compare-width announces the wait and calls the wake to release a worker parked by a 32-bit comparison. Guarded.
• The hardware reports name scePthreadAttrGet as present, and the current census does not contain it. Those OBS|sym records come from an older generation - ps5-full-run.txt says OBS|meta|1|28|521 against today's 50 sections. A presence record in a report is a fact about the census that produced it, not about the census now.

What is owed

The hardware run. Both sections are assumed/derived until a console produces the records docs/backlog/024 lists, and the futex's third register - a timeout in an unestablished unit - is deliberately not declared and not probed.