D452 - scePthreadGetthreadid answers the thread's handle, and PPSA21564 reaches 0% stubs

measured - 2026-09-01 (user-directed /loop)

With the mspace family (D451) letting PPSA21564 boot, its hot loop asked scePthreadGetthreadid 22.5k times, and unimplemented it answered the placeholder - so every thread reported one shared id, which a guest keying anything on thread identity reads as a single thread.

The representation choice. scePthreadGetthreadid (FreeBSD pthread_getthreadid_np) answers a per-thread integer id. It is implemented in orbistoun-kernel as thread::adopt("main") - the calling thread's registry handle, the same value scePthreadSelf answers. Two reasons this is right rather than lazy:

• The handle is already a u64 unique to the thread and stable for its life (NO_THREAD is 0), so it is a valid unique id. Adopting the caller gives the process's first thread - which runs guest code without ever being created - a real id rather than zero, exactly as scePthreadSelf does.
• Returning the handle rather than a small integer follows the deliberate D151 choice: a handle is the address of a real zeroed block, so a guest that treats the id as something to dereference lands on safe memory, where a small integer reproduced a read of 0x5-class fault.

The cost is that scePthreadGetthreadid and scePthreadSelf return the same value; a guest that expects the ScePthread-handle and thread-id namespaces to differ is not modelled. Recorded assumed, since the id namespace and width are inferred from the name, not verified against the target library.

Milestone. PPSA21564 now runs at 0% stubs - 499082 of 499087 calls answered by an implementation, the five remaining being singletons (_ZSt14_Random_devicev, _init_env, sceKernelGetGPI, pthread_key_create, called once or twice each). It runs to the call budget with no fault, repeatably. A title that faulted at 131 calls a session ago now boots and runs its main loop with every hot-path import satisfied; getting further now needs the budget raised or whatever the loop waits on (video/input) wired, not another missing function.

fmt/clippy/cargo test/the knowledge audit pass for the additions (the kernel crate carries unrelated pre-existing fmt drift from earlier-session work, deliberately left un-reformatted; none of it is in these additions). The change is additive - one new symbol - so other titles are unaffected.