474. The target is not the caller

2026-09-09 - directed, continuing 473

REQ-20260909T1445Z-2a90 came back with the branch the previous sweep could not reach - a raise with the leg's own scePthreadSelf() handle - and settled the whole contract in one go.

What was implemented

Delivery is synchronous. The probe's handler sets a flag; the flag reads 1 on the line after raise returns. So the handler is called here and now, nested on the calling thread, and orbistoun does it through the same call_guest path execute_once has used for call_once initialisers. The handler gets the signal number in rdi and null for the context pointer - the structure's layout is not published, and a handler that walked an invented one would read fabricated fields rather than fault at a named address.

Also in: EINVAL for signal 31, now that it was measured with a valid handle and the D648 ambiguity is gone; and StopReason::Signalled, because a raise with no handler does not return - the process takes the signal and dies, and neither Aborted nor Exited could say that without naming a call the guest never made.

And PPSA25872 did not move

DIAG raise thread=0x5e2d00000200 signum=0x1e current=0x5e2d00001dc0

The target is not the caller. The title raises on one thread while running on another - a collector suspending a mutator. Hardware delivers that; the caller gets 0 and the handler runs on the target. Orbistoun cannot: guest code runs natively, there is no loop to check a flag in, and the target is not inside an orbistoun implementation when the signal is raised.

The branch refuses. Answering 0 would move the wall today and tell the collector that a thread had acknowledged a suspension it never received - after which it reads that thread's registers and walks its stack. The failure would surface somewhere else entirely, as a hang or a corrupt heap, with nothing pointing back here (D650).

Surprises

• The trace never said "cross-thread"; one eprintln did. Every report from this title showed the raise, its arguments, and the handle - and the handle being a different thread's was invisible because nothing printed the current one beside it. Two numbers, and the whole shape of the wall changed.
• The wall got harder to move and much better understood. Before this pass it was "we do not know what raise returns". Now every return is known, the delivering call is implemented, and what remains is one specific mechanism with a named cost.

Next

• Cross-thread delivery: a per-thread pending-signal slot checked on the way into the thunk dispatch. Sketched in D650, deliberately not built in the same pass that discovered the need. Its honest limit is that a mutator spinning in pure guest code never takes the signal.
• The selfish-elf differential, still claimed and still owed.