432. The bytes were not what it wanted

2026-09-08 - directed, continuing 431

What was done

Built the delivery experiment and got a clean negative (D589). ORBISTOUN_APR_DELIVER reads the file the asynchronous path resolved into the buffer its command header names - 224,748 bytes of globalgamemanagers into 0x740009200000 - and the guest does not care.

	distinct imports
baseline	193, 193, 193
delivering	192, 192, 193

Identical fault in every run. The first delivering run said FURTHER and that was the 192-to-193 drift landing the convenient way round.

Two traps, both caught by apparatus built earlier this week

• A single run said FURTHER. Three runs each said it did not. Step::CheckRepeats measures exactly this drift, and the habit of running three is what it bought.
• The caveat did not print. APR_DELIVER was declared Effect::Intervenes and not carried in Experiments, so the run reported same with nothing saying it was propped. Declaring the effect is half the job; the conditions record is the half that gets forgotten.

What is now closed and what replaces it

"The guest needs the bytes" was a sentence in worklog 431 that would have been repeated until somebody tried it. It is a measurement now, and the next reading has to explain why the bytes did not help.

libSceAmpr exports sceAmprCommandBufferWriteKernelEventQueueOnCompletion, which says completion is reported through an event queue rather than a return value. That is the obvious next reading and nothing here has tested it.

Surprises

• Forcing all three Apr calls to success also changes nothing - measured before the delivery experiment, and it is the same shape of answer. The guest logs its complaint and carries on. So the Apr path fails and making it succeed does not help, which is a different statement from "the Apr path is the wall".
• The title prints TODO: for its own LocalFileSystemPS5::Enumerate. That is Unity's dev-build marker in a shipped binary, and it sits between the Apr wait and the fault. The fallback reading - Apr fails, the title falls back to an enumeration it never finished writing - is still the one that fits and is still unmeasured.

Next

• The event-queue completion path, which the export list points at.
• Why the guest calls neither sceAmprAprCommandBufferConstructor nor sceAmprAprCommandBufferReadFile while importing both.
• The mapping count doubling when handles clustered, carried over from 430 and still untouched.