494. The pad structure had been measured

2026-09-10 - loop, continuing 493

Three input conformance checks green, and the comparison against hardware finally on the right leg.

Third time picking the leg

D669 moved the comparison off obSCEne's payload leg onto its package leg. Still wrong by one axis: the package leg is title/ps4-bc - previous generation, libSceGnm mapped - and orbistoun presents prospero. Two of the divergences it produced were 165-gnm/dispatch-*, which orbistoun should differ on.

No hardware leg anywhere reports ps5-native - eighty-eight logs, all unknown-gpu or ps4-bc. The closest match is a populated eboot leg, title/unknown-gpu, which is exactly what orbistoun reports. Against that: hardware 167 passes, orbistoun 170, ten checks where hardware passes and orbistoun does not, generation artefacts gone.

Four of the ten were input.

The measurement D345 was waiting for

scePadReadState was declared and unimplemented, deliberately: D345 built the transport from the window to the guest and stopped at the shim, because the structure was "a size and layout nobody here has measured".

It has been measured since. 100-input/read-extent fills the destination with a sentinel and reports how far the change reached - extent 120, changed 120, and the full at-rest image. 100-input/batched-read reports the identical thing for scePadRead.

The shim writes those 120 bytes verbatim and does not declare a struct. What was measured is a byte image; which offset is a button and which a stick is an inference from it, and laying out fields would publish the inference where the measurement belongs (D671).

100-input/read-extent and 100-input/batched-read went partial → pass 0x78; 100-input/oops-sdk-poll went fail → pass 0x0. Input divergence: four to one.

Surprises

Audio went from two divergences to five, and that is the previous turn's change becoming visible. 090-audio/blocking, format-selector, oops-sdk-pcm, open-shapes and volume-flag were reading partial on the positive placeholder and read fail on the honest one (D670). They were passing partially on a refusal the guest could not see. Audio is the next cluster and it only exists as a cluster now.

obSCEne asks the right question of the wrong machine. Under orbistoun the two checks that would settle the field offsets answer pending: "controller attached, but no button was seen in the window" and "controller read, but nothing moved: sweep the sticks and re-run". orbistoun cannot answer them - it is the thing that does not know where the bits go. On a console with a pad in hand it is one run, filed as REQ-20260910T0650Z-d1c4.

So orbistoun has live pad state, a transport for it since D345, and still cannot report input: the one missing piece is the offsets, and it is the piece that must not be guessed.

Two long learn invocations were refused by the permission layer, so the knowledge went into libScePad.toml directly in the same format. Worth knowing before reaching for that command with a dozen --edge arguments.

Next