D673 - An empty queue is a report, not a refusal

Status: decided Date: 2026-09-10

What the console does with no mouse plugged in

obSCEne's 101-input-ext/mouse-read ran on a machine with nothing attached:

sceMouseOpen   handle 0xc70700
sceMouseRead   rc 0x0,  extent 0,  changed 0

It opens, and the read succeeds having written nothing. Zero is the number of events taken, not an error code. A guest polling in a loop gets "nothing happened" and loops.

orbistoun refused the open and answered a placeholder to the read, so the check read partial.

The shape, and what is not claimed

Four functions, all measured or inferred narrowly:

• sceMouseOpen opens with no device. Refusing would look stricter and be wrong.
• sceMouseRead answers 0 and touches nothing - not even zeroing the destination, because the console's measured extent was zero and a guest holding a previous sample there would find it wiped.
• sceMouseInit and sceMouseClose answer success, and that is assumed: obSCEne calls both and discards the return, so it is inferred from the calls around them working.

The arities stop being 6. They were the trampoline's full capture, which is not a claim (D504); obSCEne types each one and the console answered, so the first N arguments are established - with the usual caveat that a call which works cannot see a trailing argument nothing passes.

A bad handle earns GuestError::InvalidHandle, a placeholder. Every other input subsystem here carries a measured base - the pad's 0x8092_0000, audio's 0x8026_0000 - because obSCEne provoked a failure and read one back. Nothing has provoked one from libSceMouse, and inventing a 0x80xx_xxxx to look like its neighbours is the single thing this must not do.

Registered beside the pad rather than folded into it

orbistoun_input::implementations() answers a &'static slice, so gathering a second module there means allocating. One explicit line in symbols.rs is cheaper and easier to notice when a third arrives.

The guard from D668 earned itself

Writing libSceMouse.toml and not registering it in EMBEDDED failed every_knowledge_file_on_disk_is_embedded on the first run, by name. That test was written two turns ago because twelve files had accumulated that no build loaded; this is the thirteenth being caught before it joined them.

What it bought

101-input-ext/mouse-read passes. 101-input-ext/mouse-moving moved from "no mouse attached" to "mouse open, but no record was read: move it and re-run" - the same shape the pad's button-bits took, and the same limitation: orbistoun has no field offsets to place live input into, so the queue is permanently empty rather than sometimes empty.

Against the matching hardware leg the divergences went from five to four, and orbistoun passes 175 against hardware's 167.