A controller subsystem, and a name I talked myself out of

Asked for controllers - configurable count, real pads, keyboard mapping - and for those inputs to reach the emulator so the shell could be tested by pressing its own button.

The host side is built and works: buttons named by position, four configurable ports, a per-port key mapping held as text so no window toolkit leaks into the input contract, and a pure tap-versus-hold type that takes elapsed milliseconds rather than reading a clock. That last one made every edge an assertion instead of something found by holding a button until it looked wrong - a hold fires once, fires while still held, and a release after one is not also a tap.

Verified by pressing it rather than by reasoning about it: holding for 1.1 seconds opened the power menu, with "quit the title" correctly absent because nothing was running.

The mistake worth keeping

orbistoun-input declared six pad functions and implemented none. Before implementing, the names were checked against the symbol database and none of the six was in it, while ninety-one other names from the same library were.

From that I concluded they were previous-generation names - recalled rather than derived - and therefore hashed to NIDs no import would ever carry. I rewrote the declarations to the confirmed ...Ext siblings and wrote a confident paragraph about shims that could never be reached.

Then:

orbistoun-cli imports titles/obscene/eboot.bin

Ninety-seven pad imports, including both halves of every pair. scePadOpen and scePadOpenExt. scePadReadState and scePadReadStateExt. The library exports both and a real module asks for both.

What the error was made of: absence from a generated database was read as evidence about the platform. That database holds what the naming loop has derived - a name missing from it means nobody here has named it yet, which is a fact about this project. The direct evidence was one command away, and it is the command CLAUDE.md opens with.

Two guards earned their keep

Registering the implementations tripped every_implemented_function_is_written_down, which wanted a knowledge file - so libScePad.toml now records provenance for all nine. And orbistoun_input::implementations() had never been added to the service's list at all: the module was registered, the functions were not, and the comment directly above that list warns about exactly this - "code that looks written and never runs."

What it is worth

obSCEne already had a 100-input probe section. It now reports three passes where the recorded table had one pass and one failure, and the best of the three is close-rejects-bad-handle - a negative test, asserting a bad handle is refused.

Review from a parallel implementation

Prosperous grew a controller model for its own reasons and read this one alongside it. Three findings, all real: a keyboard could not move a stick at all (the shipped configuration is one keyboard port, so out of the box it could press seventeen buttons and drive nothing analogue), cross-port key conflicts were invisible because the seen-map was built inside the per-port loop, and consequently there was no safe way to add a second keyboard player.

Their cancellation rule is the good one and is now here: opposite pushes sum and clamp, so left and right held together mean centre - a position a stick can actually be in. Letting the first win would make the pair mean something no pad can express.

The prediction was wrong, and the reason is the finding

Expected the next names run to grow learned to about 5,592 words, now that the quadratic shapes are off and 906 million candidates fits under the ceiling. It grew by nothing.

613,445 candidates across 54 modules, 0 named. The vocabulary is fed the parts of names a run newly confirms, not the candidates it tries - and the corpus is already fully named, so nothing was offered and learn_words was never called.

The 11,842-word list came from the era when a large batch of static names was first confirmed. parts_of splits on capitals, so _ZN8Document9terminateEv yields Document9terminate - which is exactly where the mangling fragments came from, and it only happens on names being confirmed for the first time.

So the ceiling is not currently binding at all. It bites on a fresh corpus or a large new naming run, not on the next command - and the worry about 5,592 junk words landing was misplaced. D342 said the opposite and is corrected in place rather than left standing.

The instrument was right and my model of it was not, which is the reverse of the day's usual failure and worth the same amount.