D285 - The shape that spells a verb over two learned nouns

decided · 2026-08-26 · sceKernelReserveVirtualRange, and the 104 names behind it

The name that opened the image+0xafc959 wall came out of the string harvester, not the generator, and the reason is worth stating precisely: every word it needs was already in the vocabulary. Reserve is a verb, Kernel a module, Virtual and Range both learned. What did not exist was a pattern placing a verb before two learned nouns.

So this is not a vocabulary gap and no wordlist change would have helped. tests/shapes.rs already measures exactly this, and ranks the shape second of every shape the corpus wants and the grammar cannot spell:

145  +2577554%   prefix + learned + learned + learned + learned
105  +     99%   prefix + module + verb + learned + learned      <- this one
 97  +1175375%   prefix + module + verb + learned + learned + learned

105 names for +99% - roughly doubling the sweep. The one above it would unlock forty more and costs twenty-five thousand times the whole space, which is the difference between a shape worth adding and a shape that says the name has to come from somewhere else.

tail is omitted, following D261: tail holds the empty string, so a shape carrying it is a strict superset of the same shape without, at about fifteen times the candidates. The name that prompted this carries no suffix, so the cheap version is what the evidence asks for.

It does not invalidate a single existing record, and the first draft of this entry said it would. A generated record is {pattern, index}, and re-derivation resolves the pattern by name and asks it for its own index - so a new pattern adds a new enumeration and leaves every other one exactly where it was. Measured rather than reasoned: thirty-three names were unaccounted before this shape was added and the same thirty-three after it.

What does renumber is growing a vocabulary an existing pattern draws from, which is what happened earlier the same day - posix_vocabulary splitting on underscores took it from 2,923 words to 3,741, and every index built from that list moved (D213). Worth separating, because the two changes feel identical and only one carries a repair bill:

change	effect on existing records
add a pattern	none - a new enumeration beside the others
grow a vocabulary	every index in every pattern using that list

The cost a new pattern does carry is on --repair itself, which walks each pattern's whole enumeration looking for a name: another shape is another space to search before giving up.