D261 - Shapes are the binding constraint, not vocabulary, by three to one

decided · 2026-08-25 · measured against names the grammar did not find

audit says the grammar can re-derive every name in the database, which sounds like full coverage and is not. Most of those names were found by the grammar, so it can necessarily spell them - measuring coverage against them measures the search that produced them.

tests/shapes.rs measures against the names found without the grammar - read out of module strings, or seen in a trace. Those owe nothing to the pattern list, so they are a sample of what vendor identifiers actually look like. Of 183 such names:

reachable under the current pattern list	28 (15%)
unreachable, needing a shape	121 (66%)
not splittable into known words at all	34 (19%)

So the pattern list is short by more than three times what the vocabulary is short by, and buying more words is spending on the weaker lever. That confirms, with numbers, a prediction made independently from the naming side: a word only pays in the shapes that exist, and learned appears in two patterns, once each.

The first cut of this was wrong and the correction matters. Counting every independently-found name gave a 68% vocabulary gap - but 276 of those 464 names are POSIX or libc, which every vendor pattern begins with the sce prefix and cannot spell, and is not meant to. Scoped to vendor-shaped names the ratio inverts.

Three shapes are nearly free and would spell fourteen of them today:

names	cost	shape
3	+0%	prefix + learned + verb
5	+4%	prefix + module + learned + learned + verb
6	+11%	prefix + learned + verb + learned + learned

The rest are not worth having at any vocabulary size - the ranked list runs to +789%, +1862% and, for one ten-part shape, five orders of magnitude beyond the whole current space. A name needing that has to come from somewhere other than the generator.