D262
- Repeated learned became affordable when the list shrank,
and nobody noticed
decided · 2026-08-25 · a consequence of somebody else's change, found by costing it
D195 kept learned to one position per pattern because it
appears in enumerations whose size is the product of their parts, and
squaring a large list is not affordable. That was correct and is no
longer binding.
prefix-module-learned-learned-verb-tail is the shape
five of the measured names need:
learned size |
candidates |
|---|---|
| 12,255, as it stood this morning | 12,164,987,025,000 |
| 177, after the list was reduced | 2,537,649,000 - about two thirds of the whole current space |
Four thousand eight hundred times cheaper, which turns "never" into "one afternoon". The reduction was made for unrelated reasons and its most valuable consequence was invisible until something costed the shapes it unlocked.
Worth stating as a general point. A vocabulary
list's size is not only a coverage question, it decides which
shapes can exist at all - so shrinking a list can buy more
names than growing it. Nothing in the tooling said so, which is why
tests/shapes.rs now prints the cost of every shape it asks
for.