D320 - A curated list regrew to sixty-seven times its size, and only the clock noticed

decided · 2026-08-26 · found by asking why a test had been running for nine minutes

learned holds 177 words. It held 11,842, because a names run harvested strings from guest modules and merged everything is_word accepts - and two thirds of what it accepts is not vendor vocabulary at all:

Agent5pause  Agent6enable  Agent2gc  Document9terminate  Layer18accumulated

Those are Itanium C++ mangling: a length prefix and the identifier it measures. 6,451 of the 11,845 entries had that shape. They can never appear in a sce* export, so every one was pure cost.

And the cost is quadratic, because learned appears twice in two shapes. At 177 words those two are 165 million candidates each. At 11,842 they are 757 billion each, so a vocabulary round - which re-sweeps every shape using the grown slot at full size (D264) - went from about 350 million candidates to 1.5 trillion.

The visible symptom was the gate. Six tests in orbistoun-propose swept that space on every run; one of them was still going after nine and a half minutes in release and had to be killed. Nothing failed. Nothing said "the vocabulary grew" or "this round costs four thousand times what it did yesterday". The tree simply stopped finishing, which reads as a slow machine rather than a regression.

The reduction to 177 was a one-off act of curation that nothing held. D262 costed the shapes it unlocked and called that its most valuable consequence; D259 restored two words it had stranded. Neither wrote anything that would stop the next harvest putting the 11,665 back, and the next harvest did.

Three things, and the third is the one that matters:

The general shape is one this log keeps finding: a property that matters, held by nobody, noticed only when something unrelated broke. Here the property was a number, and the only instrument that reported it was wall-clock time on an unrelated test.