Lab / failure
The search that decided nothing and printed ok
An early run of Weft's composition engine spent four seconds searching, decided nothing, rendered the same plan on both arms, and reported a ratio of 0.9999 as ok.
Measurement, Negative resultsWeftWeft's composition engine prices each candidate region against a compile budget. In one early run on an x64 box, a single price took 2,251 ms of a 2,000 ms half-budget. The budget was only checked before starting a price, and a started price couldn't be interrupted.
Everything behind it was skipped: 4 of 644 candidates got priced, and the rest went back to the old pattern matchers. So the engine's plan came out byte-identical to the matchers' plan, the two arms measured the same thing, and the reducer printed ok on a ratio of 0.9999.
Three fixes came out of it. A plan the engine has decided never falls back to the matchers. The reducer refuses a comparison where both sides rendered the same kernels. And the run driver now actually runs the A/A control it had been claiming to run.