engine: compiler optimization pass (C-LEARN -33%, LTM on world3 9.8x, LTM on C-LEARN 101x) - #993

Merged

bpowers merged 10 commits into

mainfrom

engine-compile-perf

Jul 29, 2026

Commits on Jul 29, 2026

engine: parse against the cached dimension context, not a rebuilt one
bpowers
committed
engine: stop allocating four Strings per identifier in canonicalize
bpowers
committed
engine: decide "already canonical" in one scan, and stop cloning dimensions per AST node
bpowers
committed
engine: lower the parsed AST by reference instead of deep-copying it first
bpowers
committed
engine: cache the LTM variable reconstruction instead of redoing it per query
bpowers
committed
engine: build the MDL importer's dimension context once, not per array element
bpowers
committed
engine: derive the single-variable reconstruction from the cached map
bpowers
committed
engine: hash identifier-keyed maps with FxHash
bpowers
committed
test: un-ignore the heavy gates the compile speedups made affordable
bpowers
committed
engine: firewall the single-variable reconstruction so a link score is not whole-model
bpowers
committed