interesting approach! why do you think claude code doesnt create embedding of the codebase on /init and instead it just uses grep and such; despite the semantic search being clearly better? I dont see any barrier for them to implement