A local Apple Notes sync dropped from 49 seconds to 825ms by replacing N+1 Apple Events with vectorized JXA. Benchmarks against AppleScript, SQLite, file stats, and hashing show why the fastest query was not the best design.
Rust's incremental compilation cache never garbage collects. With multiple AI agent sessions running cargo across git worktrees, target/ ballooned to 45GB. Here's the root cause analysis, the empirical fixes, and what I rejected after adversarial auditing.
Head-to-head benchmark of llama.cpp vs Apple MLX for local LLM inference on M5 Max. Architecture comparison showing why fewer layers = faster inference — llama.cpp is 21-59% faster across all payload sizes.