Rust zero-cost abstractions vs. SIMD

A full-text search query on turbopuffer was taking 220ms instead of the expected ~50ms. Profiling revealed that over 60% of runtime was spent in a merge iterator, not in BM25 ranking. The root cause: Rust's zero-cost iterator abstraction, while compiling each individual call efficiently, prevented the compiler from vectorizing or unrolling across calls due to the recursive nature of `next()`. The fix was a classic database technique — batched iterators — where `next_batch()` fills an array of 512 KV pairs at once, giving the compiler a tight inner loop it can auto-vectorize with SIMD. The result: the benchmark dropped from 6.5ms to 110μs (60× faster), and the production query latency fell from 220ms to 47ms. The key lesson: 'zero-cost' means the abstraction compiles away per call, not that it has no effect on the compiler's ability to optimize across calls.