Chunk bugs hide because normal reads use friendly boundaries. I build a reusable oracle method for split matches, offsets, UTF-8, empty chunks, errors, and end-of-stream.
An initial CDC load needs a database-consistent snapshot paired with a retained log position. Buffer changes after that boundary, apply the base, then drain without stale overwrite.
Scanner benchmarks are easy to make fast by changing the work. I use a reproducible matrix for construction, chunks, matches, prefilters, cache state, and result validation.
An automaton can recognize a cross-chunk match with little state, yet a streaming API may still need bytes. The missing distinction is recognition state versus output state.
A pattern may begin in one input chunk and finish in the next. I explain the state, offsets, and split-point tests needed to keep a Rust streaming scanner correct.