tree-sitter reads UTF-8, but extractors handed it raw file bytes, so a UTF-16
file yielded zero nodes and a cp1252 name like CaféOrder was truncated. Decode
via a shared reader that returns UTF-8 bytes unchanged (ids/offsets/cache keys
for working files are byte-identical) and re-encodes UTF-16/UTF-32/non-UTF-8
using a deterministic cp1252->latin-1 fallback. The semantic whole-file and
slice readers now share the same decoder.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
