DatasetReader ignores UnmarshalOptions::max_bytes_literal_length when it parses a row #160

DatasetReader::Create(src, opts) reads rows under the default MaxBytesLiteralLength, not opts.max_bytes_literal_length. So a caller who lowers the limit still gets a row whose b"…" cell is past it, and a caller who raises it gets that cell refused.

Where

v2.0.0 (4e96aba), src/pxf/dataset_reader.cc:

auto parsed =
    ParseDatasetRow(row_bytes,
                    static_cast<int>(columns_.size()),
                    opts_.max_nesting_depth > 0 ? opts_.max_nesting_depth : kMaxNestingDepth);

and src/pxf/parser.cc:

StatusOr<DatasetRow> ParseDatasetRow(std::string_view row, int expected, int max_depth) {
  Parser p(row, max_depth, kMaxBytesLiteralLength);

Only the depth reaches the row parser. protowire-go applies the caller's limit to rows (encoding/pxf/dataset_stream.go at 903b4a7: lexer{..., maxBytesLiteral: tr.lim.maxBytesLiteral, ...}).

Measured

Through protowire-python's DatasetReader.from_bytes(data, max_bytes_literal_length=191) (trendvidia/protowire-python#36), which passes the limit to Create: on @dataset lim.H (b) with a row holding a 192-byte b"…" cell, next_or_none() returns the row. The same literal through pxf::Unmarshal at 191 is refused with bytes literal decodes to more than MaxBytesLiteralLength=191 bytes.

Done when: ParseDatasetRow takes the bytes limit, DatasetReader passes opts_.max_bytes_literal_length (0 → default), and a test lowers it on a row. protowire-python pins the current behaviour in an xfail test that flips when this lands.