qio 0.1.2

New features

write_parquet() writes a list of raw vectors as a binary (BYTE_ARRAY) column, with NULL for a missing value, and parquet_schema() declares it as "BINARY", "GEOMETRY" or "GEOGRAPHY". The geospatial types hold well-known binary (WKB) with a crs, stored as given, and for GEOGRAPHY an edge algorithm.
qio records each geospatial column chunk’s statistics as the Parquet specification defines them. The geometry types are the top-level ones only: a MultiPolygon is listed, its polygons are not. GEOMETRY chunks get a bounding box, and a NaN coordinate is left out of its own dimension only. GEOGRAPHY chunks get no bounding box, because a geodesic edge can leave the box of its vertices. A chunk holding a value that is not valid WKB records no geospatial statistics, so readers cannot skip it by mistake.
geospatial_statistics() reports the bounding box and geometry types of every GEOMETRY and GEOGRAPHY column chunk, from qio’s files or any other writer’s, so readers can choose row groups by box before reading data.

Fixed write_parquet(append = TRUE) leaving the original file unreadable when the append failed with an error after writing its first new row group (#34). An error during an append now leaves the file as it was, and "bytes"-encoded strings in columns, names, metadata or crs are refused before any file is touched.
bloom_filter_may_contain() no longer reports FALSE, “definitely absent”, for values that are present. bit64::integer64 values are hashed exactly, and unsigned columns are checked by their stored bit pattern. A fraction, an out-of-range value, or a classed value such as a factor or POSIXct is now an error instead of a FALSE, so code that relied on that FALSE must convert its values first (#43).
An explicit INT32 column no longer writes -2147483648 as a missing value. That value is R’s integer NA, so it is now refused like any other out-of-range value, and the error names the row and the value (#38).
schema() reports a GEOMETRY or GEOGRAPHY CRS whole however long it is. Files that put inline PROJJSON there, as DuckDB does, had it cut to 127 bytes.

Fixed undefined behavior reported by CRAN’s sanitizer checks (M1-SAN, clang-UBSAN and gcc-UBSAN). Batch reads no longer load integer and floating-point values through misaligned pointers, and writing a string column that contains empty strings no longer passes a NULL pointer to memcmp() or memcpy(). Results were already correct; only the undefined behavior is removed.