Function tolerant_csv_stream 

pub async fn tolerant_csv_stream(
    store: &ObjectStore,
    path: &str,
    schema: SchemaRef,
    projection: Vec<usize>,
    format: &CsvFormat,
) -> Result<SendableRecordBatchStream>

Builds a CSV stream that can skip selected record-level parse/cast errors.

This recovery path intentionally uses one-record batches. It is slower than normal CSV scanning, but keeps each parse/cast failure isolated to a single record. Arrow’s CSV decoder clears buffered rows before type parsing, so a failed multi-row flush cannot be safely retried row by row without replaying input bytes.