Binary columns containing UTF-8 text

count(), estimate_cost(), and estimate_cost_details() accept Polars Binary columns when each non-null value contains valid UTF-8. The bytes are interpreted as the exact raw text they encode: no normalization, replacement decoding, or special-token handling is applied. Counts match a String column containing the same text.

import polars as pl
import polars_tokenizer as tokens

frame = pl.DataFrame({"text": pl.Series([b"hello world", None, "你好".encode()], dtype=pl.Binary)})
counts = frame.select(tokens.count("text", tokenizer="o200k_base"))

Invalid UTF-8 in a non-null value raises a Polars ComputeError. Null values produce null output without inspecting the bytes behind their validity mask. This matters for sliced or modified Arrow arrays, where a null view can retain bytes from a previous value. The implementation does not first cast the whole column to String, because that cast can reject invalid bytes hidden under a null.

bad = pl.DataFrame({"text": pl.Series([b"\xff"], dtype=pl.Binary)})
bad.select(tokens.count("text"))  # raises ComputeError: invalid UTF-8

The native kernel borrows binary views and validates each non-null value before counting it. It keeps the same UInt32 output, null propagation, opt-in bounded cache_capacity, and byte-balanced parallel scheduling as String input. All-null and all-empty batches skip tokenization. This support is for UTF-8 text stored as Binary, not arbitrary binary data.