pandas: one missing value turns int64 IDs > 2^53 into rounded floats

pandas 2.2.3 and 3.1.0 behave the same. In a CSV id column holding 2**53+1, 1234567890123456789 and 2**63-1 plus one empty row, default read_csv gives float64, and 0 of 4 IDs survive: 9007199254740993 comes back as ...992, and 1234567890123456789 as ...768. engine='pyarrow' also gives float64 (0/4). Exact (4/4): dtype={'id': 'Int64'}, dtype_backend='pyarrow' (int64[pyarrow]) or dtype={'id': str}. The same silent conversion happens after loading, in an exact int64 Series. reindex with a new label, a left merge with an unmatched key, shift() and where() all produce float64 and round every ID above 2**53. There is no warning in any of these cases. For ID or key columns, read them as Int64 or str, or convert with astype('Int64') before the operation. An Int64 Series stayed exact through the same reindex, left merge, shift and where on 3.1.0.