[jm-rivera]

Every reading-order vector is now a table a server could serve, and the validator judges every column by one equality rule.

The SDKs replay each vector by writing it to Parquet and reading it back, as a client reads a dataset. In nested-cells-compare-by-value the flags column held [true] in one row and [1] in another, which no Arrow list column can hold, so no server could serve it and Python could not replay it. flags now holds [true] and [false], and _unorderable reports any member column pa.array cannot build (mixed list items, a struct field of mixed types). The bool-against-number case stays in the unit tests of both reference implementations, since no vector can carry it.

served_table judged flat columns with Arrow unique(), which counts 0.0 and -0.0, and two NaN bit patterns, as two values, while same_value counts each pair as one. A vector and its Parquet file could therefore group columns differently. Every column, flat or nested, is now judged by same_value in one pass, the rule table_from_rows uses.

Testing: test_a_vector_and_its_parquet_file_group_columns_alike (signed zeros, NaN payloads) fails on dev; the two new malformed-vector cases are reported. uv run tools/validate_client_contract.py passes (125 scenarios) and the tools tests pass (586).