[tae898]

What does this PR do?

I found that to_json_list() returns different Python floats on the two wheels for the same FLOAT or float32 float[] value. The native wheel returns the exact float64 widening of the float32 (NativeCodec.java widens a Float with doubleValue(), and the float32 tail of a vector is decoded with .tolist()). The JVM wheel passed each Float and each float[] to the engine's JSONObject unchanged (RowBatcher.keepNonFinite returned them as they were), and the writer prints the shortest float32 text, 0.1, which Python reads as a different double than 0.10000000149011612. The JVM wheel's own to_list() widens exactly, so on the JVM wheel to_json_list() and to_list() of the same rows disagreed, and row["vec"] == original.tolist() was false there and true on native.

The change is in RowBatcher.keepNonFinite only: a Float becomes a Double and a float[] becomes a double[] before the writer sees it, and NaN and the infinities keep their tag. The native wheel is not touched. The bridge jar is rebuilt by CI, so the new behaviour is what the JVM test jobs run.

The test is tests/test_cross_wheel_golden.py: one fixed two-row table with committed golden literals for to_list(), to_json_list() and to_columns() (FLOAT, DOUBLE, float32 float[] of 3 and of 9 values, DECIMAL, LONG beyond 2^53, DATE and DATETIME including a pre-epoch value, NULL, an embedded list). It has no branch on the backend, so the same file runs on both wheels and a literal that holds on both is the parity check. The float cases fail on the JVM wheel before the Java change. No case is marked xfail.

The JSON text of a float32 gets longer on the JVM wheel (about 17 digits instead of about 8). I have not measured what that costs in transport time.

Related issues

Refs #314. The xfail list of this PR is empty, but the audit in #314 lists cases this test does not cover yet (an integral DECIMAL in to_json_list, the text of Result.to_json(), float-list columns in to_columns and to_arrow, vector distances), so I do not close it here.

Performance changes

None claimed. The longer JSON text is not measured.

Docs

None needed: the documented contract is that both wheels return the same values.

Checklist

• A regression test fails on the old code and passes on the new
• The test suite that the change can affect passes (CI is the first run)
• uvx pre-commit run --files <changed files> passes

🤖 Generated with Claude Code