Crash (ZeroDivisionError) in _normalize() when parsing malformed or empty CSV columns

[FreshAcrobat]

Description

When running dsdiff against a malformed or empty CSV file (e.g., a file containing only headers or single-line unexpected characters that result in 0 valid data rows), the program crashes with a ZeroDivisionError: float division by zero.

Steps to Reproduce

1. Create a CSV file with no actual data rows (or malformed layout resulting in 0 row count).
2. Run dsdiff diff baseline.csv malformed.csv (or vice versa).

Traceback

  ...\src\dsdiff\drift.py:32 in _normalize
    return np.full(counts.shape, 1.0 / counts.size)
ZeroDivisionError: float division by zero

Root Cause Analysis

In dsdiff/drift.py, the _normalize function checks total <= 0 and executes a uniform fallback:

if total <= 0:
    # Uniform fallback keeps PSI finite when a side is empty.
    return np.full(counts.shape, 1.0 / counts.size)

However, if the dataset has 0 data rows, counts.size also evaluates to 0. This results in 1.0 / 0, causing a ZeroDivisionError and crashing the program.

Suggested Fix

We might want to safely intercept when counts.size == 0 inside _normalize, or validate the dataset row count earlier during the file-loading/profiling stage to provide a cleaner error message to users.