[Sad Catapella N5][Scientific integrity] CSV parser accepts NaN, infinity and impossible nonempty tail moments #46

Baseline: f784c95. Confirmed missing validation, no repo tests executed. Distinct from #33 (overflow during legitimate moment calculation).

Code: csv_io._tail checks only negative count and zero-count nonzero moments; with count > 0 it accepts nonfinite sum, nonfinite or negative sum_of_squares, or moments with physically impossible variance. HistogramTail constructor also has no checks. These moments enter pooled mean and standard deviation. Existing corrupted-tail tests test count consistency and empty-tail sums, not nonempty NaN/inf/negative moments.

Deterministic fixture: Write a valid robust-range row with underflow count >0; edit only Histogram_UnderflowSum_C0 to nan or Histogram_UnderflowSumOfSquares_C0 to -1/ inf. Count/range checks remain unchanged, and invalid statistics reach downstream computations.

Fix: For nonempty tails validate finite sum/second moment, nonnegative second moment, and the feasible relationship n * sum_of_squares >= sum**2 with justified numerical tolerance. Negative sum may be valid for signed intensities. Define how previously serialized overflowed moments are handled in coordination with #33, never silently replace with zero. Validate direct HistogramTail construction too. Add robust tests for signed negatives, normal populated tails, NaN/±inf, negative squares, impossible finite moments and high dynamic range; report row/channel/field. No Numba priority.