Figure 5: complete moderate-imbalance ordering-change map

This figure uses previously completed results only. No simulation or estimator
evaluation is performed by its plotting script.

Reproduce from the exact accompanying source tables:
  python make_figure5.py --summary primary_KSG4_summary.csv --paired paired_changes.csv --output-dir reproduced

Alternatively reproduce directly from the accompanying 64-row plotted-values file:
  python make_figure5.py

The optional --archive argument reads the two original completed tables from
submission_revision/extension/ within the original ESM_2 ZIP. All paths can be
supplied explicitly; no source archive or absolute workstation path is required.

Dependencies: Python 3, NumPy, pandas, matplotlib, Pillow.

Figure5.png: 4200 x 2820 pixels, 600 dpi, nominal width 7 inches.
Figure5.pdf and Figure5.svg: vector versions with selectable text.

The CSV contains exactly 64 KSG4_centred cells with unequal sample sizes and
nonzero population information differences. IA and IB are in nats. NA is the
smaller sample-size anchor and NB/NA is the ratio. The following three columns
are percentage points; underlying probability columns use the interval [0,1]:
  change_correct_pp
  change_correct_boot95_low_pp
  change_correct_boot95_high_pp

The paired intervals are the archived pointwise percentile-bootstrap intervals
from 2,000 resamples of independent parent-pair IDs. No raw observations or
individual overlapping subsets are resampled for those intervals. Within a cell,
601 independent parent pairs are used. Across cells, banks and nested prefixes
are reused, so cells are dependent. The color scale and numerical annotations
show point estimates; there is no significance threshold or simultaneous claim.

Each five-subset average constitutes one parent-level result. Matching subsamples
the larger B catalogue without replacement within each subset, allowing overlap
between subsets, and retains A. This figure concerns that particular procedure,
not a universal effect of matching sample sizes. Equal-information controls and
equal-size reference cells remain in the full original extension tables; correct
ordering is undefined for equal information and equal-size estimates are reused.
