Add lane-order invariance check to validation

[rabujamra]

Closes the question raised in #114. Adds validation/mr_permutation_invariance.py,
a dependency-free script (csv + numpy) that holds every observation fixed and
varies only the lane order of the combined chart, then reports which chart
results move and which do not.

Why

With by=[] the moving range runs across the whole concatenated sequence, so the
step from one organisation to the next contributes a range like any other. On the
24-ACO file that is 23 of 95 ranges. The script asks what depends on which
organisation happens to be adjacent to which.

What it reports, on aco_per_capita_expenditure.csv

  within-ACO     72 ranges   mean    933.79     invariant to lane order
  boundaries     23 ranges   mean   3037.90     3.25x the within mean

  mRbar as loaded                    1443.21
  mRbar within-only                   933.79    +54.6%

ordering              mRbar    X-limit width   X signals   mR signals
  as loaded           1443.2         7677.9          4           3
  sorted by level     1272.5         6769.9          7           0
  interleaved         1497.6         7967.4          3           6
  20k random mean     1413.4         7519.4       3.92        4.48
             range  1269-1545     6753-8221        3-7         1-9

  random orderings with zero X signals    0 / 20000  (0.00%)
  shipped-order mRbar percentile          77.0

Two results, and they point in opposite directions:

• The combined chart's conclusion is invariant. Across 20,000 random lane
  orderings the X chart never returned zero signals. "These 24 are not one
  process" survives every ordering tested.
• The boundary-signal structure is not. The count ranges 1–9; under the
  shipped order it is 3, and ordering lanes by level gives none. The flagged
  pairs differ entirely between orderings.

The script also records, descriptively, that a boundary moving range changes
organisation and resets time in one number — organisation i's last
observation against organisation j's first. On this rising series the mean
boundary range is 1.66x the mean adjacent pure-level gap. That is a property of
the lane-major sequence on trending data, not of the two organisations, and the
script says so rather than attributing a share to the time step.

Confirmed against the library

Per #114, companion=True flags exactly the three transitions the script reports
for the shipped order, at the first-year points of ACO-019, ACO-021 and ACO-024,
plus the same four X points. The script uses the classical D4 = 3.267, which
matches the library's mR limit.

What it does not test

• It does not test the lane pattern, which is order-invariant and is what the
  combined chart exists to read.
• It does not evaluate the organisation-level charts. Per Series length (T) as an axis in design-state admissibility — and a saturation question #114 those are where
  organisation-level claims belong: the recentred R5 Xbar by ACO, and the
  phased view. ⚠ Note that R5 is not wholly free of traversal — its points are
  order-free organisation means, but in ADS 2 its limits are set from R2, which
  is built from the same lane-major sequence. See Series length (T) as an axis in design-state admissibility — and a saturation question #114. This script covers only
  the combined chart.
• One dataset, one structure (T = 4, all lanes rising), 20,000 of 24! orderings.

Runtime

About one second. No dependencies beyond numpy.

python3 validation/mr_permutation_invariance.py aco_per_capita_expenditure.csv

[cnicholas]

Reviewed