Post hoc sensitivity of the two-epoch sign test and historical comparison

1. Reproduce from the root of the extracted S1. Published outputs are already
   included. The quick verification reads them without modifying any input.

   python spatial_sensitivity/verify.py

   To recompute with Linux, a C99 compiler named cc, and Python packages as
   listed in the root README, use a fresh directory for outputs:

   mkdir -p /tmp/uruana-check
   OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python spatial_sensitivity/space.py --workers 2 --out /tmp/uruana-check
   OPENBLAS_NUM_THREADS=1 python spatial_sensitivity/epoch.py /tmp/uruana-check/epoch_sensitivity.csv
   python spatial_sensitivity/height.py /tmp/uruana-check/height_sensitivity.csv
   python spatial_sensitivity/figure_s2.py /tmp/uruana-check/Figure_S2.png
   python spatial_sensitivity/verify.py --recomputed /tmp/uruana-check

2. Raw least-squares sector classification. Each draw searches all 64,800
   centers (1 degree) in all 23 eligible sectors. A sector passes if there is
   at least one center at which all 68 signed slopes are strictly positive
   (28 in 2011 and 40 in 2008). The 2008 and 2011 destination main fields
   change geographically as in the original scan. For each center the fields
   are held fixed across perturbations. The search compares the best margin
   with an exhaustive calculation for four draws for each sector and scale.

   At the native geometry, fit a separate straight line for each 2011 raw
   line/window and each 2008 channel/width/band. The 2011 response is the
   median in each occupied flight-line bin. The 2008 response is the mean
   across the occupied nodes in each northing row, giving equal weight to
   rows. Multiply each residual, divided by sqrt(1 - its OLS leverage), by
   +1 or -1 with equal probability. Add the resulting errors to the fitted
   responses. A single sign is shared by all observations in a contiguous
   northing strip within a survey/sector. This couples neighboring northings,
   flight lines, and the overlapping 2008 widths and channels. The two
   epochs use independent draws. The fixed profiles retain the observed
   coordinates, signal amplitude of residual departures, and their shared
   sign changes inside strips. Dependence between strips is not estimated.

   Strips start at the translated survey-center northing. Their widths are
   0.5 or 1.0 km in 2011 (two or four nominal bins), paired with 3.0 or
   4.5 km in 2008 (two or three grid-row spacings). There are 999 draws per
   sector per scale. Seed 20261103 is partitioned by sector index in the
   archived regional table and scale index. Since OLS is linear at fixed
   coordinates, each draw's vector of slope changes is added identically to
   every location's original 68 slopes. Survey geometry and fields stay fixed.
   The C kernel searches the entire finite grid with an exact upper
   bound that skips centers which cannot improve the best margin.

   sector_sensitivity.csv gives grid passes, native passes, the observed best
   grid margin, descriptive percentiles (2.5, 50, 97.5) of that margin and
   the count retaining all five eastern OLS sign reversals at Uruana after
   IGRF subtraction. native_slope_sensitivity.csv.gz reports each native
   line or band realization, its unperturbed slope, percentiles and sign
   count. Slopes and margins are in nT/km. The ten configurations from the
   processing study share data. This perturbation tests the primary raw-2011
   OLS score, with both 2008 channels included. It does not simulate all ten
   estimators/products, a new survey, or an error in the field models.

   The block lengths are sensitivity settings rather than fitted correlation
   ranges. Several 2008 bands have very few rows. The multipliers perturb
   lack of fit to a straight line, which includes real profile curvature.
   Geographic origins, field coefficients, choices of windows, instrument
   calibration and selection of sectors are conditioned on. Sectors are
   perturbed separately. Pass counts and percentile ranges describe this
   working model, not sampling confidence intervals, a joint distribution
   over sectors, false-discovery rates, or a probability of geological rarity.

3. Historical date. epoch.py recomputes both 2008-fitted Brazilian ensembles
   at four representative dates, compares membership, observed responses and
   frozen predictions against the archived datum-corrected point records,
   and checks the updated predictions at the reference date. The reported
   RMSE is point-weighted after independent profile centering. Changing the
   date changes the modeled observing direction, induction and intensity
   scaling. All four runs use the same observed historical residuals.

4. Model height. height.py summarizes eight backgrounds for each of the five
   eastern lines at measured GPS heights and at a constant 1.000 km ellipsoidal
   height. It fits five processing stages with each of two estimators per cell.
   Only background evaluation heights vary. Observations remain as flown;
   there is no continuation of measured profiles. To regenerate the archived
   1 km vector file from the official NOAA archive, follow core_lithosphere/
   README.txt and add --fixed-height to the rebuild_vectors.py command.

5. Figure S2. Left, original occupied raw 250 m bins in the 2011 target
   window on line 12350, with OLS and Theil–Sen lines. Right, all ranked
   pairwise slopes between distinct bins. The dashed line is their median.
   The script verifies both slopes against local_recovery/line_slopes.csv.
