AOMIC ID1000 exploratory extensions, manuscript version 1.7

Scope
-----
These analyses were added after inspecting the original AOMIC replication.
They are exploratory; the original five-part primary model is unchanged.
All original complete cases (N=926) are retained.

Inputs
------
Use the original version-controlled public AOMIC source files:
  participants.tsv
  data-subcortical_type-aseg_measure-volume_hemi-both.tsv
The original AOMIC source manifest and download provenance accompany the
main reproducibility archive. The script records input SHA-256 checksums.

Run
---
python run_aomic_extensions.py --source-dir PATH_TO_SOURCE_DIRECTORY \
  --output-dir PATH_TO_OUTPUT_DIRECTORY \
  --reference-results PATH_TO_ORIGINAL_locked_replication_results.json

The optional reference-results argument verifies the primary coefficients,
ordinary and HC3 uncertainty, global tests, and partial R-squared against the
original saved outputs. With the supplied original files, the maximum
absolute discrepancy is less than 2e-15.

Dependencies: Python, numpy, pandas, scipy, statsmodels. Exact versions are
recorded in AOMIC_extensions_results.json.

Outputs
-------
AOMIC_S3_raw_volume_comparators.csv: ordinary volume correlations and
  single-compartment regressions adjusted for age, sex and log selected tissue.
AOMIC_S4_all_pairwise_balances.csv: all ten fully adjusted balances;
  within-cohort Holm adjustment across the ten tests.
AOMIC_S5_cognitive_domains.csv: fluid, memory and crystallized outcomes;
  within-cohort Holm adjustment across the three tests.
AOMIC_S6_cerebellar_decomposition.csv: three six-part tissue contrasts and
  a six-part left/right total-cerebellum contrast; exploratory Holm adjustment
  across these four tests.
AOMIC_scale_sensitivities.csv: primary total-tissue adjustment, replacement
  by eTIV, and simultaneous adjustment for tissue volume and eTIV.
AOMIC_extensions_results.json: exact outputs, model definitions, software
  and input hashes, primary reproduction and algebraic verification.
FINDINGS.txt: numerical overview and interpretation cautions.

private/AOMIC_partial_regression_residuals.csv is generated solely to draw
the per-cohort partial-regression figure. It has no subject identifiers but
is still participant-level data; omit it from the distributed code archive.

Model conventions
-----------------
All focal balances are standardized using the within-cohort population SD,
and all remaining orthogonal composition coordinates enter their model.
HC3 intervals and p values use residual-df Student t distributions.
The six-part analyses preserve total analyzed tissue volume and include all
five ILR coordinates. Cortex/cerebellar-WM is a direct anatomical contrast;
separately standardized coefficients should not be compared by magnitude
as if their difference were itself a significance test.

The two-size model is algebraically equivalent to adjustment for log(eTIV)
and log(T/eTIV). T is the sum of five selected tissue compartments, so T/eTIV
is a selected-tissue fraction rather than a validated whole-brain parenchymal
fraction. Neither algebraic equivalence nor a positive coefficient after
dual adjustment establishes a causal explanation for the eTIV-only attenuation.
