[FIX] concordant z is the statistic, not the inverse of its corrected p

[jdkent]

The "concordant" mode reported norm.isf(p) of a p-value that had already been doubled to correct for testing both tails. That is a tail mismatch rather than a convention: it shrank the statistic to absorb a multiplicity penalty that belongs to the p-value alone.

Winkler et al. (2016), which the mode docstrings cite, define the concordant statistic as T = max(-2 sum ln p_k, -2 sum ln (1 - p_k)) -- the better of the two directed combinations, unshrunk -- and apply the two-test correction to its p-value. That is norm.isf(p / 2), and it now reproduces the winning directed statistic exactly (4e-16 over 1650 cases across both estimators).

Three symptoms had one cause. The reported z was negative wherever p > 0.5, carried the opposite sign to the effect there, and was -inf wherever the cap put p at exactly 1 -- so the least significant results had the largest magnitudes. A single input now combines to its own z; it was 2.7822 for an input of 3.0.

Thresholding the old map at any z_c selected p < sf(z_c) instead of p < 2 sf(z_c), so every concordant z-map ran at exactly half its intended two-sided alpha. The fix only raises |z|, so at a fixed threshold it can add voxels but never remove them.

p and logp are unchanged. The four pinned concordant values were not independent references: before 091b633 the tests computed ss.norm.isf(p) themselves and asserted it against a constant, and that commit moved the expression into the estimator and carried the constants across. They assert that norm.isf(p) equals norm.isf(p), and so could not detect this.

This completes #125, which gave the concordant statistic its direction but left the magnitude alone.

Adds a parametrized invariant that z and p describe the same tail in all three modes -- the check that would have caught it, and that the two right-tailed modes have always satisfied.

[jdkent]

some extra text

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.11%. Comparing base (c9b3637) to head (d0ce4a3).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #141   +/-   ##
=======================================
  Coverage   95.11%   95.11%           
=======================================
  Files          13       13           
  Lines        1985     1985           
=======================================
  Hits         1888     1888           
  Misses         97       97           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.