align Mahalanobis and DA PTP with manuscript; harmonize use_empirical_variance default; bump to 0.8.0

[settylab-dotto-bot]

Brings the implementation into agreement with the manuscript on
three statistical points and harmonizes a default-argument
inconsistency that had crept in across deprecated wrappers.

All three are intentional and minor-version-worthy. Mahalanobis
distances in the GP-only regime contract by √2, DA PTPs are
halved, and the deprecated wrappers that exposed use_empirical_variance
flip their default from True to False. Relative rankings of
genes are unchanged because the scale factors apply uniformly,
and FDR thresholds re-calibrate against the null distribution.
Absolute thresholds tuned against 0.7.0 need re-tuning.

Mahalanobis denominator — sum, not average

The manuscript defines the gene-wise distance as
D(a,b) = sqrt((μ_a − μ_b)^T (Σ_a + Σ_b)^(-1) (μ_a − μ_b)),
but the GP-posterior branch in
kompot/differential/differential_expression.py:629 was
computing combined_cov = (cov1 + cov2) / 2. This change drops
the /2 so that the combined covariance equals the variance of
the difference of two independent posterior estimators. The
sample-variance branch at the same call site was already
correctly summed, which is what made the inconsistency visible
to begin with.

Effect: D shrinks by a factor of √2 (and D² by 2)
everywhere. Same scale factor for real and null, so relative
rankings and FDR call-sets are essentially unchanged after
recalibration.

DA posterior tail probability — one-sided

The manuscript prose describes a one-sided test ("the
probability that the true density difference has the opposite
sign to the estimated change"), and the formula
PTP(x_i) = Φ(−|Δ(x_i)| / sqrt(σ_a² + σ_b²)) is one-sided.
The implementation at
kompot/differential/differential_abundance.py:607-610 was
emitting 2 · Φ(−|z|) (two-sided). This change removes the
spurious + np.log(2) so ln_ptp = log Φ(−|z|).

Effect: numeric PTPs are halved; neg_log10_fold_change_ptp
increases by log10(2) ≈ 0.301. A previously-published cutoff
of PTP < 1e-3 corresponded to |z| ≥ 3.29 (two-sided) and
now corresponds to |z| ≥ 3.09 (one-sided). Re-tune any
ptp_threshold chosen against 0.7.0.

use_empirical_variance — False everywhere

The manuscript states that empirical variance is disabled by
default. This was already true for the recommended
kompot.de() path via
GPSettings(use_empirical_variance=False) and for the
de_config_template.yaml. Four other publicly-exposed entry
points were inconsistent (default True):

• kompot.differential.DifferentialExpression.__init__
• kompot.differential.expression_model.ExpressionModel.__init__
• kompot.anndata.differential_expression.compute_differential_expression
  (deprecated wrapper)
• kompot.anndata.smooth.compute_smoothed_expression
  (deprecated wrapper)
• kompot.anndata.smooth.smooth_expression (gp=None
  fallback)
• kompot/cli/templates/smooth_config_template.yaml

All now default to False to match the recommended path and
the manuscript. Docstrings and the resource-estimation comment
for combined-covariance memory are brought in line. Code that
relied on empirical variance must now opt in explicitly via
GPSettings(use_empirical_variance=True) / the keyword arg.

0.7.0 → 0.8.0

pyproject.toml and kompot/version.py bumped to 0.8.0;
new CHANGELOG.md section under ## [0.8.0] - 2026-05-28
records the three behavior changes above with migration notes.
No release tag is cut by this PR.

Tests

tests/test_audit_fixes.py adds three focused regression
suites that pin the post-0.8.0 invariants:

• Mahalanobis denominator: monkeypatches
  compute_mahalanobis_distances to capture the covariance
  kwarg from a controlled-kernel fit (cov1 = 2I, cov2 = 3I)
  and asserts equality with 5 * I. The pre-0.8.0 code would
  have produced 2.5 * I.
• DA PTP: replicates the exact ln_ptp formula and asserts it
  equals the closed-form Φ(−|z|), both directly and through a
  full DifferentialAbundance.fit().predict() round-trip.
• use_empirical_variance default: introspects every public
  entry point's signature and asserts the default is False;
  also parses both CLI YAML templates.

tests/test_empirical_variance.py::TestAnnDataEmpiricalVariance::test_default_is_on
was the only pre-existing test pinning the old buggy default;
it has been flipped and renamed test_default_is_off.

The reference-output generator (tests/generate_reference_outputs.py)
writes to /tmp/kompot_reference/ and is developer scratch —
not checked in, not part of CI — so no committed snapshots
need regeneration. The Mahalanobis-approach tests
(tests/test_mahalanobis_approaches.py) compare in-memory vs
disk-backed implementations against each other rather than
against pinned numeric values, so they are insensitive to this
scale shift.

Full local pytest run (kompot 0.8.0 venv): 524 passed +
1 updated pinning test + new audit-fix suite (10 tests, all
green).