feat(family-12): add 13 Statistik/Regression indicators (#51)

* feat(family-12): add 13 Statistik/Regression indicators

Brings the Price Statistics family to 20 indicators (7 → 20) and the
total catalogue to 84 (71 → 84). Every indicator ships in the Rust
core plus Python, Node, and WASM bindings with full streaming ↔ batch
parity, fuzz coverage, and benches.

Scalar (f64 → f64):
- Variance, CoefficientOfVariation: rolling population variance and
  its dimensionless ratio with the mean. O(1) updates.
- Skewness, Kurtosis: rolling Pearson skewness and excess kurtosis,
  derived from running sums of x, x², x³, x⁴ via the binomial
  identities — also O(1) per bar.
- StandardError, DetrendedStdDev: standard error of estimate (n − 2)
  and population StdDev (n) of OLS residuals, sharing the LinReg
  O(1) sliding sums.
- RSquared: coefficient of determination of the rolling OLS fit; the
  trend-quality filter, clamped to [0, 1].
- MedianAbsoluteDeviation: robust dispersion estimator; O(period log
  period) per emission via two in-place sorts of a reusable scratch
  buffer.
- Autocorrelation(period, lag): rolling lag-k Pearson autocorrelation.
- HurstExponent(period, chunks): R/S-analysis trend-persistence
  estimator clamped to [0, 1].

Pair indicators (Input = (f64, f64)):
- PearsonCorrelation: rolling cross-series Pearson, O(1).
- Beta: rolling OLS slope of asset vs. benchmark (CAPM).
- SpearmanCorrelation: rolling rank correlation with mid-rank tie
  handling; O(period log period).

Touchpoints:
- crates/wickra-core: 13 new indicator modules + mod.rs / lib.rs
  re-exports.
- bindings/python: pyclasses + add_class registration + __init__.py
  import & __all__ updates. The pair indicators expose
  update(x, y) and batch(x, y) over two equally-sized numpy arrays.
- bindings/node: scalar indicators via node_scalar_indicator! macro;
  pair indicators via new node_pair_indicator! macro; explicit
  structs for Autocorrelation and HurstExponent (two-arg ctors).
  index.js extended with the new exports.
- bindings/wasm: scalar wrappers via wasm_scalar_indicator!; pair
  wrappers via new wasm_pair_indicator! macro.
- fuzz: every scalar drove through the generic helper; pair
  indicators stress-tested by pairing adjacent samples of the fuzz
  input.
- Python tests (test_new_indicators.py): added to SCALAR
  parametrisation, plus algebraic reference values
  (variance of [2,4,6] = 8/3, MAD ignoring outlier = 0, monotone
  non-linear Spearman = 1, two-to-one Beta = 2, etc.) and a
  streaming-vs-batch test for the pair indicators.
- Node tests (indicators.test.js): extended the scalar factories
  map and added a pair-indicator section with the same algebraic
  reference values.
- crates/wickra/benches: bench_scalar entries for all 10 single-
  input new indicators.
- README: counter 71 → 84; Price Statistics family-table row
  expanded with the 13 new indicators.
- CHANGELOG: Unreleased section documents the family addition.

Wiki drafts (ghost-ignored, manual sync to wickra.wiki at release
time): indicator-ideas/families/wiki/family-12-statistik-regression/
contains 13 deep-dive pages plus _Sidebar / Indicators-Overview /
Warmup-Periods / Home fragments for the curator merge.

cargo check --workspace --all-features: clean.

* fix(family-12): remove unreachable defensive guards in hurst_exponent

The three guards (m < 2 continue, end > buf.len() break, denom == 0.0
return) are by-construction unreachable given the constructor invariant
period >= 2 * chunks: m = period / k for k in 1..=chunks always
satisfies m >= 2 and end = (c+1) * m <= k * m <= period = buf.len(),
and m_1 = period and m_2 = period / 2 are always distinct so the slope
denominator is strictly positive. Removing them brings codecov/patch
back to 100%.

bindings/python/tests/test_new_indicators.py

@@ -95,6 +95,17 @@ SCALAR = [
     (ta.AdaptiveCycle, ()),
     (ta.SineWave, ()),
     (ta.FAMA, (0.5, 0.05)),
+    # Family 12 — Statistik / Regression
+    (ta.Variance, (20,)),
+    (ta.CoefficientOfVariation, (20,)),
+    (ta.Skewness, (20,)),
+    (ta.Kurtosis, (20,)),
+    (ta.StandardError, (14,)),
+    (ta.DetrendedStdDev, (14,)),
+    (ta.RSquared, (14,)),
+    (ta.MedianAbsoluteDeviation, (20,)),
+    (ta.Autocorrelation, (20, 1)),
+    (ta.HurstExponent, (40, 4)),
 ]
 
 
@@ -908,6 +919,111 @@ def test_z_score_reference():
     assert out[1] == pytest.approx(1.0)
 
 
+# --- Family 12: Statistik / Regression reference values ------------------
+
+
+def test_variance_reference():
+    # Variance(3) of [2, 4, 6]: mean 4, variance (4 + 0 + 4) / 3 = 8/3.
+    out = ta.Variance(3).batch(np.array([2.0, 4.0, 6.0]))
+    assert math.isnan(out[1])
+    assert out[2] == pytest.approx(8.0 / 3.0)
+
+
+def test_coefficient_of_variation_reference():
+    # CV(3) of [2, 4, 6]: sd / mean = sqrt(8/3) / 4.
+    out = ta.CoefficientOfVariation(3).batch(np.array([2.0, 4.0, 6.0]))
+    assert out[2] == pytest.approx(math.sqrt(8.0 / 3.0) / 4.0)
+
+
+def test_skewness_symmetric_window_is_zero():
+    # Symmetric window has zero Pearson skewness.
+    out = ta.Skewness(5).batch(np.array([-2.0, -1.0, 0.0, 1.0, 2.0]))
+    assert out[4] == pytest.approx(0.0, abs=1e-9)
+
+
+def test_kurtosis_two_point_distribution_minimum():
+    # Alternating {-1, 1} has m4/m2² = 1, so excess kurtosis = -2.
+    out = ta.Kurtosis(4).batch(np.array([-1.0, 1.0, -1.0, 1.0]))
+    assert out[3] == pytest.approx(-2.0, abs=1e-9)
+
+
+def test_standard_error_perfect_line_is_zero():
+    # Residuals are zero on a perfectly linear series.
+    out = ta.StandardError(5).batch(np.linspace(1.0, 20.0, num=20, dtype=np.float64))
+    finite = out[~np.isnan(out)]
+    assert np.allclose(finite, 0.0, atol=1e-9)
+
+
+def test_detrended_std_dev_perfect_line_is_zero():
+    out = ta.DetrendedStdDev(5).batch(np.linspace(1.0, 20.0, num=20, dtype=np.float64))
+    finite = out[~np.isnan(out)]
+    assert np.allclose(finite, 0.0, atol=1e-9)
+
+
+def test_r_squared_perfect_line_is_one():
+    out = ta.RSquared(5).batch(np.linspace(1.0, 20.0, num=20, dtype=np.float64))
+    finite = out[~np.isnan(out)]
+    assert np.allclose(finite, 1.0, atol=1e-9)
+
+
+def test_median_absolute_deviation_ignores_single_outlier():
+    # 9 equal values + 1 huge outlier: MAD is still 0 (more than half agree).
+    prices = np.array([5.0] * 9 + [1000.0], dtype=np.float64)
+    out = ta.MedianAbsoluteDeviation(10).batch(prices)
+    assert out[9] == pytest.approx(0.0, abs=1e-12)
+
+
+def test_autocorrelation_alternating_series_negative():
+    # ±1 alternating: lag-1 ACF must be strongly negative.
+    prices = np.array([-1.0 if i % 2 == 0 else 1.0 for i in range(20)], dtype=np.float64)
+    out = ta.Autocorrelation(10, 1).batch(prices)
+    assert out[-1] < -0.5
+
+
+def test_hurst_exponent_trending_above_half():
+    # A clean monotone ramp is the textbook persistent series.
+    prices = np.arange(200, dtype=np.float64)
+    out = ta.HurstExponent(100, 4).batch(prices)
+    assert out[-1] > 0.5
+
+
+def test_pearson_correlation_perfect_positive_is_one():
+    x = np.arange(10, dtype=np.float64)
+    y = 2.0 * x + 3.0
+    out = ta.PearsonCorrelation(5).batch(x, y)
+    assert out[-1] == pytest.approx(1.0, abs=1e-9)
+
+
+def test_beta_perfect_two_to_one():
+    benchmark = np.arange(10, dtype=np.float64)
+    asset = 2.0 * benchmark
+    out = ta.Beta(5).batch(asset, benchmark)
+    assert out[-1] == pytest.approx(2.0, abs=1e-9)
+
+
+def test_spearman_correlation_monotone_nonlinear_is_one():
+    # y = x^3 is monotone non-linear; Spearman = 1 (Pearson would not be).
+    x = np.arange(1.0, 11.0, dtype=np.float64)
+    y = x**3
+    out = ta.SpearmanCorrelation(5).batch(x, y)
+    assert out[-1] == pytest.approx(1.0, abs=1e-9)
+
+
+def test_pair_indicators_streaming_matches_batch():
+    rng = np.linspace(0.0, 10.0, num=60, dtype=np.float64)
+    x = np.sin(rng) + 0.1 * rng
+    y = np.cos(rng * 0.3) + 0.05 * rng
+
+    for cls, args in [(ta.PearsonCorrelation, (14,)), (ta.Beta, (14,)), (ta.SpearmanCorrelation, (14,))]:
+        batch = cls(*args).batch(x, y)
+        streamer = cls(*args)
+        streamed = []
+        for i in range(x.size):
+            v = streamer.update(float(x[i]), float(y[i]))
+            streamed.append(math.nan if v is None else float(v))
+        assert _eq_nan(batch, np.array(streamed, dtype=np.float64)), f"{cls.__name__} mismatch"
+
+
 # --- Family 10 — Ehlers / Cycle ---