fix: remove undeclared SciPy dependency from core Hurst feature #96

Core Hurst feature must not require undeclared SciPy through Numba least squares

Affected verified source: main0e7865273cd462c507de02a9d3b64de7c6cbb099, installed via the exact-tree qualified development wheel. Python3.12.11, NumPy2.5.4, Numba0.68.0, Polars2.0.0. Fresh17-package base installation has no SciPy.

Public workflow
import numpy as np
import polars as pl
from datetime import datetime, timedelta
from ml4t.engineer import compute_features
close = 100 + np.cumsum(np.random.default_rng(42).normal(0, .1, 300))
data = pl.DataFrame({'timestamp': [datetime(2025,1,1)+timedelta(hours=i) for i in range(300)], 'open': close, 'high': close+1, 'low': close-1, 'close': close, 'volume': np.ones(300)})
compute_features(data, ['hurst_exponent'], calendar=None)

Observed
ImportError: scipy0.16+ is required for linear algebra, from Numba np.linalg.lstsq used by hurst_exponent_nb. The actual fresh core environment reproduces it without mocking or changing imports. Full shared3.15 qualification independently found the same dependency problem in Hurst and registry/public property workflows. This is not specific to prereleases.

Expected and proposed focused repair
Core features work with declared core dependencies. Preserve the existing R/S definition, lag selection, warmup and intercept-inclusive OLS slope. Compute that scalar slope by centered sums instead of calling Numba LAPACK/SciPy; do not add a heavyweight runtime dependency or change the numerical definition. Keep the owning Python3.15 qualification PR93 separate, with this repair as a concrete source prerequisite.

Acceptance
Fail on the actual installed base core package; pass on a clean installed repair wheel without SciPy. Permanent eager/lazy public-workflow tests must expose missing SciPy and preserve checked numerical controls. Run focused tests, all applicable hosted supported source/wheel/platform/docs/security/package gates and ordinary reviewed-green merge. No release/deployment is requested. Resource coordination excludes heavy local full suites.

AI assistance: Codex inspected failed hosted logs, reproduced the bug in an actual minimal installed environment, and prepared the focused source/test repair and evidence. The maintainer remains accountable for the mathematical definition and checks.