feat: add multivariate temporal inference

[jousudo]

Summary

• add stdlib-only reduced-form VAR fitting with defensive coefficient, residual, and innovation-covariance accessors
• select lag order with common-sample AIC, BIC, and Hannan-Quinn scores, plus residual cross-autocorrelation diagnostics
• add conditional multivariate Granger F-tests that retain all supplied histories and remove only the proposed cause lags
• add moving-block and Politis-Romano stationary bootstraps, including context-aware and Gaussian-effect variants
• bound VAR design memory and reject adversarial maxLags before allocation
• document v0.15.0, the relevant data regimes, and the boundary between predictive VAR and structural/interventional causality

Validation

• independent base-R oracle for coefficients, innovation covariance, information criteria, F statistic, and p-value
• synthetic known-VAR(2), observed-confounder, direct-effect, lag-selection, underfit-diagnostic, serial-dependence, cancellation, and hostile-input tests
• go test ./...
• go test -race ./...
• statement coverage: 91.3% (gate: 90%)
• go vet ./...
• go build ./...
• staticcheck ./...
• govulncheck ./... (0 reachable vulnerabilities)
• 10-second fuzz smoke: 179,631 executions

Statistical scope

This is predictive reduced-form VAR/Granger inference, not SVAR identification. It controls only histories that the caller supplies. Hidden drivers, unit roots, cointegration, regime changes, nonlinear dynamics, irregular sampling, residual autocorrelation, and uncorrected searches over many pairs remain explicit limitations rather than silently asserted capabilities.