spb/metrika Public

- ZQGradientBoosting: exact-greedy trees cloning xgboost's algorithm —
  gain 1/2[GL2/(HL+l) + GR2/(HR+l) - G2/(H+l)] - gamma, leaf -G/(H+l),
  midpoint splits between consecutive distinct values, missing rows
  default left, pre-sorted feature indices; squared loss, deterministic
  (no subsampling)
- engine: 'boost y x…, rounds(#) [eta() maxdepth() lambda()]' reporting
  training R2/RMSE with an in-sample caveat; model stored in the
  estimation state (Kind.boost) so predict routes through the trees
  (missing features follow the default direction); margins and GLM
  statistics refused after boost
- validation: per-observation prediction parity with R xgboost 3.2
  (exact method, base_score = mean) at 1e-4 (xgboost is float32
  internally); a stump finds the exact midpoint split with lambda 0;
  training loss decreases monotonically in rounds
- manual entry + coverage test
- 115 tests green (swift test and xcodebuild with GPU suites)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- ZQElasticNet: cyclic coordinate descent with residual updates,
  glmnet-convention objective ((1/2n)RSS + lambda(alpha*l1 +
  (1-alpha)/2*l2)), internal predictor standardization (1/n variance),
  unpenalized intercept, coefficients reported on the original scale;
  lambdaMax helper
- matched glmnet's gaussian y-standardization quirk deliberately: the L1
  penalty is invariant to it but the effective ridge penalty scales by
  1/sd(y) — without this, alpha<1 fits diverge from glmnet by ~10%
- engine: 'elasticnet y x…, lambda(#) [alpha(#)]' and 'lasso' (alpha
  fixed at 1); missing lambda() errors with the data's lambda_max as a
  hint; predict works afterwards, margins refuses (no VCE)
- fixtures: glmnet 5.0 at thresh 1e-15 over deliberately correlated
  regressors; coefficients match at 1e-6 (documented tolerance for
  penalized iterative solvers) and the selection pattern (which
  coefficients are exactly zero) matches exactly
- ZQCoefficient gains a public initializer
- 104 tests green (swift test and xcodebuild with GPU suites)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- ZQOLSResult/ZQGLMResult expose the full covariance matrix (column-major
  k x k aligned with coefficients); IV passes it through
- EstimationState carries vce, inference df, and the estimation-sample
  design (GLMs only — AMEs need it)
- margins, dydx(varlist): OLS/IV effects are the coefficients with their
  SEs; GLM average marginal effects with analytic delta gradients
  (logit p(1-p)(1-2p), probit -xb*phi, poisson exp) over the estimation
  sample; t or z inference per model kind
- factor/interaction dydx rejected with a clear message (discrete-change
  margins later); continuous terms of factor models work
- R fixtures mirror the exact formulas at the converged coefficients;
  logit/poisson AME and delta SE match at 1e-10
- 98 tests green (swift test and xcodebuild with GPU suites)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- parser: (endog = instruments) varlist groups -> ZQIVSpec; digit-led
  sub-commands (2sls) reassembled from number+identifier tokens via
  column adjacency
- ZQFixedEffects: within estimator with Stata conventions (add-back
  means, reported _cons, df = N-K-G), within R-squared, panel-clustered
  VCE with G/(G-1) and t on G-1 df; v0.2 restriction: cluster variable
  must equal the panel variable
- ZQIV: 2SLS via thin-Q projection of the instrument matrix (Z'Z never
  formed), residuals from original regressors, Stata 'small' inference,
  classical/HC1/cluster VCE built on projected regressors
- engine: xtreg (requires xtset + fe), ivregress 2sls with dedicated
  listwise deletion across depvar/exog/endog/instruments; shared
  coefficient-table renderer extracted
- fixtures: z1/z2 instrument columns (drawn after existing draws, earlier
  golden values bit-identical), manual within/2SLS algebra in R
- 75 tests green (swift test and xcodebuild with GPU suites)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- ZQGLM: Fisher-scoring IRLS through the LAPACK QR path (X'WX never
  formed), converged to |Δll| < 1e-13; classical, robust (HC0 score
  sandwich, Stata ML convention), and cluster (G/(G−1)) VCE; LR/Wald
  chi2, McFadden pseudo-R²; information matrix re-evaluated at the
  converged beta (R's vcov carries last-iteration weights and is only
  ~1e-6-accurate by its own stopping rule — fixtures compute expected
  information at the optimum explicitly)
- Distributions: regularized incomplete gamma, chi-square CDF/p-value,
  normal quantile
- ZQCorrelate: Pearson matrix on listwise-complete data
- Engine: logit/probit/poisson tables (z, P>|z|, LR chi2, pseudo R²),
  one-way and two-way tabulate with totals, Stata-style lower-triangle
  correlate
- Fixtures: binomial/poisson outcomes drawn after existing draws (earlier
  golden values bit-identical); glm at epsilon 1e-12
- 56 tests green (12 new)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- MetrikaKit SPM package: ZQParser, ZQPlanner, ZQEngine, ZQData, ZQStats,
  ZQGPU, ZQGraphics, ZQPlugins (Swift 6, strict concurrency)
- ZQL parser: lexer, Pratt expressions, factor variables, prefix commands,
  column-cited errors with Levenshtein verb suggestions
- ZQData: DuckDB-backed load/save (parquet, csv, json, arrow)
- ZQStats: OLS via LAPACK QR, HC0-HC3 and cluster-robust SE, summarize,
  t/F distributions accurate in the far tails
- ZQGPU: Philox4x32-10 reference RNG, counter-addressable bootstrap
- ZQEngine: session actor with use/save/gen/replace/drop/keep/summarize/
  regress/count/list/graph/bootstrap/set seed/xtset/log
- SwiftUI app (xcodegen): console with history, variables sidebar,
  Swift Charts plots; sandboxed + hardened runtime entitlements
- Tests: 44 green (parser golden, R fixtures at 1e-10, Philox KAT,
  end-to-end engine); Tests/Fixtures/generate.R; Tests/Bench harness
- scripts: check_headers.sh + pre-commit hook, make_icns.sh, release.sh

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>