NEWS

---

magmaanlab 0.2.2

• Restored fit/model caches now check native pointer addresses rather than platform-specific display strings, rebuilding safely on macOS and Windows.

• Linux fallback linking to nloptr now records a relocatable library name and a sibling-library runpath instead of an absolute build-machine dependency.

• The small-N singular FIML H1 witness now has an independent unbounded- likelihood gate; its numerical stopping flag is outside live parity. Converged control moments and starts retain strict comparisons.

• Portable package checks now run on Windows, macOS and Linux with current dependencies. Live lavaan comparisons accommodate 0.7-3; the pinned lavaan-0.7.2 preset retains its own oracle.

• Windows now generates the complete vendored build configuration and locates NLopt through Rtools or explicit library flags. macOS numeric model conversion uses a locale-independent fallback where floating-point charconv is unavailable.

• Frontier PSD fitters accept psd_factor = "square" for full square covariance factors. Triangular factors remain the default; projected start covariances and endpoint audit rules are unchanged.

• Complete-data endpoint audits now support simple equality-constrained std.lv charts, including loading and latent-mean ties across factors or groups. Released marker ties preserve the constrained model dimension. Automatic constrained starts retain the existing normalized-unit fallback.

• Lab fits and prepared datasets retain read-only deletion provenance: rows supplied, used and deleted per group, exclusive reasons, and applied policy. FIML/ML2S record no listwise deletion; summary-only historical counts are marked unknown. Existing missing-data and group-label refusals are preserved.

• Complete-data GLS fits now use PORT by default, retaining the layered start and information-refined coordinates. Estimates at the same optimum are unchanged; iteration counts and runaway paths may differ. ML keeps L-BFGS.

• Complete-data ML and GLS fits with standardized extent above 10 are no longer reported converged when their supported std.lv endpoint Newton audit fails. Diagnostics record the trigger and unsupported transports; estimates are unchanged.

• FIML fit-measure calls share fit-owned saturated ingredients and baseline fits; policy points are reused by lab Wald intervals. Returned estimates and intervals are unchanged, and modified or serialized fits rebuild the cache.

• Added lab-only fit_index_intervals_misspec() for continuous ML/FIML RMSEA, CFI and TLI: discrepancy and square-root Wald variants, joint baseline influence, observed-coordinate second-order variance and typed refusals. This is an uncalibrated study candidate; ordinary defaults are unchanged.

• Fixed covariance-only complete-data ML policy fit measures in the lab: score rows now use profiled group means, so shifting observations does not change the Takeuchi correction. Ordinary magmaan fits always include mean structure and were unaffected. Fit-measure saturated EM moments now use a tighter tolerance to preserve unit invariance of the FIML trace correction.

• Continuous DWLS weight blocks now retain numeric diagonal entries through native preparation, fitting and saved-fit replay. Supplied W accepts numeric diagonals or matrices (and grouped lists); full weights keep their matrices. Weight entries are W, not whitening factors; inference recipes are unchanged.

magmaanlab 0.2.1

• Nested tests now refuse boundary_nesting only when a required tangent loses rank, rather than because the fitted factor covariance is improper or singular. PSD-boundary comparisons compute under the same interior-population assumption as global tests and retain psd_boundary; anova() also flags computed table rows. A 500-replicate weak-factor null check at each of N=100/200 gave score rejection 5.6%/6.0% and LR 6.4%/7.0% at nominal 5%, including 10.0%/1.8% improper null fits (SB and PEBA4, one restriction).

• Imported lavaan tables now reproduce the syntax-built variable inventory. This may change starts and iteration counts for models with regressions and mean structure.

• The lavaan-0.7.2 marker component now also covers FIML and ordinal/mixed DWLS, using each route's saturated H1 covariance. Switched fits replay the selected fitting attempts and revert when rejected; marker = "default" keeps the requested markers.

• Fully fixed models with latent regressions or higher-order factors and fixed means now return their implied moments without crashing in the stationarity audit.

• Explicit quadratic references (including summary/anova SB and PEBA4) now clip roundoff-negative spectrum entries using the projected-score validator; materially negative spectra still raise a numeric error.

• Ordinary magmaan(estimator = "ML") on complete all-ordinal data now uses association ML, exact Stage-1 IJ covariance with observed sensitivity, All global/nested references, and robust modification indices/equality releases. DWLS remains the ordinal default. Global tests can be liberal in finite samples; see the association-ML help caveat and decisions/08 evidence.

• Identification reports classify scale, location and rotation freedoms, enforce fixed entries and linear equalities, and expose per-direction suggested fixes. Remaining null dimensions are reported as information deficits.

• augment_model_spec(model, categories) augments ordinal specifications from declared category counts with missing threshold starts. Ordinal and mixed augmentation now builds rows column by column; prepared models use this helper.

• The lavaan-0.7.2 FIML preset now uses lavaan's scale-dependent H1 covariance ridge and retains stalled EM moments. fit$fitting$h1 records H1 convergence and update/repair counts. Converged H1 moments retain strict oracle agreement; stalled H1 endpoints are outside the compatibility contract. Native FIML H1 defaults are unchanged.

• Lab optimizer controls now accept opt-in objective and objective_downward scaling for continuous fixed-weight LS and ML2S Stage-2 LS, using the fitting GN diagonal. Defaults and audit rules are unchanged. ML/direct FIML and residual backends reject these modes as Unsupported. Ordinal and mixed routes currently ignore all coordinate scaling modes pending TASK-33.10.5.6.

• structural_identification() checks a specification or reuses a prepared model's data-free report. Prepared fits reuse construction-time identification and express cached null directions at the deciding probe point. Lab fits remain available for unidentified models, with a failed verdict.

• The lavaan-0.7.2 preset now adapts weak markers for complete-data ML. marker = "default" keeps the requested identification. Fits retain the actual fitted specification and a switch/revert table; nested inference explicitly refuses switched marker coordinates. FIML and DWLS adaptation follows separately; their preset currently reports marker = "default".

• Fits using convergence = "lavaan-0.7.2" (including the preset) report lavaan's post-estimation check separately from convergence. The six flags are available in fitting$post_check on the lab fit; ordinary fits expose them through as_lab_fit(). This adds no warning or inference restriction.

• Continuous ULS, DWLS and WLS/ADF fits now warn when within-group observed variances differ by more than a factor of 1000, with advice to rescale. The C++ diagnostic also covers ML2S Stage-2 and continuous mixed-LS moments; ML, GLS, FIML, NT/DLS weights and ordinal-only fits do not emit this advice. Lab fits expose observed_variance_ratio and numerical_scaling_message.

• A structurally unidentified model is never converged = TRUE, under any rule: a data-free rank check of the moment Jacobian (fit$diagnostics$identification) fails the verdict and names the null directions. A local Newton pass no longer certifies a scale ridge.

magmaanlab 0.2.0

Lab API and migration

• U-factor builders require explicit bread; score_quadratic() requires an explicit meat matrix. Observed bread supports pseudo-true covariance and nested restrictions; complete-data global GOF can use expected geometry.
• frontier_profile_lrt_parameter_gmm() and frontier_profile_lrt_ci_parameter_gmm() remove ij_weight and dls_a: the weight recipe comes from the fit.
• Estimated-weight inference follows recorded moment and Stage-2 weight recipes, including continuous DWLS and DLS mixing weights. Supplied-W and ordinal NT/DLS fits without a supported recipe have typed refusals; explicit fixed-weight inference remains available.
• WLS procedures use recorded fit$W; optional weight must match it or a positive multiple. ML2S RBM and case influence likewise use recorded Stage-2 weights and DLS mixing weights, and validate explicit overrides.

Inference policy and evidence

• policy_inference() and policy_nested() compose ML/FIML empirical-score, observed-bread inference, with observed score sensitivity and an expected metric. Nested LR calibration handles restricted, misspecified means. Components retain separate typed unavailable reasons.
• All-ordinal and complete mixed DWLS policy composers use exact empirical first-stage influence, estimated-weight IJ covariance and global/nested fit-function tests. Nested geometry uses the observed Hessian and supports parameter restrictions and moment-nested Wu–Estabrook threshold invariance. Nested DWLS score tests remain unavailable.
• Policy defaults are PEBA4 for ML/FIML and All for DWLS global/nested tests. SB/PEBA4 DWLS comparators remain available. Exact-first-stage and reference-law studies support the DWLS choice across ordinal, mixed, threshold-invariance and textbook designs; finite-sample size varies across settings.
• robust_ordinal_ij(..., first_stage = "exact") exposes exact empirical sampling rows and Gamma. Its default remains "opg"; the ordinary policy selects exact influence, without changing fitting weights.

Test references

• calibrate_quadratic() accepts std/sb/ss/mv/scaled_f/all/pall and eba/peba. Block counts at or above df give singleton blocks. Policy ML/FIML results label reference = "peba4" and retain p_sb/sb_scale comparators.

Modification indices

• policy_modification_indices() composes robust one-df indices and affine equality releases for ML/FIML and all-ordinal/mixed DWLS, including absent regressions among variables already in equations. policy_mi_refit() refits a candidate from its embedded null using the nested policy.
• ML/FIML robust MI and releases use observed nuisance sensitivity with an expected-information metric and EPCs. Negative observed release curvature alone no longer removes candidates; score tables retain reason and detail for numerical failures.
• Robust LS indices and releases support observed sensitivity and estimated weights for continuous ULS/GLS/DWLS/WLS/DLS and ordinal ULS/DWLS/WLS. Complete mixed DWLS/WLS adds exact first-stage estimated-weight influence. Expected sensitivity and fixed weights remain explicit comparators.
• Validated caller gamma or per-group NACOV blocks support complete ML, continuous LS and categorical LS indices/releases. Supplied Gamma requires explicit fixed-weight inference; it cannot provide casewise weight influence. Continuous-LS covariance and ML/LS profile-LRT adapters also accept Gamma.
• ML2S indices and score tests support every Stage-2 weight. mi is the naive Stage-2 statistic (mi_type = "naive_stage2"); mi.scaled uses Stage-1 covariance and, when requested, Stage-2 weight influence.

Standardized estimates

• standardized_rows(fit, vcov) provides std.lv/std.all estimates, delta-method SEs and endogenous R-squared, including fixed markers, groups and ordinal parameters. Missing covariance preserves estimates with typed missing SEs; defined standardized scales are unsupported.

Fit measures

• policy_fit_measures() provides robust point indices and per-index reasons with discrepancy, trace and df details for ML/FIML, continuous ULS, ML2S-NT and ordinal/mixed DWLS. It fits a matching independence baseline and corrects residual points for sampling bias. Policy intervals are not yet provided.
• Ordinal/mixed misspecification fit-measure wrappers accept exact first-stage influence; OPG remains their default and generalized-df TLI is retained.
• Lavaan-compatible fit measures include standard, scaled and robust families, intervals and close-fit p-values. FIML ML/MLR uses tighter H1 EM precision and is validated against lavaan.

Lavaan compatibility

• The lavaan 0.7.2 fitting preset supports continuous ML/FIML and all-ordinal or complete mixed DWLS, including delta/theta and grouped equal loadings, unrestricted covariance and supported linear equalities. Prepared estimate() accepts DWLS fitting options through the same engine as fit_model().
• convention_inference() and convention_nested() compose named bundles on retained fits with component-level unavailable reasons. Fitting presets and reporting bundles remain separate choices. Complete-data MLR uses the H1-minus-H0 trace; ordinal reporting uses per-group n minus one.
• Complete mixed DWLS supports WLSMV covariance, Wald intervals, scaled-shifted global and Satorra-2000 nested tests. Missing mixed observations, nonlinear constraints and finite bounds remain outside the mixed fitting preset.

Lab building blocks

• Association-ML covariance, global/nested tests, modification indices and linear releases use exact empirical Stage-1 influence and observed sensitivity. Tests expose All/SB/PEBA4; MI uses robust one-df laws. Ordinary association inference remains gated by policy calibration. Missing data, mixed indicators, nonlinear constraints, covariance boundaries and penalties are unsupported; moment nesting is refused.
• ML robust SE and score/MI meats default to exact casewise likelihood projections, including restricted means and group score means. moments = "auto" selects this recipe; structured/unstructured comparators remain explicit. Global GOF primitives retain their named conventions.
• refit_from_null() returns a new larger-model fit from a verified embedded restricted estimate, with native convergence diagnostics.
• Robust procedures with estimated-weight channels default to estimated-weight inference; MI selects that default by estimator so ML/FIML need no explicit FALSE. Explicit FALSE retains the fixed-weight comparator.
• Scalar profile tests and intervals default to the misspecification-scaled reference. Empirical laws require raw data; ordinary and robust-scaled comparators remain selectable.
• Global score flips use observed sensitivity for MAR FIML and expected sensitivity for complete-data correct-model GOF.
• Reliability with NULL Gamma derives empirical Gamma from raw data and errors without them. parameter_covariance() derives empirical meat from retained observations; meat = "model" requests inverse information.
• robust_nested_lrt() composes ML/DWLS policy tests and refuses unsupported routes. Explicit methods and Satorra aliases retain compatibility conventions.
• vcov(regime = "sandwich_ij") provides all-ordinal estimated-weight covariance; the default remains fixed-weight sandwich covariance.
• SAM, case-influence and FMG/semTests routes retain diagnostic/comparator conventions. infer_continuous_ls_robust() requires explicit fixed_weight = TRUE for recorded data-estimated weights.

Mplus frontend

• mplus_model() imports whole continuous single/grouped inputs, growth models, MODEL CONSTRAINT with NEW/equalities/DO loops, and MODEL INDIRECT. Specs retain source notes, numeric group order, labels and rebuildable model metadata. SQRT, PHI/pnorm and LOG10 extend expressions; inequalities remain unsupported.
• Binary/ordinal outcomes support threshold ranges/labels, DELTA scales, THETA residuals and grouped CONFIGURAL/SCALAR models. All-ordinal DWLS uses Mplus defaults; mixed, conditional and categorical ML routes report scope refusals.
• Complete observed-X moment mentions select joint random-X models. Partial mentions give the required completion statement. Group-specific role changes are refused, and unlisted group codes require filtering.
• mplus_data() reads free/fixed individual and summary data, missing flags and FILE/NGROUPS groups. It reports observation counts and dropped grouping codes; summary N-1 covariances are rescaled to N for ML. Terminal DOS Ctrl-Z markers are accepted; embedded markers give a DA01 diagnostic.
• Unsupported Mplus inputs report specific scope reasons and required edits. Auxiliary NEW coordinates are supported in barrier gradients and moment curvature; noniterative CFA and profiled SNLLS refuse them explicitly.

Fixes

• Equality-reduced ordinal parameter counts now give lavaan-compatible df for grouped threshold/loading and label constraints.
• Unequal-group all-ordinal WLSMV covariance now preserves group-weighted sandwich geometry and the reporting denominator.
• Exact ordinal/mixed first-stage influence and Gamma-diagonal movement use affected score blocks, category counts and sparse derivatives, reducing cost.
• Mixed MI now uses the equality-release criterion scale, removing an extra factor two without changing EPCs or fitting weights.

magmaanlab 0.1.0

First versioned dependency for the magmaan 0.1.0 simulation prerelease. The lab retains its methods-development API, including the EQS constructor.

Development is unfinished: remaining bugs and incomplete coverage of the main estimators and inferential procedures remain to be resolved and validated. Version 0.1.0 provides a versioned simulation snapshot, not a finished release.

High-level fit reconstruction retains defined parameter estimates using the shared C++ evaluator, independently of covariance inference.

• Generic lab covariance and nested score sensitivity now default to observed information with empirical meat; categorical and GLS/WLS covariance uses estimated-weight IJ. Explicit expected/fixed-weight choices remain comparators.
• FIML fit measures request robust inference by default. Noniterative NT defaults remain documented comparators in the postponed development area.