NEWS

fastphylosig 0.3.5

Correct the ARD ancestral-state downpass to use the child contribution P * L_child under the row-generator convention. Impossible parent states with zero posterior and zero contribution send zero messages. Shared ACE fitting, transition matrices, normalization and Delta MCMC kernels are reused.
ARD ancestral probabilities and downstream Delta estimates, local/overall summaries, tests and subsequent RNG may change. The ER downpass retains its previous arithmetic; K, lambda, D and ER Delta keep their result/RNG contracts.
Scale discrete ACE rate standard errors by the information matrix for the negative twice log likelihood. This corrects both ER and ARD standard errors; fitted rates are unchanged and Delta does not use these standard errors.
Add independent latent-state enumeration, analytic two-state transitions and information-curvature regressions. Reference-package checks remain for valid point-likelihood comparisons, not for the corrected ARD probabilities or SEs.
Cache exact eigenvalue exponent columns for eligible small three-state ARD fits. Reset entries on reconfiguration, account for hash-node storage and publish only complete columns. The normal R path configures its immutable workspace once; its earlier reconfiguration boundary issue was not triggered.
Reuse Armadillo's existing GEMV dispatch in pre-sized scratch and remove repeated vector preparation/copies. Exact exponent evaluation, optimizer steps, MCMC, corrected results and RNG contracts are retained. Cache lifecycle, memory bounds, FP controls and full-call equivalence have regression checks.
Retain public interfaces, defaults and statistical scope: K's BM reference, lambda's approximate common-local-crown reference, experimental D inference, and descriptive Delta with experimental heterogeneity tests outside the recommended workflow. Historical ARD calibration does not validate the corrected estimator. See docs/release-0.3.5 for impact and reanalysis guidance.

fastphylosig 0.3.4

Consolidate fast_hetero() clade selection, coverage, trait-wise missing values, local/overall summaries and separate signal/reference tests. K uses a BM reference; lambda uses an approximate common-local-lambda Gaussian reference. D's count-conditioned ER test remains experimental. Delta is recommended for local/overall description and MCMC diagnostics; its explicit experimental heterogeneity options remain available.
Explicit nodes override automatic clade-size bounds while min_tips continues to enforce the observed-species threshold. Automatic selection retains ordered size bounds and all node/memory safety checks.
ARD recovery regressions compare independent likelihoods at matching rate vectors, retaining the original tolerance and requiring improvement over the same run's unfinished fit.
Reuse D transition mixtures and Brent's matching best objective value in lambda calibration. Fixed-input Windows comparisons preserve statistical results and RNG; speed gains depend on workload. K kernels are unchanged.
Remove two unused private scalar encoding helpers. Refresh documentation, source/Rd agreement and executable short examples. No public API is removed.

fastphylosig 0.3.3.9021

Delta heterogeneity tests leave the recommended workflow. Local and overall Delta, their diagnostics and ordinary signal tests remain supported. Explicit experimental heterogeneity calls retain their existing results and RNG.
Define K's BM null, lambda's common-local-crown covariance null, and binary D's experimental count-conditioned symmetric ER null. Guides distinguish observed conservatism patterns from model-reference evidence and mechanisms.
Document and independently verify lambda's change of scale after crown extraction, including unequal crown ages, non-ultrametric geometry and root contraction after missing-tip pruning. Estimation and C++ kernels are unchanged.
Print/tree-plot annotations identify experimental Delta/D inference and the local scale of lambda. Recommended examples use descriptive Delta and D; all default arguments, numerical results and simulation ordering are retained.

fastphylosig 0.3.3.9020

Explicit nuisance statistic = "between_clades" tests SD(log(local Delta)), reusing the existing native dispersion kernel. Whole-tree Delta, local/global descriptions and all estimator/MCMC calculations are still retained.
Explicit nuisance_control$proposal_counts = TRUE mixes the old proposal bank with a fixed count-universal sequential proposal at equal total weights. Simulation and outward density evaluation share their actual CDF kernel; no observed counts, fitted proposal or lgamma density approximation is used. Existing Mk, mixture, likelihood and continuous-domain bound kernels are reused.
Both additions remain experimental. Count coverage has a polynomial regret bound in the independent-composition submodel; it does not cover arbitrary slow-mixing ARD correlations. Neither option changes default results or RNG.

fastphylosig 0.3.3.9019

Nuisance diagnostics distinguish a certified final bound of one, all-tail banks, unresolved envelopes and gamma clipping. Raw bounds, final width, stop reason, state support and duplicate-phenotype tail counts are retained.
Native support checks select deterministic closed-pair candidates for ARD. Existing point likelihoods must verify confidence-set membership and rank; candidate values never replace the continuous-domain upper bound. Search stops early only when a verified lower bound plus gamma already certifies a final probability of one. Reference draws and RNG are unchanged.
Proposal-density evaluation now uses effective sequential CDF thresholds, including rounded overshoots. Legal monotone thresholds retain the previous arithmetic; the sampler and default plugin/Delta/ACE/MCMC paths are unchanged.

fastphylosig 0.3.3.9018

Experimental ARD transient envelopes share four slow-edge coordinates with the existing two compact base coordinates. A positive first-jump expansion retains their noncommuting products and bounds only two-or-more extra jumps by a Poisson tail. Existing C++ stochastic-matrix and pruning kernels are reused.
Subtree-repeat indices are prepared once per nuisance workspace and reused across rectangles and boxes. Numeric messages are recalculated at each parameter rectangle, using the original multiply/add and scaling order.
Auto/transient can intersect a complete-domain mass certificate based on phenotype normalization, the observation confidence constraint and a three-state return-probability floor. Its benefit depends on tail labels; it can remain one, and adding gamma may erase a small raw improvement. Default plugin estimates and RNG are unchanged. New mass-envelope and repeated-message diagnostics are separate from optimizer/MCMC warnings.

fastphylosig 0.3.3.9017

Experimental three-state ARD nuisance bounds retain a common full transient near all six multiclass boundary families. Two compact coordinates share total base rate and branching ratio across branch lengths; a coupling error covers the other four rates without assuming a positive spectral gap or bounding the base rates. Existing C++ Taylor pruning and rank certificates are reused with size/error safeguards and separate transient diagnostics.
Fixed, membership-verified ER-direction witnesses can improve the ARD auto/transient maximization lower bound, including the complete unbounded domain. They do not certify its upper bound. Unsupported boxes retain the complete previous enclosure; default plugin estimates and RNG are unchanged.
Shared-model pruning explicitly retains nonnegative probability messages at zero-rate boundaries, avoiding negative subnormal lower endpoints before log.

fastphylosig 0.3.3.9016

Experimental ARD nuisance bounds can share the maximum-forest stationary projection across terminal branches in wide and unbounded rate boxes. A Dobrushin/reset coupling bound encloses finite-branch transient remainders; shared Bernstein polynomials certify likelihood and weighted-tail bounds.
Count-lumped positive uniformization reuses the interval kernel and caches fixed reference coefficients. Default plugin estimates and RNG are unchanged; nonmixing boxes keep the previous complete-domain enclosure.

fastphylosig 0.3.3.9015

Experimental ARD nuisance bounds can retain six shared rate coordinates through all branch transitions, pruning messages and reference likelihoods. First-order Taylor models include rigorous nonlinear/rounding remainders; signed tail/other contrasts retain numerator/denominator correlations.
Finite boxes use an automatic cost gate or explicit Taylor mode. Unbounded tails retain the existing forest/Euler enclosure. Uniformization powers are shared across branch lengths. Default plugin estimates and RNG are unchanged.

fastphylosig 0.3.3.9014

Experimental Delta nuisance tests now cover unbounded three-state ARD boxes through nonnegative forest/resolvent intervals and a uniform Euler error enclosure. Continuous search retains all infinite tails without rate caps.
Stochastic-row constraints tighten pruning bounds. ARD proposals can add six fixed, trait-independent rare-state/fast-pair components; the original ER-only proposal remains selectable. These changes affect opt-in nuisance tests only; default plugin estimates and RNG ordering are retained.
Boundary coverage does not guarantee useful power or a small numerical gap.

fastphylosig 0.3.3.9013

Delta heterogeneity gains an experimental, opt-in nuisance-domain upper bound for explicit ER (2--8 states) and ARD3 state spaces. Corrected importance ranks reuse one fixed proposal bank; conservative continuous-rate likelihood boxes and an optional finite-sample confidence set handle the common Q uncertainty. Search budgets retain an upper bound, never an unverified grid maximum.
ER covers its infinite-rate endpoint. Unresolved unbounded ARD returns one without simulating an unused bank; finite ARD domains must be specified before looking at traits. This does not establish useful power for the new method.
Existing plugin tests, Delta estimates and default RNG ordering are retained.

fastphylosig 0.3.3.9012

ACE's existing real-exponent shortcut also covers positive normal arguments. This avoids complex exponential dispatch for tiny positive real eigenvalues produced by roundoff. Complex, overflowing, underflowing and nonfinite inputs retain their previous evaluation paths; signed zeros are retained. ARD solver budgets, recovery guards, candidate selection and RNG boundaries are unchanged.

fastphylosig 0.3.3.9011

Bounded Delta ARD recovery uses the existing stable matrix exponential for positive exit rates separated beyond sqrt(machine epsilon). This prevents loss of relative spectral accuracy near an absorbing state. Ordinary fits, common reference fitting and MCMC retain their paths and RNG boundaries.
Recovery rechecks its initial objective with the same guarded evaluator; original optimizer returns and any initial numerical correction are retained. The guard is used consistently for objectives and returned ancestral rows.

fastphylosig 0.3.3.9010

Unfinished default ARD fits in Delta workspaces receive one bounded, parameter-scaled continuation, followed only when needed by one bounded L-BFGS-B attempt. Both reuse the existing C++ objective and native scratch. Successful fits, common multistart, other models and MCMC retain their paths.
Actual optimizer returns determine candidate selection and convergence. Original warnings and retry budgets are retained in optional diagnostics; retry/recovery counts are separate from numerical recovery and remaining optimizer warnings. Healthy output structures and RNG are retained.

fastphylosig 0.3.3.9009

ARD likelihoods use the existing Armadillo matrix-exponential kernel when an eigensystem is defective or ill-conditioned. Ordinary spectral arithmetic and MCMC are retained; recovery is deterministic and reported separately from optimizer warnings. Default local fitting strategy is retained.
Multistart prefers an actually converged candidate when its deviance is numerically equivalent to an unfinished minimum, with an explicit tolerance and recorded selection recovery. Materially better unfinished candidates remain unfinished. Successful ordinary selections are retained.
ACE checks every returned ancestral probability row, repairs only tiny boundary roundoff, and rejects material invalidity. Delta validates all rows before entropy endpoint jitter, including rows with interior entropy.

fastphylosig 0.3.3.9008

ACE pruning reads child likelihood rows directly into the existing complex workspace, avoiding an intermediate transposed real vector for every edge. Real exponents in a guarded normal range reuse scalar exp; complex, underflowing, overflowing and nonfinite cases retain the original path.
MCMC acceptance draws reuse R's uniform generator for fixed open-unit bounds, retaining endpoint rejection and conditional draw order. Proposals, lgamma, accumulations, saved traces, estimators and diagnostic thresholds are retained.
A predefined independent screen compares generating Q, default nlminb and multistart on 400 new inputs, then three cut scales on 300 new inputs. Larger crowns improve shuffled-crown detection in the tested 512-tip balanced trees, while fitted-Q bias, missing states and diagnostic warnings remain material. Defaults, estimators and scientific tests are unchanged; this is finite screening evidence, not a claim of general calibration.

fastphylosig 0.3.3.9007

Optional delta_control$reference_fit="multistart" improves the search for common ARD Q with deterministic starts and continuation of nlminb, reusing one prepared ACE objective. Attempts and deviance gains remain visible; the best nonconverged fit is never relabelled as converged. Observed Delta is unchanged.
Reference tail bounds retain all requested draws. Optional reference_unavailable="conservative" counts single-state reference discrepancies in the upper tail without assigning Delta values, dropping or replacing draws. Actual computation errors still make P unavailable. Default nlminb and strict rules remain; neither option corrects plug-in fitting bias.
Replicate diagnostic flags use one native pass in place of repeated R closures, preserving default estimators, diagnostics and random streams.

fastphylosig 0.3.3.9006

Delta scans and complete-tree estimates separate optimizer warnings, analytic ACE recovery, entropy roundoff recovery, other warnings and calculation errors. Reference counts flag replicates, with overlapping categories; unavailable single-state crowns remain distinct from calculation failures. The retained nsim_fit_warning column now mirrors nsim_optimizer_warning.
Reference fits reuse a scalar C++ warning evaluator instead of constructing per-fit R diagnostic tables, and accumulate replicate counts outside the result data frame. Thresholds, reason strings, estimators and random streams are retained. Custom R numeric formatting retains the existing warning formatter.
A predefined independent Delta screen uses 1,600 outer datasets, 199 reference simulations and production MCMC settings, with unique branch lengths, imbalanced crowns and rare-state ARD. Conditional null rejection is 2.8--6.0%, while ARD null unavailability is 5.5--13.0%; shuffled-crown power is 16.2--21.0%. These finite, fixed-tree results do not establish general calibration or power.

fastphylosig 0.3.3.9004

Delta rank diagnostics reuse the sorted split traces for full-trace type-7 quantiles, merging omitted middle observations for odd lengths. Folded ranks reuse this order through a monotone merge instead of another full sort. Midranks, normal scores, ESS and R-hat arithmetic remain unchanged.
The existing MCMC kernel counts down to saved iterations instead of using per-step modulo. Burn/thin semantics, proposals, acceptance, trace values, floating-point sums and random-number call order remain unchanged.

fastphylosig 0.3.3.9003

Delta repairs machine-scale entropy excursions beyond [0,1] only when the corresponding probability row is finite, bounded and normalized within a small tolerance. It then reuses the existing endpoint jitter and MCMC, warns once and reports recovery counts. Healthy entropy arithmetic, estimates and RNG remain unchanged; formerly failing calculations can continue. Material excursions are errors. The tolerance is capped at 1e-12.
Entropy calculations cache log(k) once per call while retaining division and fuse range checking with endpoint detection, preserving the complete sampler outputs and RNG of the initial roundoff-repair implementation.

fastphylosig 0.3.3.9002

ACE recovers failed ER eigensystems with the existing analytic ER transition kernel, reusing pruning, ancestral updates and prepared caches. Successful eigensystem calculations keep their original arithmetic and RNG. Recovered fits warn once and report evaluation counts; previously failed Delta analyses can now continue, changing their subsequent RNG and reference probabilities. ARD behavior and optimizers are unchanged.

fastphylosig 0.3.3.9001

An explicit experimental Delta common-ER reference fit is available through delta_control = list(model="ER", reference_fit="profile"). It reuses ACE prepared scratch, analytic ER transitions and the lambda Brent implementation, compares zero/independent endpoints, and reports fit diagnostics. Default nlminb fits, observed Delta estimators and their RNG contract remain unchanged. The alternative changes reference Q and probabilities; calibration remains exploratory. Infinite ER rates use exact independent-state transitions.

fastphylosig 0.3.3.9000

Delta adds rank-normalized/folded R-hat, bulk/tail ESS and relative MCSE for alpha, beta and their ratio. Ratio diagnostics enter automatic flags. Historical diagnostics remain available. Reference MCMC and ACE warnings have separate counters; diagnostics never resample or change estimates.
The common lambda reference in fast_hetero now defaults to pooled REML, independently checked against dense GLS and held-out null simulations. Local/overall ML estimates remain unchanged. lambda_reference="ML" retains the historical reference. The new default intentionally changes reference fits, null statistics and probabilities; boundary calibration is approximate.
Delta/ACE analyses reuse native scratch across fits of the same immutable tree, state count and model, with a bounded analysis-owned cache and explicit cleanup. PSOCK processes use owned scratch. Final ancestral likelihoods use the same prepared kernel; modern diagnostic normal scores are cached locally and column FFTs reuse preparation with bounded dominant buffers. These performance paths preserve the corrected scientific baseline and RNG.
Expanded whole-tree calibration reports unavailable results, Monte Carlo uncertainty and model sensitivity. High-state Delta and sparse ARD reference tests remain exploratory; successful numerical checks do not establish general nominal error control. See the installed heterogeneity guide.

fastphylosig 0.3.2.9000

Scientific correction: Delta is now the posterior mean of beta/alpha, rather than the ratio of posterior means. Its batch-means MCSE and new ESS_Delta/split_Rhat_Delta summarize that same ratio trace.
Corrected lognormal random-walk MH acceptance for independent exponential priors on the raw alpha and beta parameters. Invalid proposals are rejected once, and chain initialization has explicit RNG sequencing. Delta values, tests and RNG trajectories intentionally change from 0.3.1.9000.
Repeated-length ACE exponent caching now uses its actual 64 MiB byte budget, removing the fixed 4096-group cap. Delta caches logs of accepted raw shapes so rejected proposals do not repeat logarithms. These performance changes preserve the corrected estimator's numerical results and RNG contract.

fastphylosig 0.3.1.9000

ACE fits now combine the exact repeated-length exponent cache with bounded parallel batches for uncached branches. A compact immutable worklist avoids computing cached exponents again and enables large mixed-length fits; the threshold depends on uncached work rather than total tree size. Batch widths may be odd, avoiding a separate parallel region for one leftover branch. Fully cached and small workloads keep their serial path. Pruning, optimizer and RNG order remain unchanged.
Large ACE likelihood-only fits with distinct branch lengths now evaluate independent complex exponent vectors in bounded 2 MiB OpenMP batches, using the existing ACE expression and serial pruning/reduction order. The default uses at most two threads; options(fastphylosig.ace_threads = 1L) selects serial execution. Small fits and builds without OpenMP retain the serial path; Delta PSOCK workers use serial inner fits. Full objective trajectories, Delta outputs and RNG streams are checked against the preceding development build.
Delta now uses a native V2 canonical graph/UTF-8 byte ordering path while retaining source IDs, descriptors and attributes. Rate fits reuse a private ACE workspace, transform repeated tip states once per objective, and cache exponent vectors for exactly equal branch lengths with bounded storage. MCMC caches accepted-state log-gamma values. The optimizer, floating-point expression order and random draw order are unchanged; strict result/trace/RNG checks accompany the optimization.
K crown signal tests now call the production C++ permutation kernel on the scan's compiled NA-mask groups, removing repeated tree preparation. Automatic cuts jump whole selected/undersized traversal intervals and reserve extraction buffers. The shared ACE preflight uses the native topology validator; its ancestral downpass lazily evaluates repeated row normalizations, including floating-point fixed points/two-cycles, instead of scanning the entire matrix after every sibling pair. Paired checks retain estimates, tests and RNG streams.
Local K/lambda signal tests now respect crown eligibility (min_tips and estimator status). Lambda returns local LRT probabilities in the first scan, avoiding a second fit. Delta tests reuse scan NA/ACE contexts and shuffle crown-sized vectors without unused permutation trace diagnostics. Shared C++ reference summaries now use the same plus-one Monte Carlo error for all methods; Mk simulation pools exactly equal branch lengths to reduce cache memory and transition preparation. Estimators and RNG order are retained.
Added multistate Delta to fast_hetero(), reusing ACE likelihood workspaces, native entropy/MCMC and diagnostics. Local/overall contrasts, categorical state counts/frequencies and optional local signal tests accompany tree maps.
Added an optional new common-Mk reference check of RMS local/overall log Delta deviations, with native cached full-tree simulation, uniform-root likelihoods integrating NA tips, complete replicate accounting and fitted reference metadata. This tests signal deviations, not rate shifts; broad statistical calibration remains future work. Descriptive scans remain the default. Single-state crowns and failed replicates produce explicit status.
Tree plots use log Delta and reverse the default D palette so stronger signal has a consistent colour direction. D local signal tests reuse its normalization simulations without additional RNG draws.
Shared prepared D contrast factors across the public scalar/batch engines and heterogeneity; removed repeated branch adjustments and duplicate traversals. Batch columns use one OpenMP region and smaller scratch buffers.
Reused lambda likelihood scratch buffers in the specialist optimizer and heterogeneity bootstrap. D reuses conditional-likelihood/count workspaces, fitted transitions and sampling buffers; both bootstraps reuse branch standard deviations while retaining their existing RNG order.
Added lambda heterogeneity calibration with a fitted shared-lambda Gaussian reference and binary D scans/calibration with a conditional shared-rate ER reference. Both bootstrap engines run in C++ without dense covariances.
Added P_heterogeneity, Monte Carlo error, reference summaries and fitted common parameters. K retains its P_BM fields as compatible aliases.
D preserves each crown's observed state count using polynomial pruning and downward conditional sampling, reusing random/Brownian normalization (d_nsim) and count probabilities across draws. Plots distinguish overall D signal probabilities from ER heterogeneity.
Named the heterogeneity API fast_hetero(); the earlier development name has been removed.
Added optional complete-tree K/lambda estimates and signal tests through overall/overall_test, reusing the existing native engines. Plots annotate the complete-tree signal test separately from the K BM dispersion test.
Added plot(h, tree = tree) to colour crown-clade K/lambda estimates on the phylogeny, with source-node labels, logarithmic K colours, grey unavailable regions, and deepest-crown precedence for nested selections. Tree annotation uses a linear native traversal and checks selected descendant identities.
Added fast_hetero() for continuous-trait K/lambda scans on explicit crown clades or automatic disjoint clades with bounded sizes. Reports original node identities, per-trait missingness/failures and coverage.
Added an optional K dispersion calibration against Brownian motion, preserving local topology and missingness, using chunked tree simulations and the existing compiled K engine without dense covariance matrices. BM reference envelopes and the upper-tail probability have explicit model interpretations; nested scans remain descriptive.
Added print/plot methods, dense numerical and simulation oracles, and a Chinese guide separating signal heterogeneity from PVR prediction quality.
Accelerated clade scans with a lightweight checked tree path, grouped missing-data kernels and reuse of compiled crown subtrees. BM branch propagation now runs in C++, preserving R's random draws and chunk semantics.
Moved clade indexing, selection, extraction, exact NA-mask grouping and batch signal orchestration into C++. Fused BM generation, prepared K evaluation and dispersion calibration retain serial R RNG ordering, reuse K factors/workspaces and cap dominant block buffers at 64 MiB. NA pruning retains ape's established crown/root/unary behavior.
Removed obsolete native/R heterogeneity entry points, keeping independent reference implementations in test helpers. Removed unused K workspaces and duplicate shape checks, and omitted unused BM fit diagnostics. Complete-data scans defer loading ape.

fastphylosig 0.3.1 (2026-10-27)

Aligned read-only tree readiness with V2 requirements for character tip labels and a declared positive Nnode. Invalid representations no longer appear ready immediately before preparation rejects them.
Preserved explicit D plot P-value column selection, including differing aliases and custom table columns under the automatic null selection.
Strengthened scalar K dispatcher tests to compare actual fields and simulation values, and corrected lambda's serial-execution documentation.
Preserved explicit numeric species names through both preparation layers; positional matching now applies only to missing names or automatic data-frame rownames.
Bound D plot probabilities to the selected null model and preserved fitted probabilities and strict contrast-scale tail counts in calibration plots.
Aligned R and C++ lambda boundaries with ape's scaled-range ultrametric tolerance. Approximately ultrametric trees use the exact terminal-edge feasibility limit, which can be slightly below phytools' max-height ratio; exactly ultrametric and non-ultrametric domains are unchanged.
Indexed long packed-mask identities by exact equality before using them as environment names. Existing short keys and V2 context persistence remain supported. The 50K regression covers preparation, subsets, persistence, fallback grouping and K without permutations; it does not qualify every method or simulation workload at that size.
Corrected D's type-7 Brownian threshold interpolation between the first two order statistics. A first factor level represented by one tip now produces one Brownian state per replicate for distinct simulated values, instead of an all-zero state vector. Strict threshold and tail rules, RNG draw order, and estimator definitions are unchanged.

fastphylosig 0.3.0

Replaced the Blomberg's K permutation evaluator's per-permutation tree traversals with an adaptive hybrid fused kernel. Where fusion is selected, one traversal serves a block of permutations; hybrid evaluation obtains the GLS offset from a tree-only dot product instead of a full upward/downward pass pair.
Added an internal strategy selector. It maps the workload (number of tips, number of traits, number of permutations, memory budget) onto the evaluator variant and the block size. The selector is not part of the public API and adds no arguments to any exported function.
The permutation stream is unchanged. Blocks are a storage unit, not a drawing unit: permutations are still generated serially in global order under R's RNG. Fixed-seed validation found identical p-values and exceedance counts; K values may differ from 0.2.0 at floating-point rounding scale.
Public API, estimator definitions, the pointwise p-value rule, and the threading policy are unchanged. The 0.2.0 evaluator remains in the source as an internal reference; the production selector uses hybrid evaluation.

fastphylosig 0.2.0

Integrated Canonicalization Contract V2 for deterministic, representation-safe tree preparation. Tip IDs and labels are preserved; internal nodes and edge rows are canonicalized without using source internal-node IDs, source edge order, locale collation, delimiters, or a probabilistic hash as ordering truth.
Replaced the legacy delimiter-based prepared-tree identity with versioned, exact structured encodings of tip.label, Nnode, edge, and edge.length. A package-owned protected record is checked before cached evidence or a numerical kernel can be used.
V2 prepared contexts support saveRDS()/readRDS() reuse across R sessions. Missing, legacy, or unknown context schemas are rejected with an instruction to run prepare_tree() again.
Added representation, heterogeneous-branch, locale, mutation, persistence, and K/lambda/D/Delta/ACE parity gates. Estimator definitions, RNG streams, threading policy, numerical tolerances, and public APIs are unchanged.

fastphylosig 0.1.0

Consolidated repeated trait-table validation and C++ quantile code, removed obsolete ACE development timing hooks, and reorganized the English and Chinese guides around a one-command start and an explicit advanced workflow.
Declared compatibility with R 4.1.0 and newer. Windows source builds use the matching Rtools toolchain; OpenMP remains optional with a serial fallback.
Added the method-explicit fast_signal(tree, data, method, ...) dispatcher for K, lambda, D, and Delta, with public fast_k() and fast_lambda().
Kept fast_d(), fast_delta(), and fast_ace() as specialist entry points; methods are never inferred from trait values and no public numerical-engine selector was added.
Made match_tree_data() the documented matching API and retained match_phylo_data() only as a deprecated compatibility alias.
Added shared tree/data preparation, compact workflow metadata, packed NA-mask grouping, retained-subtree validation, and bounded prepared-tree caches while preserving specialist result classes and statistical contracts.
Restricted automatic tree normalization to representation-only changes. It never invents roots or branch lengths, jitters zero branches, resolves a biological polytomy, ultrametricizes a tree, or silently removes tips.
Added consistent per-trait status, message, and note output for non-finite, insufficient, single-state, constant, and undefined traits.
Added a shared tree-requirement registry, actionable method-specific errors, direct ACE topology/branch guards, root-sensitivity fixtures, and explicit zero-terminal/zero-internal branch contracts.
Added controlled K permutation and D random/Brownian null fixtures, lambda fixed/optimized pruning tests, unified-versus-specialist equivalence tests, large-offset stability tests, and deterministic raw/prepared comparisons.
Clarified that fast_delta() implements a single-tree statistic with local ACE and RNG paths. Cross-language parity and the multiple-tree extension are outside the package contract.
Replaced the large-tree NA-subset cache key with the packed mask key, so 3000/5000-tip analyses no longer hit R's 10,000-byte environment-name limit.
Set Delta defaults to mcmc_sim = 10000, thin = 10, and burn = 100. Added one aggregated retained-sample warning and one aggregated ESS/R-hat/ MCSE warning per batch; warnings do not rerun or lengthen chains.
Reused permutation-invariant ACE workspaces inside Delta and skipped unused post-fit Hessian/SE work only in the private Delta ancestral-state helper. Public fast_ace() still computes and returns standard errors.
Removed generated build artifacts and obsolete development-only benchmark records from the working tree; retained only concise release and reference notes needed to explain the current package contract.
Final release-contract patch: ambiguous simultaneous x/X inputs now fail for D and Delta, ACE rejects invalid method values clearly, and unused Delta nsim input is ignored when test = FALSE. Prepared-tree fingerprint validation now also covers Nnode; targeted regressions and the full local check were rerun without changing numerical kernels.

fastphylosig 0.0.4

Added check_tree() for read-only, method-specific readiness checks for K, lambda, D, and Delta. Diagnostics distinguish global tree defects from the topology and branch-length contracts of individual statistics.
Added conservative resolve_tree(). It standardizes edge order and internal node numbering, rechecks the result, records provenance, and refuses unsafe repairs such as random polytomy resolution, branch jitter, or tip deletion.
Added method-level progress messages and stable elapsed-time metadata to fast_signal(), fast_d(), fast_delta(), and fast_ace().
Added explicit R-boundary topology guards for D and ACE. D now rejects unary nodes or noncanonical roots instead of silently discarding a branch; ACE now rejects mixed unary/polytomous topologies that previously could pass a node- count-only check.
Added boundary and numerical-contract tests for ACE transition evaluation; production tolerances and public results remain unchanged.

fastphylosig 0.0.3

Made the validated tree-pruning implementations the automatic production paths for Blomberg's K and Pagel's lambda. Public callers no longer select dense, tree, spectral, or validation engines.
Added streaming, chunked K randomization with Monte Carlo uncertainty and optional null retention, without allocating a full tip-by-simulation matrix.
Added a tree-linear fixed-lambda likelihood and a batched C++ Brent optimizer returning lambda, log-likelihood, GLS mean, sigma2, LR, P, and optional likelihood profiles without production VCV/eigendecomposition work.
Added explicit Monte Carlo uncertainty and successful/failed simulation counts for both Fritz-Purvis D null hypotheses.
Retained Delta ESS, split R-hat, MCSE, iteration, and permutation diagnostics.
Profiled ACE and Delta before optimization. ACE branch-transition arithmetic and repeated Delta permutation calculations are the demonstrated hotspots; small PSOCK workloads can be slower than serial execution.
Added production and component benchmarks plus expanded deterministic, stochastic, boundary, offset, topology, and cache regression tests.

fastphylosig 0.0.2

Tree preparation and cache

prepare_tree() now compiles O(n) topology and traversal data eagerly while creating VCV, Cholesky, and lambda spectral payloads only on demand.
Added separate structural and numerical caches, a byte-budget numerical LRU, packed C++ grouping of trait-wise missing-data masks, and cache_info().
D, Delta, ACE, and tree-engine K workflows no longer request dense VCV data.

Continuous signal

Kept dense/Cholesky K as the default compatibility engine.
Added opt-in engine = "tree" for O(n p) observed Blomberg K and engine = "validate" for explicit tree-versus-dense comparisons.
Lambda remains on the validated spectral engine with a dense numerical fallback; no unvalidated tree-lambda implementation is exposed.

D and Delta

Added bounded, fused D null simulation and traversal with configurable chunk_size and keep_null.
Added Delta chain means/SDs, ESS, split R-hat, covariance-aware MCSE for Delta, permutation P MCSE, successful-count reporting, and safe RNG metadata.

Validation and documentation

Added deterministic golden contracts, compiled-tree, packed-mask, streaming, tree-K, cache, and Delta-diagnostic tests.
Added a unified environment-controlled benchmark matrix and rewrote the user and expert documentation around explicit compatibility boundaries.