cluster_network() alias for summarize_network(). It
collided with Nestimate::cluster_network() — a completely different
function (PAM clustering on sequence data, one network per cluster) —
and the two silently masked each other depending on package attach
order, producing confusing unused arguments (k = ..., cluster_by = ...) errors. Use summarize_network() (or its remaining short form
cnet()) for matrix-to-cluster aggregation in cograph.LICENSE..smooth_blob() (used by plot_simplicial() and
overlay_communities()) now guards grDevices::chull() against
non-finite anchor coordinates. Previously a node lacking layout
coordinates (NA/Inf) aborted the blob with "finite coordinates are
needed"; such anchors are now dropped before the convex-hull step.motifs() / subgraphs() roxygen documentation with the
post-audit behavior shipped in 2.3.2 (census type_summary counts,
min_count handling, and corrected plot legend descriptions).type_summary now holds
real MAN-type counts in census mode, min_count is honored in census
mode, and the swapped source/target color description in
plot.cograph_motif_result() is corrected.motifs() and plot_simplicial() on Nestimate-backed
workflows (HON / HYPA sequence inputs).panel_layout(): tightened dimension validation and made the
restoration claim honest — it now restores only the par() settings
it actually changed.combined argument (default TRUE) on every multi-panel plot
function: splot() group-cascade, plot_netobject_group(),
plot_netobject_ml(), plot_net_bootstrap_group(),
plot_group_permutation(), plot_compare(), splot.net_mlvar(),
plot_network_evolution(), plot.cograph_motifs(),
plot.cograph_motif_result(), plot.cograph_motif_analysis(), and
plot.tna_disparity(). With combined = FALSE these functions draw
panels into the active device without calling graphics::par(mfrow=...),
so callers can drive their own layout (e.g. graphics::layout() or
the new panel_layout() helper). Default TRUE preserves prior
behavior — every existing call site renders identically.panel_layout() helper sets up a custom multi-panel device layout
for use with combined = FALSE. Accepts either a uniform-grid
c(nrow, ncol) or a graphics::layout() matrix for non-uniform
layouts (e.g. one wide panel + two narrow ones). Returns a par()
snapshot for restoration via on.exit().test-coverage-splot-{41,42}.R: bumped n_nodes from 4 to 10 in
seven per-edge attribute tests so the seed=42 sampler does not
produce duplicate (1,2) pairs that trip cograph's
undirected-duplicate-edge detector.test-coverage-class-network-41.R: aligned the
set_layout_coords() mismatched-row-count test with the strict
input validation already enforced by R/class-network.R.test-overlay-communities.R: prefixed two communities() calls
with cograph:: to avoid tna masking when both packages are
loaded in the suite (per CLAUDE.md "namespace masking" gotcha).cr_color #D4820A -> #D4829A in plot-forest.R;
show_value default FALSE -> TRUE in splot-nodes.R), corrected
dataset dimensions in data-hai.R (302 -> 429 x 287), corrected a
reference to the nonexistent igraph::is_bipartite() (now
bipartite_mapping()), expanded centrality() @param measure
lists for mode, cutoff, invert_weights, and membership to
match the implementation, dropped baked-in measure counts that rot
on each addition, and removed nonexistent themes from sn_theme
documentation. No runtime behavior changes from the documentation
pass itself.plot_simplicial() now warns when anomaly is set on an input that
has no anomaly concept (HON, association rules, link prediction,
character pathways, method = "hon" / "rules"). Previously the
argument was silently dropped, so calls like
plot_simplicial(hon, anomaly = "over") and
plot_simplicial(hon, anomaly = "under") produced byte-identical
plots. anomaly is honored only for net_hypa inputs and
method = "hypa" auto-builds.centrality() gains an umbrella argument tna_network (logical or
NULL). When TRUE (or auto-detected from a tna/group_tna/ctna/
ftna/atna input), all measures shared with tna::centralities()
match byte-for-byte: loops = FALSE, invert_weights = TRUE,
diffusion_method = "power_series", transitivity_type = "onnela".
Side-by-side audit confirms zero divergence on OutStrength,
InStrength, ClosenessIn/Out/All, Betweenness, Diffusion,
Clustering (max|diff| = 0). Any per-argument override the user
passes explicitly always wins over the umbrella.centrality() (and centrality_diffusion()) gain a
diffusion_method = c("kandhway_kuri", "power_series") argument. The
default NULL auto-detects: "power_series" for tna inputs (matches
tna::centralities(., measures = "Diffusion") byte-for-byte when
loops = FALSE), "kandhway_kuri" (the existing 1-hop binary-degree
formula, Kandhway & Kuri 2014) for everything else. Previously
cograph's diffusion silently disagreed with tna's because cograph used
an unweighted neighborhood-degree sum while tna uses
rowSums(P + P^2 + ... + P^n) on the diagonal-zeroed weighted matrix
— the same name covered two different statistics. Set explicitly to
override the auto-detect.tests/testthat/test-validate-nestimate-bootstrap-permutation.R
asserting that centrality() on a Nestimate netobject agrees with
centrality() on its $weights matrix when the diagonal is
non-zero. Locks in the upstream Nestimate fix to
.extract_edges_from_matrix() (Nestimate >= 2026-05-02) which now
preserves self-loops in $edges. Without that fix, loop-bearing
netobjects (e.g. Nestimate::build_mcml() |> Nestimate::as_tna())
silently under-counted node degree by 2.edge_label_size is now coupled to the node label cex at a
fixed 0.55 fraction (edge_cex = 0.55 * mean(node_label_cex)) so the
node-to-edge-label ratio stays a stable ~1.82x across canvases. This
replaces the previous EDGE_LABEL_SCALE_CAP-based compensation, which
let the ratio drift from 2.5x at reference to 3.6x at poster canvases
because edge labels were clamped to a tighter 1.6 ceiling while node
labels scaled freely to 2.3. The visible effect: edge weight
annotations are now readable at poster sizes instead of shrinking
relative to node labels. User-explicit edge_label_size still wins
and receives the same (capped) visual-scale compensation as before;
only the default path changed.render_edges_splot()
into splot.R so the final cex is produced in a single place.splot() now applies device-dependent compensation to text, line, and
point sizes so visual ratios (label-to-node, legend-to-plot, edge
thickness) stay consistent when the output device changes. This fixes the
long-standing "labels too big at high DPI" and "legend desynchronised from
the plot" issues when saving PNGs at res = 300 or res = 600 with
pixel-default width/height, and when resizing the RStudio plot pane.
Implementation: a single compute_visual_scale() reads the active
device's canvas size (dev.size("in")) and returns multipliers keyed off
a 5.9-inch reference (matching the default RStudio 7×5" pane so
backward-compatible behaviour at the default canvas is preserved).
Multipliers are clamped to [0.55, 1.9] to keep thumbnails and posters
legible. See the new R/visual-scale.R.scaling = "fixed" mode on splot() — and corresponding global
option options(cograph.visual_scaling = FALSE) — disables device
compensation for reproducibility-sensitive workflows that calibrated
against the previous behaviour.splot() return value now carries two attributes for downstream tooling:
cograph.visual_scale (the multiplier list) and cograph.node_diam_in
(the representative node diameter in inches at the rendered device).render_legend_splot() plus the new shared
.render_legend_base() (R/render-legend-shared.R) replace the ad-hoc
legend cex/pt.cex handling with a single compensated path. plot_htna,
plot_mtna, plot_mlna, plot_mcml still use their historical scale
multiplier arguments; Phase 2 will migrate them to the shared helper.splot.netobject now routes on the Nestimate $method slot rather than
just direction. Undirected sequence-based networks from build_cna() and
wtna(method = "cooccurrence") get oval TNA-family styling (layout,
palette, donuts) with arrows and dotted edge starts automatically dropped
because the matrix is symmetric. Glasso / cor / pcor / ising networks
still get psych_styling = TRUE (spring layout, Okabe-Ito palette).from_tna() auto-detects integer-valued weight matrices (ftna, ctna, raw
counts) and sets weight_digits = edge_label_digits = 0 so edge labels
render as 2304 rather than 2304.00. Fractional weights still format
to two decimals. Explicit user-supplied weight_digits still wins.psych_styling = TRUE is now exported as a first-class styling preset
(undirected counterpart of tna_styling) — Okabe-Ito palette, spring
layout, no arrows — applied by default to splot.netobject on
correlation-family input and to the $contemporaneous / $between
constituents of net_mlvar.splot() dispatch coverage across the tna and Nestimate class
hierarchies, ensuring tna, ftna, ctna, group_tna, tna_bootstrap,
group_tna_bootstrap, tna_permutation, group_tna_permutation,
netobject, netobject_group, netobject_ml, net_mlvar, wtna_mixed,
net_bootstrap, net_permutation, boot_glasso, mcml, net_hon,
net_hypa, and simplicial_complex all reach the correct renderer.detect_duplicate_edges(), aggregate_duplicate_edges(),
simplify.cograph_network(), and the internal check_duplicate_edges()
helper now respect directed vs undirected semantics. Previously the
canonical (min/max) endpoint key collapsed A -> B and B -> A into one
edge even on directed graphs, matching igraph::simplify() ground truth..compute_modularity() replaces a nested for loop with cluster-wise
vectorization (sum(A[idx, idx]) - sum(k_out[idx]) * sum(k_in[idx]) / m),
per the project "no for loops" rule. Results verified bit-exact against
igraph::modularity().is_directed() now recognises CographNetwork R6 objects — previously
only the cograph_network list format dispatched correctly.compute_layout_for_cograph() uses layout$get_type() instead of the
removed $name field on CographLayout.network_small_world() returns 0 (valid: no triangles means
definitively not small-world) instead of NA_real_ when the observed
clustering coefficient is zero but path length is finite.simplify.cograph_network() threads the directed flag through to edge
aggregation so directed multigraphs collapse correctly.simplify() performance refactor for large networks plus a cleaner
title-composition path.motifs(), extract_motifs(), and plot.cograph_motif_analysis
examples reworked to use n_perm = 10L (or significance = FALSE) and
promoted from \dontrun to CRAN-runnable (optional tna branches stay in
\donttest). Retires 320 seconds of latent CRAN timing risk — every
example now runs in under 4 seconds.test-audit-fixes.R — ground-truth regressions for the directed edge
semantics, modularity vectorization, and small-world behaviour changes.test-integer-weight-labels.R — locks from_tna() integer-weight
auto-detect behaviour and precedence of explicit weight_digits.test-equiv-{assortativity, cluster-quality, communities, disparity, edge-centrality, network-summary, robustness, standalone-measures}.R —
numerical equivalence against igraph, sna, centiserve, brainGraph,
influenceR, tidygraph, and NetworkX. Gated by
skip_coverage_tests() + skip_on_cran(), so they do not run on the
CRAN pipeline.These measures don't fit the per-node centrality() data frame, so they live as standalone functions:
estrada_index() — graph-level spectral invariant: \eqn{EE(G) = \sum_i e^{\lambda_i}}, equal to the trace of the matrix exponential of the adjacency. Equivalently, the sum of subgraph_centrality() across all nodes. Matches networkx.estrada_index at machine epsilon (max relative diff ~5e-15 across random test graphs).trophic_incoherence() — graph-level food-web stability measure (Johnson et al. 2014). Defined as the population standard deviation of per-edge trophic differences \eqn{s_v - s_u} where \eqn{s_i} is the trophic level of node \eqn{i}. Zero for perfectly coherent DAGs (e.g., a pure chain). Matches networkx.trophic_incoherence_parameter at machine epsilon. Directed-only; reuses the existing trophic_level calculator.group_centrality(x, nodes, measure = c("betweenness", "closeness", "degree")) — Everett-Borgatti (1999) group centrality for a set of nodes. Returns a scalar. Supports mode = "in"/"out" for directed-degree variants. Group closeness and group degree match networkx.group_*_centrality bit-exact. Group betweenness implements the textbook Everett-Borgatti / Puzis 2008 definition (fraction of shortest paths passing through at least one node in the group), which diverges from networkx.group_betweenness_centrality on some graphs due to a known quirk in NetworkX's Puzis-Yahalom-Elovici iterative algorithm. Verified via an independent Python brute-force: cograph matches the textbook definition; NX produces larger values on graphs with many overlapping shortest paths. Documented in the roxygen "Divergence from NetworkX" section.dispersion(x, u = NULL, v = NULL, normalized = TRUE, alpha = 1, b = 0, c = 0) — Backstrom-Kleinberg (2014 Facebook) pair-level measure of tie strength. Counts the number of "well-dispersed" mutual friends of u and v (pairs of common neighbors that are not directly connected and share no common neighbor inside u's ego network other than u and v). Matches networkx.dispersion bit-exact across all 156 edges on the karate club graph. Returns a scalar, named vector, or data frame depending on which of u, v are specified.Added the five Gould-Fernandez (1989) brokerage role counts, a foundational measure in social network analysis (~1500 citations). Each role is a separate per-node measure requiring a membership argument (following the same pattern as participation, within_module_z, gateway), and counts open directed 2-paths a -> v -> c through broker v:
centrality_brokerage_coordinator() — all three in broker's group (w_I)centrality_brokerage_itinerant() — endpoints same group, broker different (w_O, "consultant")centrality_brokerage_representative() — broker + source same, target different (b_IO)centrality_brokerage_gatekeeper() — broker + target same, source different (b_OI)centrality_brokerage_liaison() — all three in different groups (b_O)Bit-exact match against sna::brokerage$raw.nli for all five roles across 20 random directed graphs. Implemented natively (no runtime dependency on sna). Key implementation detail: the Gould-Fernandez counting rule requires open 2-paths only — triads where a direct edge a -> c already exists are excluded. This matches sna's C implementation exactly and was derived empirically (sna's .C("brokerage_R", ...) has no R-level source).
Directed-only; warns and returns NA on undirected input.
centrality_prestige_domain() — directed-graph prestige measure: for each node \eqn{v}, the number of other nodes that can reach \eqn{v} via a directed path. Classical Wasserman-Faust (1994) measure from sna::prestige(cmode = "domain"). Bit-exact match against sna, implemented natively via igraph::distances(mode = "out") + colSums(is.finite(D)) - 1 (no runtime dependency on sna). Directed-only; returns NA with a warning on undirected input.centrality_prestige_domain_proximity() — distance-weighted variant: R_v^2 / (D_v * (n - 1)) where R_v is the number of reachers and D_v is the sum of their geodesic distances to v. Bit-exact match against sna::prestige(cmode = "domain.proximity") on strongly connected directed graphs. On graphs with any unreachable pair, sna has a known bug (FALSE * Inf = NaN collapses the denominator, producing all-zero output); cograph's is.finite()-masked formula produces mathematically correct values on any directed graph. Directed-only.centrality_katz() — Katz (1953) status index. Bit-exact match against centiserve::katzcent (cograph mirrors centiserve's exact LAPACK call sequence). Also matches igraph::alpha_centrality(exo = 1) and networkx.katz_centrality_numpy at machine epsilon. New katz_alpha parameter (default 0.1).centrality_hubbell() — Hubbell (1965) input-output centrality. Bit-exact match against centiserve::hubbell (cograph mirrors centiserve's full-inverse LAPACK call path). Note: centiserve's default (weights = NULL) silently ignores E(g)$weight; to reproduce cograph's behavior with centiserve on weighted graphs, pass weights = igraph::E(g)$weight explicitly. New hubbell_weight parameter (default 0.5).centrality_information() — Stephenson-Zelen (1989) information centrality. Bit-exact match against sna::infocent on connected undirected graphs (cograph mirrors sna's exact construction and solve() call sequence).centrality_pairwisedis() — Pairwise disconnectivity (Potapov et al. 2008). Directed-only; fraction of reachable ordered pairs that become unreachable when each node is removed. Bit-exact match against centiserve::pairwisedis. Warns and returns NA on undirected input, matching the convention used by salsa, leaderrank, and trophic_level.centrality_reaching_local() / reaching_global() — Local and global reaching centrality (Mones, Vicsek & Vicsek 2012). Bit-exact match against networkx.local_reaching_centrality across the directed unweighted, undirected unweighted, and weighted branches. Undirected unweighted LRC coincides with igraph::harmonic_centrality(normalized = TRUE) (documented). reaching_global() is a graph-level hierarchy statistic in [0, 1].plot_simplicial() now accepts tna, netobject, net_hon, and net_hypa objects directly — higher-order pathways are auto-built and visualized with proper state labels, no manual extraction needed. New parameters: method ("hon" / "hypa"), max_pathways, ncol. Dismantled mode uses gridExtra grid layout with scaled nodesprint.cograph_network() now shows a structured summary: node/edge counts, density, reciprocity, weight range, and top-degree nodes — replacing the minimal R6 default outputmcml S3 class with as_mcml() generic for type-safe handling of Markov Chain Multi-Level models — enables print(), plot(), and method dispatch on MCML objects%||% operator for R 4.1 compatibility (no longer requires R 4.4+)$between → $macro, $within → $clustersas_tna() on MCML objects now returns a flat group_tna list instead of a nested structureplot_mcml() now suppresses zero-weight edges instead of drawing invisible lines, and strips leading zeros from edge labels (.32 instead of 0.32)cluster_summary() are now preserved in the macro diagonal, reflecting intra-cluster retention ratesoverlay_communities() for drawing community blob overlays on any network plot — accepts method names, membership vectors, or pre-computed community objectsplot_simplicial() for higher-order pathway visualization, rendering simplicial complexes as smooth blobs with flexible separators and a dismantled view optionvalue_nudge parameter to plot_transitions() for controlling the distance between flow labels and nodesbundle_legend_size, bundle_legend_color, bundle_legend_fontface, bundle_legend_positionlabel_size, label_color, label_fontface, label_hjust) to plot_transitions(), plot_trajectories(), and plot_alluvial()cluster_summary() for aggregating network weights at the cluster level, producing between-cluster and within-cluster matrices from raw transition databuild_mcml() for constructing Markov Chain Multi-Level models from edge lists or sequence data with automatic cluster detectioncluster_quality() for modularity-based cluster quality metrics and cluster_significance() for permutation-based significance testingas_tna() to convert cluster summaries to TNA objects for bootstrapping, permutation testing, and plotting with splot()simplify() for pruning weak edges from networks, with configurable weight threshold and aggregation methoddisparity_filter() for backbone extraction (Serrano et al. 2009), with methods for matrices, tna, igraph, and cograph_network objectsrobustness() for network robustness analysis with targeted (betweenness, degree) and random attack strategies, plus ggplot_robustness() for faceted ggplot2 outputtemporal_edge_list() for converting sequence data to timestamped edge listssupra_adjacency(), supra_layer(), supra_interlayer() for multilayer supra-adjacency matrix constructionlayer_similarity(), layer_similarity_matrix(), and layer_degree_correlation() for comparing layers in multilayer networksaggregate_weights() and aggregate_layers() for weight aggregation across layersverify_with_igraph() for cross-validating cograph centrality and network metrics against igraphmotifs() / subgraphs() as a unified API for triad census (node-exchangeable counts) and instance extraction (named node triples), with auto-detection of actor/session columns, rolling/tumbling window support, and exact configuration model significance testingplot_mcml() for Markov Chain Multi-Level visualization showing between-cluster summary edges alongside within-cluster detail, with pie charts, self-loops, and 22 customization parametersplot_chord() for native chord diagrams with automatic weight-based arc sizingplot_time_line() for cluster membership timeline visualizationplot_htna() orientations: "facing" (tip-to-tip columns) and "circular" (two semicircles), plus intra_curvature for drawing intra-group edges as dotted bezier arcsthreshold parameter to all plot functions for filtering edges/cells below a minimum absolute weightvalue_fontface, value_fontfamily, and value_halo parameters to plot_heatmap() for text styling controlscale_nodes_by: indegree, outdegree, instrength, outstrength, incloseness, outcloseness, inharmonic, outharmonic, ineccentricity, outeccentricityscale_nodes_scale parameter to splot() for dampening (< 1) or exaggerating (> 1) centrality-based node sizing differencessplot(): when plotting tna objects, qgraph-style parameters (vsize, asize, edge.color, lty, shape) are automatically mapped to cograph equivalentsnode_label_format (e.g., "{state} (n={count})") for showing counts on transition plot nodesbundle_size for aggregating individual trajectories into weighted summary lines in large datasetsshow_values / value_position for displaying transition counts on flow lineslabel_position consistency across ALL columns (first, middle, last) in trajectory plotsgamer_data, group_engagement, srl_dataset_node_groups() / get_node_groups() for managing cluster assignments on cograph_network objects$meta with getter/setter functionsgroup_tna support to splot() for direct plotting of grouped TNA modelscentrality_* wrapper its own focused help pagesplot() viewport calculationsplot()'s signature and silently dropped when dispatchingplot_heatmap() so high values get dark colorsbuild_mcml() density method crash when weight vector had no names.collect_dispatch_args() helper to replace 6 copy-paste dispatch blocks, using match.call() + mget() for reliable argument capturecentrality() with 23 measures and individual wrappers: degree, strength, betweenness, closeness, eigenvector, pagerank, harmonic, authority, hub, alpha, power, kreach, diffusion, percolation, eccentricity, transitivity, constraint, coreness, load, subgraph, leverage, laplacian, current-flow betweenness, current-flow closeness, voterankedge_betweenness() for edge-level centralitydetect_communities() with 11 algorithms: louvain, walktrap, fast_greedy, label_propagation, leading_eigenvector, infomap, spinglass, leiden, optimal, edge_betweenness, multilevel — plus com_* shorthand aliasescluster_significance() for permutation-based validationnetwork_summary() and summarize_network() for computing comprehensive network-level statistics (density, reciprocity, transitivity, diameter, components, degree distribution)plot_transitions() for alluvial/Sankey flow diagrams, with plot_alluvial() and plot_trajectories() wrappersplot_bootstrap() and plot_permutation() for significance-styled visualization of bootstrap and permutation test results — significant edges rendered solid on top, non-significant edges dashed behindplot_mixed_network() for overlaying symmetric (undirected, straight) and asymmetric (directed, curved) edges on the same networkplot_heatmap() for adjacency matrix heatmaps with optional hierarchical clustering and plot_ml_heatmap() for multilayer 3D perspective heatmapsplot_compare() for difference network visualization showing edge-weight changes between two networkssplot() S3 methods for tna_bootstrap and tna_permutation objectsmotif_census(), triad_census(), and extract_motifs() for triad motif analysis with pattern filtering, significance testing, and network diagram visualizationfilter_edges(), subset_edges(), select_nodes(), select_edges() for flexible network subsettingset_groups() for storing cluster assignments on cograph_network objects with automatic dispatch to plot_htna() / plot_mtna()cograph_network objects as input, in addition to matrices, igraph objects, and tna objectslayout_spring and layout_gephi_fr algorithms: vectorized attraction forces, edge aggregation for dense networkspar(pin) error on exit when plot device state was corruptedx across all plotting functions:
plot_tna(): input → xplot_htna(): input → x (was model)plot_mtna(): input → x (was model)splot() already used xtplot() default margins causing tiny plots compared to splot()vignettes/qgraph-to-splot.md)The following parameters have been renamed for consistency. The old names still work but emit deprecation warnings:
| Old Name | New Name | Reason |
|----------|----------|--------|
| esize | edge_size | Add edge_ prefix, expand abbreviation |
| cut | edge_cutoff | Add edge_ prefix, clarify meaning |
| usePCH | use_pch | Fix camelCase to snake_case |
| positive_color | edge_positive_color | Add edge_ prefix (matches theme storage) |
| negative_color | edge_negative_color | Add edge_ prefix (matches theme storage) |
| donut_border_lty | donut_line_type | Expand lty abbreviation |
edge_label_fontface now accepts string values ("plain", "bold", "italic", "bold.italic") in addition to numeric valuesmlna() for multilevel network visualization with 3D perspectivemtna() for multi-cluster network visualization with shape-based cluster containersplot_htna() for hierarchical multi-group network layouts with polygon and circular arrangementstplot() as a qgraph drop-in replacement with automatic parameter translationarrow_angle parameter for customizable arrowhead geometryedge_start_dot_density parameter for TNA-style dotted edge starts indicating directionfrom_tna() — no manual matrix extraction needednetwork and qgraph objects as inputpie_values vector to donut_fill when all values are in [0,1]splot() when other parameters were specifieddonut_shape validation rejecting custom SVG shapesfrom_qgraph() when a layout override was providednormalize_coords()from_qgraph() by using a matrix intermediary for per-edge vector reorderingnrow(el) crash: qgraph's Edgelist is a list, not a data.framedonut_empty parameter for rendering unfilled donut nodesfrom_qgraph() for converting qgraph objects to cograph format, reading resolved graphAttributes for accurate parameter extractionlayout_info guard causing errors on certain device configurationssoplot() for grid/ggplot2-based network plotting — full feature parity with splot() using a different rendering backendlayout_oval() for oval/elliptical node arrangementslayout_scale parameter to expand or contract the network layout, with "auto" mode for node-count-based scalingedge_start_style parameter for visually indicating edge direction via styled start segments (dashed, dotted)soplot() curve direction and edge defaults diverging from splot() behaviorrescale_layout distorting oval aspect ratios by switching to uniform scalingpar(pin) restoration error on plot device exitsplot() — a base R graphics engine for network visualization using polygon(), lines(), and xspline(), providing better performance than grid-based rendering for large networkssn_save() with configurable DPIdonut_color API to accept 1 color (fill), 2 colors (fill + background), or n colors (per-node)