countryatlas 3.0.0

The whole of the roadmap tracked in #19, in one release. 46 new exports and two new datasets take the package from “join World Bank data to a map” to “join anyone’s data, to the map as it was in 1950, and say honestly what the picture does and does not support”.

The version is 3.0.0 rather than 2.1.0 because three things change behaviour for existing code: world_map(projection = "mercator") now produces a different (and usable) map, morans_i() returns two more columns, and geom_country_labels() takes data as its second argument. Details under Breaking changes.

Breaking changes

• geom_country_labels() gains data as its second argument, matching the geom_*(mapping, data, ...) convention every other ggplot2 geom follows. A call that passed repel positionally – geom_country_labels(aes(...), FALSE) – now binds that FALSE to data and must name it. Named calls are unaffected.

• morans_i() returns n_excluded and excluded. Code that assumed a five-column result will see seven. The statistic itself is unchanged.

• world_map(projection = "mercator") draws a different map – see the bug fix below. It was previously unusable, so this is a fix rather than a regression, but the output does change.

• style = "binned" draws different bins in world_map(), globe_map() and value_by_alpha_map(). n_bins was silently ignored there (see the bug fix below), so the bin count and boundaries change for any existing call. style = "quantile" and "jenks" are unaffected.

• classification_report = TRUE returns NULL for style = "continuous" instead of a row-per-distinct-value table, and warns. Code that read the attribute after a continuous map got a table that described nothing; it now gets nothing, explicitly.

• world_map()’s new arguments were deliberately appended after recenter, so no existing positional call to it changes meaning.

• country_join() and country_join_all() gain warn, defaulting to TRUE. They now warn about country names that resolve to nothing, so code that joins on messy names will start emitting a warning it did not before. The join result is unchanged; pass warn = FALSE for the old silence.

• world_map(engine = "tmap") failed opaquely on an older tmap. The engine builds its scales with tm_scale_intervals(), tm_scale_continuous() and tm_scale_categorical() – the tmap 4 API. tmap 3 configured scales through arguments on tm_polygons() and exports none of them, and because DESCRIPTION pins no version on any Suggests package, the need_pkg("tmap") gate was satisfied by any tmap at all. An older install then died on R’s own “‘tm_scale_intervals’ is not an exported object from ‘namespace:tmap’”, which names neither the cause nor the cure. The engine now checks for the constructors it needs and says so, pointing at either an upgrade or engine = "ggplot2". The check is on the capability rather than a version number, as as_ggsql_source() already does for duckdb’s shared_home: that is the thing actually required, and it stays correct whichever release introduced it.

• locate_country() measured its snap-back distance from the wrong point. A point falling just outside the coarse coastline is snapped to the nearest country when it is within tolerance_km, and that step paired points with countries using st_nearest_points(..., by_element = TRUE). The argument is pairwise; by_element is not a formal of it, so sf absorbed the name into ... and ignored it, returning the full n×n cross product instead of n pairs – after which each unmatched point was measured against a different point’s country. Cuba’s centroid sits 10.8 km off the 110m coastline and should snap; in a 17-point call it was measured against American Samoa’s point, came out at 10167 km, and was dropped to NA. One unmatched point was safe, because 1x1 is the same thing as pairwise – which is why every single-point example was correct and only multi-point calls were wrong. The error ran both ways: a bogus distance landing under the tolerance would have attributed a point to a country it is not in. Fixed, with a length check so it cannot silently regress to a cross product again.

• spatial_lag() did not say which countries its weights excluded. A country with no neighbour gets NA, which the documentation states – but that NA is indistinguishable from one caused by a missing input value, and nothing in the result said which countries the weights had dropped. morans_i() and gearys_c() answer this by returning excluded; getis_ord() and local_morans() omit the row, so a setdiff() recovers it. spatial_lag() returns the caller’s frame, so it now attaches the codes as the "countryatlas_excluded" attribute – the same channel this package already uses for "countryatlas_cartogram" and "countryatlas_clubs". Not a warning: on real geography some country always lacks a land neighbour, so a warning would fire on every ordinary call. On a panel the attribute holds the union across years, since the weights are geography rather than time.

• neighbors() could not tell a typo from an island. An unresolved name becomes NA, and the %in% filter simply never matches it, so a misspelling returned no rows – indistinguishable from a country that genuinely has no land border. The function’s own example makes that collision explicit (“Japan has no land border”). distance_between() and convert_country() take the same input and both report an unresolved value; neighbors() now does too, and gains warn = TRUE. Iceland still returns zero rows in silence, because that zero is real.

• country_timeline() said nothing about a name it could not resolve. An unrecognised input came back as a row of NA, which reads as a country with no recorded history rather than a name nobody matched – in a package whose whole premise is that unmatched names are the problem worth reporting. dissolve_country() takes the same input shape and has warned about exactly this since 1.0.0. country_timeline() now gives the same warning, with the same check_country_match() pointer, and gains warn = TRUE to match its sibling. A historical name is still silent: "USSR" is meant to fail the ISO lookup and be resolved by the historical spine, which is why that lookup’s own warning stays suppressed.

• Three panel verbs did nothing, silently, on a cross-section. Handed one year per country, growth_rate(), lag_by_country() and diff_by_country() each added a column that was NA in every row and said nothing – the verb accomplished literally nothing, and the result read as a computation that had run. They now report it (countryatlas_all_na_result), stating what the verb needs (“a lag of 3 needs 4 years for the same country”) and pointing at complete_years(). The notice fires only when the whole column is NA and the source column had data, so the ordinary panel case – where just the first year per country is NA – stays silent.

• country_network() validated top_n after building the network. It was the one verb here that did its whole job and then rejected an argument it could have rejected at the door. top_n only trims the finished edge list, so the check is now made first.

• share_of_world() returned a column of NA without saying why. A zero or non-finite total is a dead end – the division would give NaN or Inf – and the guard that turns it into NA was right. Its silence was not: per_capita() and to_ppp() both report an unusable denominator, under the same two condition classes, while this one returned nothing but NA and read as “these countries have no share” rather than “there was no total to take a share of”. It now reports it, naming the years affected on a panel (only the bad years go NA; the rest keep their shares).

• world_map(engine = "tmap") now projects. It took projection and recenter and drew in the frame’s own CRS regardless (see the bug fix below), so every existing tmap map changes: the default is Equal Earth, as it already was on the ggplot2 engine. A bad projection name is now an error there rather than being ignored.

• Five verbs now return a tibble rather than the class they were handed. to_ppp(), smooth_rates(), interpolate_missing(), spatial_lag() and rank_countries() returned whatever arrived: a data.frame stayed a data.frame, and a grouped frame stayed grouped (see the bug fixes below). They now normalise, as their sibling verbs always did. Code that relied on df[, "col"] dropping to a vector, or on an inherited grouping surviving the call, changes.

• complete_years() returns an sf frame when given one. It previously dropped the class while leaving the geometry column in place, so a pipeline that repaired the result with st_as_sf() is now re-wrapping something that is already sf – harmless, but no longer necessary.

• A duplicated column name is now refused. per_capita() and to_ppp() accepted a frame with two gdp columns and silently computed from the first; ten other verbs failed with a message about internals in tibble. All of them now reject it up front (see the bug fix below).

• region is applied when geometry = "none". world_data() and join_world() ignored it on that branch and returned every country in the world (see the bug fix below), so a call passing both now gets fewer rows – the ones it asked for. A bounding box, having nothing to clip against, is refused rather than silently dropped.

• Several backends now warn where they used to accept an argument and drop it – scale and projection on the polygon backend, the ggplot2-only arguments on the tmap and mapgl renderers, title/subtitle on world_table(engine = "tibble"), and ... on two of interactive_map()’s engines. The drawings are unchanged; only the silence is. Code running under options(warn = 2) will now stop where it previously carried on.

• Argument values are no longer partially matched. Every exported function that takes a fixed set of choices now validates with rlang::arg_match() instead of match.arg(), so style = "quant" errors rather than resolving to "quantile". The error names the argument, the function and the valid choices, and suggests the one you meant – match.arg() reported only 'arg' should be one of ..., naming neither. Spelled-out values are unaffected.

• world_tiles places 171 of its 239 countries in a different cell. The break in the placement scan left only its inner loop, so a country whose own cell was taken was re-placed at the last free cell in the search square rather than the first – and every cell claimed along the way stayed marked occupied for nobody. That wasted 129 of 368 cells and pushed later countries further out still: mean displacement from a country’s true position was 1.9 cells with a worst case of 7.1, against 0.9 and 2.8 now. tile_map() layouts change accordingly. The grid is still 40x24 and still has one country per cell.

• dispute_policy() returns the policy it replaced, not the one you set. R’s convention for a setter – options(), par(), sf::sf_use_s2() – so on.exit(dispute_policy(dispute_policy("neutral"))) now restores the old value, where before it was a no-op. Reading the current policy with no argument is unchanged.

• world_geometry("coastline") and world_geometry("ocean") name their geometry column geometry. It was x, from st_as_sf() on a bare sfc, making these the only two what values whose column was not called what the return contract says. Code that referred to $x must change; code written against the other four now works on all six.

• gini() and theil() warn instead of returning a bare NA when the input is empty, or when na.rm = FALSE and values are missing. The value is unchanged; code running under options(warn = 2) will now stop where it previously carried an unexplained NA forward.

Bug fixes

• globe_map() built globes it could not draw. On the default sf backend, 63 of 216 sampled viewpoints (29%) failed when the plot was printed, with grid’s “Invalid graphics path”: a country on the horizon lost its far-side vertices in the orthographic transform and kept whatever was left (at lon 120, lat 20 Chad became a single point). The plot object built fine, so nothing noticed until it was drawn, and a 60-frame spin_globe(backend = "sf") could not get through one rotation. The viewpoints that did draw closed every horizon country with a straight chord. Geometry is now cut at the horizon on the sphere before it is projected, in globe_map(), world_map(projection = "orthographic") on either engine, world_geometry(), projection_compare() and tissot_map(), CShapes’ historical borders included; every sampled viewpoint draws, and the colour scale and coverage are the same from every side.

• world_map(na_style = "hatched") built maps it could not draw where sf does not load. ggpattern draws the stripes through gridpattern, which clips them with sf when the plot is printed. sf is always installed alongside, but where its system libraries (udunits, GDAL, GEOS, PROJ) are missing from the library path it cannot be loaded, so the map built fine and then failed on print with “unable to load shared object units.so”. Hatching now checks that sf loads and otherwise falls back to grey with a message, as it already did when ggpattern is absent.

• world_map(na_style = "hatched") took half a minute to print on the polygon backend. ggpattern clipped a fresh set of stripes to every polygon, and the no-data countries of a world map are many polygons (169 for co2_per_capita in world_snapshot, most of them islands, plus Antarctica). The stripes are now clipped once, to all of them together: the same map, with the same stripes, prints in about a second.

• Countries with no data vanished from animate_world() and got their own panel in facet_map(facet = year). attach_geometry() gives a country the data does not cover an NA year, so it belonged to no frame: the animation never drew it (gganimate warned twice while dropping it) and the small multiples drew it in an extra panel labelled NA. It is now drawn, as no data, in every period.

• value_by_alpha_map() and the value-suppressing palette ranked polygon vertices, not countries. The polygon backend repeats a country’s values down every vertex of its outline, so the default transform = "rank" weighted each country by how complex the outlines below it are: Chile, at the 69th percentile of countries by population, was drawn at the 30th, and the mean opacity error across 210 countries was 0.25. world_map(uncertainty = ) ranked the same way and put 101 of 189 countries in the wrong value x uncertainty cell. Both now rank one value per country, as the quantile breaks always have.

• attach_geometry() let a column of the caller’s replace the map geometry. A shared name keeps the caller’s version, so a group, lat, long or order column (a treatment group, a capital’s latitude) displaced the polygon backend’s own, and every map drawn from the result was nonsense: a two-value group drew the world as three polygons. That is now refused, naming the column; the sf backend is unaffected.

• historical_geometry() gave the Soviet Union Russia’s code. Gleditsch and Ward use one code for both, so the 1980 USSR came back as iso3c = "RUS" and attach_geometry(year = 1980) painted all fifteen republics with Russia’s value, the anachronism audit_time_coverage() reports as "before_existence". A code now stays NA until historical_codes says the state existed. ?historical_geometry also said owner is NA for a sovereign state (it is the state’s own code) and named the Gold Coast as an entity with no code (CShapes calls it Ghana and it carries "GHA").

• A non-breaking space turned the USSR into Russia. The historical-name lookup trimmed only ASCII whitespace, so "USSR" pasted from a web page missed its alias and fell through to countrycode, which resolves it to Russia alone: dissolve_country() returned one row instead of fifteen, check_country_match() did not flag it historical, and country_timeline() answered for Russia.

• lisa_map() reported two different coverages. Its caption counted the countries it coloured while map_provenance() counted every country with a value, including the ones the weights could not connect, which are drawn as no-data: 142 against 189 on the bundled snapshot. Provenance now counts what is drawn.

• od_map() dropped flows the basemap cannot draw without saying so. Hong Kong, Macao, Tuvalu, the British Virgin Islands and Gibraltar have no polygon on the polygon backend, so China’s largest destination simply vanished from China’s panel. They are now named in a warning.

• convergence_club() dropped countries without a complete series. A country with one missing year disappeared from the result altogether; it now comes back as club = NA, the documented “not classified”, with a warning naming it.

• smooth_rates() let a negative count into its model. The pooled rate and excess variance went off their domain, every _shrinkage came back above 1 (the help promises 0 to 1) and rates were pushed away from the global one. A negative count is now unusable and reported. The unusable-denominator warning also counted rows whose numerator was missing; it counts denominators only. rate_check() reported a standard error of exactly 0 for a negative rate and now gives NA, with a warning.

• world_map(uncertainty = , n_bins = ) said n_bins was ignored while using it, for the value classes of the palette. The notice is gone there; n_uncertainty without uncertainty, which really does nothing, now says so instead.

• join_world() replaced a caller’s region, continent or iso2c column in silence, where standardize_country() warns. It now warns too, when a value actually changes.

• One unmatched name in a region vector was dropped in silence. region = c("France", "Germny") drew France alone; the name that matched nothing is now reported, and a continent or group name inside a vector, which region accepts only on its own, is explained rather than called a failed match.

• growth_rate() read the row after an infinity as -100%. Anything finite over Inf is 0, so type = "yoy" gave -1 after an infinite value and type = "cagr" gave -1 for every year of a series starting at one. Both are NA now, with a warning.

• audit_coverage() counted an infinite value as present, while coverage_map(), which both help pages call the same report as a map, shows it as missing. It counts as missing in both now.

• world_map(engine = "tmap", style = "binned") drew pretty breaks. The ggplot2 engine draws n_bins equal intervals; the tmap engine now does too.

• A blank iso3c was reported as an excluded country by morans_i(), gearys_c() and the other spatial statistics, inflating n_excluded.

• A custom weights frame naming one link twice kept only the last weight. It is now refused, with advice to sum repeated flows first.

• standardize_subnational() refused a lowercase or padded ISO 3166-2 code ("de-by", "DE-BY "), under a message promising case-insensitive matching. Codes are upper-cased and trimmed first.

• A bad countryatlas.cache_max_age or countryatlas.cache_max_size was blamed on the cache directory. "a" or NA switched persistent caching off with “Cannot write to the cache directory”; a negative value was accepted. Both options are now validated when read, naming the option.

• bivariate_map() did not validate dim, so "a", NA, c(2, 3) or 2.5 failed in base R or in biscale; a custom biscale palette also turned the provenance style into one string per colour.

• interactive_map() crashed on a fill with nothing to scale. The leaflet engine died on “Wasn’t able to determine range of domain” for an all-missing numeric column, and the mapgl engine refused an all-missing column and failed on a constant one or one with a single country’s value (“values and stops must have the same length”). Such a map now draws as no-data, as the ggplot2 engines always did.

• interactive_map(engine = "plotly", projection = "orthographic") failed inside plotly (“number of columns of matrices must match”), whose converter cannot take an orthographic view. It now says so and names the engines that can.

• country_factsheet() misreported neighbours it had not computed. With the lookup unavailable (no sf), France’s sheet blamed the 110m basemap for “excluding” Andorra and Monaco, and Andorra’s said “None found”. The sheet now says the neighbours were not computed, and why.

• ?share_of_world promised a proportion in [0, 1] and ?index_to said nothing about a negative base year; both now say what a negative value does.

• compare_sources() names a source listed twice rather than failing on duplicated column names, and a non-finite year passed to world_data() or country_data() is refused as such rather than leaking a coercion warning.

• cartogram_map() and dorling_map() failed in the orthographic projection, inside cartogram (“all sizes are missing and/or non-positive”), because the far side’s vertices were dropped rather than cut at the horizon. They now draw the visible hemisphere. A contiguous cartogram in Winkel Tripel also failed when printed, on a graticule the theme never shows.

• growth_rate(type = "cagr") gave a country with a zero base year NA without saying so, and when that was every country, blamed the series for being shorter than two years. It now names them.

• deflate() left a row NA in silence when its year’s index was zero, infinite or missing, where to_ppp() and per_capita() report the same thing. It now says how many rows.

• compare_sources() crashed on a year no source covers, with base R’s “subscript out of bounds”; it now returns the empty comparison with a warning. An infinite value made the relative difference NaN, which disagrees counted as agreement; it is left out of the comparison and named.

• recenter = 360 or -360 failed on the sf backend with sf’s “polygons require at least 4 points”, as did an out-of-range value such as 500, which is now refused by name before any geometry is touched.

• bubble_map() ignored projection on its default polygon backend without saying so, as world_map() already reports; it now names backend = "sf" as the way to project.

• Errors named functions the caller never called. A bare column passed to one of the string-taking arguments was reported as “Error in value[[3L]]()” (a tryCatch() handler), a missing column in country_join_all() as “Error in FUN()”, and a bad origin, weights matrix, scale, region, footnote or override table under an internal helper’s name such as wdj_to_iso3c() or weights_custom(). Each now names the verb. A bad global option is reported without a call.

• The on-disk cache could delete files it had not written. With options(countryatlas.cache_dir = ) pointed at a folder that held anything else, the first cached fetch deleted every file in it without an .rds extension (a sweep meant for entries left by 2.0.x), cachem’s 30-day expiry and size cap applied to every .rds file there, and clear_wdi_cache(disk = TRUE) removed the folder and everything below it. Entries now carry their own .countryatlas extension, pruning and clearing touch only files the cache wrote, and the folder itself is removed only when that leaves it empty.

• local_morans() did not permute conditionally, as it documents. It shuffled all n values with one permutation shared by every country, so a country’s own value could land among its neighbours. For the extreme values a hot-spot map is about, that inflated the pseudo p-value: Monaco came out at 0.028 where the conditional reference distribution (and spdep’s localmoran_perm()) gives 0.0035. Each country’s neighbours are now drawn from the other n - 1 values.

• convergence_club() depended on row order. pivot_wider() lays its columns out in order of first appearance and the log-t test reads them by position, so a shuffled panel, or one whose first country lacked the first year, formed different clubs from the same data. The years are sorted first.

• smooth_rates() shrank every year of a panel toward the earliest year’s global rate, under a warning that said the other years were dropped though every row came back. The empirical-Bayes prior is now estimated per year, the design spatial_lag() already uses for a panel.

• bubble_map() drew a negative total as a bubble of its absolute value (France at -1.4e9 as large as China) and an infinite one as an infinite bubble. bubble_map() and spike_map() now set negative and infinite values aside with a warning and count them as missing; spike_map() used to drop them and blame a missing centroid. A missing size on the polygon backend no longer leaks ggplot2’s “Removed 1 row” at print time.

• An infinite fill was counted as shown. Every fill scale draws Inf as no data, but the footnote = "auto" caption, map_provenance() and coverage_map() counted it as present: “189 of 240 countries shown” over a map showing 187. It now counts as missing everywhere, na_style = "omit" and "hatched" treat it as the no-data it is drawn as, and the map verbs name the countries holding one. interactive_map(engine = "leaflet") no longer fails on one with leaflet’s “Wasn’t able to determine range of domain”.

• flow_map() with a missing or infinite weight leaked “Removed 50 rows”, or failed at print time with grid’s “‘lwd’ must be non-negative and finite”. Those flows are dropped with a warning, as flow_matrix() drops them. cartogram_map() and dorling_map() likewise name an infinite weight instead of handing it to cartogram.

• growth_rate(type = "yoy") returned Inf and NaN after a zero, where every sibling verb gives NA for a zero denominator. It now gives NA, with a warning.

• A row with a missing year was treated as the latest year. lag_by_country(), diff_by_country() and growth_rate() gave it a value computed from the last real year, beta_convergence() dropped its whole country, sigma_convergence() and share_of_world() reported a phantom NA year, and interpolate_missing() crashed on it, and could have overwritten an undated observed value with NA. Undated rows are now left out of anything that needs a position in time.

• complete_years(method = "locf") carried values in row order, not time order, when the data held years outside years, and returned the frame out of order.

• The United Kingdom left the EU a day early. country_groups_history reads to as the first day of non-membership (Austria left EFTA and joined the EU on 1995-01-01), but the UK’s exit was recorded as 2020-01-31, its last day as a member. It is now 2020-02-01, and the column’s help states the convention.

• projection_distortion() measured a WGS84 projection against a sphere. Equal Earth read between 0.9955 and 1.009 instead of 1, and Mercator showed up to 0.38 degrees of angular distortion, contradicting both checks its documentation offers. It now uses the ellipsoid’s radii of curvature and central differences: Equal Earth, Gall-Peters and the Lambert azimuthals read 1, Mercator 0, and the 0.7% that Mollweide and Eckert IV show (PROJ applies their spherical formulas to the ellipsoid) is documented.

• theil(na.rm = FALSE) broke its own decomposition on a missing group. The row stayed in total and fell out of both components, so total no longer equalled between + within. It now returns NA, as it already does for a missing value under na.rm = FALSE.

• Reports that said something false. country_timeline() warned that “Tanganyika”, “Zanzibar” and the “United Arab Republic” matched nothing while returning their history. audit_time_coverage() returned a phantom all-NA row for an unreadable year, and now names the dissolved codes it flags. compare_sources() kept unresolved keys as iso3c = NA rows, counted them in only_x/only_y, and called two of them “a duplicate country”. index_to() and deflate() reported two unidentified countries as “1 country: NA”. world_table() gave tied values different ranks. country_join() and country_join_all() warned that iso3c was “replaced” when it was the join column itself. subnational_map(projection = ) printed raw cli markup and advised geometry = "sf". audit_coverage() left a blank code out of unmatched, and check_dispute_coverage() listed NA and repeated values as bad codes.

• Inputs accepted and then failed somewhere else. A numeric region that is not a four-number box (region = 250) drew an empty map, and a reversed box was accepted. Custom country_weights() codes were taken verbatim, so lowercase ones made every statistic report “Not enough connected countries” with advice about islands; they are normalised, and codes that differ only in case are now named as the key problem they are. spatial_lag() on a panel aborted entirely when one year was too sparse. deflate(), to_ppp() and rate_check(rate = ) did not check that their value column was numeric, and per_capita(), to_ppp() and add_indicator() leaked dplyr’s join-type error for a character year. value_by_alpha_map() drew a character value as categories labelled “quantile”. classify_compare(), projection_compare() and od_map() died on “factor level [2] is duplicated” for a repeated method, projection or origin. country_join_all(by = iso3c) gave “object ‘iso3c’ not found”. historical_geometry(Inf) failed inside as.Date(), locate_country() on zero points leaked four base R warnings, and as_ggsql_source(format = "parquet") left its connection open when the write failed.

• interactive_map(engine = "leaflet") hard-wired a numeric colour scale. A discrete fill reached leaflet::colorNumeric() and died inside leaflet with “Wasn’t able to determine range of domain” – the same defect auto_fill_scale() was written to fix for the ggplot2 engines, and that the mapgl engine handles with match_expr(). ?interactive_map documents no per-engine restriction on fill, so the leaflet engine now branches on the column’s type and uses colorFactor() for a discrete one.

• The leaflet engine could be broken by a column named pal. Its fillColor = ~ pal(get(fill_name)) formula was evaluated by leaflet against the data as an environment, so a column of that name was found before the palette function and then called. The values are computed directly now, which also removes the only get()-in-a-formula column read in the package.

• The leaflet and ggiraph engines read iso3c without checking for it. check_cols() covered fill and tooltip, and check_map_geometry() does not require a key, so an sf frame without one failed at render time from inside leaflet or rlang. Both now name the missing column.

• bubble_map(backend = "sf") carried two geometry columns. as_tibble() strips the sf class but leaves the live sfc column, so the st_drop_geometry() further down saw a plain tibble and returned it unchanged – and the join then produced geometry.x / geometry.y, renaming the active column out from under coord_sf(). It now uses sf_drop() first, as four other verbs already do.

• index_to() returned an all-NA column in silence, three ways. It and deflate() take the same base_year and match it the same way, and deflate() reports every case where the rebasing cannot happen while this reported none. A Date or POSIXct year was the worst: == coerces the number to that class, so as.Date("2000-01-01") == 2000 compares against 1970-01-01 plus 2000 days and is false for every row – every country indexed to NA, and read.csv() with a date-parsing reader produces exactly that column. That shape is now refused by name. A base_year the panel does not cover, and a country whose base-year value is missing or zero, still come back NA as documented – but the verb now names the countries it could not index, so an all-NA column is readable rather than indistinguishable from a source that had no data. A character year still works, deliberately: this verb only matches on the year, and "2000" == 2000 is true.

• deflate() aborted on a zero-row panel, reporting “Years present: Inf and -Inf” from range() of nothing, and would have asked the World Bank for the year range Inf..-Inf. It returns early with the promised column, typed.

• rate_check() warned that a zero-row frame had no usable denominator. quantile() of an empty vector is NA too, so the guard for an all-NA denominator fired for a frame that simply had no rows.

• beta_convergence() and convergence_club() leaked base R’s “essentially perfect fit: summary may be unreliable”. It named neither the verb nor the column, and from convergence_club() it described an internal log-t regression the caller does not know exists. beta_convergence() now says so in the package’s own voice – naming which columns are meaningless as a result – and the internal fit is silent.

• simplify_geometry() skipped its documented MULTIPOLYGON cast for an sfc. The validator accepts sf and sfc; the cast gated on sf alone, so half the accepted inputs came back as a mix of POLYGON and MULTIPOLYGON.

• world_table(engine = "gt") rendered year as “2,020”. year now joins rank in the columns fmt_number() leaves alone – it is an identifier, not a measurement.

• tile_map()’s provenance called a numeric fill “categorical”. auto_fill_scale() picks the scale from the column’s type, so a numeric fill was drawn continuously and recorded as categorical, in the record whose job is to say what was drawn.

• split_antimeridian() wired up only the last crossing. A carry row was computed and stored for every crossing but only the last was read, so with two or more crossings the middle segments began at their first raw vertex instead of at the antimeridian edge, leaving a visible break on the re-entry side. The code read as though it were general; it handled one.

• A region that resolves to NA codes selected uncoded geometry. resolve_region()’s continent branch reads countrycode’s codelist, where a few rows carry a continent and no iso3c, and %in% treats NA as a value – so NA %in% c("FRA", NA) is TRUE and a subset keyed on the result pulled in every geometry row whose own iso3c was NA. The sf backend guarded this; the polygon backend did not. The codes are now filtered at the source, so all callers are covered.

• country_overrides() accepted any string as an iso3c value. wdj_to_iso3c() whitelists every override value as a legitimate code, so country_overrides(c(Freedonia = "1")) put "1" in the iso3c column and every join keyed on it. Values must now look like an ISO 3166-1 alpha-3 code. Separately, any(!nzchar(nms)) missed an NA name – nzchar(NA) is TRUE – while wdj_to_iso3c() rejected one, so the two validators disagreed about what a valid override table is.

• world_query() interpolated identifiers into SQL without escaping them. Only title was escaped, and it is the one value that lands inside quotes; every other name went in as bare SQL. A column name with a space produced invalid SQL with no diagnostic, and a name derived from untrusted input reached a string that ggsql::ggsql_execute() then runs. fill, source, size, facet, projection, palette and transform are now checked against a plain SQL identifier.

• nuts_geometry(countries = "Germany") said “NUTS covers the EU, EFTA and candidate countries only” – true, and not the problem. countries is read as iso3c, so a country name resolves to nothing; the error now names the origin that would have worked.

• od_map()’s origin and origins are one letter apart and adjacent in the signature, and mean unrelated things. Both mix-ups used to fail somewhere else; each is now diagnosed where it was made.

• geom_country_labels() did not validate size or mapping, which feed ggplot2 arithmetic and modifyList() respectively; world_table() checked title but not subtitle; and ggsql_wkb_frame() overwrote an existing column of the geometry column’s name in silence, where eleven other column-adding verbs warn.

• world_query() silently ignored arguments its layer does not use. size belongs to layer = "bubble" and n_bins to layer = "binned", and the other layers took them without a word: n_bins was dropped, since only the binned layer emits a BIN clause, while size still went into the VISUALISE list as pop AS size on a choropleth, which has no size channel to put it on. Both are now reported, together when both are given. The emitted query is unchanged – a pass-through builder passing a clause through is defensible – but the silence was not, since the abort beside it already treats a layer/argument mismatch as worth naming.

• globe_map() crashed at some viewpoints with an error from the geometry engine. globe_map(sf_data, lon = 90, lat = 30) died on IllegalArgumentException: point array must contain 0 or >1 elements, which names neither the projection nor the viewpoint. coord_sf() draws a graticule, and on the one genuinely hemispheric projection those lines are clipped at the horizon – a clipped line can reduce to a single point, which GEOS rejects. winkel_tripel already had datum = NA for its own graticule trouble; "orthographic" now does too, and all 65 sampled viewpoints draw. The other three azimuthal projections show the whole globe, so nothing is clipped and they are unaffected.

• country_weights() silently ignored the arguments its scheme does not use. k, cutoff_km, w and scale each belong to exactly one scheme, and the other three dropped them without a word. The costly case is country_weights("knn", w = my_matrix): the caller’s own adjacency matrix was discarded and nearest-neighbour weights returned instead, so the result looks entirely reasonable and is not what was asked for. cutoff_km passed to "knn" reads as a distance cap that was never applied, and scale only ever affected "contiguity". All eleven combinations now report which arguments were ignored and which one the chosen scheme is built from; each scheme’s own argument, the shared style, and a default passed explicitly stay quiet.

• Arguments that do not apply were silently ignored by the map verbs. The package reports an inert argument rather than pretending to honour it – that is what warn_projection_ignored() and warn_recenter_ignored() are for, and attach_geometry(), world_geometry(), world_data() and join_world() all call them. world_map() did not, though it takes both arguments and is the verb people actually reach for: on the polygon backend, which draws in unprojected longitude/latitude, projection = "mollweide" and recenter = 180 looked honoured and changed nothing. n_bins was inert in the same way under style = "continuous" (a colourbar has no classes) and style = "categorical" (the classes are the values) – the same complaint the 3.0.0 fix for style = "binned" answered. All three are now reported, under countryatlas_projection_ignored, countryatlas_recenter_ignored and countryatlas_n_bins_ignored, and value_by_alpha_map() gets the two that apply to it since it carries its own copies of those arguments. facet_map() and coverage_map() inherit them by passing ... through.

• attach_geometry(year = ) warned about the caller’s frame size, not the geometry. Its threshold was matched < nrow(geom) * 0.5, so an ordinary three-country frame – 3 of the 97 entities CShapes has for 1960 – was told “only 3 of 97 historical entities matched” and advised to join on gwcode “for full historical coverage”, which is not what someone asking for three countries wants. What the message claimed is a property of CShapes rather than of the request: a few entities never had an ISO code (6 of 97 in 1960, 3 of 174 in 2019) and no iso3c join can reach them. That is what it now reports, independently of the frame, and the nothing-matched case is left to the same helper every other geometry verb uses. The whole historical branch had no test coverage.

• A missing-column error named an argument the function does not have. check_cols() wrote `data` into its message unconditionally, which is right for the twenty-odd verbs whose frame argument is called that and wrong for the three whose is not: country_weights(w = ) reported Columns "iso3c" and "neighbor" not found in `data``` and `cartogram_diagnostics(x = )` reportedColumn “nope” not found in `data```. The helper now takes the name to use, and the duplicated-column error beside it does too.

• A provider renaming its time column turned the years into the values. adapter_reshape() – the shared step that puts a provider’s response into the package’s shape, behind fetch_owid(), fetch_eurostat(), fetch_oecd() and fetch_comtrade() – checked that the entity and value columns exist but not the year column. The value column is auto-detected as the first numeric column other than the year, excluded by name, so with the year column missing the exclusion removed nothing and the provider’s own time column was picked as the indicator: the mapped values came back as 2000, 2001, 2002 while year was entirely NA, with nothing said. A missing year column is now reported like the missing value column beside it.

• locate_country() gave a raw sf error for a points object with no CRS. sf::st_as_sf(df, coords = c("lon", "lat")) leaves the CRS missing whenever crs is forgotten, and the object then reached st_transform() and came back with sf’s “cannot transform sfc object with missing crs” – naming neither the argument nor the fix. The non-sf case one line above had already been given a proper message; this is the same failure one step further in. It now says which argument is at fault and how to set the CRS.

• subnational_map() joined against the wrong column when the caller’s code column collided with the geometry’s. The geometry side is meant to be nuts_id “whatever the caller’s column is called”, but the guard was if (!by %in% names(geom)) – false exactly when by names one of the geometry’s own columns. by = "name" therefore joined the caller’s NUTS codes against the geometry’s region names and matched nothing, and by = "iso3c" silently joined at country granularity instead of regional. by is documented as “the code column in data”, so the geometry side is now always nuts_id.

• subnational_map() blamed the codes for an all-NA indicator. It counted matches as sum(!is.na(fill)), so a panel whose indicator is entirely NA – a real thing to map, and one this package draws with an na.value and reports in the caption – was refused with “no rows matched the geometry”, sending the reader to check NUTS vintages for a mismatch that never happened. The count is now taken on the join key. Both of these sat in code that no test reached, because the path needs a GISCO download: R/subnational.R was at 66% coverage where every other file was above 90%. It is now exercised with a mocked nuts_geometry().

• correlate_indicators() let a tidyselect error through for a missing column. Every other verb reports one as `Column "x" not found indata``; this one selected throughdplyr::select()and surfaced thevctrs_error_subscript_oobraised by tidyselect -- "Can't select columns that don't exist" -- which was the last error at this boundary not carrying the package's class. The selection is a full tidyselect expression, so it cannot be checked by name up front; the failure is caught and re-worded instead, andstarts_with(),where()`, ranges, bare columns and the no-argument form all still work.

• Rows whose iso3c did not resolve were treated as one country. dplyr::group_by() puts every NA in a single group, so a panel carrying two unmatched names was one country to every verb that reads a neighbouring row within it: growth_rate() reported the change from one unmatched row to the next – 899% between two unrelated countries – and lag_by_country(), diff_by_country(), index_to(), interpolate_missing(), deflate() and beta_convergence() read across them the same way, silently. That frame is exactly the one a user is most likely to hold, because standardize_country() has already warned that those names did not match and left their iso3c as NA. The grouping now falls back to whatever does identify the row – the same country, then group order distinct_countries() uses for its uncoded branch – so an appended aggregate row (no iso3c, country = "World") still groups as one series, while rows nothing identifies each get a key of their own and are no longer read across. complete_years() had the same conflation twice over: the two unresolved rows shared one completed year grid, and the geometry carry that gives invented rows their country’s shape matched on iso3c – where match() treats NA as equal to NA – so it copied one unidentified country’s polygon onto the other’s invented rows. Both are keyed on the unit now, and each unresolved row gets its own grid and its own shape.

  A blank code is treated the same way. "" is not NA, so every is.na() guard missed it – but read.csv() without na.strings = "" gives a blank for every empty cell, and standardize_country("") already resolves to iso3c = NA, so a blank identifies no country anywhere else in the package. Two blank-coded rows produced the same fabricated 899%. Whitespace-only codes count as blank too, using the same Unicode [\h\v] class standardize_country() uses rather than trimws()’s ASCII-only default.

  The two panel guards were keyed on iso3c while the verbs grouped by the unit, which made them disagree: four single-year unidentified rows were reported as one country with gaps between them, and two unidentified rows in the same year as a duplicated country-year. Both now key on the unit, so they describe what the verb actually did; a genuinely duplicated country-year and a genuinely irregular panel are still reported.

• The unparseable-year warning printed its own markup and padded its list with NAs. Two defects in one message. cli does not re-interpolate a substituted value, so {source_label} emitted the label verbatim: three of the six call sites showed users `{.arg data}`: 2 time values are not a year and, from the public source extension point, Source {.val {source}}: – where the braces also referenced a variable that exists only in the caller’s frame. Separately, the list of offending values was sliced to min(4, sum(bad)), sizing it by the number of bad rows while the values themselves were de-duplicated, so a column of one repeated placeholder – "N/A", "..", "-", the common case – read "N/A", NA, NA, and NA. The label is now pre-rendered in the caller’s environment and the list is taken with head(), so the warning names its source and shows only values that are really there.

• audit_time_coverage() invented history for a factor iso3c. It looked the dissolution and succession dates up by indexing a named vector with the frame’s iso3c column – and indexing a named vector with a factor selects by the factor’s integer codes, not its labels. A frame from read.csv(stringsAsFactors = TRUE) therefore matched whichever rows of historical_codes happened to sit at those positions: a clean panel of France, Germany, Italy and Spain came back reporting France dissolved in 1993, Italy in 1992, Spain in 1991 and Germany as existing only from 2010, from the one verb whose whole job is catching that kind of mistake. The year key had already been hardened against the same class of bug (a Date year became a day count, hence read_year()); the iso3c key had not. The column is now coerced before the lookup, so a factor iso3c gives the same answer as a character one – verified across 27 verbs, none of which now differ on a factor iso3c or an integer value column.

• A stray 60 KB Rplots.pdf shipped inside the source tarball. Printing a plot with no device open makes R open the default one, which in a non-interactive session writes Rplots.pdf into the working directory – for the test suite, tests/testthat. .Rbuildignore carried ^Rplots\.pdf$, anchored at the package root, so it matched the root copy and not the nested one, and R CMD build included the artefact. The pattern is now unanchored, and a test setup file points the default device at the null file so the file is not written in the first place – plots still render, so a drawing error still fails a test.

• A factor year column silently produced wrong numbers. dplyr::arrange() and order() sort a factor by its level index, not by the label, and stats::approx() coerces a factor the same way. A year column arrives as a factor more often than it looks – read.csv(stringsAsFactors = TRUE), some importers, and any deliberate factor(year) for plotting – and every verb that reads a neighbouring row then read the wrong neighbour. lag_by_country() and diff_by_country() took the value from the wrong year, growth_rate() reported -0.75 on a series that had doubled, and interpolate_missing() left the gap NA because the targets fell outside the anchors it had mis-placed. Nothing warned: the panel came back the right shape with the wrong values. warn_irregular_years() had the same trap from the other side – as.numeric() on a factor returns level indices, so it reported a perfectly regular annual panel as having gaps. The nine year-ordering sites, the linear interpolation path and the gap check now go through one year_sort_key() helper, so a factor or character year gives the same answers as a numeric one, and the gap warning is now classed countryatlas_irregular_years so a deliberately decadal panel can silence just that one. earliest_per_unit() had carried this coercion since 2.0.0, which is why the map-drawing path was unaffected and this survived. A year column that is not numeric at all – an ordered factor of period labels – still sorts by its own order, and complete_years() still asks for a numeric year, since it has to generate the grid rather than merely order it.

• A bare column name was reported as “object not found” by six arguments. A handful of arguments take a column name as a string while the verbs around them take a bare column through tidy eval, so the package converts that predictable slip into a message naming the argument and showing the quoted form to write. The conversion had been applied to the value, indicator, columns, codes and by arguments of interpolate_missing(), complete_years(), audit_coverage(), world_table(), country_codes() and aggregate_regions(), and not to the join-key and query arguments of the same shape: aggregate_regions(by = region) explained itself while attach_geometry(by = iso3c), subnational_map(by = nuts_id), register_country_source(key_col = iso3c), world_query(facet = year), world_query(size = pop) and as_ggsql_source(geometry_col = geom) all reached the user as base R’s “object ‘iso3c’ not found”. All six now report the argument and the fix. A genuine error inside such an argument still surfaces as itself.

• The plotting verbs rejected a column passed as a string. Every unquoted column argument in the package goes through one helper, whose own error hint tells the user to “pass the column unquoted (fill = my_col) or as a string”, and the column check then accepted the string form happily. But world_map(), tile_map(), bubble_map() and interactive_map() spliced the captured expression straight into aes(), which honours only the unquoted form: fill = "value" mapped the constant string "value". The failure then arrived at build time, from ggplot2, as “Discrete value supplied to a continuous scale” (or “Binned scales only support continuous data” under style = "binned") – naming neither the argument, nor the column, nor the fix, which is the exact failure the helper exists to prevent. world_map(d, "value") – the package’s most common call – errored while world_map(d, value) drew. The mapping is now built from the validated column name, so both forms produce the same plot; coverage_map(), spike_map() and the quantile, jenks and categorical styles were already correct, which is why this survived.

• interpolate_missing() invented data across decades. max_gap is documented as “the longest run of consecutive missing years to fill … because interpolating across a decade is not interpolation”, and it counted missing rows. On any panel that is not annual the guard was therefore defeated: a decadal panel of 2000, 2010, 2020 with 2010 missing is one missing row, so the default max_gap = 3 filled it – placing an invented value ten years from either anchor, the exact thing the parameter exists to refuse. Five-yearly data was interpolated across 15 years the same way. A run is now refused if it is longer than max_gap in rows – which is what an annual panel means by consecutive missing years, so annual behaviour is unchanged – or if any filled year sits further than max_gap from its nearest observation, which is what makes the decadal case wrong. Measuring the bracketing span instead would have been too blunt: on 2000, 2001, 2005 the filled 2001 is one year from an anchor and interpolating it is sound, even though the span is five. The documentation was already right; the code has caught up with it.

• interpolate_missing(method = "linear") leaked a dplyr internal for a year it could not read. Linear interpolation places the filled value along the year axis via approx(), so a labelled year column arrived as NA and surfaced “need at least two non-NA values to interpolate” wrapped in an across() error, naming neither the column nor its type. It is now a classed error naming the unreadable values and pointing at method = "locf", which carries the last value forward in row order and needs no arithmetic. It tests whether the year can be read as a number rather than whether it is numeric, deliberately: approx() reads "2000" happily, so a character year still works here – the difference from deflate() and beta_convergence(), which do the arithmetic themselves.

• gearys_c() and getis_ord() listed their count columns without saying what they count. morans_i() defines all three – n (countries used), n_excluded (countries with data the weights could not reach), n_links (non-zero weights) – and the other two named the same columns and left them undefined, so a reader had to go to a third function’s page. The counts are consistent: verified across all six spatial verbs on one graph, n is 187 everywhere it appears, n_excluded 2, n + n_excluded is the 189 countries supplied, n_links matches the non-zero weights, and the local forms return exactly n rows. All of that is now stated where the columns are listed.

• audit_coverage() did not say which way round its counts run. Its @return described the three elements but none of their columns, and both na_rates$n and by_group$n_countries are denominators – everything counted, not everything present. That is the opposite orientation from map_provenance(), whose n_countries is the numerator, and the two share that column name: on 215 countries with 24 missing, audit_coverage() gives n = 215 where map_provenance() gives n_countries = 191. The package had already spelled this out for map_provenance() (“the numerator, not the denominator, which its name does not say on its own”); the verb whose whole purpose is reporting coverage had not. Every column is now defined, with the contrast stated. The numbers themselves were correct throughout – verified against the data, including that the group counts sum to n and each na_rate is the within-group share.

• as_ggsql_source() accepted things that are not data, and gave a bare error for a closed connection. data is documented as a map-ready frame and was never checked, so as_ggsql_source(1:5) wrote an integer vector out as a table and returned a connection – a “world source” containing no countries at all – while NULL surfaced Data frame with at least one column required, a message from DuckDB rather than from here. A con the caller had already closed reached dbWriteTable() as base R’s unclassed Invalid connection, naming neither the argument nor its state. Both are now classed errors, checked at the one point that knows whose connection it is – which is also why the existing bookkeeping there closes a connection the function opened and leaves one passed in alone.

• A negative or infinite custom weight was accepted and quietly changed the answer. country_weights("custom", w = ) validated the matrix’s dimnames, squareness, type and NAs, and the type message even promises “any non-negative number” – but nothing enforced it. Row standardisation divides by the row sum, so one negative weight moved Moran’s I from 0.714 to 0.497 with no warning, and an all-negative matrix cancelled to exactly the all-positive answer, discarding the caller’s signs entirely. An infinite weight normalised to NaN, and the verb then reported not enough connected countries with data – diagnosing connectivity when the cause was the weight. gini(), theil() and the global G all refuse a negative input for the same reason, so both are now classed errors naming how many entries are at fault. The long-frame form of w gets the same two checks, and all three built-in constructors satisfy them.

• A partly named custom_match silently ignored the unnamed entries. The check tested is.null(names()), which catches a wholly unnamed vector but not c(Freedonia = "FRA", "DEU"), whose names are c("Freedonia", ""). An entry named "" or NA can never match a spelling, so it sat there doing nothing while the caller believed it was overriding something – the same silent-but-useless entry the surrounding comment was written about, and invisible in the output, unlike a mistyped code which at least shows up in the iso3c column. Every element must now be named. Values are still not checked against the ISO list, deliberately: the package’s own overrides map Kosovo to XKX, a World Bank user-assigned code that is not in countrycode::codelist$iso3c, so validating that way would reject the very case overrides exist for.

• A custom source that failed, or was written with the wrong signature, reported neither. register_country_source()’s boundary checks what an adapter returns thoroughly – a non-data-frame, a missing key column and a duplicate key are all named and pointed at the fetch contract – but not the call itself. An adapter that raised its own error re-raised it bare, so "provider is down" reached the caller with nothing to say which source produced it; and one defined as function(indicator) surfaced R’s unused arguments (countries = countries, years = years), naming neither the source nor the documented arity. Both are what someone developing an adapter hits, and both are the one moment the package knows which source it just called. They are now classed errors that name the source, preserve the provider’s own message verbatim (braces included – a cli template would have interpolated them), and point at the contract.

• A flow between near-antipodal countries was drawn as a streak across the top of the map. great_circle() sampled a fixed 50 points along the arc, and a near-antipodal arc passes within half a degree of a pole, where longitude turns almost arbitrarily fast: Belgium to Tonga stepped 131 degrees of longitude between two consecutive points, Greenland to Japan 121, Tonga to the Netherlands 136. split_antimeridian() only cuts a step wider than 180, so none of these were cut and geom_path() joined them literally – the same failure that function was written to fix for trans-Pacific flows, caused by the pole rather than the antimeridian. 16 of 8010 real ordered country pairs hit it, every one at latitude 87 to 89. The path itself was right; 50 points is simply too coarse where it turns fastest, and the step shrinks in proportion to the sample count. The offending segments are now bisected until every step is under 20 degrees, so all 8010 pairs come in under that, at a cost of at most 63 points instead of

  50. An ordinary arc never trips the threshold and keeps its n points exactly.

• gridded_cartogram() now documents that crowded neighbours overlap. Each country’s block is centred on its own centroid with no collision avoidance – a deliberate choice, since a global packing solve would push countries away from where they belong, and it was recorded in a code comment. It was not in the documentation, though, and it is not marginal: at the defaults about a third of the cells are drawn on top of a cell belonging to a different country, across some sixty countries, rising to roughly two thirds at cells = 2500. A partly hidden block cannot be counted or compared, which is the one thing this verb asks the reader to do. A new section says so, quantifies it, and names cell_size as the lever – dropping it from 2.5 to 1.5 cuts the overlap at cells = 1000 to about a tenth – with a pointer to dorling_map(), which resolves collisions by displacement instead.

• beta_convergence() returned a silent NA for speed and half_life on an unbalanced panel, and the documented reason did not apply. Turning beta into an annual rate inverts beta = -(1 - exp(-lambda T))/T, which needs 1 + beta * T > 0. That holds whenever every country spans the same T, but beta is fitted on growth already annualised per country, so a mix of spans leaves the mean span irreconcilable with the fitted slope. On a panel where half the countries were observed for 10 years and half for 50 – ordinary coverage for real data – beta came back at -0.041 with p = 9.5e-14 and r_squared = 0.96, unmistakable convergence, while the two most interpretable columns were blank and the documentation said NA meant beta >= 0, which it plainly was not. The guard itself was right: log() of a negative is not a speed. It now warns, naming the range of spans and saying that beta and its inference are unaffected, and the documented NA conditions list both cases.

• local_morans() returned a lag column that was not the neighbour average it is documented as. It reported the lag of the centred value, which is what the statistic and the quadrants are built on, beside a value column that is raw – so the two columns in the same tibble were on different scales, and neither matched spatial_lag()’s column of the same name. That breaks the Moran scatterplot the pair exists for: with value centred on 17384 and lag on 0, the quadrant boundaries fall in different places on the two axes, so the usual reference lines disagree with the cluster column beside them. The reported lag is now the plain neighbour average, identical to spatial_lag(). The centred lag is still what ii and the four quadrants use internally, which is what Anselin (1995) defines them on, so no statistic, p-value or cluster label changes.

• A value-suppressing uncertainty palette drew nothing when only one country was usable. The two axes are ranked with percent_rank(), which is (rank - 1)/(n - 1) and therefore NaN for a single observation, so cut() returned NA and the row got no colour – a country whose value and uncertainty were both present, drawn as though neither were. This is not a one-row-input curiosity: a mostly-missing uncertainty column with a single usable country blanked the entire VSUP layer. A lone observation has no rank position relative to others, so it now lands in the middle of each ramp, claiming neither extreme, which is also where a maximally uncertain value sits. Rows that are genuinely missing still get no colour, and a palette with more than one usable row is unchanged.

• rate_check() ranked a zero-count row as the most reliable in the table. The output is documented as “sorted with the least reliable first”, and it sorted on expected_se, which is the Poisson standard error sqrt(y)/d – exactly 0 when the count is 0. So a country with no events out of 251 people sorted last, presented as the most trustworthy row on the page, while another with one event out of the same 251 sorted first as the least trustworthy. Observing nothing is not evidence of precision; it is the small-number problem this verb exists to name, and nothing else in the output caught it either, because the default threshold is a quantile of the denominators. The ordering now uses the standard error a single event would imply, sqrt(max(y, 1))/d, which is identical to expected_se for every row with at least one event – so no other row moves – and expected_se itself is unchanged, still the plain Poisson standard error.

• convergence_club() found no clubs at all on an indexed panel. The log-t statistic refused to compute whenever H_1, the cross-sectional dispersion in the first period, was zero – and it is exactly zero whenever every country starts equal, which index_to() guarantees by construction. So the natural pipeline, convergence_club(index_to(panel, value, base_year = ...), value_index), silently placed nobody: the same panel in levels found seven clubs covering 16 of 20 countries, and indexed found none.

  H_1 cannot affect the answer. log(H_1/H_t) is log(H_1) - log(H_t), and log(H_1) is one constant across the regression, so it lands entirely in the intercept while the statistic returned is the t on the slope – verified bit-identical for H_1 of 1e-6, 1 and 1e6, and for omitting it altogether. The guard now requires only the periods inside the regression window to be positive and finite, which also stops earlier periods being over-checked when they never enter the fit. A window with genuinely zero dispersion is still NA, since there is no decay rate to estimate.

• repair_country_names() promised something it does not do. The documentation said the no-stringdist fallback “repairs a subset of what Jaro-Winkler would … and never picks a different country”. The second half is false, and not marginally: the two metrics choose the nearest known name independently, so "Libia" is repaired to Liberia with stringdist and to Libya without it, both accepted at the default threshold. Across a 42-misspelling sweep the two disagreed on the candidate five times, once reaching the user as a different substitution. No code change can make two different string metrics agree on a nearest neighbour – and normalising the candidate gate in the fallback, which was the obvious repair, agrees less often (35/42 against 37/42), so the metric is left alone. The documentation now says plainly that the engines can land on different countries, recommends installing stringdist, notes that neither metric is uniformly better (Jaro-Winkler repairs "Maroco" to Monaco, not Morocco), and points at the reported substitutions and the "repairs" attribute as the thing to check. The same paragraph’s aside about transpositions was ambiguously worded and is now explicit with the numbers: the fallback charges two edits for a transposition, so "Germny" is repaired (0.14) and "Frnace" is not (0.33), where Jaro-Winkler scores the latter 0.06 and repairs it.

• getis_ord(local = FALSE) returned a silent NaN in three cases. The global G is a ratio of cross-products, and the denominator is zero when every value is zero (0/0) or non-finite when the products overflow or underflow at extreme magnitudes (Inf - Inf, or an underflow to 0). The branch immediately above it already warns and returns NA for negative values, citing gini() as the precedent for “outside the domain is NA plus a word about why”; these three slipped past it. They now warn and return NA too, and the overflow message says that the statistic is unchanged by a positive scale factor, so rescaling the column fixes it.

• The permutation p-values did not say which tail they used. morans_i() documented its own (one-sided, upper); local_morans() and gearys_c() documented neither their tail nor that they consume the RNG. That matters for reading them: local_morans() is two-sided, and gearys_c() is one-sided on the lower tail – the opposite way round from Moran’s I, because Geary’s C runs from 0 (neighbours identical) up through 1 (no autocorrelation), so a small c is the evidence of clustering. All three now give the formula, note that the p-value can never be exactly zero (its floor is 1/(n_perm + 1)), and say to set a seed for reproducibility.

• A categorical map’s colours depended on the machine’s collation locale. A character fill column reaches ggplot2 unfactored, and its discrete scale derives the level order by sorting – which consults LC_COLLATE. The same script on the same data gave a legend of Belgium, Chad, Zambia, aland, Åland under C collation and aland, Åland, Belgium, Chad, Zambia under en_US, so every category was drawn in a different colour and two people running identical code got different maps. The level order is now pinned byte-wise (order(method = "radix"), which is precisely what a plain sort() does not do), so it is the same everywhere. An incoming factor is left alone – the caller has already chosen an order. Same fix applied to interactive_map()’s mapgl branch, where the sorted values are paired positionally with the colour stops, and to the classification_report fallback, whose rows came out in a locale-dependent order.

• classify_compare() drew half a panel as grey, and which half depended on row order. It took the class breaks from one arbitrary row per country – the same flaw as the break computation behind world_map() – but here every row is then cut() against those breaks, so the rows the chosen year did not cover fell outside the range and came back NA. On a two-year panel one year classified correctly and the other vanished into na.value, and swapping the input’s row order swapped which. It now de-duplicates (country, value) pairs, which still collapses the polygon-backend vertex rows the de-duplication exists for and spans the panel otherwise.

• distinct_countries() disagreed with itself. The branch for rows that resolved to an ISO code picked the earliest year explicitly; the branch for rows that did not, three lines below, still took distinct()’s first row. A panel carrying an unmatchable country name in two years therefore kept the earliest year for every country that resolved and an arbitrary one for the country that did not. Both branches now use the same rule.

• compare_sources() dropped duplicate rows in silence. A provider answering twice for one country-year hands back whichever row came first, order-dependently. fetch_wdi() reports exactly this and its comment argues why – “this is the response of a third party, which is more reason to” – and this path said nothing. It now warns with the count and says the first row was kept.

• A map’s colour classes depended on the caller’s row order. The break computation de-duplicated to one row per country, which is right for a polygon-backend frame – one row per vertex, hundreds per country, all carrying the same fill, and quantiles must not be weighted by how complex an outline is. But it picked an arbitrary row when a country’s rows genuinely differ, i.e. a panel: the same panel handed over in a different order gave quantile breaks of 10-100 or of 1000-10000, so the same data drew a different map and nothing said so. It now de-duplicates (country, value) pairs, which collapses the vertex rows exactly as before and spans the whole panel otherwise – what facet_map(facet = "year") wants from a shared scale, and order-independent either way.

• Coverage counts depended on row order too. na_coverage() – behind the world_map() caption, map_provenance() and the honest-maps verbs – took the first row per country, so the same panel reordered reported 2 of 4 countries missing or 0 of 4. It now counts a country as shown if any of its rows has a value, matching imputed_count(); identical on the map-ready cross-section it is documented for. This mattered in practice because facet_map(facet = "year") deliberately hands world_map() the whole panel, so that arbitrary number was the caption on a plot showing every year.

• classification_report = TRUE reported a different table for the same map. It counts countries per class, so one row per country is the right shape, but it too used distinct()’s first row. It now takes the earliest year, the rule distinct_countries() already promised, which is also now shared between the two rather than written out twice.

• map_provenance() reported n_imputed = 0 for every data frame, no matter how many values were interpolated. The data-frame branch never set the field, and the fallback below it is %||% 0L – so unlike the other unset fields, which default to NA and claim nothing, this one asserted that nothing had been imputed. A map-ready frame carrying two interpolated values reported zero of them, in the function whose whole job is reporting provenance. n_imputed is a data-side fact and is now computed from the frame; the plot path is unchanged.

• imputed_count() picked an arbitrary row per country. It counts once per country, because a map draws one polygon per country, but it did so via distinct(.keep_all = TRUE) – the first row. Identical on the map-ready cross-section it is documented for, but on a panel the first row is an arbitrary year, so a value interpolated in any other year counted as nothing imputed at all. Now “imputed in any row for this country”.

• Running interpolate_missing() twice turned the imputation flag off, which the documented hard rule says cannot happen. The flag was recomputed from scratch on each call as “was NA before, is not now” – and after the first call the filled cells are no longer NA, so every TRUE became FALSE. world_map() reads that column to avoid drawing imputed values as though they were observed, so a second defensive call – or a re-run notebook cell – silently produced a map with no provenance caption. “This value was imputed” is a property of the data, not of the call that produced it, so an existing logical flag is now carried forward and the operation is idempotent. Carried as a column rather than a vector, because the pipeline re-sorts by iso3c and year. A flag column that is not logical is somebody else’s, and still warns that it is being overwritten.

• One infinite value silently voided a whole z_score column. scale() drops NA but runs an infinity straight through the mean and the SD, so a single Inf made every z_score in rank_countries() NaN – beside a rank and a percentile that were still correct, because both are rank-based and untroubled by an infinity. That is the easiest kind of corruption to miss. z_score is now NA with a warning saying which two columns are unaffected.

• sigma_convergence() returned a silent NaN for a year containing an infinity, through the same !is.na() filter that beta_convergence() had: Inf passes both the NA test and > 0, so it reached sd(log(x)). The filter now tests is.finite(), and the reported n already says how many observations survived.

• An infinite base year made index_to() report a country as having collapsed to zero. The guard already returned NA for a base that was missing, NA or 0, but not for one that was infinite – so every other year of that country became finite / Inf, a plain and entirely plausible 0, while its neighbours indexed correctly. The guard now tests !is.finite(), which covers NA and NaN exactly as before.

• An infinite value crashed beta_convergence() with a base R error. The filter kept observations that were !is.na() and > 0, and Inf passes both, so an infinity reached log() and lm() died with NA/NaN/Inf in ‘x’ – unclassed, naming nothing. gini() and theil() already treat an infinity as unusable; the filter now tests is.finite(), dropping it alongside the NA and non-positive values it already dropped.

• A character year broke the verbs that do arithmetic on it, each in a different way. read.csv() yields "2000" readily, and most verbs here tolerate it because they only sort or group. The four that actually compute with a year did not: beta_convergence() died with base R’s non-numeric argument to binary operator, growth_rate(type = "cagr") surfaced a dplyr mutate error quoting an internal expression, and deflate() exposed its internal join as Can’t join x$year with y$year. All three now use the same guard complete_years() already used, so the message names the column and the conversion. growth_rate(type = "yoy") is unaffected – it never does year arithmetic – and neither are lag_by_country(), index_to(), sigma_convergence() or convergence_club().

• locate_country() crashed on an NA coordinate and went quiet on an out-of-range one. A missing longitude or latitude – what a failed geocoding step produces all the time – reached sf::st_as_sf() and came back as base R’s missing values in coordinates not allowed, an unclassed error naming neither argument. Those rows now return NA, which is already the documented way this function says “no country here”, and the rest of the points are located as usual. Separately, a longitude outside [-180, 180] or a latitude outside [-90, 90] returned a silent NA, indistinguishable from open ocean, so a column still in 0-360 degrees – or with the two swapped – looked like a world of water rather than a mistake; it is now a classed error naming the offending pairs and the conversion.

• A gap in year made lag_by_country(), diff_by_country() and growth_rate(type = "yoy") compare rows that are years apart, silently. These read the previous row, not the previous year, so on a panel of 2000, 2002 and 2005 the 2005 growth rate came back as 27.3% – the change since 2002 – in a column the docs call year-on-year and the argument calls "yoy". The arithmetic is unchanged, because a quinquennial panel is a legitimate design and re-keying the lag would move everyone’s results; but the three verbs now warn and name the countries whose years are not consecutive, pointing at complete_years(). type = "cagr" divides by the real year span and so is silent, as before.

• convergence_club() died with a bare simpleError on a repeated country-year. pivot_wider() folds the repeat into a list-column and as.matrix() then failed with base R’s invalid ‘type’ (list) of argument, naming neither the verb nor the offending rows. A repeat is fatal to the log-t test rather than merely inaccurate, so it is now a classed error that lists the country-years and points at check_panel_unique().

• theil() returned NaN for both components when one group’s weights summed to zero. sw == 0 was guarded for the whole vector, but not per group: the group mean sum(w[i] * x[i]) / swg was 0/0, and the NaN poisoned between and within while total came back a perfectly good number – with no warning to say which group did it. A group with no weight has no share of the population, so its contribution to both components is exactly zero, and the decomposition identity total == between + within now holds in that case too.

• interpolate_missing() stripped a column’s class when, and only when, the series had a gap. The fill went through ifelse(), which drops attributes, so a Date column came back as bare numbers – 2020-01-01 returned as 18262. A series with no NA took an early return that preserved the column, so the same column changed type depending on its data. The fill now assigns into a copy of the input, which keeps the class.

• Passing ISO codes to a verb that expects country names gave no usable advice. origin defaults to "country.name" throughout, so country_factsheet("FRA") – the canonical ISO3 code – was refused and the message suggested check_country_match(), which has nothing to say about a code. When every unresolved value is a valid code under some other origin, the message now names it: It resolves under origin = "iso3c" – try that instead. Applies to country_factsheet(), neighbors(), distance_between() and the join verbs. A genuinely unknown name still gets the close-name suggestion and no misleading hint.

• Accented country names written in NFD resolved to NA. The same accent can be encoded as one precomposed code point (NFC) or as a base letter followed by a combining mark (NFD) – identical on screen, different strings to match(). countrycode’s tables are NFC, so "Turkiye", "Sao Tome and Principe", "Aland Islands", "Curacao" and "Reunion" in their accented NFD spellings matched nothing, silently, even in a UTF-8 locale – which is precisely where ?country_overrides promises accented names do work. macOS returns NFD for filenames and several export paths emit it, so this is input a user gets without choosing it. Cote d'Ivoire was unaffected and hid the problem: countrycode’s regex for it keys on an ASCII substring that decomposition leaves alone. Names that resolve to nothing are now retried with their combining marks stripped, which turns an NFD spelling into the ASCII spelling the package already documents as resolving in any locale. Only the unresolved values are retried, so the pass can add a match but never move one: across the 314 names the package and countrycode ship, none changed.

• The README said common_indicators “keeps the 20 you actually use” after this release grew it to 22 with the two price-conversion series. Corrected, and the count is now asserted against nrow(common_indicators) so the prose cannot drift from the data again.

• The global getis_ord(local = FALSE) gave the same answer for a variable and its negation. The general G is a ratio of weighted to total cross-products, so every x_i * x_j term survives negating the whole variable: g(x) and g(-x) came back as bit-for-bit the same double, and the statistic could not tell a coldspot pattern from a hotspot one. Getis & Ord (1992) define it for a variable with a natural origin and no negative values, and gini() – the nearest analogue here, a global index with a positivity domain – already warns and returns NA in this case. The global branch now does the same, naming how many values are negative, instead of handing back a plausible-looking number computed outside its domain. Zero is still allowed, positive columns are numerically unchanged, and the local Gi* branch standardises and is untouched.

• growth_rate(type = "cagr") returned a bare NaN for a negative value, silently. The compound rate is (v_t / v_0)^(1/n) - 1, and the code guarded v_0 > 0 – the base of the ratio – but not v_t. A fractional power of a negative ratio is NaN, so a single negative year in an otherwise positive series put a NaN in the column with nothing said. Every neighbouring measure reports this: theil() drops non-positive values and says how many, gini() warns about negatives, sigma_convergence() warns when no value is positive, beta_convergence() errors, and growth_rate() itself warns when every row comes back NA. The partial case was the one silent spot, and it is the case real data hits – a deficit or a net flow dipping below zero for a single year. Negative values now give NA with a countryatlas_cagr_negative warning naming how many rows are affected. A value of exactly 0 is unaffected and still gives -1, an annualised -100%, which is the right answer for a series that went to nothing; type = "yoy" is a plain ratio change, is defined for negatives, and is untouched.

• index_to() documents what it does when base_year is not in the data. A country whose series does not cover the base year indexes to NA rather than stopping the call, because the rebasing is per country and a partial answer is still a real one – so a base_year no row anywhere carries (including any non-integer value, which no year can equal) gives an all-NA column rather than an error. That was pinned by a test and explained in a code comment, but nothing in ?index_to said it, and the sibling deflate() refuses such a year instead. Both behaviours are unchanged; the contrast is now stated, with an example.

• A units value crashed the numeric guards. sf’s st_area() and st_distance() return units objects, so a threshold computed from geometry and passed to a numeric argument is an easy mistake – and is.numeric() is TRUE for one, so it sailed through the type check. The range comparison then raised the units package’s own “both operands of the expression should be units objects”, naming neither the argument nor this package. check_number() and check_top_n() now ask whether the comparison is possible before making it, and say what to do: “cutoff_km must be a plain number. Got … Drop the unit first, e.g. as.numeric(cutoff_km).” A classed numeric that does compare, and a named scalar, are still accepted as before.

• Three more arguments consumed a function before checking it. A sibling of the validator fix below, with a different mechanism: the value was used before any check ran, so the crash came from the consuming call rather than the message. check_choice() reached as.character(x) on its default-detection line, which is how world_map(projection = mean) died; country_groups() reached setdiff(group, valid), which coerces (and a formula got as far as “duplicated() applies only to vectors”); and complete_years() called anyNA(years) inside the condition of the guard meant to catch bad input, which errors outright on an environment. All three now refuse a function or environment up front and name the class.

• Passing a function to a validated argument crashed the validator itself. check_string(), check_bool(), check_number() and check_top_n() build their message with {.val {x}}, which coerces the value to character – and that fails outright for a closure, with base R’s “cannot coerce type ‘closure’ to vector of type ‘character’”. The error path therefore crashed instead of reporting the bad input, and every caller inherited it: register_country_source(mean, ...), clear_country_cache(source = mean), as_ggsql_source(name = mean) and world_query(source = mean) all gave a bare simpleError from base R rather than naming the argument. length() was no defence – a closure has length 1, so the length branch never fired, and length(globalenv()) counts bindings, so an environment reported “Got 5 values”. All four now name the class instead: “Got <function>”. Ordinary values, including the value itself and the wrong-length counts, are unchanged.

• A numeric argument turned an “unknown name” error into a cli crash. Several messages put the pluralisation marker ahead of the value – "Unknown group{?s}: {.val {bad}}". With no quantity set, cli reaches for the most recent interpolation to find one, and a numeric vector there is read as the quantity itself, which must be length 1. So country_groups(c(1, 2)) died on cli’s own “length(object) == 1 is not TRUE” rather than reporting the bad input, and the same held for check_cols(), world_table(columns = ) and classify_compare(methods = ). Character vectors were fine, which is why it went unnoticed. The five affected messages now set the count with cli::qty(length(x)), the form reporting.R already used. Note qty(x) alone is not enough for a numeric x – it hits the same trap. Singular and plural now read correctly for both types, and character messages are unchanged.

• The spherical-area helper mis-measured rings at the antimeridian. ring_area_km2() picks a country’s largest polygon piece so the label lands on the mainland rather than in an ocean. Longitudes arrive wrapped into [-180, 180), so a raw edge difference jumps by ~360 degrees at the antimeridian, and the two failure modes pointed opposite ways: a ring crossing 180 measured 179 times too large (a 2-degree equatorial square came out at 8.85e6 km2 instead of 49447), while one encircling the pole cancelled to about zero (a cap at latitude -80 measured 7e-11 instead of 3.87e6). Either can hand the “largest piece” choice to the wrong ring – exactly the mislabelling the helper exists to prevent. Edges are now differenced modulo 360, which agrees with the analytic area for an equatorial square, a wrapped square and a polar cap alike. No centroid changes: the bundled polygon data splits pieces at the antimeridian, so all 239 still match country_meta. The point is that the result no longer depends on the data being split that way.

• tissot_map() failed for many ordinary radii. Its circles are built by walking out along azimuths 0..2*pi and wrapping longitude into [-180, 180). az = 0 and az = 2*pi are the same point, but the wrap could send the two ends to -180 and +180 – the same meridian, opposite signs – so st_polygon() saw an unclosed ring and sf reported “polygons not (all) closed”, a bare simpleError naming neither the argument nor the function. Whether it happened depended on the radius, and the default hid it: at radius_km = 500 no centre in the grid produced such a ring, at 1000 six did and at 5000 fifteen. The failures were not monotonic – 500 and 10000 worked while 1000, 2000, 5000, 8000, 12000 and 20000 did not. The ring is now closed by construction rather than relying on the arithmetic to reproduce its first vertex, and every radius from 100 to 40000 km draws. The antimeridian guard is unchanged, so a larger radius still drops the circles that would smear across the map.

• An enormous k in dorling_map() surfaced a GEOS exception. k was bounded below but not above. check_number() already refuses Inf, so k = Inf errored cleanly, but a merely enormous finite value passed and then overflowed the coordinate arithmetic inside GEOS, which reported an illegal-argument exception about encountering non-finite numbers in an orientation test – a bare simpleError from a C++ library, naming neither k nor the function. It is now capped at .Machine$integer.max, the bound the counting arguments elsewhere use, so the refusal comes from the package and names the argument. Realistic values are unaffected: k = 1e6 still draws.

• top_n past integer range leaked a base R error. Inf is the documented “no limit”, and world_table() and country_network() both gate on is.finite(top_n) to detect it – so a finite value beyond integer range passed the gate and then broke on the coercion behind it. as.integer(1e18) is NA, so utils::head(df, NA) surfaced base R’s “invalid ‘n’ - must contain at least one non-missing element, got none”: a bare simpleError naming neither top_n nor the package. Asking for at most 1e18 rows of a 191-row table is the same request as Inf, so it is now normalised to Inf rather than rejected. Representable limits are unchanged, .Machine$integer.max still works, and 0 / negatives / NA / non-numeric are still refused with a countryatlas_error. Found by fuzzing argument values, looking for errors the package had not classified.

• The two hand-curated datasets are pinned to the scripts that build them. country_groups_history and disputed_territories are maintained by hand in data-raw/, so the script is the source of truth and drift can happen in either direction – editing the .rda without the script, or the script without rebuilding. Both are now re-run during the tests and compared to the shipped data, the same guard data-raw/overrides_snapshot.R already had. Both reproduce byte for byte today.

• interpolate_missing() no longer carries a second, unreachable duplicate-column guard. check_panel_unique() grew one of its own earlier in this cycle, and it runs five lines earlier, so the copy inside interpolate_missing() could never fire – its stale comment still claimed the function “did not” reject duplicate names. Line coverage showed the block never executing. The rejection itself is unchanged, still raising countryatlas_duplicate_columns, and is now covered by a test.

• audit_time_coverage()’s three console messages are covered by tests. None of those lines had ever executed, which meant the pluralisation in “{n} row{?s} fall{?s/} outside the country’s existence” had never been checked at both counts – and {?s} keys to the most recently interpolated value, a construct this package has got wrong before. Both readings are correct (“1 row falls”, “2 rows fall”) and are now pinned, along with the no-findings message and the empty-input column contract.

• rate_check() reported “no countries flagged” when it meant “unknown”. When no small-denominator threshold can be computed it warns that flagged is NA throughout – but is.finite(den) & den < thr yields FALSE, not NA, for a non-finite denominator, because R short-circuits FALSE & NA to FALSE. An all-NA denominator therefore produced FALSE for every row, so sum(out$flagged) returned 0: a confident “nothing is flagged”, which is precisely the misreading the warning was written to prevent. An all-zero or all-negative denominator did yield NA, so the cases disagreed with each other as well as with the message. With no threshold, no row can be compared to one, so flagged is now NA for every row and the warning fires exactly when that is true. The normal path is unchanged, including its FALSE for a missing denominator among usable ones, which keeps sum(out$flagged) working as intended.

• getis_ord() computed its spread with the unstable one-pass formula, so a tightly clustered column produced Inf or NaN z-scores. The spread was sqrt(sum(x^2)/n - mean(x)^2), which subtracts two nearly equal large numbers and loses every significant digit once the mean dwarfs the variation. Measured on 8-country vectors: 1e9 + 1:8 drove the spread to exactly 0, so the denominator was 0 and every z_score came back Inf with p_value 0 – reading as “every country is a significant hotspot”; 1e10 + 1:8 gave a spread 90 times too large, so nothing was ever significant; and 1e12 + 10 * 1:8 went negative under the sqrt, returning NaN z-scores and leaking base R’s warning about them. The zero-variance guard caught none of it, because such a column is not constant, only nearly so. Centring before squaring is algebraically the same quantity and unconditionally stable, and it is how local_morans() already computed its second moment. Well-conditioned columns are unaffected to floating-point noise. Gi* z-scores are shift and scale invariant by construction, which is what the old form violated, so that invariant is now asserted rather than any particular number.

• A trailing space on region silently drew an empty map. resolve_region() trimmed in its iso3c branch and nowhere else, so one trailing space gave three different answers: "FRA " resolved, "Europe " fell through to country-name matching and errored, and "EU " was accepted as a three-letter code – nchar("EU ") is 3 and it is already uppercase, so it reached the branch that passes an unrecognised code through untouched and came back as the literal string "EU ". Every public caller of that helper – world_geometry(), world_data(), attach_geometry(), join_world() and country_borders() – therefore subset to nothing and reported nothing, which is the outcome that branch exists to prevent. Every branch now reads the same trimmed value, so all 15 shipped groups, the continents, codes and names agree across padding. The error message still quotes the caller’s value untrimmed.

• Padding was tolerated for a code but not for an overridden spelling, and a non-breaking space defeated the trim entirely. trimws()’s default class is [ \t\r\n], so the iso3c branch’s deliberate trim missed every non-ASCII space – "FRA " resolved and "FRA\u00a0", what a code pasted from a web table or a Word document carries, did not. Separately the override lookup keyed on the raw value while the whitelist was built from the trimmed one, so with custom_match = c(Somaliland = "SOM"), "Somaliland" gave SOM and "Somaliland " gave NA; an ordinary trailing space was enough. Both now use one Unicode-aware trim. A BOM or a zero-width space is a format character rather than whitespace and is still rejected rather than silently accepted.

• top_n went unvalidated in world_table() and country_network(). Inf is the documented “no limit”, which check_number() rejects, so the guard was a bare is.finite() – and everything is.finite() rejects then skipped validation altogether. top_n = "5" and top_n = NA silently returned every row instead of five, and top_n = NULL failed on if with R’s “argument is of length zero”. Both now validate up front and still take Inf.

• world_table(subtitle = ) was dropped unless a title came with it. gt draws the subtitle inside the header block that a title opens, so a lone subtitle had nowhere to go; it now says so instead of vanishing.

• value_by_alpha_map() drew a half-lit map when opacity meant nothing. The alpha scale had no limits, so it rescaled to whatever spread the frame happened to have; an equalize column with nothing usable collapsed to one value and ggplot2 placed it at the midpoint of alpha_range. The result read as “every country equally weighted” – the one impression this verb exists to prevent. Opacity is now absolute (limits = c(0, 1)), so a frame with no usable equalising variable is drawn at the floor and says so, and two maps of different subsets are comparable.

• rate_check() returned an all-NA flagged column in silence. With no positive finite denominator anywhere, the tenth-percentile threshold is NA and every comparison against it is NA too – so sum(out$flagged), the obvious next step, came back NA rather than a count. It now says why.

• deflate() returned an all-NA country without saying why. A country with no usable deflator in base_year has nothing to rebase against, so every one of its values is NA – correct arithmetic, but in the output indistinguishable from a country the source never covered. It now names them and suggests a base year the panel actually spans.

• od_map() dropped a named origin without saying so. A country that appears in the OD table only as a destination has no outflow to draw, so it was filtered out – silently, leaving the caller to notice that they asked for four panels and got three. Named origins now warn (and point at direction = "in"); trimming the origins = <n> top-N list stays quiet, which is what top-N means.

• distance_between() returned a silent NA for a name that is not a country. Two causes produce an NA distance and only one is the documented gap: a country with no bundled centroid is expected – ?distance_between says so, and country_weights() already reports it – while a value that resolves to no country at all is a mistake, usually the wrong origin, and handed back a column of NA with nothing said. The second is now reported, naming the values and the origin in force; the documented gap stays quiet.

• simplify_geometry() and theme_world_map() never validated their first argument. Both check their second one carefully. simplify_geometry() handed a non-spatial object straight to rmapshaper, leaking “no applicable method for ‘ms_simplify’ applied to an object of class NULL” – rmapshaper’s generic rather than the argument – and failed differently again through the sf::st_simplify() fallback, so the message depended on which optional package the caller happened to have. theme_world_map() got base R’s bare “non-numeric argument to binary operator”. Found by calling all 102 exports with degenerate input and classifying which errors were the package’s own.

• Three verbs joined on an unstandardised key and said nothing. Lowercase, mixed-case and padded iso3c values match nothing, so attach_geometry() drew every country as no-data, tile_map() drew every tile grey, and add_indicator() attached a column of pure NA – each of which reads as a coverage problem, or as the provider having no data, rather than as a key problem. An unmatched code stays quiet, because the basemap genuinely holds fewer countries than the snapshot; matching nothing now says so and points at standardize_country(), which normalises case and whitespace. attach_geometry(year = ) already reported its match rate; the ordinary branches did not.

• complete_years() also leaked one on a missing year. It infers the span with seq(min(year), max(year)), so a single NA produced base R’s “‘from’ must be a finite number”, naming neither the column nor the package. The years argument has been checked for NA all along; the column it defaults from had not, and one blank cell in a CSV is enough to hit it.

• complete_years() leaked an internal error on a non-numeric year. Its years argument was checked carefully but the year column was not, so a character one surfaced as dplyr’s “Can’t join x$year with y$year due to incompatible types” – naming dplyr’s internals rather than the column – and a factor got base R’s bare “‘min’ not meaningful for factors”. Both come straight out of a CSV read.

• country_factsheet() printed NA as the country name for any code with no country_meta row. The fallback was first_or_na(row$country) %||% iso, but %||% only replaces NULL and first_or_na() returns NA_character_, so it never fired: country_factsheet("Kosovo") headed its output NA (XKX) while listing four real land neighbours underneath. It now falls back to the name the caller used.

• country_meta had five blank capitals alongside 34 NAs. WDI_data uses "" for an unknown capital and it passed through unchanged, so is.na(capital) was wrong for Gibraltar, Hong Kong, Israel, Macao and the Palestinian Territories, and the factsheet printed capital: with nothing after it. Blank strings are now NA, in the dataset and in the script that builds it.

• Suggests: testthat understated its minimum. The suite has used expect_no_error() (testthat 3.1.5) for some time and now uses expect_no_match() (3.2.0), while DESCRIPTION still asked only for 3.0.0. Bumped to >= 3.2.0.

• rank_countries() and share_of_world() said nothing about a repeated country-year. interpolate_missing() and complete_years() already report that shape, but the two verbs that aggregate across rows did not – and their output is the harder to reconcile. Given a frame with USA-2020 duplicated, rank_countries() returned the same country holding ranks 1 and 3, and share_of_world() gave it shares of 0.1 and 0.7 against a world total that counted it twice. Both now report it, each naming its own consequence rather than the lag family’s.

• Four numeric-column verbs did not check that the column was numeric. country_network() validates its weight with check_numeric_col(); the verbs shaped like it did not. bubble_map() and flow_map() reached ggplot2’s bare Discrete value supplied to a continuous scale, and only at build time – so the call returned happily and the failure surfaced when the plot was printed, naming neither the argument nor the column. spike_map() blamed the join (No rows with a non-negative <col> joined to a centroid) and convergence_club() blamed the panel (Not enough countries with a complete series), when in both cases the column simply was not a number. All four now refuse at the call, naming the column.

• fetch_indicator() trusted a registered source’s key_col claim. When a source declared key_col = "iso3c" its codes were used verbatim, but the declaration is the source’s claim rather than a guarantee – and this is the package’s public extension point. Lowercase codes ("usa") passed through unchanged, a factor stayed a factor, and numeric UN M49 codes (840) sailed through as numbers, each producing rows that silently joined to nothing and read as “the provider has no data”. The key is now standardised whatever the source claims (a no-op on codes that are already right), and values that are not usable as the declared key become NA and are named (countryatlas_bad_key).

• A factor value column was read as its level indices. adapter_reshape() coerced the provider’s value column with as.numeric(), which on a factor returns level indices: a column of factor("10", "20") became 1, 2. check_numeric_col() rejects a factor outright with precisely this advice (as.numeric(as.character(x))), and its comment notes how easily such a column happens – but the adapters take theirs from a third party, so they cannot reject it and must not misread it either.

• Every source path misread the provider’s time column. ... forwards to the client, so the caller chooses its type: eurostat’s time_format = "num" returns a numeric year and "raw" a character one, while the default is a Date. Each adapter assumed one shape. fetch_eurostat() failed on the other two with base R’s opaque invalid 'trim' argument – format() reading "%Y" as its trim argument. fetch_oecd() was worse, failing silently: as.integer() on a Date returned 18262, the day count, as the year, and a quarterly "2020-Q1" became NA. Both now read the year from a date, a number, or a string leading with one, and say so when a value is not a year at all.

  The same bare as.integer() reached six sites, not two: fetch_comtrade(), whose period is YYYYMM for monthly data (so "202001" became the year 202001); adapter_reshape() itself, which fetch_owid() reaches without preprocessing; and fetch_indicator() – the public extension point, where the year is whatever a third-party fetch function returned. All six now share one reader.

  Two more sites outside the adapters took a year from the caller and read it the same wrong way. audit_time_coverage() turned a Date year column into day counts (1990-01-01 -> 7305) and then flagged both USSR rows as post-dissolution – silently wrong output from the one verb whose job is catching that class of mistake. And deflate() reported a Date base_year back as `base_year` 11323 is not in year, a number the caller never supplied; it now accepts a date, a number or a string, and names what it was actually given when it cannot.

• The source adapters returned an empty frame in silence when no entity resolved. adapter_reshape() resolves the provider’s entity column with suppressWarnings() on purpose – every Our World in Data or Eurostat response carries aggregate rows like "World" and "EU27" that are not countries and never resolve – but that also swallowed a provider renaming its entities, or the wrong column being named. The result was zero rows and no explanation, sending the reader to check their own indicator code, while the same function takes care to explain an empty input a few lines earlier. Nothing resolving at all now aborts (countryatlas_no_entities), naming the entities it could not match. An empty result from the countries/years filters is unaffected – those entities did resolve.

• bivariate_map() leaked classInt’s “single unique value” for a column with no variation. classInt needs two distinct values per axis to cut classes from, and a constant column reached it as a bare simpleError from a third-party package, naming neither the column nor the function nor anything to do about it. Both axes are now checked, naming the column and its distinct count.

• tile_map() overstated its coverage, the same way bubble_map() and spike_map() did. The bundled equal-area grid does not cover every code – Hong Kong and Macao have snapshot data and no tile – so counting the input’s coded countries as shown claimed 191 where the map could draw 189. Coverage is now measured against what the grid can actually place, and the countryatlas_no_centroid warning names the countries it cannot.

• per_capita() returned Inf for a zero population, silently. deflate() and to_ppp() were fixed for precisely this – there is a test named “an unusable deflator or PPP factor gives NA, not Inf” whose comment records that Inf “propagated silently into every scale and summary downstream” – but the most used function of the family never got the fix. A zero or missing population now yields NA and says so, matching its siblings. A negative population still passes through unchanged and silently, because that is pinned deliberately elsewhere in the suite – negative values are the caller’s business and the arithmetic stays honest.

• country_weights() built an edgeless graph without comment. country_weights("distance", cutoff_km = 1) returned 239 countries and zero links, as did an all-zero custom matrix. Every statistic then refused to run with Not enough connected countries with data – an error about the data, raised far from the cutoff_km or matrix that actually caused it. Building weights that link nothing now warns (countryatlas_empty_weights), naming the cause specific to the scheme.

• smooth_rates() and to_ppp() returned all-NA columns in silence – the same bug rate_check() already had fixed, kept by its two siblings. With no finite positive denominator every rate is NA and the smoothed column with it; with no finite positive conversion factor every converted value is NA. Correct arithmetic either way, but the output looked like a computation that had run rather than one with nothing to run on. Both now say so, and report a partial loss with a count.

• country_weights("custom") validated only the row and column names, so bad input leaked a bare base-R error from deep downstream instead of being refused where the caller could act. A character matrix reached rowSums() as 'x' must be numeric; an NA entry was accepted, then killed any statistic built on it with subscript out of bounds; an NA endpoint in a long frame surfaced as NAs are not allowed in subscripted assignments; and an NA weight was accepted outright, silently turning every result into NA. All four are now refused at construction, naming the argument and the count. Numeric and logical matrices and long frames are unaffected – and this is the path vignette("honest-maps") recommends for a non-geographic adjacency (trade volume, migration, shared language).

• One dplyr::filter() referenced its columns bare. The bounding-box branch of world_geometry(geometry = "polygon") filtered on long/lat rather than .data$long/.data$lat – the only such site against twenty .data$ uses elsewhere. A bare reference falls back to a variable of that name in the calling scope when the column is absent, so it can filter on the wrong thing where .data$ errors plainly. With it converted, utils::globalVariables() no longer needs to declare long and lat; only animate_world()’s NSE default time = year remains.

• spin_globe()’s example was never executed by anything. It sat in \dontrun{}, which R CMD check skips even under --run-donttest, so the one example of the GIF pipeline was free to rot unnoticed. It needs no network – only bundled data and gifski/magick – so it is now \donttest{}, guarded on those packages and cut to six frames, which runs in about five seconds and is checked from here on.

• as_ggsql_source(format = "duckdb") did not say who owns the connection it returns. duckdb keeps its in-memory database alive until the handle is released, and neither @return nor @param format told the caller to close it – while the "parquet" branch quietly closed the connection it opened. ?as_ggsql_source now spells out all three lifecycles, and a connection the function opened itself is released if the write throws, so a failure cannot leave a handle nobody holds.

• sigma_convergence() returned an empty or blank series in silence. Its positive-value filter is documented – n counts what survived – but two of its outcomes were not. A column with no positive values came back as a 0-row tibble, and a year with a single country got sigma = NA from sd(), so an empty or blank convergence series was indistinguishable from a real one. Both now warn (countryatlas_no_positive, countryatlas_thin_year), the latter naming the years affected. An ordinary panel is unchanged and silent.

• gini() and theil() returned NA without saying why. Both carried a comment stating the package convention – “NA plus a word about why (as for zero weights)” – while the line beneath returned NA in silence for exactly those cases. An all-zero column and all-zero weights were indistinguishable from a missing input. Both now warn (countryatlas_undefined_index) naming which it was. Perfect equality still returns 0, silently, because that is a value and not an undefined one.

• A constant column made the spatial statistics return NaN in silence. morans_i(), gearys_c(), local_morans() and getis_ord() all divide by the cross-sectional variance, so a column with no variation between countries is 0/0 – and getis_ord()’s z-score came back Inf where the numerator was non-zero. For a statistic that is worse than an error, because it reads like a computed result. All four now warn (countryatlas_zero_variance), explaining that these measures compare variation between neighbours and there is none, and return NA rather than NaN. An all-zero column also stops getis_ord() dividing gi_star by its own zero sum. Real data is unaffected.

• map_provenance() gained n_total. n_countries holds the countries actually drawn with a value – the numerator – but the name reads like the map’s country total, which is n_countries + n_missing. Anyone taking it as the denominator understated their own coverage. The denominator is now a field of its own, and the three counts are spelled out in ?map_provenance.

• The coverage caption did not pluralise. footnote = "auto" is built with sprintf() and lands on a published map, where a single-country frame read All 1 countries shown. and an empty one All 0 countries shown. The same noun appears in coverage_map()’s caption and map_provenance()’s print block. All three now agree with the count, an empty frame says No countries to show., and a frame with no coverage to report gets no caption rather than one full of NA.

• spatial_lag() gave every year of a panel the first year’s neighbour average. It is the one spatial verb that returns a column aligned to the caller’s own rows, and it matched on iso3c alone. France’s gdp_per_capita ran 39,683 -> 158,734 -> 277,784 across three years while gdp_per_capita_lag sat at 63,409 for all of them, so value / lag – the Moran-scatterplot axis this column exists for – silently compared 2002 against 2000. A panel now gets a lag computed per year; a cross-section is unchanged.

• The spatial statistics silently computed on an arbitrary year of a panel. align_weights() – shared by morans_i(), gearys_c(), getis_ord(), local_morans(), spatial_lag() and lisa_map() – reduced to one row per country with a bare distinct(iso3c, .keep_all = TRUE), so a panel collapsed to whichever row came first in the frame. Moran’s I on the same data returned 0.47 or 0.29 depending only on row order, and unlike the map verbs these said nothing at all about having chosen. They now go through the shared reduction: the earliest year, deterministically, with the countryatlas_panel warning. A genuine cross-section is unaffected.

• “Only the earliest year of each country is used” was not true. The one-row-per-country reduction behind rate_check(), audit_coverage(), correlate_indicators(), world_table() and every map verb warned that it keeps the earliest year, but the code was distinct(iso3c, .keep_all = TRUE), which keeps whichever row comes first in the frame. That is the earliest year only for a caller who happened to sort by year: shuffle the same panel and rate_check() returned a different numerator for France (20 sorted, 30 shuffled) and the map verbs drew a different year, while the warning went on promising “earliest” either way. The earliest year is now selected explicitly, so these verbs are reproducible for a given frame regardless of row order, and the survivors keep their original relative order.

• standardize_country() silently destroyed columns the caller never asked about. add defaults to c("iso3c", "iso2c", "continent", "region"), so the ordinary call – standardize_country(d, country), to get iso3c – also replaced any continent, region or iso2c already in the frame. A user’s own regional classification vanished without a word, while the eleven other column-adding verbs all report this through warn_overwrite(). It now warns (countryatlas_unasked_overwrite) naming the columns, only when add was left at its default: passing add yourself still means you asked for those columns, as documented. add = "iso3c" adds just the code.

• country_join() silently multiplied rows when standardising collapsed two names onto one code. wdj_to_key() maps distinct inputs to the same code – "France" and "FRANCE ", or "Congo" and "Congo-Kinshasa" – so a y whose names all looked distinct could join one country twice. dplyr says nothing: with unique keys on the other side that is an ordinary one-to-many, not the many-to-many it flags. Two rows became three, France appeared twice with different values, and a downstream sum() double-counted it. Both country_join() and country_join_all() now warn (countryatlas_key_collapse), naming the code and the inputs that collapsed onto it. Only a collapse is reported: duplicates already present in the input under one name, and country-by-year panels, are legitimate one-to-many joins and stay silent. warn = FALSE silences it.

• Subnational maps counted countries instead of regions, and binned on them too. na_coverage(), classification_table(), apply_binned_fill() and imputed_count() de-duplicate before counting – the polygon backend repeats a country’s value down every vertex – but keyed on iso3c. A subnational frame carries iso3c as well as a region code, so every NUTS region of a country collapsed to one row: a 280-region map reported “27 of 27”; because distinct() keeps the first row per key, blank regions inside a country whose first region had data were reported as zero missing, claiming complete coverage of a map with visible holes; and the quantile breaks were computed from 27 values rather than 280, so the colour scale itself was wrong. All four now key on the most specific unit column present (nuts_id, then iso_3166_2, then iso3c, then group).

• country_borders() silently omits five whole countries at its default scale, and country_factsheet() reported the resulting short count as fact. Adjacency comes from Natural Earth, and the default scale = "small" (110m) has no polygon at all for Andorra, Liechtenstein, Monaco, San Marino or the Vatican – countries for which a land border is the entire geography. They contributed no rows, so each reported zero neighbours and France reported 8 instead of 10, under a heading stating the count as fact. ?country_borders now has a Which countries the default leaves out section naming them and pointing at scale = "medium" (which has all five), and country_factsheet() names what its count excludes – only for the countries actually affected. (The Brazil and Suriname borders it reports for France are real, via French Guiana.)

• Passing something that is not a column to a column argument leaked a raw rlang error. quo_arg_name() called rlang::as_name() directly, so world_map(data, 5) failed with Can't convert a double vector to a string. and world_map(data, gdp_per_capita + 1) – a natural thing to try – with Can't convert a call to a string. Neither named the argument, the function, or what was expected, and the leak reached every one of the package’s ~66 unquoted column arguments. The error now names the argument and the offending expression, and suggests the fix: compute the column first for an expression, drop the pronoun for .data$x, or pass a bare name or string for a literal.

• world_table(value = NULL) labelled an unsorted slice with a rank column. With nothing to rank on, the frame keeps whatever order it arrived in and top_n takes an arbitrary slice – but the result was still numbered 1..n under a heading that told the reader these were the top n by something. world_table(world_snapshot$countries, top_n = 5) presented Afghanistan as “rank 1” next to an empty GDP cell. The column is now added only when a value was given, and truncating an unsorted table warns.

• A failed World Bank download was cached to disk as if it had succeeded. WDI::WDI() answers a failed request by warning and returning a zero-row frame; memoise cached that, and the World Bank cache is persistent – so one call made while the network was down poisoned every later session until someone ran clear_wdi_cache(disk = TRUE) by hand. An empty response is now never cached, and says why (countryatlas_no_data).

• register_country_source(cache = ) did nothing. The flag was stored in the registry and reported by country_sources(), but nothing ever read it: the documented per-session memoisation never happened, and cache = FALSE was equally inert. Results are now memoised per source, keyed on the indicator, countries, years and any extra arguments, and cleared by clear_country_cache(source = ). Empty results are not memoised.

• world_map(engine = "tmap") could not draw in its default style. tm_scale_intervals() is tmap’s interval scale, and "cont"/"cat" are not interval styles – they name different constructors. So the default style = "continuous" failed outright with tmap’s Invalid style. Style should be one of "fixed", "sd", "equal", "pretty", ..., and style = "categorical" warned that an interval scale was being applied to non-numeric data. Each style now reaches the constructor tmap has for it.

• flow_map() drew every trans-Pacific flow across the whole map. A great circle Tokyo -> Los Angeles has longitudes running ...178, 179, -179, -178..., and geom_path() under coord_quickmap() joined those two points literally, so the arc streaked back across Africa instead of crossing the ocean. Arcs are now split at the antimeridian and land exactly on the edge.

• flow_map() titled its legends weight whatever the caller’s weight column was called, because the internal arc frame’s column is literally named weight. Both scales now carry the real name – which also merges what were two legends of the same variable into one.

• geom_country_labels(mapping = ) could only name four columns, and cancelled flag = TRUE. The centroid reduction returned iso3c/long/lat/flag and dropped the rest, so aes(colour = continent) – the ordinary reason to pass a mapping – died on “object ‘continent’ not found”. Separately, the caller’s mapping replaced the defaults instead of adding to them, so supplying one removed the label aesthetic entirely and silently ignored flag.

• animate_world(title = ) was discarded. The frame marker was written straight into title, overwriting anything passed through ... to world_map(). The title now stays and the frame label moves to the subtitle.

• style = "binned" ignored n_bins. It was passed to ggplot2 as n.breaks, which is only a hint – scales::extended_breaks() snaps to round numbers – so n_bins = 5, 6 and 7 all drew five bins and 3 drew four. world_map(), globe_map() and value_by_alpha_map() now compute explicit equal-interval boundaries, so n_bins means the same thing under "binned", "quantile" and "jenks" as documented – and in every verb that takes it, not just world_map(). globe_map() and value_by_alpha_map() also record the boundaries in provenance, which only world_map() did before.

• classification_report = TRUE fabricated classes for continuous fills. With no break vector the report fell back to one row per distinct value: a continuous choropleth of GDP per capita produced a 189-row table of n = 1 that looked like a classification and described nothing. A continuous colourbar now reports NULL and warns (countryatlas_no_classes); "binned" gets a real table, because it now has real breaks.

• bubble_map() and spike_map() failed on data already carrying centroid_lon/centroid_lat – including world_geometry("centroids")’s own output. dplyr suffixed both sides of the join to .x/.y and the aes() referring to centroid_lon then found no such column. The incoming columns are dropped in favour of the bundled ones.

• bubble_map() and spike_map() overstated their coverage. Both join the bundled centroid table, which does not cover every code in the codelist. The five countries with snapshot data and no centroid – Hong Kong, Macao, Gibraltar, the British Virgin Islands and Tuvalu – were dropped at draw time (bubble_map() left-joined and let ggplot2 mutter “Removed 5 rows”; spike_map() inner-joined and said nothing), but provenance was computed on the frame as it arrived. A population map that never drew Hong Kong reported “215 of 215” with n_missing = 0. Coverage is now measured against what is actually placed, and a new countryatlas_no_centroid warning names the countries that could not be.

• na_style = "hatched" and disputes = "mark" silently discarded the projection. ggpattern::geom_sf_pattern() and the geom_sf() inside the dispute layer each return list(<layer>, <CoordSf>) – a default coord_sf(crs = NULL) – and ggplot2 replaces the plot’s coordinate system unconditionally when a coord is added. Both layers go on after wdj_coord_sf(), so they threw away the requested CRS and its latitude clip: world_map(projection = "mercator", na_style = "hatched") and the same call with projection = "robinson" drew byte-identical maps. Two honesty features quietly reprojecting the map is the opposite of the point.

• clear_wdi_cache(disk = FALSE) deleted the on-disk cache, and any other file beside it. memoise’s cache_filesystem()$reset() is file.remove(list.files(dir, full.names = TRUE)), so forget() on a disk-backed memo was never an in-memory operation – the call the examples label “forget the in-session memo” wiped the persistent cache and every unrelated file that happened to share the directory. The disk flag gated nothing on the way in; it only decided whether the directory was additionally removed. Dropping the memo reference is what an in-session clear means here, and the next fetch reads the existing entries straight back.

• years meant something different for the World Bank than for every other source. It is documented as a numeric year vector, and the four adapters honour that, but the WDI path used only min()/max() as a request range and never filtered the result – so years = c(2000, 2020) asked for two years and got twenty-one.

• A mistyped classify silently added no classification columns. world_data() and country_data() filtered the argument with intersect(), so classify = "incomes" matched nothing and quietly produced a frame with no income, continent or region column rather than saying the value was not recognised. language went straight to WDI, where a length-2 value surfaced as “the condition has length > 1”. Both are now checked before the network call, so the error does not depend on having a connection.

• A mistyped region drew an empty map in silence. region accepts a continent, a group name, iso3c codes, country names or a bounding box, and anything matching none of those fell through to name-matching, resolved to NA, and produced an empty subset – so region = "Europ" returned a blank map with no explanation, and region = NA reached the internal tests as base R’s “missing value where TRUE/FALSE needed”. Both are now reported. An explicitly-uppercase unknown code still yields an empty subset, which is deliberate and documented in the code.

• add reported its problems under somebody else’s argument. standardize_country(add = ) and locate_country(add = ) name attributes to derive from iso3c, and neither validated the argument: an unknown name surfaced as countrycode’s complaint about its destination argument, a bad one in locate_country() as convert_country()’s to, and add = NA as base R’s bare “missing value where TRUE/FALSE needed”. Both now check it, and the shortcut table they accept is shared with the validator so the two cannot drift apart.

• A malformed custom_match could put a non-code in iso3c. origin was validated and the override table was not, although every value in it lands in the iso3c column – and the iso3c branch whitelists those values as valid by construction, so nothing downstream rejected them either. custom_match = c(Freedonia = 1) therefore produced iso3c == "1", and every join, geometry lookup and group test after it keyed on that. The table must now be the name-to-code map country_overrides() returns.

• wdi_search() searched for whatever it was given. Only field was checked, so a non-string pattern went straight into the regex and matched something plausible: wdi_search(1) returned 10,125 indicators and wdi_search(NA) the entire 29,495-row catalogue, both without complaint, while an empty one leaked base R’s “invalid ‘pattern’ argument”. A cache that was not a WDIcache() object died on “$ operator is invalid for atomic vectors”. Both are now checked.

• spin_globe(file = ) was the one argument its validation block missed, so a non-string path leaked base R’s “invalid ‘path’ argument” from inside the writer – the block’s own comment says a bad argument is the caller’s bug and the message should not depend on which optional packages are installed.

• audit_coverage()’s group breakdown measured one indicator and did not say which. by_group$na_rate is computed on the first indicator – with the default that is whichever numeric column comes first – while the printed heading reads “Coverage by group”, so a region’s GDP coverage of 0.25 read as its overall coverage even though population and life expectancy were complete. The table now carries an indicator column naming what was measured.

• compare_sources() could report two providers as agreeing perfectly when it had never compared them. An unnamed indicator of more than one code was silently truncated to its first element and broadcast to every source, so the verb compared a code against itself and found no disagreement. Its own error message elsewhere already told callers to “pass a single unnamed code”; nothing enforced it. Now rejected, with a pointer to the named form.

• deflate() overwrote an existing column in silence. to_ppp() and smooth_rates() are the same shape – take a panel, write one derived column back into it – and both announce a clash before destroying the caller’s column. deflate() did not, so a frame that already carried gdp_real lost it without a word.

• A repeated country-year made interpolate_missing() rewrite observed values. stats::approx() collapses tied x-values to their mean, so two rows sharing a year were both overwritten with the average – 20 and 999 became 509.5 apiece – and the _imputed flag reported FALSE for them, since it compares “was NA” against “is not NA” and neither ever was. The safeguard the code documents, that a filler changing an observed value shows up as a flagged cell, cannot catch this case. The malformed input is now reported, and approx()’s own tie notice – which reached the caller as an uninformative “There was 1 warning in dplyr::mutate()” – no longer leaks.

• A repeated country-year corrupted every verb that reads across rows. lag_by_country(), diff_by_country(), growth_rate(), index_to(), complete_years(), deflate(), to_ppp(), convergence_club() and beta_convergence() all read neighbouring rows, so a duplicated country-year silently shifted them: with France’s 2019 present twice as 20 and 999, 2020 was lagged against 999, giving a difference of 979 and 4895% growth. The malformed input is now reported, and it names the offending key.

• The one-row-per-country verbs collapsed a panel silently. world_table(), rate_check(), audit_coverage(), gridded_cartogram(), bubble_map(), spike_map(), globe_map(), interactive_map(), tile_map() and correlate_indicators() reduce to one row per country – which is for repeated geometry rows, not for time. Handed a panel they kept whichever row sorted first and presented that year as the answer: France’s 2018 value of 10, out of 10, 20 and 30. They now say so. tile_map() was worse than silent – its grid holds exactly one cell per country, so a panel fanned 239 cells out into 659 overlapping ones.

• A panel handed to world_map() was drawn silently. attach_geometry() joins a panel deliberately – facet_map() and animate_world() are built on it – so a multi-year frame reaching a single static map drew each country once per year and let whichever row came last win, with a caption that still counted each country once, so nothing looked wrong. world_map() now says so and points at the two verbs meant for a panel. Faceting by year and animate_world() resolve the panel and stay quiet; faceting by anything else does not – each panel still stacks every year – so there it still warns.

• Three counts lacked the upper bound their coercion needs. world_query(n_bins = ), od_map(origins = ) and convergence_club(min_size = ) validated only a lower bound, then coerced with as.integer() – which past 2^31-1 returns NA with R’s bare “NAs introduced by coercion to integer range”. The query came out reading BIN fill INTO NA, od_map() reached seq_len(NA), and every size comparison in convergence_club() became NA. compute_breaks() has carried this bound since 2.0.0; these three now do too.

• subnational_map() dropped unmatched regions in silence. Its join keeps the geometry and discards data rows that match none of it, so codes from a different NUTS vintage simply vanished – and only a total join failure was reported, even though that error names the vintage problem exactly. Partial mismatches are now named too.

• check_country_match() suggested a country for a blank cell. The guard skipped NA and "" but not " ", and Jaro-Winkler finds spurious similarity between a two-space string and a name containing spaces – so a whitespace-only cell, the commonest thing a CSV import produces, came back confidently suggesting “Congo - Kinshasa” at a distance of 0.29, well inside the threshold.

• audit_coverage()’s n_missing and na_rate are no longer named vectors. vapply() names its result after indicator, so a$na_rates$na_rate handed back c(gdp = 0.1) instead of 0.1 – the same wart country_network() already calls unname() on.

• Three verbs folded every uncoded country into a single row. The de-duplication that stops a polygon frame counting a country once per vertex ran distinct() on iso3c, and distinct() treats NA as a value – so countries the codelist could not resolve collapsed together. audit_coverage(), the verb the package points at for “which countries are missing”, named one unmatched country out of four and divided every na_rate by a short n; rate_check() and world_table() quietly returned three rows for a five-row input. Coded rows are still collapsed; uncoded ones now de-duplicate on whatever else identifies them.

• country_join() and country_join_all() never reported a failed reconciliation. Reconciling both sides to a common key before joining is the entire premise of these two verbs, yet a name that resolved to no country was dropped in silence – join_world() has warned about this all along, and wdj_to_key() only speaks up when a name resolves to iso3c but has no COW/GW code, which on the default key = "iso3c" is never. Both now name the unresolved values, per side and per table, and both take warn = FALSE to opt out.

• aggregate_regions(fun = "weighted_mean") returned NaN for a zero total weight. The unweighted branch turns an empty group into NA precisely because 0, NaN and +/-Inf all read as real figures for a region there is no data for; the weighted branch checked only for missing values, so weights that were all zero – or that cancelled – divided by zero and leaked the NaN the documentation promises never to produce.

• add_indicator() suffixed the column it said it was overwriting. When the fetched indicator shared a name with a column already in data, the warning read “Overwriting … rename them first to keep the original values” while the left_join underneath produced val.x and val.y – so the caller got neither the column they asked for nor the one they had, advised by a message describing a mechanism that was not running. It now overwrites, as every other warn_overwrite() caller does.

• An uncoded geometry row was counted as a country. The bundled sf basemap carries one row with no iso3c; coverage counted it, so it sat in the denominator and in n_missing while missing_iso3c – which sorts, and so drops NA – listed one fewer. footnote = "auto" read “159 of 176 countries shown; 17 missing” where provenance could name only 16 of them. Coverage now counts coded countries only, and the two agree.

• gridded_cartogram() did the same, and mentioned only half of it. It drops countries with no positive value and countries with no bundled centroid; only the second was ever reported, and provenance counted the survivors, so a grid covering 94 of 215 countries reported “94 of 94”. It now names both drops and keeps the whole input as the denominator.

• A cartogram reported the world it kept, not the one it started from. cartogram_map() and dorling_map() drop every country without a positive weight – necessarily, since a cartogram’s area is the weight – but they did it silently, and provenance was then computed on the survivors, so n_total shrank to match. A map showing 69 of 175 countries reported “69 of 73”. Coverage is now measured against the frame as it arrived, and the countries that could not be sized are named.

• bivariate_map() overstated its coverage the same way. It classifies two variables jointly, so a country holding only one is drawn as no-data – but provenance counted the x column alone. On a half-covered second variable it reported 159 countries shown where 69 were coloured. It now counts both, and names what it dropped.

• A VSUP map overstated its own coverage. world_map(uncertainty = ) colours a country only where it has both a value and an uncertainty, but coverage was counted from missing fill values alone. On a frame whose uncertainty column was sparser than its value column – an uncertainty join that half missed, say – footnote = "auto" reported 159 of 175 countries shown when 69 were actually drawn, and map_provenance() agreed with it. The footnote that exists to stop a map overstating what it covers was doing exactly that. Coverage now counts both columns, and the countries dropped for a missing uncertainty are named.

• interpolate_missing() flagged the wrong rows as imputed. The “was missing” vectors were captured from the input and compared against the output, but the pipeline between them arranges by iso3c and year – so the two lined up only when the caller happened to pass an already-sorted frame. On anything else the _imputed flags landed on entirely different rows, marking observed values as imputed and imputed ones as observed. Since world_map() reads that column to write its caption, an unsorted panel produced a map that misreported exactly which countries had been filled in. The flags now travel with their rows.

• flow_matrix(fill = ) corrupted the flows it was meant to leave alone. fill is documented as the value for pairs with no flow, but the matrix was initialised to it and the observed weights accumulated on top, so fill = -1 turned a flow of 10 into 9 – and symmetric = TRUE then added the fill to itself. Flows are now accumulated into zero and fill is applied only to pairs no flow ever reached. The default fill = 0 is unaffected.

• compare_sources() reported disagreement between the wrong sources. Every column of the pairwise summary is computed for its own pair except n_disagree, which read the row-wise spread across all sources: comparing three providers, a pair that agreed exactly was still counted as disagreeing wherever some third provider was the outlier. A source that reports one value for every country also leaked cor()’s “standard deviation is zero” warning; the correlation is NA either way, so the case is now decided explicitly.

• fetch_comtrade() fetched only the first indicator. It read indicator[[1]] and ignored the rest, so a two-commodity call quietly returned one column – against the documented “one column per indicator” that its three sibling adapters honour. It now makes one request per commodity and joins them.

• The source adapters validate indicator before doing anything else. Called directly with an empty vector, fetch_owid(), fetch_eurostat() and fetch_oecd() returned a silent NULL, and fetch_comtrade() leaked comtradr’s “subscript out of bounds”. All four now give the same message fetch_indicator() always gave, and give it whether or not the provider’s client package is installed.

• A provider that changes shape is named as such. adapter_reshape() checked the entity column but not the value column, although fetch_eurostat() hard-codes "values" and fetch_oecd() guesses between two spellings; a missing one became as.numeric(NULL) and surfaced as a recycling error several frames away. fetch_comtrade()’s year column had the same gap.

• country_sources() no longer writes to your home directory. It reported which of the five backing packages were installed by calling requireNamespace() on each, which loads them – and comtradr creates ~/.cache/R/comtradr in its .onLoad. Merely asking which sources were available therefore created a directory in the user’s file space, which CRAN policy forbids and R CMD check reports as a new file in another directory. Availability is now tested with system.file(), which does not load.

Five defects found by auditing 2.0.1; three produced a wrong or unusable picture rather than an error.

• cartogram_map(), dorling_map() and tile_map() rejected a categorical fill. All three hard-wired scale_fill_viridis_c(), so a discrete fill column – which their fill argument documents no restriction on – was accepted at the call and then died at print time with ggplot2’s bare “Discrete value supplied to a continuous scale”. The scale is now chosen from the column’s type, as world_map() has done since 2.0.0.
• geom_country_labels(data = ...) was unusable: data was hard-wired in the layer call while ... was documented as passing to that same call, so the ordinary idiom for labelling a subset failed on R’s formal argument "data" matched by multiple actual arguments.
• world_map(projection = "mercator") produced an unusable map. Natural Earth’s Antarctica reaches -90 degrees, where Mercator’s y goes to infinity; PROJ clamps rather than erroring, so the panel came out three times taller than the world is wide, with the inhabited world a sliver above one grey rectangle of smeared Antarctica. Mercator is now clipped to +/-85.05113 degrees, as Web Mercator has been since it was defined.
• locate_country(points = ) leaked an sf internal (“no applicable method for ‘st_transform’”) for anything that was not already sf. A plain lon/lat frame is the commonest thing to pass, so it now gets a named error saying how to convert it.
• morans_i() could not reveal what it had dropped – see below.
• A wrong coding scheme died inside countrycode. origin is user-facing on seventeen exported functions – neighbors(), country_join(), standardize_country(), flow_map(), country_timeline(), in_group() and the rest – and was validated only as “a string”. Anything else reached countrycode::countrycode() and failed there, so origin = "country" raised a forty-item list of accepted values attributed to an argument the caller had never passed. Invalid schemes now raise a countryatlas error naming the one they probably meant: "country" and "name" both suggest "country.name", "iso3" and "ISO3C" suggest "iso3c".
• convert_country()’s own from and to were the two that guard missed. Every scheme other than "country.name" and "iso3c" skips the iso3c hop, so from reaches countrycode() directly and was still blamed on origin; and to was never checked at all, so a typo produced “the destination argument must be … one of the column names in the conversion directory” – an argument, and a directory, the caller never mentioned. Both now name themselves and suggest the nearest real value, so to = "contnent" offers "continent". The shortcut destinations (flag, currency, tld, calling_code, name_fr-style localised names) are unaffected.
• register_country_source(key_col = ) was used as a coding scheme. key_col names the column fetch returns, and is documented and indexed as such – but it was also handed to countrycode() as origin, so the only registrations that worked were those whose column happened to be named after a scheme. Every bundled source uses the "iso3c" default, which short-circuits before countrycode() is reached, so nothing in the package exercised it and the first source registered with key_col = "country" died. How to read the column is now its own argument, key_type (default "iso3c"), validated at registration and reported by country_sources().
• “5 countries … has no cowc code.” The COW / Gleditsch-Ward key warning interpolates which table it refers to between the count and the verb, and cli keys an agreement marker to the most recent interpolated value – a length-1 string here – so the verb stayed singular however many countries were lost. cli::qty() now re-keys it to the count. (Only literal markup such as {.field iso3c} is safe to sit between a count and its agreement.)
• world_map(engine = "tmap") ignored most of its arguments. The engine received all of world_map()’s arguments and used ten of them, dropping the rest in silence: projection and recenter (so the documented default, Equal Earth, was dropped just as quietly as an explicit request), plus na_style, footnote, classification_report, uncertainty and disputes. projection and recenter are now honoured – wdj_crs() resolves both and tm_shape() takes the result, which also means a bad projection name is finally rejected instead of ignored. The remainder are genuinely ggplot2-specific, so they are now named (countryatlas_engine_ignored) rather than quietly not happening.
• globe_map(interactive = TRUE) accepted styling it could not use, and validated nothing. The hand-off to MapLibre returned before arg_match() and check_label_args() ran, so style = "nonsense" and a length-3 title were taken without a murmur, and style, palette, n_bins, borders, title, legend, na_label and backend were all dropped. The arguments are now validated before the hand-off, and the ones MapLibre cannot carry are named.
• world_map(engine = "tmap") ignored na_label. The argument was passed into the tmap backend and never used, so a caller who named the missing-data key got tmap’s own default instead, with nothing to indicate the label had been dropped – while the ggplot2 engine honoured it. Every tmap scale takes label.na, so it is now passed through, and both engines share one normalisation of what the argument means (first element; a length-1 NA or NULL leaves the engine’s own formatter alone).
• projection was accepted and silently ignored by the polygon backend. It is documented for the sf backend, and recenter already warned when it could not be honoured, but projection did not: the polygon backend returns unprojected longitude/latitude, so projection = "mollweide" looked honoured and changed nothing. It now warns (countryatlas_projection_ignored). world_data(geometry = "none") fetches no geometry at all and so warns for scale, projection and recenter alike, where before it took all three and dropped them.
• world_data(region = ..., geometry = "none") returned every country in the world. region is documented as a plain “Optional subset”, but it was only ever applied inside attach_geometry() – which the "none" branch skips – so the argument was dropped in silence and a request for one continent came back with all of them. It now subsets the table directly. A bounding-box region has nothing to clip against without geometry, so that combination is refused rather than guessed at.
• A scale of more than one value produced an error from base R. The sf backend builds its cache key with paste0("scale_", scale) before validating, so a length-2 value vectorised into a two-element key and [[ on an environment failed with “wrong arguments for subsetting an environment” – naming neither the argument nor the package. A typo, a number, NA and NULL all reached the real check and reported properly; only the multi-value case escaped it. scale is now checked before it becomes a key. Affects country_borders(), neighbors(), morans_i(), world_geometry(geometry = "sf") and everything else routed through get_world_sf().
• scale was accepted and silently ignored by the polygon backend. It selects a Natural Earth resolution, which only the sf backend fetches; the bundled polygons come at one resolution. world_geometry(scale = "large") and attach_geometry(scale = "large") therefore returned small polygons as though the request had been honoured, and scale = 2 was not even rejected. Both now warn (countryatlas_scale_ignored) and point at geometry = "sf", matching the existing notice for recenter.
• cartogram_diagnostics() leaked the geometry engine’s error. One invalid ring makes s2 refuse st_area() with “Loop 0 is not valid: Edge 0 crosses edge 2”, which names neither the country nor the package. It now reports which geometries are invalid and points at sf::st_make_valid(). Unlike country_borders(), it cannot simply fall back to the planar GEOS engine: an area is the number this function reports, so switching engines would quietly change the answer.
• complete_years() gave every year it invented an empty geometry. The function exists so “animations do not flicker on missing years”, but on an sf panel it produced exactly that: tidyr::complete() fills an invented row’s geometry with an empty geometry rather than NA, so the tidyr::fill() that carries static columns forward saw nothing missing and skipped it, and the completed years rendered blank. Each country’s own shape is now carried across its invented rows.
• complete_years() also dropped the sf class. tidyr::complete() returns a plain tibble while leaving the geometry column intact, so the result still held a live sfc column that geom_sf() and st_bbox() would not accept until the caller re-ran st_as_sf(). Fifteen other frame-returning verbs preserve sf; this one now does too.
• interactive_map() discarded ... for two of its five engines, one of them against its own documentation. @param ... promised the dots reach world_map() for the "plotly" and "ggiraph" engines – but the "ggiraph" branch assembles its own ggplot instead of calling world_map(), as a comment in it says, so interactive_map(engine = "ggiraph", style = "quantile") returned a default continuous map with no hint that the classification had been dropped. The "leaflet" engine builds its own map too and was not documented at all, and "mapgl" forwards to mapgl::maplibre(), which was also undocumented. Both self-assembling engines now name what they were given, the @param says where the dots actually go for each of the five, and "plotly" still forwards them (so an unknown argument is still an error there).
• world_table(engine = "tibble") dropped title and subtitle in silence. A tibble has no header, so neither can be drawn – but the gt path already refuses to drop a subtitle quietly (“subtitle needs a title”), and the same objection applies here: the caller got back a table they believed was titled. Both are now named, and when gt is simply not installed the notice points at installing it rather than at an engine the session cannot reach.
• index_to() with no base_year gave base R’s error. Omitting the argument reached check_number() and produced ‘argument “base_year” is missing, with no default’. deflate(), the sibling with the same argument, has said “base_year is required.” since 2.0.0.
• disputed_territories$administered_by and $claimed_by mixed ISO codes with placeholders, and said nothing about it. The columns read as iso3c and mostly are, but six parties are entities ISO assigns no code to and are written as ABK, CYP-N, OST, PMR, SAH and SOL – so treating the column as ISO-keyed produced silent NAs in a package whose premise is that ISO codes are the join key. The iso3c column had always documented its own NA convention; these two now document theirs – including that five of the six administer the like-named territory (which has no ISO code either) while SAH is a claimant only, of a territory ISO does code (ESH) – and a test pins the exact set so it cannot drift or absorb a typo.
• Three more character column arguments died on base R’s “object not found”. aggregate_regions(by = region), world_table(columns = gdp) and country_codes(codes = iso3c) all take column names as strings while sitting next to an argument that takes a bare column, which makes the slip an easy one – and it failed while the argument was being evaluated, before any validation could run, so the message named neither the argument nor the package. complete_years(), interpolate_missing() and audit_coverage() already caught it; these three now do too. (rank_countries(within = ) deliberately accepts either form and is unchanged.)
• A repeated column name was ambiguous everywhere but one verb. What read.csv(check.names = FALSE) gives you for a sheet with two gdp columns makes every by-name reference ambiguous. interpolate_missing() refused it; nothing else did. Ten verbs leaked the tibble error “Column name gdp must not be duplicated. Use .name_repair to specify repair” – a message about internals in tibble, not about the caller’s data – and two, per_capita() and to_ppp(), silently succeeded: they computed from whichever column [[ reached first and dropped the other without a word. The guard now lives in the validators every verb already calls (check_cols(), check_panel_cols(), check_panel_unique()), and in the two functions that reach as_tibble() before validating anything (countryatlas_duplicate_columns).
• rank_countries() returned a data.frame where its siblings return a tibble. It ended in dplyr::ungroup(), which strips a grouping but does not normalise a class – and because it deliberately ungroups its input first, mutate() left a plain frame plain. Four sibling verbs ended the same way and were safe only because group_by() had already made their intermediate a tibble; all five now state the contract explicitly.
• Four verbs leaked an incoming grouping. to_ppp(), smooth_rates(), interpolate_missing() and spatial_lag() ended in a bare data, so they returned whatever class arrived: a grouped frame stayed grouped and the caller’s next mutate() silently computed per group, and a plain data.frame never became a tibble. Their sibling verbs all normalise on the way out. The leak also hid on early-return paths of otherwise-fixed verbs – smooth_rates(method = "none"), interpolate_missing(method = "none") and both branches of spatial_lag() – so a mode of the same function returned a different class from its default.
• join_world(region = ..., geometry = "none") returned every row, for exactly the reason world_data() did: the "none" branch returned before any of the geometry arguments were applied, though region is documented as a plain region subset. It now subsets, refuses a bounding box that has nothing to clip against, and warns for scale, projection and recenter.
• standardize_subnational() never consulted the regions crosswalk, and did not say so. It looks for a geo_name, name or region_name column in regions::nuts_lau_2019 or regions::all_valid_nuts_codes. As of regions 0.1.8 the first exposes lau_name_national and lau_name_latin, and the second has no name column at all – so neither matched, the lookup was skipped, and the caller was told only that their regions “did not resolve”, with a hint about European coverage that had never been consulted. It now reports once per session that the installed regions offers no crosswalk it can use, so the real reason is visible. The behaviour is otherwise unchanged, deliberately: the regions datasets that do pair names with codes (nuts_changes, google_nuts_matchtable) key NUTS codes such as DE2 rather than ISO 3166-2 codes such as DE-BY, and filling an iso_3166_2 column from them would break this function’s promise that an unresolved region gets NA, never a guess.
• Four messages read wrongly when they described exactly one thing. subnational_map() said “1 value … match no geometry and are dropped”, check_dispute_coverage() said “1 have no ISO code” and “1 … territories appear”, and gridded_cartogram() said “there are 1 countries to place”. Each now agrees with its own count. cli keys a {?...} agreement to the most recent interpolated number, so the count and every agreement have to sit together – an interpolation in between silently re-keys them, which is why the totals in two of these messages moved to the end of the sentence.
• An unreachable GISCO reported itself as an sf method error. giscoR::gisco_get_nuts() answers a failed download with NULL rather than an error – the same shape as owidR’s blank result, which fetch_owid() names explicitly and for exactly this reason. nuts_geometry() had no such guard, so the NULL reached sf as “no applicable method for st_as_sf applied to an object of class NULL”, which says nothing about GISCO being unreachable and sends the reader to check their arguments. An empty response now says so, naming the level and vintage. A response carrying no NUTS_ID column is diagnosed too: previously the derived iso3c came back zero-length and base R failed with “replacement has 0 rows, data has 2”.
• A changed World Bank response shape was reported as a failed download. countrycode() is handed raw$iso2c directly to derive iso3c, so a response carrying neither key raised its own “sourcevar must be a character or numeric vector” – which the fetch wrapper then relabelled “Could not fetch indicator … from the World Bank API”. That blames the network for a change in the provider’s response, and attaches advice about an argument the caller never passed. The response shape is now diagnosed by name, listing the columns that did arrive, and the wrapper passes that diagnosis through instead of overwriting it.
• A non-numeric provider response became a column of NA in silence. as.numeric() turns text that is not a number into NA without complaint, so a provider answering with "n/a" or ".." – or renaming a column so the value column now holds a label – handed back pure NA. That reads as “the provider has no data for these countries”, which is a very different claim from “the response was not numeric”. fetch_eurostat() and fetch_oecd() were exposed because they name the value column outright, where fetch_owid() auto-detects it and so refuses a non-numeric column up front. The count is now reported, with the offending values shown. A value the provider itself reported as missing is already NA and is not counted.
• The provider adapters discarded duplicate rows in silence. fetch_owid(), fetch_eurostat(), fetch_oecd() and fetch_comtrade() all end by keeping one row per country-year, which is the contract the downstream joins rely on. But they kept whichever row came first and said nothing, so a provider answering with two different values for one country-year handed back an arbitrary one – order-dependently, and invisibly. The count is now reported, as it is everywhere else this package drops rows. A genuine panel is untouched, and a clean response stays silent. country_data()’s own collapse is deliberately left quiet: there the cause is two iso2c codes mapping to one iso3c, which is structural to the World Bank country list and would fire on essentially every call.
• A source whose keys collapsed repeated the caller’s rows in silence. Standardisation merges keys as well as failing on them – “United States” and “USA” both reach USA – and add_indicator()’s join then matched twice, so a two-row frame came back with three rows and one country holding two different values. dplyr only warns on many-to-many relationships, not one-to-many, so nothing was said. join_world() has warned about this since it gained that check; the source adapters now do too. The standardised key is also checked directly, because the collapse check by construction only sees a code reached from more than one raw value – a source that simply returns USA twice collapses nothing and so said nothing, while the join still turned two rows into three. A genuine panel is untouched: the check keys on iso3c and year together wherever a year column exists. The documented fetch contract now states the uniqueness requirement, which it never did – it specified the columns and said missing data should arrive as NA rather than a missing row, but never that the key must not repeat.
• A cross-section joined to a multi-year fetch lost the year that distinguished the rows. Dropping the fetch’s year column is right for a single-year fetch, where it broadcasts the one value across the frame. Done unconditionally, a two-row cross-section joined to a three-year source came back as six rows: the same gdp three times, against values whose year had just been deleted, so nothing recorded which year any of them belonged to. The column is now kept when it is doing work, and add_indicator() says why the frame grew.
• add_indicator() handed back the source’s own key column as though it were requested data, because the columns to add excluded only "iso3c". A source keyed on anything else left a stray column – country next to iso3c – in the result.
• Re-registering a source kept serving the old one’s answers. The session cache keys on the indicator, countries and years but not on fetch, so correcting a broken adapter and registering it again returned the broken result – which is what writing an adapter looks like. Registering a name now drops that source’s cached answers.
• An optional column argument was not validated like a required one. quo_arg_name() gives every unquoted column argument its “must name a column, not gdp + 1” error, with the advice to compute the column with mutate() first – and for the required arguments it did. Twelve optional ones called rlang::as_name() directly, so an expression reached the caller as rlang’s own “Can’t convert a call to a string”, naming neither the argument nor what it wanted. In several functions the required argument was checked and the optional one beside it was not: bubble_map() validated size but not color, flow_map() validated from and to but not weight, cartogram_map() validated weight but not fill. Now covered: per_capita(pop), aggregate_regions(weight), rank_countries(within), flow_matrix(weight), flow_map(weight), bubble_map(color), cartogram_map(fill), gridded_cartogram(fill), cartogram_diagnostics(weight), interactive_map(tooltip), animate_world(time) and join_world(country_col). within’s deliberate character-vector form is unaffected.
• Three optional columns were never checked for existence either. per_capita(pop = ) left the population vector NULL and failed with base R’s “replacement has 0 rows, data has 4”; aggregate_regions(weight = ) and rank_countries(within = ) failed from inside dplyr. All three now say which column is missing, as their required arguments already did.
• A data frame was accepted where a country vector belongs. as.character() on a data frame deparses each column into a string, so neighbors(my_df) came back with the two “countries” c("USA", "FRA") and c(1, 2) – silently, because those are just strings that match nothing and every row then reads as “country not found”. Nearly every other verb here takes data as its first argument, so handing a frame to one that takes a vector is the natural mistake, and eight of them took it: neighbors(), convert_country(), country_timeline(), dissolve_country(), in_group(), distance_between(), check_country_match() and repair_country_names(). gini() and theil() already refused a data frame; all of them now do. A data frame is only the common case – as.character() deparses any list element that is not a single value, so list(c("FRA", "DEU")) collapsed to the one string c("FRA", "DEU") and returned a single NA where two codes were asked for. A flat list of scalars, a matrix and a factor all coerce correctly and are left alone.
• aggregate_regions() crashed on the package’s own sf output. An sf frame is the one geometry shape this verb can aggregate correctly – it carries a single row per country, so the totals are right – but dplyr’s sf-aware summarise() unions the geometries per group, and the bundled Natural Earth polygons include two invalid ones (Sudan and Mozambique). So join_world(geometry = "sf") |> aggregate_regions(population) died with the raw GEOS error TopologyException: side location conflict, while the same call on a plain frame worked. It also drew the “aggregating counts each country once per geometry row” warning, which is true of a polygon frame and false of an sf one. The geometry is now dropped up front – the documented return was always “a tibble of by plus the aggregated value” – giving results identical to dropping it by hand, and the warning is confined to the polygon case where the hazard is real.
• Four verbs reorder your rows and did not say so. growth_rate(), lag_by_country(), diff_by_country() and interpolate_missing() read each country’s series in time order, so they return rows sorted by iso3c then year – but documented only “data with a column added”. A caller holding a vector alongside data and relying on position would have had it silently stop lining up. The sorting is required and unchanged; the documentation now states it.
• ?interpolate_missing described the wrong mechanism. It said the "countryatlas_imputed" attribute is what “the map verbs read”. They do not: provenance counts the *_imputed columns by name, which is why the provenance chain survives verbs that drop attributes, as dplyr does. The page now points at the columns and says the attribute is a convenience nothing in the package consumes.
• An sf frame stopped being sf halfway through a pipeline. per_capita(), share_of_world() and standardize_country() all document their result as “data with the requested columns added”, and rank_countries() honours that – but these three ended with tibble::as_tibble(), which strips the sf class. The geometry column survived, so nothing looked wrong until the next verb: join_world(geometry = "sf") |> share_of_world(population) |> world_map() died on “data has no map geometry”, while the same pipeline through rank_countries() worked – which is what gave it away. sf now survives all three. Every other class is still normalised to a tibble and grouping is still dropped, both of which as_tibble() had been doing at the same time and only one of which was wrong.
• Three map verbs broke on a frame that already had the columns they join in. Each attaches coordinates under a fixed name, and when the caller’s frame already carried that name dplyr suffixed both sides to .x/.y, leaving the code reading the unsuffixed name to fail on a column that had been renamed out from under it – always as an internal vctrs or ggplot2 error that says nothing about countries:
  • gridded_cartogram() on a frame carrying centroid_lon/centroid_lat – which country_meta supplies, so joining it for capitals or area first is ordinary – failed with “Can’t subset rows with is.na(df$centroid_lon) | ...”.
  • tile_map() on a frame with a row or col column, names common enough to collide by accident, failed with “Problem while computing aesthetics”.
  • flow_map() on a frame with x0, y0, x1 or y1 – exactly what a caller who geocoded their own endpoints has – failed with “Can’t subset columns that don’t exist”.
  bubble_map() and spike_map() already dropped the caller’s centroid columns before the same join; all three now drop the colliding copy, and the output is identical either way. Where the clashing column is one the verb actually reads – a fill that is also a tile coordinate, a weight that is also an arc endpoint – it is refused by name rather than silently dropped.
• deflate() and to_ppp() did not guard the join collision per_capita() does. All three fetch a series into a .wdj_* column and join it on iso3c/year. A caller whose own frame already had that column made dplyr suffix both sides to .x/.y, leaving the fetched column unreachable and the arithmetic operating on nothing. per_capita() has dropped the colliding column since it hit this; its two siblings now do the same.
• audit_coverage() reported perfect coverage of nothing. tibble::as_tibble() turns a bare vector into a one-column tibble called value, so a character vector of country codes – the mirror of the mistake above – came back as a coverage object whose three tables were all empty. That reads as “no missing data” when nothing had been examined at all. standardize_country() already refused a non-frame; this now does too.
• world_table() returned an empty table without saying why. Rows whose value is NA cannot be ranked, so they are dropped – but when that removed every row, a frame that went in with countries in it came back with none, and nothing distinguished “the column is empty” from “there was nothing to report”. It now says so, while a partial drop (the normal case) and a 0-row input stay silent.
• recenter was silently dropped on the polygon backend. The polygon backend returns lon/lat vertices, and recentring them means re-splitting every ring at the new antimeridian – which is what sf::st_break_antimeridian() does on the sf backend. recenter was simply ignored on the polygon side, so join_world(recenter = 150), attach_geometry(recenter = 150) and world_geometry(recenter = 150) returned byte-identical coordinates to recenter = NULL and drew an Atlantic-centred map for someone who had asked for a Pacific-centred one. They now say so and point at geometry = "sf", in the same shape as the existing bounding-box warning on that backend. ?world_geometry was also the one help page that did not already describe recenter as an sf option.
• n_perm = 0 returned undocumented NA p-values. morans_i() documents the escape hatch (“use 0 to skip the test”), and all three statistics validate n_perm with a lower bound of zero – but local_morans() and gearys_c() described the argument only as “permutations for the pseudo-p-value”, so passing 0 handed back a whole column of NA with nothing to say the test had been skipped. All three now document it, and say that p_value comes back NA.
• Two bundled datasets did not name every column they have. ?country_meta listed sixteen of its seventeen columns, omitting income – the World Bank income group that audit_coverage(by = "income") and rank_countries(within = "income") read – and ?world_snapshot called continent, region and income merely “classifications” without naming them.
• interpolate_missing() silently renamed a duplicated column. complete_years() and audit_coverage() reject a frame with two columns of the same name; this one filled the first and handed the second back renamed – v and v in, v and v.1 out, without a word.
• index_to() indexed a whole country to NA depending on row order. year == base_year is NA for a missing year, and x[c(NA, TRUE)] returns an NA element before the real match, so taking the first element picked up the NA. The same three observations gave 100/150/50 with the missing year last and NA/NA/NA with it first. The base year is now chosen from the rows that have one.
• compare_sources() invented a country from an unparseable year. read_year() deliberately records NA for a time value it cannot parse, and subsetting a data frame by a condition containing NA appends a row of all NA – a phantom country with no iso3c, which then survived the join into the comparison table.
• A brace in borrowed text replaced the message it was reporting. A cli bullet is a template, so text taken from somewhere else – a parallel worker’s error, an on-disk cache read failure, countrycode’s own complaint – had its braces interpolated. A worker failing with bad json {"a": 1} reported “Could not evaluate cli {} expression” and the real failure was gone, which is the worst possible moment to lose it. Four sites now pass the borrowed text through as a value.
• A bare column in a string-taking verb gave object 'v' not found. Nine of these verbs take an unquoted column through tidy eval; three take strings – interpolate_missing(value), complete_years(value) and audit_coverage(indicator). Writing the bare column that works everywhere else produced base R’s “object not found”, naming neither the argument nor the string it wanted. They now say what to write – value = "v" – while leaving a genuine error in the argument untouched.

Spatial statistics: weights first

morans_i() shipped with one hard-wired weight scheme, land-border contiguity. An island has no land border, so Japan, the United Kingdom, Australia, Indonesia, Madagascar, New Zealand, the Philippines, Iceland and every small island state carried no weight and left the analysis – 49 of the 191 countries with data in world_snapshot. The omission is systematic, not random.

• country_weights() builds the weights as a first-class object: "contiguity", "knn" (every country gets neighbours, islands included), "distance", or "custom" – which is how a non-geographic neighbourhood, trade volume or migration flows, goes through the same API.
• morans_i(weights = ) accepts it, and now reports n_excluded and an excluded list-column whatever the scheme. The two answers differ and both are defensible: contiguity gives I = 0.61 on 142 countries, k-nearest gives 0.47 on 189.
• local_morans() and lisa_map() give the LISA cluster classification and its map; gearys_c(), getis_ord() and spatial_lag() complete the set. No spdep required – at ~200 countries the dense arithmetic is trivial.

Data sources beyond the World Bank

A contract, not N bespoke fetchers.

• register_country_source() teaches the package a new provider. A registry rather than more Suggests, so a source with no CRAN package – V-Dem, the IMF, a proprietary internal feed – is a first-class citizen without the package depending on anything.
• fetch_indicator(), add_indicator() and country_sources() are the user-facing verbs; fetch_owid(), fetch_eurostat(), fetch_oecd() and fetch_comtrade() ship as built-in adapters, each gated on its own client package.
• compare_sources() is the one that earns its place. Anyone can call owidR; what nobody does is tell you that OWID and the World Bank disagree about GDP per capita for fourteen countries because of different vintages, PPP bases or territorial definitions. On the ISO spine that comparison is one join.
• clear_wdi_cache() generalises to clear_country_cache(source = ); the old name still works.

Time

• historical_geometry() draws the world as it was, from CShapes 2.0 (1886-2019) – states and colonies, which is what makes a pre-decolonisation map possible at all. world_geometry(year = ) and attach_geometry(year = ) route to it.
• The ISO spine does not reach back, and the package now says so. ISO 3166 was published in 1974 and never covered colonies, so country_join(key = ) and country_join_all(key = ) gain "cowc", "cown" and "gwn" as alternate spines, and warn – naming which table – about the dependencies COW/GW cannot carry. historical_geometry() is keyed on gwcode, with iso3c as a best-effort extra that is NA for every entity that never had one.
• country_groups_history (new dataset) dates membership for twelve groups, so country_groups(as_of = ) and in_group(as_of = ) answer the question a panel actually asks. The United Kingdom was in the EU in 2016 and not in 2021; EFTA had eight members in 1965 and has four now. Commonwealth, G20 and OPEC are deliberately not dated – their histories involve suspensions and contested dates, and a fabricated date is worse than an absent one. The table is validated at build time against country_groups_tbl.
• country_timeline() reads the crosswalk both ways: what the USSR became, and what Estonia was part of. audit_time_coverage() catches the rows a successful join leaves wrong – South Sudan with 1995 data, Czechoslovakia with 2001 data.

Honesty as a feature

• world_map(na_style =) chooses how missing countries are drawn: "grey" (the default, and ambiguous – grey reads as “low” to many people), "hatched" (diagonal hatching via the optional ggpattern, unmistakable and greyscale-safe), "outline" or "omit".
• world_map(footnote = "auto") generates the coverage line – “174 of 195 countries shown; 21 missing” – so a map cannot quietly overstate what it covers. A string is used verbatim.
• coverage_map() maps data availability itself, the cartographic counterpart to audit_coverage()’s table.
• world_map(classification_report = TRUE) attaches the breaks, the method and the count of countries per class to the plot. A map whose top class holds one country and whose bottom holds ninety is misleading, and the counts say so at once.
• classify_compare() draws the same choropleth under several classifications with that table for each. On world_snapshot’s GDP per capita, equal-interval and pretty breaks put 173 of 189 countries (92%) in a single class while quantile spreads them 38/38/37/38/38 – the difference between a map that says something and one that says nothing. ?world_map now states Brewer & Pickle’s (2002) finding that quantiles read best and Jenks materially worse, with the citation, since that is the reverse of the common GIS default.
• value_by_alpha_map() encodes the value in colour and an equalising variable (usually population) in opacity, over a neutral background. It is the answer to the small-number problem – an eleven-thousand-person country’s rate shouting as loudly as a billion-person country’s – and unlike a cartogram it solves it without distorting geometry (Roth, Woodruff & Johnson 2010).

Honesty, continued

• disputed_territories (new dataset) records that a territory is contested and who the parties are. It does not adjudicate, rank claims or imply any claim is better founded. dispute_policy() records which map convention you are using – and warns that selecting "de_jure" does not change a single shape, because the package ships no claimed-boundary geometry. check_dispute_coverage() and world_map(disputes = "mark") stop a contested area passing unremarked.
• world_map(uncertainty = ) switches the fill to a value-suppressing uncertainty palette (Correll, Moritz & Heer 2018): the value range contracts as uncertainty rises, so an uncertain estimate cannot claim an extreme colour, and the legend becomes the value x uncertainty grid.
• rate_check() flags rates computed over denominators too small to trust, and smooth_rates() shrinks them toward the global rate by empirical-Bayes. On a five-country example Tuvalu’s rate shrinks from 2.7e-4 to 5.9e-5 at a weight of 0.08, while China’s is untouched at 0.9999.
• interpolate_missing() fills panel gaps and flags every value it invents, non-optionally. Gaps longer than max_gap are left alone, because interpolating across a decade is not interpolation. world_map() reads the flag and notes it in the caption.
• projection_distortion() measures what a projection does, numerically. It cross-validates: every equal-area projection reads exactly 1.000 areal distortion, Mercator reads 4.4 mean and 132 max, and Mercator alone reads ~0 degrees angular.

Projections you can interrogate

• projection_info() returns the property table for all thirteen projections: construction family, equal-area, conformal, the PROJ string, and a note on what each is good for. subset(projection_info(), equal_area) is the short answer to “what may I safely use for a choropleth”.
• projection_compare() draws one choropleth under several projections at once, holding the data and the classification fixed so only the CRS varies.
• tissot_map() draws Tissot’s indicatrix – circles of equal ground radius projected with everything else. On Mercator they stay round and grow enormously; on Equal Earth they hold their area and shear. It makes the package’s honesty claim visible rather than asserted.
• Equal Earth is now documented as the recommended default for world thematic maps, with the Savric, Patterson & Jenny (2019) citation. The default itself is unchanged.

Provenance

• map_provenance() reports what went into a map – package version, snapshot year, geometry backend, projection, classification and its breaks, fill column, and countries shown versus missing. Every one of those is already known at plot time; this makes it readable. Every map verb records it, not just world_map(): a provenance feature that covers some of the verbs is worse than none, because the gap only shows up when someone relies on it.
• inst/CITATION credits the package and the sources it reconciles: countrycode (Arel-Bundock, Enevoldsen & Yetman 2018), the World Bank, Natural Earth, and the papers behind Equal Earth, the classification guidance, value-by-alpha and Moran’s I. citation("countryatlas") now produces the full set.

Flows, reporting and subnational

• flow_matrix(), country_network() and od_map() turn a bilateral table into a matrix, a network summary, or small multiples – the answer to the arc map’s problem, which is that past a few dozen flows flow_map() is a plate of spaghetti.
• country_factsheet() and world_table() are the last mile for the audience that wanted a table. world_table() uses gt when it is installed and a tibble otherwise.
• standardize_subnational(), nuts_geometry() and subnational_map() go one level below the country, scoped to where a maintained code system and free geometry both exist. There is no admin2, no bundled boundary data, and no attempt to pretend a French departement and a US state are the same kind of object.

Cartograms, projections and renderers

• gridded_cartogram() allocates one cell per N people by largest remainder, so the reader can count. cartogram_diagnostics() reports the residual area error, because cartograms fail quietly – on the bundled snapshot the default contiguous cartogram leaves Greenland 371x too large.
• cartogram_map(type = "flow") routes to cartogramR’s Gastner-Seguy-More algorithm. Each type now gates only on the package it actually needs.
• interactive_map(engine = "mapgl") renders through MapLibre GL, and globe_map(interactive = TRUE) turns the static orthographic globe into one you can spin with the mouse.
• world_map(engine = "tmap") is an alternative static renderer for people already working in tmap. The package stays ggplot2-native; this is a door, not a second front door.
• world_query() gains layer, facet, size and n_bins, so the database-side path can express bubbles, binning and small multiples.

Deprecations

• wdj_overrides() now warns. It has been soft-deprecated since 2.0.0, but an interactive-only note never reaches the scripts still calling it. Use country_overrides(); the two return the same table.
• ?wdj_overrides now says it is deprecated. The code started warning in this release, but the shared help page still described wdj_overrides() as “a backward-compatible alias” and its examples called it twice – so the one place a reader goes to check was describing the previous release’s behaviour, and running the documented examples emitted the deprecation warning. The examples now use country_overrides().
• options(countryatlas.gdp_compat = TRUE) now warns. The gdp_per_capita_2015 alias dates from 1.0.0 and will be removed.

Other changes

• New remove_country_source(). register_country_source() had no counterpart, so registering was permanent for the session: anything that registered a source – an example, a test, an exploratory script – left country_sources() reporting different rows for the rest of the session with no way back. Removing a source also drops its memoised answers, so re-registering the same name cannot serve results from the function that was just removed. The five built-in sources are protected.

• The persistent World Bank cache now expires and is size-capped. It used memoise::cache_filesystem(), which has neither, so the directory grew without bound for the life of the installation – while CRAN policy allows a cache under tools::R_user_dir() only if “sizes are kept as small as possible and the contents are actively managed (including removing outdated material)”. It now uses cachem::cache_disk(): entries expire after 30 days and the least-recently-used go once the directory passes 50 MB, both adjustable with options(countryatlas.cache_max_age = ) and options(countryatlas.cache_max_size = ). Expiry also matters on the merits, since World Bank observations are revised. Entries written by earlier versions are swept once, and a corrupt entry is now silently re-fetched rather than warned about and left in place. cachem is an unconditional dependency of memoise, so this adds nothing to install.

• clear_country_cache() can now release the geometry caches. Called with no source it also drops the Natural Earth sf layer held per scale (tens of megabytes at scale = "medium") and the memoised map_data("world") tibble (~99,000 rows per override set). Those are the largest things the package keeps in memory and there was no API to release them; in a Shiny app or a plumber process they were held for the life of the process. Naming a source leaves geometry alone. Relatedly, world_polygons is memoised in .onLoad() rather than at build time, which is both memoise’s own guidance and what makes the cache reachable.

• add_indicator() no longer claims “only what you need is fetched”. True for comtrade, partly true for wdi, and false for owid, eurostat and oecd, which download the full dataset and filter locally – so add_indicator(one_row, "owid", "life-expectancy") transfers every country and year to keep one value. ?add_indicator now says which is which.

• ?map_provenance listed 15 of the verbs that carry provenance. od_map(), subnational_map(), projection_compare() and world_map(engine = "tmap") also carry it; tissot_map() is the one map verb that carries none, and now says so. A test loops over every verb and asserts the provenance names the column the caller asked about.

• ?historical_geometry did not document owner or capname, which it returns when the installed cshapes supplies them. owner is what makes dependencies = TRUE legible, so it is worth naming.

• ?country_groups_tbl gives the membership date. It pointed at the package NEWS for the reference date, where two different dates were on record; the Rd now states 2026-06-01 and documents the as_of attribute that carries it in code. An Rd should not delegate a fact to a changelog.

• ?disputed_territories says that nothing in the package reads administered_by or claimed_by. dispute_layer() and dispute_note() key on iso3c alone; the party columns are for your own filtering, and are documented precisely so such a filter is writable.

• The README computes the count it quotes. “42 of 215 countries silently vanish” was hard-coded prose next to a figure built from a live fetch, so it drifted as the World Bank revised its coverage – by the time anyone checked it was 37 of 210. Both now come from one computation.

• stats, utils, tools, grDevices and parallel are declared in Imports. They were called with :: and not declared. Base-priority packages are exempt from the undeclared-:: check, so this was never a check finding; declaring them is conventional and removes any dependence on how that exemption is treated in future R.

• Two reads of ggplot internals are isolated. ggplot2 4.0.0 moved ggplot to S7 with @ accessors and a compatibility layer over $. animate_world() now reads the title with ggplot2::get_labs() where available, and cartogram_map() stores its frame in an attribute so cartogram_diagnostics() need not reach into the plot’s data slot at all.

• Build-time validation that only reported. ?country_groups_history says the table “is validated at build time against country_groups_tbl”; the script compared, message()d a mismatch, and wrote the .rda anyway, so a disagreement scrolled past in a build log – and NEWS records that this drift has already bitten once. It is a hard stop() now, accumulating every mismatch first, and it reads the snapshot from the working tree rather than from the installed package (which would compare the new table against the previous release’s snapshot). world_snapshot’s assembly gained relationship = "one-to-one" and a uniqueness assertion, and disputed_territories now validates administered_by / claimed_by against its documented six placeholders in both directions.

• country_meta$area_km2 is anchored to real areas. The formula that produced it lived in data-raw/ as a copy of the package’s ring_area_km2() that had silently missed the antimeridian fix – and nothing pinned the column numerically, the only assertions being > 0 and “>90% non-NA”, both of which a 179x-inflated value passes. The copies are reunited, the build script self-checks against the analytic area for an equatorial square, a wrapped square and a polar cap, and the shipped values are checked against ten known country areas. (The shipped values were correct; nothing had established that.) data-raw/overrides_snapshot.R’s copy of the override table is likewise pinned to the package’s.

• The CI matrix gained a _R_CHECK_DEPENDS_ONLY_ leg. All five existing legs install every Suggests, so none of them ran the configuration CRAN actually runs – which is how a broken \link{} to a Suggests-only topic reached a release. The Linux legs also generate tr_TR.UTF-8, so the Turkish-dotless-i hardening that broke 2.0.0 on CRAN’s Fedora flavours is actually exercised rather than skipped.

• The silence policy covers every offline verb. It enforced “a correct call to any verb is completely silent” for 39 of 102 exports, and every warning bug found in the pre-release review sat in the gap. It now covers 32 more, plus the cross-product of world_map()’s na_style / disputes / style arguments and a zero-row leg for the rate verbs – which is what found the deflate() and rate_check() bugs above.

• The Honest maps vignette named two countries that contiguity weights do not drop. Its island list included the United Kingdom and Indonesia, but the UK keeps its land border with Ireland and Indonesia keeps its borders with Malaysia, Papua New Guinea and Timor-Leste – so neither is excluded. A vignette about maps quietly misleading, quietly misleading. Corrected, and it now uses the two as the illustration that you cannot tell from the finished map who dropped out.

• Two more vignette claims corrected. Getting started called world_snapshot$countries$income “an ordered factor”; it is a plain factor whose levels are in income order, built that way deliberately in data-raw/, and is.ordered() is FALSE. countryatlas and ggsql said DRAW spatial arrived at ggsql 0.4.0 while three places in R/ said 0.4.1 – both are true of different components, but read together they were a contradiction, so the engine and the R package are now named separately.

• Every vignette’s prose claims are pinned by tests. Three were already; the four added or rewritten since then were not, which is how the errors above survived. The README’s are pinned too, including that its verb table covers every non-deprecated export.

• as_of documents what a bare year means. It resolves to 1 January, so in_group("Croatia", "EU", as_of = 2013) is FALSE – Croatia joined that July. The convention was always this; only the documentation was silent, and the examples all sat far from a boundary. Pass "YYYY-MM-DD" when the month matters.

• Every figure in the vignettes now carries alt text, so the pkgdown site and the rendered articles are readable with a screen reader.

• The geom_country_labels() example labelled all 188 countries at once, which made ggrepel drop every one of them; it now shows the two idioms that work – selecting a subset through data, and zooming in.

• Documentation and vignettes zoom with coord_quickmap() rather than coord_cartesian(). Both replace the map’s coordinate system, but only the former keeps the latitude-dependent aspect ratio, so the Europe inset was being drawn without any fixed aspect at all.

• DESCRIPTION declares Language: en-GB, and inst/WORDLIST records the package’s technical vocabulary, so the CRAN spell check reports real typos instead of 300 false positives.

• common_indicators gains the two price-conversion series deflate() and to_ppp() reach for by default.

• inst/CITATION credits the new methods: Anselin, Geary, Getis-Ord, Correll et al., Phillips-Sul, Gastner-Seguy-More and Schvitz et al. alongside the existing set.

• The mean Earth radius is one shared constant rather than three literals.

• A bad projection or scale now says which argument it means. Both were validated by match.arg() inside a shared helper, which produces R’s anonymous 'arg' should be one of ... – naming neither the argument nor the function the user called. Seventeen exported functions take projection and nine take scale, so one poor message was reachable a great many ways.

• Suggests grows to 41 packages. That is heavy, and it is the price of a light core: nothing here is needed to install the package or to draw a choropleth.

countryatlas 2.0.1

A patch release. The only behaviour change is to a test and to where an optional DuckDB connection keeps its state; nothing in the package’s own API moves.

CRAN check failures

• test-standardize.R’s de-accenting test failed on CRAN’s two r-devel Fedora flavours, which run in a latin1 locale. The test asserted that outside a UTF-8 locale iconv(x, to = "ASCII//TRANSLIT") cannot produce a resolvable spelling – generalising from LC_CTYPE=C, which is the case ?country_overrides actually documents. That is not what glibc does: the \u00e7 escape makes the input UTF-8 marked in every locale, so iconv reads it as UTF-8 and only the target charmap matters. Latin-1 has transliteration data and still yields "Curacao"; only C/POSIX, which has none, degrades to NA or "Cura?ao". The test now asserts the invariant that holds in every locale – de-accenting may or may not resolve, but it never resolves to a different country – and passes under C, latin1, latin9 and UTF-8.

Housekeeping

• as_ggsql_source(format = "duckdb") now opens its connection with duckdb(shared_home = FALSE) where the installed duckdb supports it (1.4 and later). By default duckdb keeps downloaded extensions and secrets in ~/.duckdb, which a throwaway in-memory table has no business creating in the user’s home; R CMD check also reports a new ~/.duckdb among “new files in some other directories”.
• The test suite no longer writes into the checking account’s file space. Rendering one girafe() widget – which interactive_map(engine = "ggiraph") does, and which no 1.0.0 test did – makes gdtools copy 90 Liberation font files into tools::R_user_dir("gdtools", "data"). R CMD check snapshots that tree and reports anything new, which is the NOTE CRAN raised against 1.0.0 for our own WDI cache (fixed separately, in wdj_cache_dir()). A new tests/testthat/setup-user-dirs.R points the R user directories at the session temp directory before any test runs, so the widget still renders, just somewhere disposable.
• ?clear_wdi_cache claimed nothing was written until a World Bank fetch succeeded, so an offline session “never creates” the cache directory. The directory is in fact created the first time a cached fetch is attempted, because that is when the location has to be proved writable – a failed fetch leaves it behind empty. The help page now describes what happens.

countryatlas 2.0.0

A major release that wires countryatlas into the database-rendering world via ‘ggsql’, widens the map vocabulary, and fixes several correctness issues found by auditing 1.0.0. The version is bumped to 2.0.0 because the bug fixes change the output of world_map() (quantile binning), bubble_map() / flow_map() (de-duplicated symbols), geom_country_labels() (label placement) and convert_country() (override-only entities), so code that depended on the old behaviour may see different maps or values.

New: database-side rendering with ggsql

• as_ggsql_source() exports a curated, ISO-reconciled, WDI-joined table (with sf geometry WKB-encoded) as a ggsql source: a DuckDB connection, a Parquet file, or a nanoarrow stream. countryatlas does the reconciliation ggsql’s static bundled world can’t; ggsql does the database push-down and Vega-Lite output countryatlas doesn’t.
• world_query() emits a ggsql spatial query (VISUALISE … DRAW spatial PROJECT TO … SCALE … LABEL …), a dependency-free string builder.
• interactive_map(engine = "ggsql") registers the data and renders the map in DuckDB, returning a Vega-Lite widget.
• ggsql, duckdb, DBI and nanoarrow are optional Suggests. See the new countryatlas and ggsql vignette.

New: maps, projections and helpers

• globe_map(): an orthographic globe choropleth, with backend = "sf" (smoothest limb) or backend = "polygon" (needs only maps + mapproj).
• spin_globe(): a rotating-globe animated GIF (one globe_map() frame per central longitude, assembled with gifski or magick).
• facet_map(): small-multiple choropleths (the static counterpart to animate_world()).
• wdj_crs() gains eight projections (mercator, winkel_tripel, eckert4, gall_peters, orthographic, azimuthal_equal_area, north_polar, south_polar); world_map() / world_geometry() accept them all.
• locate_country(): point-in-polygon lookup tagging lon/lat with iso3c.
• repair_country_names(): the “act on it” companion to check_country_match(): auto-applies confident string-distance fixes.
• country_join_all(): reduce-join many messy country tables on the ISO spine.
• growth_rate(), index_to(), share_of_world(): panel analysis helpers.
• country_overrides(): preferred name for wdj_overrides() (kept as an alias) after the rename to countryatlas.
• country_groups_tbl gains Mercosur, GCC, Nordic and Visegrad.
• country_borders(): a tidy adjacency edge list built from polygon topology (sf::st_touches()), with neighbors() for a vectorised per-country lookup.
• distance_between(): great-circle (haversine) distance between two countries’ centroids; needs neither sf nor the network.
• dorling_map(): the Dorling cartogram promoted to a first-class verb, with k/itermax tuning; cartogram_map() itself gains ... passthrough to the underlying cartogram::cartogram_*() call.

New: historical entities, inequality and spatial statistics

• historical_codes: a curated, dated crosswalk of dissolved entities (Soviet Union, Yugoslavia, Czechoslovakia, East Germany, Netherlands Antilles, North/South Yemen, pre-2011 Sudan, United Arab Republic, Tanganyika/Zanzibar, North/South Vietnam, Serbia and Montenegro) to their successor states, with retired ISO codes where they existed. Kosovo is included among the Yugoslav successors on a territory basis (documented).
• dissolve_country(): resolve a mixed vector of historical and modern names to successor iso3c rows (one-to-many, dated); modern names pass through as single rows, so a whole messy column pipes in unchanged.
• check_country_match() gains a historical column. It flags dissolved entities even when countrycode “matches” them. The headline case is "USSR", which countrycode silently resolves to Russia’s RUS, so Soviet-era data becomes Russian data with no warning.
• correlate_indicators(): pairwise indicator correlations on the spine (pearson/spearman, pairwise-complete, per-pair n), tidy long output.
• beta_convergence() / sigma_convergence(): the two standard convergence diagnostics: the growth-on-initial-level regression (with implied convergence speed and half-life) and per-year cross-country dispersion.
• gini() and theil(): inequality across countries, population-weightable; theil() decomposes exactly into between/within components when a grouping (continent, income) is supplied.
• lag_by_country() / diff_by_country(): panel lag and difference grouped by iso3c and ordered by year, completing the panel toolkit around growth_rate() / index_to() / complete_years().
• morans_i(): global Moran’s I with a permutation pseudo-p-value, computed on the row-standardised country_borders() adjacency. No spdep dependency: the weights come from the package’s own curated topology.
• spike_map(): triangular spikes at country centroids (height ∝ value), the overplotting-resistant cousin of bubble_map(); needs only maps.
• convert_country() accepts to = "name_<lang>" ("name_fr", "name_es", "name_zh", …) for localized country names via countrycode’s CLDR tables.
• world_map(style = "binned") legends now show SI-formatted breaks (4M, not 4e+06) when scales is installed; the continuous scale uses the same formatter.

Bug fixes

• bivariate_map() errored on every call (“the condition has length > 1”, pre-dating 2.0.0). The two fill columns were injected into biscale::bi_class() with !!rlang::sym(), but bi_class() reads them with as.character(substitute(...)) rather than tidy eval, so the injection deparsed into a multi-element vector inside biscale. The happy path is now covered by a test (the old one only checked that the function errors cleanly when sf is absent).
• as_ggsql_source() and interactive_map(engine = "ggsql") errored on any sf input – the whole point of the ggsql bridge. sf::st_as_binary() returns a classed WKB object, which tibble rejects (“all columns must be vectors”); the geometry column is now the plain list of raw vectors that nanoarrow encodes as binary and DBI writes as a BLOB.
• projection = "winkel_tripel" errored on every render – one of the eight projections this release adds. The CRS built fine and the geometry projected fine, but ggplot2::coord_sf()’s graticule collapses to a degenerate single-point segment under PROJ’s Winkel Tripel, which GEOS rejects (“point array must contain 0 or >1 elements”). The graticule is now skipped for that projection only; [theme_world_map()] blanks panel.grid anyway, so nothing visible changes. All 13 projections are now covered by a full-render test.
• world_geometry("coastline", geometry = "sf") errored with a GEOS TopologyException in every projection except "plate_carree": a couple of Natural Earth rings are self-intersecting and sf::st_union() (unlike the spatial predicates) refuses them. The geometry is repaired before the union.
• world_geometry(region = c(xmin, ymin, xmax, ymax)) – and world_data() / attach_geometry() with a bounding-box region – errored on the sf backend (“Loop 0 is not valid”), because sf::st_crop() runs under the strict S2 engine on unprojected geometry. It now clips with the GEOS planar predicate, as country_borders() / locate_country() already did.
• convert_country()’s warn argument was documented but silently ignored (every internal countrycode() call is wrapped in suppressWarnings(), because countrycode also warns on intermediate hops that convert_country() goes on to recover). It now reports inputs that match no country, like standardize_country() does. A recognised country whose destination value is genuinely missing still returns NA quietly.
• world_map() / globe_map()’s na_label was accepted and silently ignored. The "quantile", "jenks" and "categorical" legends now label their missing-data key with it (the continuous and binned colourbars have no NA key to name, which the documentation now says).
• Kosovo’s XKX resolves for country and flag from from = "iso3c", not just from its name. It has no row at all in countrycode::codelist, so everything derived from the code was NA – which surfaced as country_borders() / neighbors() returning NA names for Kosovo’s four land borders, locate_country(add = "country") returning NA for points inside it, and standardize_country(add = c("country", "flag")) doing the same. The curated fallback table now carries the name and flag too.
• per_capita() without an explicit pop column died with an opaque vctrs error (“Can’t subset columns that don’t exist: .wdj_pop”) when the World Bank population fetch failed or timed out – fetch_wdi() deliberately degrades to a keys-only tibble in that case. It now reports the failed fetch and points at the pop argument.
• theil() returns NA shares (not NaN) for a perfectly equal distribution, where the total is 0 and the shares are undefined – matching how gini() and share_of_world() treat a zero denominator.
• country_join() / country_join_all() no longer cross-join rows whose iso3c is NA: unmatched countries used to collapse to a single NA key and fan out into a Cartesian product. The joins now pass na_matches = "never" (#4).
• country_join_all() validates the length of origin (must be 1 or one per table) instead of failing with a cryptic “missing value where TRUE/FALSE needed” error (#16).
• join_world()’s auto-detection (detect_country_col()) honours the candidate priority order instead of picking the first column by data-frame position, so a region column no longer shadows a real country column (#6).
• standardize_country(add = ...) accepts any raw countrycode destination (e.g. "iso3n") again instead of erroring with “subscript out of bounds” (#5).
• standardize_country(origin = "iso3c") now validates codes: strings that are not real ISO 3166-1 alpha-3 codes become NA (and are flagged by warn) rather than passing through uppercased and unchecked (#12).
• country_data(latest = TRUE) / world_data(latest = TRUE) for a single year now returns each country’s most recent non-NA value: the fetch window is widened so an earlier observation can actually be found (#7).
• fetch_wdi() keeps iso2c / country for a country that appears only in a non-first indicator (they are coalesced across indicators) instead of leaving them NA (#8).
• world_map() / globe_map() with style = "quantile" / "jenks" no longer error on a constant, single-country, or all-NA value column; degenerate breaks now fall back to a single bin (#9).
• flow_map() returns the base map (instead of erroring) when no origin-destination pair resolves to a centroid (#10).
• aggregate_regions(fun = "min"/"max") returns NA for an all-NA group instead of Inf / -Inf (#11).
• share_of_world() returns NA (not NaN/Inf) when the (per-year) total is zero or non-finite (#13).
• gini() returns NA with a warning for negative input rather than a value outside the documented [0, 1] range (#14).
• spike_map() no longer produces NaN spike coordinates when every height is zero (#15).
• world_query() honours transform even when palette = NULL, emitting a standalone SCALE fill VIA <transform> clause (#17).
• world_map(style = "quantile"/"jenks") computed breaks over polygon vertices, so a country’s geometric complexity biased the quantiles and the bins held unequal numbers of countries. Breaks are now computed on one value per country.
• bubble_map(backend = "sf") placed bubbles in projected metres on a degrees base map (off the map). The base map and bubbles now share one projected CRS via coord_sf().
• Polygon centroids returned more than one row for ten iso3c codes (overrides map several names, such as Azores/Madeira → PRT, to one code), fanning out joins in bubble_map() / flow_map(). Centroids are now one antimeridian-safe row per country (the largest piece).
• geom_country_labels() placed labels at the bounding-box midpoint over all of a country’s pieces, so the US / Fiji / NZ labels drifted into the wrong ocean. Labels now sit on each country’s largest piece.
• projection = "plate_carree" built an incoherent PROJ string (+proj=longlat … +units=m); it is now true equirectangular (+proj=eqc).
• convert_country() only applied wdj_overrides() for to = "iso3c", so override-only entities (e.g. “Canary Islands”, “Azores”, “Bonaire”) returned NA for every other destination (continent, region, iso2c, flag, currency, country name, …). It now resolves the override-corrected iso3c first and derives every other destination from that.
• Kosovo’s XKX needed extra care: it has no row at all in countrycode::codelist, so deriving destinations purely via the iso3c round-trip above is NA for everything, which would have regressed flag/region/country, since 1.0.0 already resolved those via direct name matching (verified against the actual 1.0.0 code). convert_country() now recovers from the original name when the iso3c round-trip comes back empty, and fills iso2c/continent (which neither path classifies) from the same curated fallback standardize_country() uses. Net effect versus 1.0.0: zero regressions, plus newly-working continent/iso2c for Kosovo, which also fixes locate_country(..., add = "continent") for points inside it.
• interactive_map(..., tooltip = ) was accepted but silently ignored by every engine (pre-dating 2.0.0). The "ggiraph" and "leaflet" engines now use the supplied tooltip column, defaulting to fill as before when omitted.
• world_data(overrides = ) (and attach_geometry(overrides = )) accepted a custom name -> iso3c override set but silently ignored it (pre-dating 2.0.0) – the geometry backend always matched with the default wdj_overrides(). The override set now flows through to both the polygon and sf matchers, so a custom mapping actually changes which polygons a country claims.
• repair_country_names() no longer records a no-op “repair” when a dissolved entity’s own name (e.g. “Yugoslavia”, which exists in the codelist but has no ISO code) comes back as its closest suggestion; dissolve_country() is the right tool there and is what the report now points to.
• A mistyped column name is now reported by countryatlas rather than leaking out of ggplot2 as a bare “object ‘x’ not found” from inside a layer, or out of vctrs as a subscript error. world_map(), globe_map(), facet_map(), tile_map(), bubble_map() (both size and color), spike_map(), flow_map() (from, to and weight), interactive_map() (fill and tooltip) and morans_i() all validate up front, matching the message per_capita() / rank_countries() / bivariate_map() already gave. morans_i() in particular used to blame the geometry (“not enough bordering countries with data”) for a column that simply wasn’t there.
• audit_coverage(indicator = ) silently reported n_missing = 0 and na_rate = NaN for a column name that isn’t in data; it now errors.
• world_map() / globe_map() errored (“‘length = 2’ in coercion to ‘logical(1)’”) when na_label was longer than one element. The first element is used to label the single NA key, and a NULL / NA label still leaves the default formatter alone.
• per_capita() sent start = Inf to the World Bank when data had a year column that was entirely NA; it now falls back to last year, as it already did for a frame with no year column at all. Its degraded-fetch guard also covers the join keys now, so a partial population fetch produces the actionable “pass a population column” error rather than a raw vctrs subscript error.
• theil() returned NaN when every weight was zero; it now returns NA, matching gini().
• world_map() / globe_map() with style = "categorical" and a numeric fill column let ggplot2 raise “Continuous value supplied to a discrete scale” at build time, naming neither the column nor the style. They now error at the call, name the column, and point at "quantile" / "jenks" / "binned".
• bivariate_map() no longer leaks biscale’s “var has missing values, omitted in finding classes” warning, which fired on essentially every call because real indicators always have gaps (the classes were valid either way). Any other biscale warning still passes through.
• A region given as lowercase iso3c codes silently lost countries. Falling through to name matching resolved some codes by accident (countrycode’s country-name regex is case-insensitive, so "usa" matched) but not others ("can" did not), so region = c("usa", "can") subset to the USA alone. Codes are now recognised in any case, matching what standardize_country(origin = "iso3c") already accepted; an all-uppercase unknown code is still taken at face value rather than reinterpreted as a country name.
• country_codes() silently dropped a column name it did not recognise, so a typo returned a table quietly missing that column; it now errors and lists the available shortcuts.
• Every sf-backed call printed three or more lines of sf internals to the console – "Spherical geometry (s2) switched off", st_intersection’s "although coordinates are longitude/latitude ... assumes that they are planar", and the matching "switched on". The source was sf::st_break_antimeridian(), which toggles the s2 engine and runs an intersection internally, and which sits on the path of every sf call: a plain attach_geometry(geometry = "sf") emitted them, as did world_map(), world_geometry(), country_borders(), neighbors(), morans_i(), locate_country() and simplify_geometry(). It is now wrapped in the same quietly_sf() helper the other sf calls already used, so those paths are silent. (The notices bypass R’s condition system, so suppressMessages() could not have caught them.)
• cartogram_map() / dorling_map() never validated their weight or fill column, the one place the rest of the package’s existence checks were missed. A bad weight reached cartogram as "missing value where TRUE/FALSE needed" (or, for the Dorling variant, a warning about max() and then a wrong picture), and a bad fill was not caught at all.
• index_to() likewise never checked its value column, so a typo produced a dplyr error from inside mutate(); base_year and to are validated too.
• Scalar arguments are validated, so a typo or an NA names the argument instead of surfacing as "missing value where TRUE/FALSE needed", classInt’s "n less than 2", or a PROJ complaint about lat_0. Covers n_bins (world_map() / globe_map(), on the binned path as well as the quantile/jenks one), lon/lat/recenter/lat0 (globe_map() on both backends – the polygon one goes to coord_map() and previously accepted a nonsense orientation silently), n (flow_map()), max_height / width / alpha (spike_map()), max_size / alpha (bubble_map()), keep (simplify_geometry()), threshold (repair_country_names()), n_perm (morans_i()), and n_frames / fps / width / height (spin_globe()).
• Two of those scalar arguments were not merely reported badly – they drew the wrong thing in silence. A negative max_height drew spike_map()’s spikes upside down, and globe_map(lat = ) beyond +/-90 built a CRS PROJ rejects, which only surfaced later as coord_sf()’s "crs not found: is it missing?".
• geom_country_labels(repel = TRUE) silently drew plain labels when ggrepel was not installed – the one degraded optional backend the package did not announce, where classInt, gganimate and rmapshaper all report theirs. It now says so once per session (the argument defaults to TRUE, so reporting on every call would be noise), and stays quiet when repel = FALSE was asked for.
• The number of bins no longer depends on whether classInt is installed. For a fractional n_bins, classInt truncated internally while the base-quantile fallback passed the fraction to seq(length.out = ) and produced one break more, so the same call binned differently in different environments. n_bins is now truncated to a whole number of bins before either backend sees it.
• simplify_geometry(keep = 0) errored under rmapshaper but was silently accepted by the sf::st_simplify() fallback, so the same call behaved differently depending on which optional package the caller had installed. A proportion of zero keeps no vertices, and both paths now reject it.
• Count arguments are bounded above as well as below. The scalar checks required a finite number, but several call sites then coerce with as.integer(), which returns NA past 2^31-1 – so n_perm = 1e10 or n_bins = 1e10 produced “NAs introduced by coercion” or, worse, “missing value where TRUE/FALSE needed”. n_bins, n (flow_map(), lag_by_country(), diff_by_country()), n_perm and n_frames now name the range. morans_i() also dropped a max(0L, ...) clamp that the validation had made unreachable.
• lag_by_country() / diff_by_country() clamped n <= 0 up to 1, so a lag of 0 quietly returned a lag of 1; it now errors.
• complete_years(value = ) silently ignored a column name that wasn’t in data under the default method = "none", while erroring from all_of() for "locf" / "linear"; it now errors consistently.
• interactive_map(engine = "ggsql") now gates on ggsql >= 0.4.1 rather than mere presence. DRAW spatial – the clause world_query() emits – arrived in the ggsql engine at 0.4.0, while the ggsql R package is still 0.3.3, which accepted the call and then failed inside its own SQL front end on a clause it did not know. The gate now refuses with an actionable message instead. ?world_query records that the clause has shipped in the engine but not yet in the R bindings, and that PROJECT TO additionally needs a spatial backend (for DuckDB, its spatial extension); world_query() itself remains a dependency-free string builder.
• R CMD check no longer writes to the checking user’s persistent cache. The \donttest{} examples fetch from the World Bank, so the memoised on-disk cache was being populated under tools::R_user_dir() during a check; under check it now lives in the session temp directory instead. Normal use is unchanged, and options(countryatlas.cache_dir = ) still overrides both. ?clear_wdi_cache now documents where the cache lives and how to disable it.
• An unmatched country in your data could be drawn as a real country. dplyr joins default to na_matches = "na", so an NA ISO code matched another NA ISO code – and Natural Earth carries Somaliland as a polygon with no ISO code. Any row whose country failed to resolve therefore joined onto Somaliland’s geometry and was plotted there; with two or more unmatched rows the join also fanned out many-to-many, duplicating that polygon once per row so the visible fill was whichever happened to be drawn last. Affected attach_geometry(geometry = "sf") (and so join_world() and world_map() downstream of it) and bubble_map(backend = "sf"). All country-keyed joins in the package now pass na_matches = "never", which country_join() and country_join_all() already did; the keyless polygon is still drawn, now correctly as a no-data feature. A test asserts the invariant across the whole namespace so a new join cannot reintroduce it.
• world_query() emitted a silently malformed query for any argument that was not a single string. sprintf() vectorises, so projection = c("a", "b") produced two PROJECT TO clauses, source = character(0) deleted the FROM line entirely, and title = NA became the literal text 'NA' – each of which surfaced only later, as a parse error inside ggsql’s SQL front end. world_query() and as_ggsql_source() now validate their string arguments up front and name the offending one. NULL still omits an optional clause, and an empty title is still allowed.
• as_ggsql_source(format = "parquet") built its COPY ... TO '<path>' statement by string interpolation, so a path containing an apostrophe – legal in a filename – closed the SQL literal early and broke the statement. The path is now quoted with DBI::dbQuoteString(), matching the dbQuoteIdentifier() treatment the table name already had.
• suffix = character(0) made the whole computation vanish. suffix is paste0()-ed onto the value column’s name, and dplyr’s "{character(0)}" := is a silent no-op – so growth_rate(x, g, suffix = character(0)) returned x unchanged, with no growth column and no error. suffix = NA produced a column named gNA, and suffix = "" overwrote the source column in place. per_capita(), growth_rate(), index_to(), share_of_world(), lag_by_country() and diff_by_country() now require a single non-empty string (lag_by_country()/diff_by_country() still accept NULL for the default suffix).
• Several arguments produced an error that named nothing rather than the argument at fault:
  • convert_country(to = c("country", "continent")) – a plausible attempt at two destinations – raised “the condition has length > 1”, and a zero-length to or from raised “argument is of length zero”.
  • origin did the same across every function that resolves country names. It is now validated once in the shared internal, so standardize_country(), country_join(), country_join_all(), join_world(), flow_map(), neighbors(), distance_between(), in_group(), repair_country_names() and check_country_match() are all covered.
  • attach_geometry(by = character(0)) raised “argument is of length zero”.
  • aggregate_regions(by = character(0)) raised nothing at all: it grouped by no columns and silently collapsed the world into a single row. by remains documented as plural, so multiple grouping columns still work.
• Logical arguments are validated too, closing the same gap in two forms. The borders argument of world_map() and globe_map() fed a bare if (), so a bad value raised one of four opaque base R errors (“missing value where TRUE/FALSE needed”, “argument is of length zero”, “the condition has length > 1”, “argument is not interpretable as logical”) – none naming borders. Elsewhere the value went through isTRUE(), which never errors but silently turns anything that is not TRUE into FALSE, so the caller got the opposite of what they asked: rank_countries(x, v, desc = "yes") ranked ascending, putting the lowest value at rank 1, and gini(x, na.rm = "yes") kept the NAs and returned NA. All 23 logical arguments across the package now require TRUE or FALSE and name themselves when they do not get it.
• gini() and theil() returned a wrong number for a wrong-length weights vector. Both recycled it with rep_len(), which accepts any length silently, so gini(1:10, weights = c(1, 2)) returned 0.2902 – computed from an alternating 1,2 pattern – where the correctly-weighted answer is 0.3. Someone weighting by population and mistakenly passing a vector of the wrong length got a plausible figure and no indication anything was wrong. weights (and theil()’s groups) must now be length 1 or the length of x, and weights must be numeric. complete_years() likewise rejects a years vector that is non-numeric, empty, or contains NA, all of which it previously coerced to NA behind base R’s warning.
• Omitting a required argument now names it. Every affected function resolved its column argument with rlang::as_name(), which raises argument "x" is missing, with no default for a missing value – naming rlang’s own parameter, and none of these functions has an argument called x. world_map(), tile_map(), facet_map(), spike_map(), bubble_map(), globe_map(), interactive_map(), rank_countries(), growth_rate(), per_capita(), aggregate_regions(), index_to(), share_of_world(), lag_by_country(), diff_by_country(), beta_convergence(), sigma_convergence(), morans_i(), world_query(), country_join(), bivariate_map(), cartogram_map() and flow_map() now report e.g. `fill` is required. Optional tidy-eval arguments are unaffected.
• distance_between() paired the wrong countries when a and b had mismatched lengths. It combined them through vectorised arithmetic, so R’s recycling applied: 2 countries against 3 returned a[1]-b[1], a[2]-b[2] and a[1]-b[3], behind only base R’s “longer object length is not a multiple” warning, and 2 against 4 recycled cleanly with no warning at all. Equal lengths, or a length-1 side for one-against-many, are now required – the same rule locate_country() has always enforced for lon/lat.
• Three documented contracts did not match their code: ?locate_country said lon/lat were “recycled together” when the function has always required equal lengths, and ?gini / ?theil promised that weights was “recycled against x the usual R way”, which is precisely the behaviour removed above. All three now describe what the functions do. locate_country()’s length error also read “or an points sf object”.
• Offline safety is now covered by tests rather than assumed. Checks run \donttest{} examples and rebuild vignettes, and CRAN policy does not allow either to fail for want of a network connection. world_data() and country_data() degrade a failed fetch to a warning and a metadata-only frame, and wdi_search() reads WDI’s bundled indicator list rather than the API; all three are now asserted, so a change that made any of them require a connection would break the suite.
• An unwritable cache directory silently cost you your data. memoise::cache_filesystem() does not validate the directory it is given: it constructs successfully and only fails when something is written, which happens deep inside the fetch. So with a read-only or otherwise unusable cache location, country_data(cache = TRUE) reported Could not fetch indicator "..." from the World Bank API and returned the country spine with every indicator NA – blaming the API for a local permission problem, while the same call with cache = FALSE returned the data perfectly. (The tryCatch() that was meant to fall back never fired, because constructing the cache never errored.) The directory is now checked before use, with a fallback to session-only caching and a one-time message naming the real cause.
• Projected maps failed outright under options(OutDec = ",") – the ordinary setting in comma-decimal locales. The PROJ strings are built by pasting numbers, so recenter = 48.9 became +lon_0=48,9, which PROJ rejects; the invalid CRS then surfaced as sf’s opaque “crs not found: is it missing?”. Every number destined for a machine-readable string is now formatted with an explicit decimal mark.
• Two more of the same kind, triggered by options(scipen = -10), which formats a double in scientific notation: sf::st_crs(4326) became EPSG:4.326e+03 and yielded an NA CRS (surfacing later as st_crs(x) == st_crs(y) is not TRUE from locate_country()), and Natural Earth’s scale 110 became 1.1e+02, so world_geometry(geometry = "sf") failed with 'countries1.1e+02' is not an exported object. Every EPSG code and Natural Earth scale is now an integer literal, which scipen does not affect.
• simplify_geometry() and world_geometry("graticule") are insulated from two upstream bugs of the same family, both reproducible without this package: rmapshaper serialises keep for V8, which rejects the 0,1 that options(OutDec = ",") produces, and sf::st_graticule() overflows the node stack under options(scipen = -10). Both calls now run with those two options normalised, and the caller’s settings are restored immediately afterwards.
• rank_countries() silently ranked within groups when handed a grouped frame. Its mutate() honoured the caller’s group_by(), so the same data ranked 4, 1, 3, 2 ungrouped and 2, 1, 2, 1 after an incidental group_by(region) upstream in the pipe – with within = NULL in both cases, which documents a global ranking. rank, percentile and z_score were all affected. within is now the only thing that sets the ranking scope, matching every other function here, which imposes its own grouping rather than inheriting the caller’s. A test asserts that a grouped input changes no answer, and that nothing leaks grouping into its return value.
• A non-numeric value column now errors by name instead of producing nonsense. A factor column is easy to acquire – read.csv() on a column with one stray non-numeric entry gives you one – and arithmetic on it failed four different ways: growth_rate() and per_capita() returned a column of NAs behind base R’s “‘/’ not meaningful for factors”; share_of_world(), rank_countries() and aggregate_regions() raised an opaque error from inside dplyr::mutate(); gini() managed “missing value where TRUE/FALSE needed”; and morans_i() quietly returned a plausible-looking statistic. These, plus index_to(), diff_by_country(), beta_convergence(), sigma_convergence() and theil(), now name the column and its actual type. lag_by_country() is deliberately unchanged: it does no arithmetic, so lagging a factor or character column remains legitimate.
• aggregate_regions() reported a figure for groups it had no data for. Values are dropped before aggregating, so a group whose every value is missing had nothing left – and each base function got that wrong differently: "sum" returned 0, "mean" and "weighted_mean" NaN, and "min"/"max" -Inf/Inf plus a warning. “This region’s total is 0” is a claim, not an absence, which matters in a package built around honest missing-data handling. All six now return NA, as "min"/"max" were already meant to; groups that do have data are unaffected, and a partially-missing group still aggregates the values it has. ?aggregate_regions documents this.
• complete_years() failed on a zero-row panel, where every other panel helper returns zero rows. With no years it reached seq(min(numeric(0)), max(numeric(0))) and died on base R’s “‘from’ must be a finite number”; with years supplied it died on tidyr’s “Can’t recycle year (size 3) to size 0”. Neither message names anything the caller did. It now returns the empty frame, columns intact, for all three method values – while still reporting a bad years or value argument.
• bivariate_map() and cartogram_map() (and so dorling_map()) failed inside their optional dependency when no row carried the values they need. This is easier to hit than it sounds: attach_geometry() joins geometry-on-the-left, so a frame with nothing in it arrives at the plotting verb as full-length columns of NA. biscale then indexed sVar[1:(length(sVar) - 1)], which becomes 1:-1, and reported “only 0’s may be mixed with negative subscripts”; cartogram compared NA in if (meanSizeError < maxSizeError) and reported “missing value where TRUE/FALSE needed”. Neither mentions the data. Both now say which columns are empty, as spike_map() already did, and both reject a non-numeric column by name. Partly-missing columns still draw from the rows that do have values.
• gini() could kill the R session. It computed the weighted mean absolute difference with outer(), an n-by-n matrix – fine for the ~200 countries it is written for, but it is exported and accepts any numeric vector. A geometry-joined column is 99,338 rows, needing about 79 GB, and the process was killed outright: no error, no message, no result. The kernel is now the sorted cumulative form, O(n log n) in time and O(n) in memory, which agrees with the pairwise definition to floating-point noise (verified across ties, zero weights, single values and 340 random cases) and handles a million values in well under a second. Every documented figure is unchanged.
• aggregate_regions() silently multiplied its answer when given a frame with map geometry attached. The polygon backend expands each country into hundreds of vertex rows, so a row-wise total counts it once per vertex: for the bundled snapshot a regional total of 497,265 came out as 280,951,373. It is reachable directly off world_data(geometry = "polygon"). It cannot de-duplicate on iso3c, since by = c("region", "year") roll-ups legitimately repeat a country, so it now warns and says what to do instead. Country-level tables and panels are unaffected.
• ?audit_coverage described the raw list it returns without mentioning that the object is classed and has a print() method, so what you actually see at the console is a formatted report rather than the list. Both are now documented.
• world_geometry("ocean") drew nothing at all, in every projection. Under the S2 engine – sf’s default since 1.0, so everywhere – st_as_sfc(st_bbox(-180, -90, 180, 90)) collapses to a two-point, zero-area polygon, and the collapse is invisible because st_bbox() reports the stored extent instead of recomputing it from the (empty) coordinates. The rectangle is now built by constructing the ring explicitly, which S2 never gets to reinterpret, and its edges are densified so a curved projection has points to bend: the layer comes out at Earth’s true surface area (5.1e14 m2) and covers 98-100% of the countries layer across all nine world projections. st_break_antimeridian() is no longer applied here at all – with lon_0 = 0 it cut the outline at +/-180, its own edges, taking it down to two thirds of the globe and, under Mollweide, to nothing.
• Where an ocean background cannot be drawn, world_geometry() now says so instead of returning an invisible layer: the four hemispheric projections ("orthographic", "azimuthal_equal_area", "north_polar", "south_polar") show half the globe, and a whole-globe rectangle cannot be recentred without covering only part of the map.
• world_geometry("coastline") and world_geometry("ocean") returned a bare sfc rather than the sf object ?world_geometry promises (and that the other four what values deliver), so dplyr verbs failed on exactly those two. Both are now sf; the geometry is unchanged.
• ?world_geometry’s @return was a single line naming no columns. It now lists what each what returns, and warns that the sf backend’s centroid_lon/centroid_lat are in the object’s own CRS – projected metres, not degrees, so centroid_lon for France is 174097, not 2.1. Use country_meta$centroid_lon for degrees.
• sf::st_coordinates() failed on world_geometry("countries", geometry = "sf"), in every projection including the default. Natural Earth supplies 177 uniform MULTIPOLYGONs, but st_break_antimeridian() runs an st_intersection() internally that collapses a single-part MULTIPOLYGON to a POLYGON, leaving 148 POLYGON + 29 MULTIPOLYGON – an sfc_GEOMETRY column, which st_coordinates() does not support. Extracting vertices from the package’s own geometry, an ordinary thing to want, therefore errored with “not implemented for objects of class sfc_GEOMETRY”. The column is cast back to MULTIPOLYGON, which is a type change only: row count, codes and land area are unchanged. ?world_geometry also now notes that a hemispheric projection returns empty geometries for the far side, where the same st_coordinates() limitation applies for a different and correct reason.
• world_map() accepted a frame with no geometry, returned a ggplot object without complaint, and then failed only when the plot was printed – with ggplot2’s “Problem while computing aesthetics … Caused by error in .data$long”, which names nothing the caller did. It validated the fill column but never the long/lat/group columns the polygon path needs. Forgetting attach_geometry() is the easiest mistake to make here, and it is easy precisely because the other plotting verbs do not need it: tile_map(), bubble_map(), spike_map() and globe_map() all take a country-level frame, so world_map(snap, gdp) looks like it should work too. It now says so at the call, and names the fix. facet_map() delegates to world_map() and is covered by the same check; the four country-level verbs are deliberately unchanged, and a test pins that asymmetry.
• interactive_map(engine = "ggiraph") reported a missing geometry differently from the other engines. It assembles its own ggplot rather than calling world_map(), so it bypassed that check and failed at render time on .data$long, while engine = "plotly" named the problem properly. Both now give the same message. ?interactive_map also documents that the "leaflet" engine attaches geometry itself if handed a country-level table, which the others do not.
• The numeric fill styles now require a numeric column. style = "continuous" and "binned" reached ggplot2 and failed only when the plot was printed (“Discrete value supplied to a continuous scale”, “Binned scales only support continuous data”), neither naming the column; "quantile" and "jenks" did not fail at all – the break computation returns early on a non-numeric column, so the fill fell through to the discrete scale and drew a plausible map whose legend claimed quantile bins it had never computed. The reverse direction, style = "categorical" on a numeric column, was already guarded, so this closes the pair. animate_world() and facet_map() inherit it.
• attach_geometry() no longer warns when given a panel. Joining one row per country-year against polygon vertices is legitimately many-to-many – it is what animate_world() and facet_map() are for – so dplyr’s “unexpected many-to-many relationship” warning was noise. The relationship is now declared, as the cache merge already did.
• ?repair_country_names now says which way the stringdist fallback errs. The threshold argument already noted that the metric changes when stringdist is absent (Jaro-Winkler versus a length-normalised edit distance); it now adds that the fallback is the more conservative of the two, repairing a subset of what Jaro-Winkler would – mainly missing transposed letters, as in “Frnace” – and never choosing a different country. Measured over 120 single-typo names: 98 repaired with stringdist, 77 without, none repaired that stringdist did not, and no wrong repairs either way. A test now holds the package to that.
• simplify_geometry() undid the geometry-type fix above. Both simplifiers collapse a single-part MULTIPOLYGON to a POLYGON, so a homogeneous input came back as a mixed sfc_GEOMETRY column and sf::st_coordinates() failed on the result – the same defect as world_geometry("countries"), restored one step downstream. The output is cast back; row count and geometry are unchanged.
• simplify_geometry()’s keep argument now means roughly the same thing with and without rmapshaper. The sf::st_simplify() fallback was given a fixed dTolerance of (1 - keep) * 10000, i.e. metres whatever the coordinate system: 9 km on a projected frame, which barely simplified anything (79% of vertices kept at keep = 0.1), and 9000 degrees on a lon/lat frame, which is meaningless and only survivable because preserveTopology keeps a husk. The tolerance is now scaled to the object’s own extent, so the fallback behaves the same on either coordinate system and responds monotonically to keep. ?simplify_geometry says that only rmapshaper honours keep as a true proportion.
• gini()’s negative-value warning said Returning "NA", which reads as the two-character string rather than the missing value. It now renders as NA, matching the same correction already made elsewhere.
• The test suite no longer fails on R >= 4.6. Several tests set options(scipen = -10) to exercise the scientific-notation bugs fixed above, but R 4.6 clamps scipen to a minimum of -9 and warns (“invalid ‘scipen’ -10, used -9”), so an exact round-trip assertion failed and three warnings were raised – an ERROR under R CMD check on current R, even though the package code itself was correct. The tests now use -9 and compare against the value R actually stored.
• The wdj_overrides() soft-deprecation notice told the wrong people. It lived in the shared function body, so in an interactive session it fired for country_overrides() – the very replacement it recommends – and for every public function that takes the override table as a default argument (standardize_country(), convert_country(), attach_geometry(), check_country_match(), repair_country_names(), world_data() and the geometry backends). Callers were advised to stop using a function they had never written, and the advice was unactionable. The notice now fires only for a direct call to wdj_overrides(); the table itself is unchanged. The documented default is now country_overrides(), so ?attach_geometry and friends name a function the reader can actually look up.
• ?country_borders’s whole-world example runs again. It was wrapped in \dontrun{} on the grounds that the whole-world adjacency is expensive, but it takes about a quarter of a second from a cold session – only 2.7 times the region = "Europe" subset that already ran live. CRAN discourages \dontrun{} for code that can be executed, so it is now a guarded \donttest{} and is actually exercised by the check. That leaves six \dontrun{} topics, each genuinely unrunnable: a live World Bank fetch, an HTML widget, a DuckDB connection, a 60-frame GIF, and a call that deletes files.
• ?country_overrides’s advice for accented names in a non-UTF-8 locale did not work in that locale. It offered de-accenting with iconv(x, to = "ASCII//TRANSLIT") as an alternative to running under UTF-8, but //TRANSLIT is itself locale-dependent: under LC_CTYPE=C it returns NA, or replaces each accent with ? when given an explicit from = "UTF-8", so nothing resolves either way. The section now says de-accenting has to happen while still in a UTF-8 locale, and that the ASCII spellings the override table carries are what work everywhere.
• A bad options(countryatlas.workers) reached mclapply(). The option is advertised in this file, so a stray value is reachable: "abc", NA and Inf all became NA workers and surfaced as “missing value where TRUE/FALSE needed” from deep inside a parallel fetch, while c(2, 4) silently used the larger of the two. It is now checked, and the message names the option. Values below one are still clamped to one, as before – that path was never the problem.
• A bad options(countryatlas.cache_dir) did the same thing one layer down. The option is documented in ?clear_wdi_cache, and a stray value reached dir.exists()/dir.create(): NA, a number and TRUE each gave “invalid filename argument”, character(0) gave “argument is of length zero”, and a two-element vector gave “the condition has length > 1”. It is now checked and the message names the option. An empty string is still accepted and still degrades to session-only caching; NA_character_, which used to degrade silently, now errors like the other bad values.
• An empty cache directory failed on R 4.6 but not on R 4.4. dir.create("") warns and returns FALSE on R 4.4, so the fallback to session-only caching worked; on R 4.6 it errors with “zero-length ‘path’ argument”, which escaped the surrounding suppressWarnings() and propagated. An empty path is now recognised as “no disk cache” before the filesystem is touched, so the behaviour is the same on every R version.
• ?theil now says that a row with a missing group is dropped along with rows whose value is missing, so the decomposition’s total is computed over the grouped subset and can differ from the ungrouped theil(x). On the bundled snapshot that difference is entirely Puerto Rico, which has no region. theil() also gains numeric anchors on the bundled data, which gini() already had.
• Every sf-backed verb leaked sf’s internal chatter as message() conditions. The console was already clean, but silencing it by redirecting the message stream leaves the conditions themselves travelling to whatever handler the caller installed, so purrr::quietly(), testthat::expect_silent() or a plain withCallingHandlers() around attach_geometry(), neighbors(), country_borders(), locate_country() or morans_i() still saw three to nine “Spherical geometry (s2) switched off” / “assumes that they are planar” notices. They are now muffled as well as redirected. (The comment claiming these notices bypass R’s condition system was simply wrong – they are ordinary message()s, and only a few GDAL diagnostics need the stream redirect.)
• ?world_geometry called all four azimuthal projections “hemispheric” and said the far side comes back as empty geometries. That is true of "orthographic" alone; "azimuthal_equal_area", "north_polar" and "south_polar" are Lambert equal-area and image the whole globe, with the far side stretched around the rim and nothing dropped. The page now distinguishes them and points at region for a genuine polar view, and the error "ocean" raises in those projections no longer gives “it shows one hemisphere” as the reason.
• ?world_geometry now documents the Natural Earth features that have no ISO code and so come back with iso3c NA – Somaliland at every scale, plus the Indian Ocean Territories and Ashmore and Cartier Islands from "medium" on.
• ?theme_world_map said the theme is used by all the package’s plotting functions. bivariate_map() is the exception – it applies biscale::bi_theme() so the map matches biscale’s own legend, and its axis titles, panel grid and background differ as a result. The page now names the exception.
• ?simplify_geometry documented keep as a proportion “(0-1)”, but keep = 0 is rejected on both simplifier paths (it would leave nothing to draw). The range now reads “greater than 0 and at most 1”.
• complete_years(value = ) fabricated data in the columns it was not given. A numeric column left out of value was classified as a static attribute and carry-filled, so naming fewer columns invented more figures – and even method = "none", which exists to complete the grid and fill nothing, produced a carried-forward value for the missing year. Measure columns are now excluded from the attribute carry whether or not they are named; an unnamed one stays NA in the rows complete() adds. value = NULL behaves exactly as before.
• aggregate_regions() silently ignored weight for every fun except "weighted_mean", returning the unweighted figure – on European GDP per capita, 38,323 where the population-weighted answer is 29,896. Passing weight with any other fun now errors, mirroring the existing abort when fun = "weighted_mean" is given no weight.
• flow_map() dropped a flow whose endpoint it could not resolve without a word, and when nothing resolved it returned a bare world map with no arc layer at all. It now warns, naming the values it could not place and pointing at origin – feeding it iso3c codes while origin still defaults to "country.name" is the usual cause, and it drew a blank map.
• ?in_group now says that a value origin cannot resolve answers FALSE, indistinguishable from a country that is genuinely outside the group, and points at check_country_match() for telling the two apart.
• world_map() / globe_map() passed palette, title, legend and na_label to viridisLite and ggplot2 unchecked, so a mistake came back in their vocabulary rather than the package’s: palette = c("magma", "viridis") reached a bare switch() and reported “EXPR must be a length 1 vector”, and a numeric palette was accepted without a word. A length-2 title or legend was accepted too, and ggplot2 then drew both strings on top of each other. All four are now checked; world_query() already validated the two of them it takes (palette and title). na_label keeps its documented tolerance – there is one NA key, so the first element wins and a length-1 NA still means “leave the default formatter alone” – but it now says so instead of doing it silently. A number is still a perfectly good label.
• country_data() / world_data() resolved a conflict between latest and the shape arguments silently, and with opposite precedence depending on the year: a multi-year year overrode latest = TRUE, while a single year had latest = TRUE override panel = TRUE. The winner is unchanged – both were already the documented behaviour – but the call now warns, naming the argument being dropped, instead of returning a shape nobody asked for.
• locate_country(tolerance_km = ) was unvalidated, and a character value did not merely give an opaque message – it produced a wrong answer. R compares dkm <= tolerance_km as strings when the tolerance is character, and "2650" <= "a" is TRUE, so every unmatched point snapped to its nearest country however far away it was: a mid-Pacific point came back as Fiji, where the documented behaviour is that open ocean stays NA. Now checked, before the sf gate.
• Three more scalars the validation sweep had missed now name themselves instead of failing in a dependency’s vocabulary: correlate_indicators(min_n = ) (an NA gave “missing value where TRUE/FALSE needed”, a length-2 value “the condition has length > 1”), and dorling_map(k = ) / dorling_map(itermax = ) (which reported cartogram’s “all sizes are missing and/or non-positive” and an assertion naming its internal maxiter).
• interactive_map(engine = "ggsql") reported a missing ggsql >= 0.4.1 for a frame that simply had no geometry – sending the caller after a package that has not shipped in the R bindings at all, only to meet the real error afterwards. The sf check now runs ahead of the package gates. spin_globe() had the same inversion for its fill column: its scalars were moved ahead of the animation gate in an earlier pass, but the column check was not, so a mistyped column still asked for gifski.
• join_world() could not read a column of ISO codes without being told to. Its fallback detector tried each character column with origin = "country.name", which does not match most alpha-3 codes, so a column of them was rejected outright – and a column named iso3c was found by name but still read as country names, so join_world(tibble(iso3c = c("FRA", "JPN"))) warned that nothing matched and returned all NA. Detection now tries the code schemes as well and carries the one that worked through to the conversion, so iso3c, iso_a3, iso2c and an unrecognised name like code all resolve. An explicit origin still wins, and a column whose name implies a scheme is only read that way if the scheme actually resolves it.
• A bounding-box region on the polygon backend only filters vertices; it cannot clip a polygon, so a country crossing the edge keeps a truncated ring that geom_polygon() closes with a straight chord (France loses 202 of 605 vertices and the two ends sit 15 degrees apart). Nothing in the returned tibble showed it, and the vignette presented the box as clipping the shapes. It now warns and points at geometry = "sf", where the clip is a real sf::st_crop(); ?world_geometry and the vignette say which is which.
• attach_geometry() on a frame that already had geometry multiplied the rows instead of refusing. The join is by country, and a polygon frame holds one row per vertex, so re-attaching joins a country’s vertices against themselves: the bundled snapshot went from 99,338 rows to 310,977,360. The call declares relationship = "many-to-many" – correctly, since one country really does have many vertices – which switches off dplyr’s own guard against exactly this. It now errors, and says to pass the country-level table.
• share_of_world() on a grouped frame with no year column returned a share of the group, not of the world: sum() inside mutate() is per group, so a frame grouped by continent came back with a column that summed to 5 instead of 1, under a name and a help page that both say “world”. A panel was already safe by accident, because the function regroups by year and that replaces the caller’s groups. It now ignores the grouping in both cases, and says so where it would have mattered.
• The verbs that add a column now say when they replace one the caller already had. rank_countries() overwrote rank, percentile and z_score, and per_capita(), share_of_world(), growth_rate(), index_to(), lag_by_country() and diff_by_country() overwrote their target column, all in silence – a user’s own rank column simply vanished. It is a warning, not an error, because re-running a verb on its own output is legitimate. standardize_country() is deliberately exempt: add names the columns literally, so replacing an existing continent is what was asked for.
• geom_country_labels() on an sf map failed with rlang’s internal “Column long not found in .data”. The layer reads the polygon backend’s long/lat columns, and its own aes() was evaluated against the sf frame before the guard inside could run. It now errors with its own message and points at ggplot2::geom_sf_text(aes(label = iso3c)), which is documented on the help page too.
• geom_country_labels() put every country that crosses the antimeridian on the far side of the planet when the frame had no group column. group is what identifies a country’s separate pieces, and the label belongs on the largest; without it the fallback averaged the raw longitude range, so measured against the largest-piece centroid Fiji was 177.8 degrees out, New Zealand 169.6, and even the USA 96.6 – its Aleutian tail dragging the mid-range to the Gulf of Guinea. Averaging in wrapped coordinates cuts those to 0.2, 6.9 and 31.4. It remains an approximation, and the help page now says that placement is exact only while group is present.

Housekeeping

• gini() and theil() returned a silent NaN when the input contained an infinity. Inf is not NA, so it passed na.rm and (for Theil) the non-positive filter, then made the mean infinite and every share Inf/Inf. Every other verb propagates an infinity visibly – Inf in, Inf out, which the caller can see – but an inequality index has no such value to report, so both now warn and return NA, as they already did for a zero total weight. Infinite weights are caught too. NaN is still treated as NA and dropped.

• ?theil promised a tibble whenever groups is supplied, but every degenerate path – nothing left after na.rm, a zero total weight, an infinity – returns a single NA instead. The help page now says so.

• dorling_map(k = 0) passed validation and then failed inside cartogram with “all sizes are missing and/or non-positive”. check_number()’s bounds are inclusive, so lo = 0 admitted a value the next layer cannot use – the same hole the integer ceilings were added to close, at the other end of the range. It is now rejected with a message naming k; anything above zero still works.

• ?countryatlas gains an Options section. Two of the three options the package reads – countryatlas.workers and countryatlas.gdp_compat – were described only in this changelog, which is not reference documentation, so a reader of the help pages had no way to find them. (wdj_workers()’s own comment noted that the option was “advertised in NEWS”, which is how a bad value became reachable.) All three are now documented where they are looked for, and a test fails if a future option is read without being listed.

• The test suite runs in a quarter of the time (R CMD check’s test phase went from 389s to 92s). One block asked neighbors() for each country in turn and then again for each of that country’s neighbours, to confirm the reverse edge – and neighbors() recomputes the whole world’s sf::st_touches() adjacency on every call, so that was ~465 rebuilds and 292 seconds, four fifths of the suite. neighbors() is vectorised, so one call does the same work. The same properties are asserted (irreflexive, no repeated pair, every edge symmetric), and a failure now names the offending countries rather than only counting them.

• ?neighbors now says to pass a vector rather than loop, since that cost is invisible from the outside: every call rebuilds the whole world’s adjacency, so asking about one country costs the same as asking about all of them, and adding countries to a single call only adds the filtering. Measured here, one country takes about as long as 153 of them, which makes a loop over them roughly two orders of magnitude more work. The note names country_borders() as where the cost comes from. A test pins the fact the advice rests on – one country_borders() call per neighbors() call, whatever the length of x.

• ?attach_geometry claimed that Gibraltar, Hong Kong, Macao, Tuvalu and the British Virgin Islands have geometry in “no backend at any scale”. Only Gibraltar does not: the other four are carried by the sf backend at scale = "medium", which is the very fix the preceding sentence recommends for microstates. It was the no-tile list from ?world_tiles – coincidentally the same five names – pasted into a paragraph about geometry. Corrected, and the section’s coverage counts (215 snapshot countries; 210, 169 and 214 carried) are now pinned by a test, so an upstream Natural Earth change surfaces as a failure rather than as silently wrong advice.

• as_ggsql_source(format = "parquet") wrote into the working directory when no path was given: the default was the bare relative path "<name>.parquet". CRAN policy is that a package writes nowhere but the session’s temporary directory unless the caller says otherwise. The default is now a file of that name under tempdir(), and since the function returns the path the workflow is unchanged; an explicit path still writes exactly where you point it. spin_globe() already defaulted to tempfile(), and no other export writes at all.

• ?country_meta and ?world_snapshot now say that their country columns disagree, and why. country_meta carries countrycode’s English names and world_snapshot the World Bank’s, so 38 of the 215 shared countries are labelled differently (“South Korea” against “Korea, Rep.”). Each table is faithful to its own source and neither is wrong, but joining the two on iso3c leaves you holding two country columns with nothing to explain the difference – the very reconciliation country_join() advertises, using the same example. The count is pinned by a test, along with the referential consistency of all five code columns against country_meta.

• ?world_data and ?country_data now say where the country label comes from, because the same call produces two different spellings: a successful fetch carries the World Bank’s names (“Korea, Rep.”), while the country spine used when the fetch returns nothing carries the countrycode names (“South Korea”) – as does every other function in the package. Nothing can reconcile that offline, since the World Bank spelling only exists in the response, so both pages now point at iso3c as the stable key and at convert_country(iso3c, to = "country") for one consistent set of labels.

• utils::globalVariables() declared 29 names where 7 are needed. Emptying it and reading what R CMD check actually reports showed the rest were covered by the .data$x idiom the code uses throughout, which needs no declaration at all; three of them (subregion, NY.GDP.PCAP.KD, gdp_per_capita_2015) never appeared as bare symbols anywhere, only in a comment or as string literals. A stale entry is worse than clutter: it silences the “no visible binding” NOTE for a new bare use of the same name, which is the warning that would otherwise catch a typo. A test now fails if a declared name is not a real bare symbol in R/.

• Six exported functions failed when the package was loaded but not attached – countryatlas::dissolve_country(), distance_between(), country_groups(), in_group(), tile_map() and world_geometry(region = <group name>) all died with “object ‘historical_codes’ not found” or similar. They referred to the bundled datasets by bare name, and a bare name resolves only while the package is on the search path: under countryatlas::fn() in a script with no library() call, the lazy-data objects are not reachable. They are now countryatlas::-qualified. Every test in the suite attaches the package, so nothing caught this; a static check now fails if a bare reference reappears.

• per_capita() failed when the caller’s frame already had a column named .wdj_pop, the internal name used for the fetched population. The join suffixed both sides to .wdj_pop.x / .wdj_pop.y, so the column the division reads came back NULL and base R reported “replacement has 0 rows, data has 2”. Any pre-existing column of that name is now dropped before the join. Only the branch that fetches population was affected – passing pop explicitly never touched it.

• Two verbs leaked someone else’s message on an empty frame. facet_map() gave ggplot2’s “Faceting variables must have at least one value”, which names neither the argument nor the package; it now says the frame has no rows to facet, and notes that the other map verbs draw an empty panel instead. geom_country_labels() ran the centroid summary over nothing, where range() warns twice and dplyr adds a deprecation note on top – it now returns early, silent as it is on a full frame. Every other plotting verb already handled a zero-row frame cleanly, either drawing an empty panel or naming the reason it cannot.

• ?attach_geometry now says that geometry is attached once per row, not once per country. A panel wants exactly that – one row per country-year, each carrying the shape – but a frame that repeats a country by accident draws it more than once, and only the last one painted is visible. dplyr’s own many-to-many warning is suppressed by the relationship the join declares, so nothing signals it.

• ?country_data’s example quoted a retired World Bank indicator. The bank replaced the EN.ATM.CO2E.* carbon series with the AR5 greenhouse-gas series, and the bundled common_indicators table had already been updated, but the example still asked for EN.ATM.CO2E.KT – so anyone copying it got a warning and an all-NA column. It now uses EN.GHG.CO2.MT.CE.AR5, which returns data. R CMD check reports examples “OK” without failing on the warning, so nothing surfaced this; a test now checks every indicator code quoted in R/, man/ or the vignettes against the bundled table.

• country_groups_tbl was out of date by two years in four places, while carrying an as_of stamp of 2026-06-01 that claimed otherwise. Sweden was missing from NATO (acceded 7 March 2024; Finland had been added, so the table had been maintained to 2023 and no further), Angola was still in OPEC (left 1 January 2024), BRICS still held only its original five (Egypt, Ethiopia, Iran and the UAE joined in January 2024, Indonesia in January 2025), and The Gambia was missing from the Commonwealth (rejoined 2018). Corrected, so the counts are now NATO 32, OPEC 12, BRICS 10 and Commonwealth 56. Saudi Arabia is deliberately still absent from BRICS: it was invited in the 2024 round but has never confirmed accession. in_group("Sweden", "NATO") returned FALSE before this.

• options(countryatlas.cache_dir = ) was ignored once a cached fetch had happened. The memoised fetcher was built on first use and kept for the rest of the session, so relocating the cache afterwards silently kept writing to the original directory – and ?clear_wdi_cache offers that option as the way to relocate the cache without saying it has to be set first. It only ever took effect because clear_wdi_cache() happened to reset the state. The fetcher is now rebuilt when the directory changes, and the “cannot write to the cache directory” notice is once per directory rather than once per session, so a second unwritable location is not swallowed.

• A corrupt cache entry was reported as a World Bank outage. An interrupted write leaves a truncated or empty .rds, and readRDS()’s “unknown input format” surfaced under “Could not fetch indicator … from the World Bank API”, sending the caller off to debug a connection that was fine – the same misattribution already fixed for an unwritable cache directory, now fixed on the read side. The warning names the cache and gives the recovery command, which matters because the bad entry persists: every later call degrades to the country spine until clear_wdi_cache(disk = TRUE) is run. A genuine network failure still blames the network.

• When the cache directory was unwritable, caching stopped working altogether instead of falling back to the session. The in-memory memo that stands in for the disk cache lives in one process, but multiple indicators are fetched with parallel::mclapply(), so each worker warmed a memo and then exited with it: every call re-fetched every indicator, hitting the World Bank API again and again with nothing to show for it. Fetching is now serial when the memo is memory-only – the repeated round-trips cost far more than the one-shot parallel speedup – and unchanged when the disk cache is available, since a disk memo is shared by every worker.

• A failed indicator was dropped from the result without a word, whenever more than one indicator was requested. fetch_one_safe() degrades gracefully and warns – “Could not fetch indicator … from the World Bank API”, or the corrupt-cache variant – but it runs inside parallel::mclapply(), which brings back a worker’s value and discards the conditions it signalled. Since having several indicators is exactly what makes the fetch fork, and parallel = TRUE is the default, the common case was the silent one: a column simply missing from the table with no explanation. A single indicator, which never forks, warned correctly – which is why this went unnoticed. Conditions are now carried back and re-signalled in the calling process, on the serial path too so both report identically, and one problem is reported once even if two entries name the same series.

• Country lookups silently returned NA in Turkish, Azeri and Crimean Tatar locales. toupper() and tolower() follow LC_CTYPE, and in those locales i and I are not a case pair: toupper("idn") returns a dotted capital I, not "IDN", and tolower("ISO3C") returns a dotless i. Five places folded an ASCII identifier that way and then compared it against plain ASCII, so every ISO code containing an i (IDN, IND, IRL, IRN, ISL, ISR, ITA, BIH, CIV, FIN, …) failed to resolve. The failures were quiet and the blast radius uneven: world_geometry(region = c("ind", "chn")) returned Ivory Coast, Indonesia, Isle of Man and India – one unfoldable element made the whole vector fall through to name matching, which then matched on substrings; dissolve_country("SOUTH VIETNAM") stopped finding its alias; and join_world() on a frame with ISO3C and geo columns picked geo, joining on the wrong column entirely. Identifier folding is now done with an explicit ASCII table (ascii_upper() / ascii_lower()) and does not consult the locale.

  This fixes countryatlas’s own folding, which covers every path keyed on an ISO code, a column name or an alias. It cannot fix matching on a country name: that goes through countrycode, whose regexes are themselves locale-sensitive (countrycode("Ireland", "country.name", "iso3c") is NA under tr_TR). So a user with LC_COLLATE=tr_TR still sees name-keyed gaps – notably the polygon backend, which labels its geometry by joining region names to codes, so world_geometry(region = "IND") comes back with 21 rows of Siachen Glacier instead of India. Working around that would mean forcing the C locale around every countrycode call, which risks mangling accented names for everyone else; it is left for upstream.

• The two-core cap that CRAN policy requires of a check was applied only when _R_CHECK_LIMIT_CORES_ held the exact string "TRUE". R CMD check --as-cran does set it to that – but only when it is not already set, so the value that actually arrives is whatever the check flavour or CI exported, and R’s own parser for these variables reads "true", "True", "T", "1", "yes", "Yes" and "YES" as true as well. Under any of those spellings the cap did not apply and a multi-indicator fetch forked detectCores() - 1 workers in the middle of a check. The test is now inverted: a value that is set and does not explicitly parse as false means “limit”, which also covers "warn". An explicit "false", "F", "0" or "no" is still honoured as a deliberate opt-out.

• ?world_data’s example called world_data(2020) unconditionally, and the default geometry = "polygon" backend comes from the suggested maps package. R CMD check runs \donttest{} blocks, so on a check flavour configured without suggested packages – CRAN runs one – that example failed with “The package "maps" is required for the polygon geometry backend”, taking the whole examples step down with it. Writing R Extensions requires code that uses a suggested package to be conditional, examples included; the call is now guarded with requireNamespace("maps"). The second call in the block passes geometry = "none" and needs nothing beyond the hard dependencies, so it is left to run unconditionally.

• The beyond-the-choropleth vignette failed to build wherever rnaturalearth was absent. Its chunk guard was has_sf <- requireNamespace("sf"), but the sf geometry backend gates on three packages – sf, rnaturalearth and rnaturalearthdata – so on a machine with sf but without the Natural Earth data the guarded chunk evaluated to TRUE, ran, and stopped R CMD build with “The packages "rnaturalearth" and "rnaturalearthdata" are required for the sf geometry backend”. The other two vignettes already tested for rnaturalearth; all three now test for the same trio the code itself gates on. Three tests had the same incomplete guard and errored rather than skipping in that configuration; they now share a skip_if_no_sf_geometry() helper.

• New hex logo, drawn by the package itself (data-raw/hex_logo.R): an orthographic globe (globe_map()’s projection) carrying a viridis choropleth of world_snapshot GDP per capita on Natural Earth geometry joined by attach_geometry(), with spike_map()-style population spikes rising off the horizon and the binned-legend swatches under the wordmark.

• The gdp_per_capita_2015 compatibility alias (a one-cycle deprecation shim from 1.0.0) is now opt-in: set options(countryatlas.gdp_compat = TRUE) to restore it. The default is FALSE, so world_data() no longer emits a duplicate column.

• world_snapshot refreshed to year 2024 (was 2022) and rebuilt with the latest WDI data and curated overrides.

• country_groups_tbl membership date bumped to 2026-06-01 (was 2024-01-01).

• ?world_snapshot was out of sync with the rebuilt data (missing the “Snapshot year: 2024” note); regenerated.

• Fixed a stray orphaned code fence at the end of the countryatlas and ggsql vignette that broke its markdown structure.

• .Rbuildignore now excludes the session-local .claude/ directory, which git ignores but R CMD build does not, so it was shipping in the tarball and tripping R CMD check’s “hidden files and directories” NOTE.

• Comments in R/overrides.R are ASCII-only, so no source file carries non-ASCII characters outside a deliberate \U escape.

• ?world_snapshot no longer splits a code span across two source lines, which had left the checked-in .Rd disagreeing with what roxygen2 regenerates.

• beta_convergence() failed with a bare "subscript out of bounds" when the initial levels had no spread across countries. A constant predictor makes lm() return an NA coefficient, which summary() then drops entirely, so the lookup for it fell off the end. It now says that the initial levels have no spread and why that matters.

• The forking path is now tested. fetch_wdi(parallel = TRUE) is the default for a multi-indicator request, so wdj_lapply()’s mclapply branch runs on one of the package’s busiest code paths, yet no test reached it – every other test used a single indicator or passed parallel = FALSE. Confirmed: a parallel fetch is identical to the serial one, forking preserves order, ... reaches the workers, wdj_workers() honours options(countryatlas.workers) and CRAN’s two-core limit, and an error inside a fork is surfaced rather than left as a try-error for downstream code to trip over.

• The World Bank fetch and assembly path is now tested offline. It needs the network, so it had no coverage at all despite holding the least obvious logic in the package: fetch_wdi()’s multi-indicator reduce-merge (shared keys are coalesced rather than suffixed, and values stay aligned per country-year), its degradation when one indicator of several fails, country_data(latest = TRUE)’s “most recent non-NA” collapse (which opens the window at 1960 and skips a missing latest year), the panel key, the duplicate-key case where two iso2c codes map to one iso3c, and that cache = TRUE really short-circuits a repeated fetch. No defects were found; the tests pin the behaviour.

• spin_globe() validates n_frames / fps / width / height / lat before gating on gifski / magick, matching how globe_map() orders the two: a bad argument is the caller’s bug and the message should not depend on which optional packages happen to be installed.

• tile_map() phrased the missing-iso3c error differently from the three other verbs performing the identical check; all four now read the same.

• Half the examples that were marked \dontrun{} now actually run: it covered 14 of the 55 documented topics and now covers 7. Nine verbs gained executable examples – locate_country(), country_borders(), neighbors(), morans_i(), simplify_geometry(), bivariate_map(), cartogram_map(), dorling_map() and the polygon-backend globe_map() – having been unrunnable only because their examples made a live World Bank fetch. Driven by the bundled world_snapshot instead, and guarded with requireNamespace(), they are \donttest{} examples that execute in under half a second each. The safe form of clear_wdi_cache() is now a live example too. \dontrun{} remains only where the code genuinely cannot run in a check (a network fetch, an HTML widget, a written GIF, ggsql >= 0.4.1, or deleting files).

• The tests guarding this release’s fixes were verified by mutation: each fix was reverted in a scratch copy and the suite had to fail. 35 mutations, all now detected – but four were not at first, and each pointed at a real gap:

  • is.na() is TRUE for NaN as well, so the checks on theil()’s zero-weight result and its shares at perfect equality passed whether the value was the fixed NA or the NaN the bug produced. Both now assert the exact value.
  • Nothing verified that quantile breaks are computed on one value per country rather than per polygon vertex – the fix that forced this major version bump. Removing the de-duplication, or flipping the flag that controls it on either world_map() or globe_map()’s polygon backend, broke no test. There are now checks on the helper, on both call sites, and on the property that matters: roughly equal numbers of countries per colour.
  • interactive_map(tooltip = ) was unprotected: the existing tests assert the returned object’s class, which passes whether the argument is honoured or silently dropped – exactly the pre-2.0.0 bug. Both the "ggiraph" and "leaflet" engines are now checked on the column they are actually handed.

• Degenerate-but-valid input is now covered by tests: perfect equality (gini() / theil() return 0, not NaN), zero-variance columns (NaN z-scores and NA correlations rather than errors), duplicate (iso3c, year) panel rows, poles and antipodal great circles, collinear rings, all-NA and all-zero fill columns, and single-country frames.

• ?distance_between and ?country_meta now state which countries have no bundled centroid. country_meta is assembled from countrycode::codelist, which has no Kosovo row, so distance_between("Kosovo", "Serbia") is NA even though neighbors("Kosovo") and country_borders() know about it – and ten small or dependent territories have a row but no centroid. The @return already said NA was possible; it did not say which countries, and the asymmetry with the geometry backends was surprising.

• The same double-counting affected three more places, all gated on the presence of a group column – which polygon frames have and sf frames do not:

  • correlate_indicators() reported n and r over geometry rows rather than countries. That column exists precisely so a correlation computed on a few countries cannot masquerade as a world fact, so an inflated n defeated the point.
  • audit_coverage() reported the wrong country count and a wrong NA rate for every indicator.
  • bubble_map(backend = "sf") drew two bubbles for a divided country, where the polygon path already guaranteed one per country. All three now reduce whenever an iso3c column is present.

• Quantile and jenks breaks on the sf backend double-counted divided countries. The de-duplication added in this release skipped the sf path on the assumption that Natural Earth is one row per country, but it is not: Cyprus occupies two rows sharing one iso3c at 110m, as do Cyprus and India at 50m. Breaking on the raw column shifted the cut points enough to move real countries into the wrong bin – Saudi Arabia and Libya changed colour in the bundled snapshot at n_bins = 5. Both backends now de-duplicate on the key, so “one value per country” holds exactly rather than nearly.

• convert_country(x, to = "calling_code") returned alpha-3 country codes instead of telephone calling codes – "FRA" where 33 was meant. The shortcut was mapped to countrycode’s genc3c column, which is an ISO-style three-letter code, not a dialling prefix; it now maps to the telephone column, so France gives 33, the USA 1 and Japan 81. A source comment claimed the limitation was “documented as best-effort”, but ?convert_country never mentioned calling_code at all; the shortcut is now listed there.

• The Description field – the text CRAN renders on the package page – advertised nine map idioms while the package ships eleven: spike_map() and facet_map() (small multiples) were missing from the vocabulary list, and the analysis-helper examples predated this release’s inequality and convergence statistics. Every idiom it names now corresponds to an exported verb.

• ?countryatlas listed morans_i() under “Core data assembly” while _pkgdown.yml listed it under “Analysis helpers”, so the two navigational indexes described the same function as two different kinds of thing. It is a spatial statistic, so it now sits with gini(), theil() and the convergence measures on both surfaces.

• The @seealso cross-references are reciprocal. All three the package had pointed one way only: a reader of ?gini was sent to theil() but a reader of ?theil was sent nowhere, and likewise for beta_convergence()/sigma_convergence() and dissolve_country()/check_country_match(). ?theil did not link to gini() at all – it named Gini in prose without a cross-reference. The four related diagnostics (check_country_match(), repair_country_names(), dissolve_country(), plus historical_codes) now all reference each other.

• Every page taking a projection argument now says where the valid values are. Only world_map() enumerated the 13 projections and only world_geometry() pointed at it; the other eight entries said no more than “Projection.” (bivariate_map()) or “Projection options for the sf backend”, leaving a reader with nothing to go on. Relatedly, the three pages that document projection and recenter together described only the projection, so recenter’s meaning – a central meridian – was missing from world_data(), join_world() and attach_geometry().

• The rnaturalearthhires requirement is now documented on every page that takes a scale argument, not just ?world_geometry. Seven topics – world_data(), join_world(), attach_geometry(), locate_country(), country_borders(), neighbors() and morans_i() – described scale without mentioning that "large" is unobtainable from CRAN, so a reader of any of those pages met the gate with no warning.

• scale = "large" was offered as a plain option but needs the rnaturalearthhires package, which is not on CRAN and is not in Suggests. Left ungated, rnaturalearth responded by trying to install it into the user’s library from a non-CRAN repository and then failing obscurely. It is now gated with a message naming the package and the repository to get it from, and pointing at scale = "medium" (50m) as the option that needs nothing extra. ?world_geometry documents the requirement, and the sf & projections vignette no longer demonstrates the scale most readers cannot run.

• ?country_borders recommended a graph recipe that produced nonsense. igraph::graph_from_data_frame() treats the first two columns as the edge endpoints, and country_borders() returns iso3c_a, country_a, iso3c_b, country_b – so columns 1 and 2 are both endpoint A, and passing the whole tibble built edges from each country’s code to its own name (56 vertices instead of 37 for Europe, every French edge running FRA to "France"). The documented call now passes only the two code columns.

• ?neighbors and ?country_borders now warn that igraph also exports a neighbors() – taking a graph and a vertex rather than country names – so whichever package is attached later wins. ?country_borders recommends igraph for turning the adjacency into a graph, which walks users straight into the clash, so both pages now say to qualify the call as countryatlas::neighbors(). It is the only collision between this package’s exports and any of dplyr, ggplot2, tidyr, tibble, sf, maps, WDI, countrycode, scales, leaflet, plotly, igraph, raster, terra, purrr, stringr, forcats, readr or the base packages.

• ?country_overrides now documents why every name in the override table is plain ASCII: ASCII spellings match in any locale, whereas accented spellings rely on countrycode’s own matching and resolve to NA under a non-UTF-8 locale (LC_CTYPE=C). The note points at iconv(x, to = "ASCII//TRANSLIT") for input that may carry accents.

• The test suite is now green under CRAN’s noSuggests configuration (_R_CHECK_DEPENDS_ONLY_=true), which runs with every optional package absent. Four tests called globe_map(backend = "polygon") or morans_i() without guarding on mapproj / sf, so they errored on the dependency gate instead of exercising what they were written to check.

• The quantile/jenks binning that world_map() and both globe_map() backends perform lived as three near-copies kept in step by hand; it is now one internal helper. Verified behaviour-preserving by comparing the rendered fill of every style x n_bins x backend combination before and after (50 fingerprints, ~2.6M values, all identical).

• Documented \value claims are now asserted as executable contracts, so Rd prose cannot drift from the code in silence. The 36 exports whose \value makes a specific structural promise – named columns, a single row, an attached "model" object, a length matching the input – are covered; the rest return a ggplot, a layer or a widget and are checked by their own tests. Every claim audited was already accurate; the tests keep it that way.

• The four exports that had no test call site at all – wdi_search(), clear_wdi_cache(), animate_world() and cartogram_map() – are covered, including cartogram_map(type = "contiguous") (the default type, previously never exercised) and wdi_search()’s zero-match and single-match paths.

• The README’s “optional features at a glance” table is corrected against what the code actually gates on: spin_globe() was listed as needing only maps + mapproj when it also hard-requires gifski or magick; the ggsql row overstated the requirements of as_ggsql_source() (which never needs ggsql) and understated the version interactive_map(engine = "ggsql") needs; and locate_country(), flow_map() and bubble_map() were missing.

• The countryatlas and ggsql vignette said DRAW spatial “was added in 0.4.1” as plain fact; it now says that version is newer than what CRAN ships, which is why the query-executing chunks are shown but not evaluated.

• The package-level overview (?countryatlas) was missing simplify_geometry() and clear_wdi_cache() from its section list, though _pkgdown.yml had both.

• The set of ISO codes the package treats as countries was computed in two places (name matching and the World Bank aggregate filter); it now comes from one internal helper, which region resolution uses as well, so the three callers cannot drift apart.

• New offline test suites pin the things a structural test cannot: closed-form anchors for the hand-rolled numerical kernels (haversine distance, spherical polygon area, great-circle interpolation, Gini/Theil, sigma and beta convergence, Moran’s I against an independently built weights matrix) and internal-consistency checks on every bundled dataset (no duplicate or unknown iso3c, coordinates in range, one country per world_tiles cell, historical_codes in step with its alias table, and country_meta centroids still agreeing with polygon_centroids()).

• README and vignettes now demonstrate every exported function: wdi_search(), country_codes(), complete_years(), growth_rate() / index_to(), repair_country_names(), country_join_all(), locate_country() and facet_map() gained worked examples, and the vignettes prefer country_overrides() over the soft-deprecated wdj_overrides(). The README’s rendered output and figures were stale (pre-dating the quantile-breaks fix and the gdp_per_capita_2015 opt-in) and have been re-rendered from the 2.0.0 code.

• geofacet is dropped from Suggests: no code ever used it, and ?tile_map / the README claimed a geofacet-backed small-multiples feature that did not exist. Facet a tile_map() like any other ggplot, or use facet_map() for choropleth small multiples.

• The README’s figures are shipped in the tarball again, so the images on the CRAN package page resolve. .Rbuildignore excluded the generated .pngs but not the (much larger) .gif, which left six of the seven images broken.

• ?world_snapshot’s @format said “two elements” while listing three, and advertised an sf element that is NULL in the released package. It now documents what actually ships and points at attach_geometry() for geometry.

• ?attach_geometry documents which countries each geometry backend actually carries. Rows with no matching geometry are dropped silently, and the sf backend’s scale changes which countries exist rather than only how detailed they are: of the 215 countries in world_snapshot, the default scale = "small" (110m) maps 169 and scale = "medium" maps 214. Five territories (Gibraltar, Hong Kong, Macao, Tuvalu, the British Virgin Islands) are in no backend at any scale.

• ?world_tiles and ?tile_map said “one square per country” without saying how many. The grid is the 239 country_meta rows that have a bundled centroid, so the 10 without one have no tile and tile_map() silently drops data rows keyed on them. Both are now documented, and a test pins the grid to that definition.

countryatlas 1.0.0

A single, comprehensive release that takes the package from a one-function proof of concept to a complete toolkit for joining world data to maps. The spirit is unchanged (ISO codes as the universal join key, one call to a map-ready table) but pushed to its full potential.

Breaking-ish changes

• world_data() is generalised but backward-compatible: world_data(2020) still returns the classic polygon-backed, GDP-per-capita tibble. The only visible change is the column name gdp_per_capita_2015 → gdp_per_capita. A one-cycle deprecation shim keeps gdp_per_capita_2015 available as an alias (toggle with options(countryatlas.gdp_compat = FALSE)).
• The 16 regions the previous version silently dropped (Kosovo, Micronesia, the Virgin Islands, Saint Martin, Bonaire/Saba/Sint Eustatius, the Canary Islands, Madeira/Azores, …) are now matched via [wdj_overrides()] instead of deleted, so they appear on maps. Diffs of map output will show increased coverage.

New: core data assembly

• world_data() gains indicator (one or many WDI codes; named vectors drive clean column names), multi-year panels, an sf backend (geometry = "sf"), region subsetting, latest, projections and caching.
• country_data(): the lightweight, one-row-per-country analysis table.
• world_geometry(): projected, region-subset geometry (countries, centroids, coastline, borders, graticule, ocean).

New: the join engine (exposed for your data)

• standardize_country(), join_world(), attach_geometry(), country_join().

New: diagnostics

• check_country_match(), wdj_overrides(), audit_coverage(): never lose a country silently.

New: reference data & translation

• convert_country() (flags, currency, tld, research codes), country_codes(), country_groups() / in_group(), wdi_search().
• Bundled datasets: world_snapshot, country_meta, common_indicators, country_groups_tbl, world_tiles.

New: analysis helpers

• per_capita(), aggregate_regions(), rank_countries(), complete_years().

New: visualization

• world_map() (continuous / binned / quantile / jenks / categorical), bubble_map(), bivariate_map(), cartogram_map(), tile_map(), flow_map(), animate_world(), interactive_map(), geom_country_labels(), theme_world_map().

Performance & offline

• WDI fetches are memoised with an optional on-disk cache; multiple indicators are fetched in parallel (parallel::mclapply) where supported. See clear_wdi_cache().
• The bundled world_snapshot lets every example, test and vignette run offline and deterministically.

Engineering

• Namespace hygiene (targeted @importFrom instead of blanket @import).
• Input validation with friendly cli / rlang errors.
• A testthat (3e) suite; network calls are skipped offline and on CRAN.
• Algebraic invariants alongside the closed-form anchors. An anchor pins one value, which a wrong divisor can still satisfy for a single input; these pin the relationships instead – shares summing to one within each year, per_capita() multiplying back to the original value, a deflator equal to the base year and a PPP factor of one being identities, compounding growth_rate() walking the series back out, index_to() landing exactly on to at the base year, ranks forming a permutation, and gini()/theil() reaching their analytic bounds and staying scale-free.
• The test run is kept out of the checking account’s file space, so R CMD check cannot report “new files in some other directories” – the NOTE CRAN raised against 1.0.0. The redirect now sets R_USER_CACHE_DIR as well as XDG_CACHE_HOME: tools::R_user_dir(pkg, "cache") reads the former first, so setting only the fallback left the redirect at the mercy of the machine, and wherever R_USER_CACHE_DIR happened to be set a suggested package’s cache still escaped to the real user directory.
• Vignettes and a pkgdown site.
• Refreshed CI: R-CMD-check, test-coverage and pkgdown workflows.
• Heavy spatial dependencies (sf, rnaturalearth, cartogram, biscale, geofacet, gganimate, leaflet, …) are all in Suggests and gated by rlang::check_installed(), so the base install stays light.

Group memberships in country_groups_tbl are point-in-time as of 2024-01-01.

countryatlas 0.1.0

• Initial experimental release with a single world_data(year) function.