data: add the Americas — 55 territories, 95 datasets, 16,195 features

[andresgmg]

Country outlines from Natural Earth 10m (public domain); first-level and municipal divisions from geoBoundaries gbOpen under permissive licences only; Chile from IDE Chile.

162 MB of data, largest file 14.2 MB, everything inside the CDN ceiling. The catalog now has 55 country pages grouped by UN M49 sub-region.

The pipeline from #11 was reused unchanged in shape — but four things had to be fixed once it met 57 countries instead of one.

1. Silent data loss

The municipal split filtered out features whose parent could not be resolved. Switching from a centroid test to largest-overlap cut Argentina's unmatched count from 13 to 8 — but the remainder are not a bug in the join:

geoBoundaries publishes 23 ADM1 units for Argentina. The country has 24. The Autonomous City of Buenos Aires is missing, so its comunas overlap no province at all.

Those now land in an unassigned part instead of vanishing, and the split asserts it holds every input feature. 12 features across the continent are in this state.

2. Resolution measured against the wrong file

The size budget was applied to the whole-country file, which is optional. Brazil was on its way to a 400 m tolerance because its 5,570 municipalities do not fit in one file — while every one of its 28 per-state files fits at 100 m with room to spare.

Split levels now simplify at the standard tolerance, the budget is checked per part, and the combined file is dropped when it does not fit.

3. Licence claims that were not true

The catalog labelled every country CC-BY-4.0, including territories whose only dataset is a public-domain Natural Earth outline.

Licence varies per level, not per country. Brazil alone: outline public domain, states CC BY 2.5, municipalities CC BY 3.0 IGO. Each dataset now records its own licence, upstream provider and vintage; the country-level value is derived, becoming mixed with an explicit list where levels differ. validate_data.py checks every dataset, not just the country.

Resulting distribution: 33 territories pure public domain, 22 mixed.

4. Toolchain

mapshaper is pinned in package.json and installed locally. npx -y re-resolves the package on each of the several hundred invocations a continent needs, and concurrent runs collide on the npm cache — that was the intermittent Brazil failure. make_previews.mjs invokes the entry script with node directly, since execFileSync cannot spawn the .cmd shim on Windows.

Upstream quirks — documented, not smoothed over

• admUnitCount disagrees with the files geoBoundaries serves. Suriname reports 62 ADM2 units and ships 58; Honduras 299 and ships 298. I verified source and built counts match exactly in every case — the loss is upstream metadata, not this pipeline.
• Ecuador and Peru each ship an ADM1 whose shapeISO contains an asterisk (EC-D*, PE-LMA*), illegal in a Windows filename, which broke the split with ENOENT halfway through writing. Split keys are now sanitised.
• Jamaica (827 ADM2 against 14 parishes) and Saint Lucia (547 against 10 quarters) look mis-tiered and were left out rather than published wrong.
• Brazil's ADM1 is a 2006 vintage. Shown on its catalog page rather than hidden.

Coverage gaps

Fifteen countries have no first-level divisions here — ODbL for Colombia, Costa Rica, Cuba, Guatemala, Guyana, Honduras, Haiti, Nicaragua, Panama, Suriname, Trinidad and Tobago, Uruguay and Saint Vincent and the Grenadines; CC-BY-SA for Grenada and Greenland. Several still have a permissive municipal tier, so they appear with an ADM2 and no ADM1 above it.

Bonaire/Sint Eustatius and Saba and Bouvet Island ship nothing at all.

This is the cost of the permissive-only rule and it is named explicitly on the roadmap. Gaps close by finding a national SDI or HDX release, not by relaxing the rule.

Verification

• mkdocs build --strict clean; validate_data.py 0 errors, 0 warnings across all 55
• Split parts sum exactly to their level total for every country
• Source vs built feature counts identical for every dataset checked
• Maps verified in-browser on Brazil: 4 / 30 / 5,573 SVG paths for 1 / 27 / 5,570 units
• site/ is 18.3 MB with no file over 5 MB — the CI leak check was changed from a total-size cap to per-file, since the site legitimately grows with each continent while a leaked full-resolution dataset is always far larger than any preview

🤖 Generated with Claude Code