Replace the two cell-polity crosswalks with the polycell, and move the area denominators onto it by eduaguilera · Pull Request #619 · eduaguilera/whep

eduaguilera · 2026-08-07T04:07:24Z

Replaces the two incompatible cell↔polity crosswalks with one unit, and moves every area denominator on the extent side of #458 onto it.

Draft — the identity half is not here, and two known gaps are called out at the bottom. Please read those before the diff.

The unit

The polycell is the intersection of a 0.5° cell with a polity over that polity's validity interval, polycell_id = f(cell_id, polity_code). It carries territory decomposed into three separately addressable categories:

polity_area_ha = land_area_ha + inland_water_ha + ice_area_ha

Per-hectare biogeochemical rates apply to land_area_ha; polity_area_ha is what reconciles with official national areas. Inland water and ice are territory that receives real fluxes, not rounding error — deposition to freshwater is a live eutrophication term.

What was wrong

A per-hectare rate multiplied by whole-cell area over-counts by 11.0% — global terrestrial area is 12.9931 Gha against 14.3195 Gha of whole cells. polity_frac is a valid partition of the land; it was simply multiplied by the wrong base.
Two crosswalks were live in one package. The carbon path read a centroid grid where cell_area_frac defaults to 1, so a whole border cell went to one polity; the nitrogen path split fractionally. No test detected it, because each path was internally self-consistent.
Border splits were quantised to 1/36 of a cell (subcells = 6L), not the 1/144 often assumed.
Two live end_year fenceposts read the boundary inclusively, so on a succession year both epochs activated and the first-sorting row took every cell — and the dissolved predecessor sorts first in 315 of 315 cases, so the successor got no row and cells outside its smaller extent were dropped. That lost mass, not just labels.

Measured movement, per consumer

Each consumer moved on its own commit, against a characterisation baseline pinned first, so the movement is attributable rather than aggregate.

consumer	movement
deposition, split key	mass conserved exactly — new − old = 0 t; rate `identical()` over 63,065 cells
deposition, categories	scope selectable, defaults to whole territory → ledger bit-identical
synthetic	per-polity totals moved by exactly 0; weighted share movement 1.99%
carbon	+1.32% on masses (stock, C input, SON)
spatialize	zero — enforcement only
whole ledger	−0.504% on N input, of which 95.6% is the identity gap, not geometry

Carbon is the one to look at. Per-hectare densities are invariant under a pure re-split, so a density check reports ~0 whether or not the migration is correct — and here the global mean stock density falls while the mass rises, so a density check would have reported the wrong sign. All carbon movement is therefore measured on sum(density × area_ha).

Two decisions worth disagreeing with, if you do

The ledger keeps the whole territory. Restricting deposition to the land share would remove 0.89 Tg N (−1.44%), concentrated on lake and glacier states. We measured it, then didn't take it: N deposited on a lake or glacier still drives indirect N₂O and still reaches the eutrophication pathway, so the terrestrial-only scope is wrong for those terms. Scope is an explicit argument defaulting to "territory", recorded in an output column.

The table keys on polity_code, not the bucket. polity_area_code folds 544 codes into 201, with 206 holding Sudan and South Sudan — the fold #591 caught corrupting published series. Consumers convert at their own boundary, so the fold happens visibly at a join rather than irrecoverably in the grid. This is why #579's grid_area_code as an added column was the right shape.

Evidence

The producer went through five adversarial verification rounds; four found a real defect, two of them blocking, and every one was invisible to a green test suite. Silent territory deletion (Greece losing 10% while reporting healthy status); an abort that made the production call impossible at any year; five live mutants; and dead gap code whose only guard could detect duplication but never omission.

So the acceptance check here is mutation testing, not the suite. Every commit carries a baseline-gated sweep in a sandbox that cannot write the repo, with sources checksummed either side. Surviving mutants are either killed or proved equivalent by measurement — one over 228,175 comparisons, another over 102,988.

Every commit also proves "no new failures" against a git archive of its own parent, in the same environment, diffed at block level rather than by totals — two offsetting changes would cancel in a total. One earlier control was invalidated because the two sides ran in different environments; the tell was a skip-count mismatch, and it is now caught by printing environment fingerprints on both sides.

rcmdcheck run with tests 0 errors / 0 warnings / 1 note; devtools::test() FAIL 0 | PASS 8459 in the offline job's own configuration; lint_package() 0; pkgdown complete (302 topics, 302 _pkgdown.yml entries, 0 missing and 0 extra).

One disclosure about that evidence, because it is the reason this PR first went red on seven checks. Every local check on this branch ran rcmdcheck(args = c("--no-tests", …)), which skips the suite entirely — so a set of census assertions pinned against an older polity vintage stayed green locally and failed everywhere in CI. whep-polities #551 flipped GRC-1830-1913 to wiki_status = "superseded"; the live filter then drops it, build_polycell_support() returns 0 rows with no error and no warning, and 13 assertions ran against an absent frame. They are re-pinned on the live successor GRC-1881-1913 against an engine-independent reference (planar clip measured in Lambert azimuthal equal-area, 7.4e-05 relative), so the pin constrains neither the clipping nor the area engine. Two of them were passing vacuously on empty frames and now cannot: under a mutant that kills the fixture the same way #551 did, they still pass while the new guards fail.

Known, and deliberately not fixed here

8 of those 9 polities still lose their deposition, and Fix polity_area_crosswalk for regions.csv's 9 present-day-invalid area codes #459 did not fix it. Re-measured after the fourth reconciliation with main, against the deployed cell_polity_fraction.parquet (191 area codes). The codes that still receive no polycell are 61 Equatorial Guinea, 153 New Caledonia, 154 North Macedonia, 209 Eswatini, 212 Syria, 276 Sudan, 277 South Sudan, 299 Palestine. Only Fiji is fixed — the Refresh the polities snapshot: geometry now loads under s2 #662 snapshot refresh leaves FJI-1800-2025 s2-readable (0 s2-invalid on the 751-row table) and the crosswalk maps area 66 onto it. Fix polity_area_crosswalk for regions.csv's 9 present-day-invalid area codes #459 is closed and Model the Rest-of-World reporting members in their own right (#459) #628 is merged, but they promoted these territories on the reporting_polity_code axis only: polity_area_crosswalk still carries polity_code = ROW-1850-2025 and polity_area_code = 999 for all six, and carries no row at all for SYR-1967-2025, PSE-1948-2025, SWZ-1894-2025 or GNQ-1968-2025. The support keys on polity_code → polity_area_code, and the producer drops polity_type == "aggregate", so the promotion never reaches it. Sudan and South Sudan stay folded on 206 — Bucket 206 sums Sudan + South Sudan but is attributed to SDN-2011-2025: the value and its polity describe different territories #414 is now closed, and the open decision is costed in Un-folding FAOSTAT 276/277 out of bucket 206 loses 4.2% of the region's tonnage and breaks the ISO3 lookup #680 and lbm364dl/whep-polities#139.
What carbon drops is larger than this PR claimed, and two different defects were being summed as one. At the carbon support year 2015, 32 clipped polities carry no area_code and are deleted by .carbon_drop_unkeyed(). The six named above are 28.2 Mha of territory (Syria 18.70, Equatorial Guinea 2.70, North Macedonia 2.54, New Caledonia 1.88, Eswatini 1.71, Palestine 0.62) — the old 26.6 Mha over seven, minus Fiji, on a different basis. The other 26 are 252.5 Mha, almost all of it Greenland 213.2, Western Sahara 26.8 and French Guiana 8.4. None of those 26 resolves to a reporting bucket at all, so they are territory outside the reporting vocabulary rather than a crosswalk gap, and no deposition row ever existed for them to lose. Basis warning: the GLWD water layer and the Natural Earth ice layer are unreadable in this environment (WHEP_LPJML_INPUT_DIR and WHEP_NATURALEARTH_DIR unset), so every figure in this bullet is territory, an upper bound on land. That is a tight bound for the six (Syria 18.70 territory against 17.91 land measured previously) and a poor one for Greenland, most of which is ice. Measured on a polity-subset build, proved subset-invariant by a control that rebuilds the same targets with five neighbouring polities present (max |land diff| = 0 ha). The "single largest lever left in the ledger" claim is withdrawn rather than restated: it was never measured against the ledger's other terms, and these two populations should not be added into one lever.
Water and urban are not migrated. Both are gated on data layers on your side — water on what the LPJmL depths are referenced to (see WHEP_LPJML_RUN_DIR and lpjml_hydrology.R both still name the 1901-2009 run, while the complete 1901-2023 run sits beside it #598: the wiring and the docstring both still name the 2009 run while the complete 2023 run sits beside it), urban on HYDE. The decisions are recorded: regenerate landfrac and rerun for water, key at HYDE's native 5 arcmin for urban.
feed_lpjml.R is frozen. Its cell_area_ha multiplication inverts a normalisation applied upstream, so substituting land area there would under-count global grass — the mirror image of the defect this PR fixes.
Six no visible binding lines are deliberate — three in water_balance.R, one in feed_intake_redistribute.R, one in the frozen feed_lpjml.R, and, as of the fourth reconciliation with main, one in main's own new .read_fraction_country_grid(). Dropping the polity_frac global was the mechanism that proved no consumer was missed: with it present a planted missed consumer is invisible, with it removed the planted site is named — which is exactly how main's new fractional-grid reader announced itself here. The note count is unchanged at 1.
The great-circle border displacement (cshapes straight borders render as great circles: 123,276 km2 of Canada is booked to the USA, invisible to every conservation check #529, ~123,000 km² of Canadian prairie booked to the USA) is detected and reported, not corrected — the fix belongs in the polygons.

Also in here

Fixes for two live epoch fenceposts (1,912 polity-years move, 1,103 off a dead label); a status-aware tie-break so a retired interval can no longer outrank a live one (Montenegro 1913–14 was resolving to a 37.6% smaller polygon); the open-interval rule so 2025 resolves at all; and inst/scripts/reconcile_polity_areas.R, which reconciles every polity against FAOSTAT land + inland waters and attributes all 45 material discrepancies to polygon quality, identity, or convention.

Small polities are the headline there: Kiribati drew 34.3× its official area under the old crosswalk and now draws 1.18×.

@lbm364dl — review please. The data$cell_polity contract is measured to the byte if you want to sequence the build_cell_polity() re-key against it.

The polycell spatial-support work rests on numbers measured interactively (LUH2's terrestrial mask, the whole-cell over-count, the cell-polity crosswalk's shape and the border-cell cost). None of them were reproducible. This script regenerates all of them from source, prints each against the recorded value, and flags any that moved. It also settles two things the migration depends on. HaNi's per-cell mass is referenced to the WHOLE grid cell inside a land-masked domain, not to the land fraction: coastal cells divided by whole-cell area sit at a median 0.9937 of their interior neighbours, where a land-referenced mass would sit near 0.5. And the zero-terrestrial crosswalk cells are not marine at all -- every one of them carries Natural Earth country land; they are LUH2 coastline/small-island resolution misses and cells LUH2 classes wholly as ice or water. Every input is resolved from an environment variable and the script aborts naming it, so no machine-specific path is committed. Lives in inst/scripts alongside the other diagnose_* scripts, which is .Rbuildignore'd. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The polycell producer does not exist yet. These 18 tests are its contract, written first so the implementation is measured against a stated spec rather than against itself. Every fixture is a synthetic sf geometry, which is what the injectable geometries/water/ice arguments exist for -- no pins, no network, no rasters. They pin the eight cases the plan names plus four the preflight added: a coastal polycell at 5% land (the whole-cell base is never the answer); the four zero-land island states with EA6's polycell counts, recovering polygon area even when LUH2 is injected as zero land, since a producer that masked by LUH2 would silently delete four countries; two-, three- and four-polity cells; a geodesic border split that whole-subcell counting gets wrong in both of its ways; a border change across an exclusive end_year; the pig fixture, where the two polities' national totals are deliberately incompatible and nothing may cross the border; the three area categories as separately addressable non-land territory; unclaimed land emitted rather than renormalised away; polity_code never parsed for dates, with a code whose embedded years disagree with its columns and a hyphenated-prefix code; NA-explicit wiki_status and polity_type filtering; polygon_status passed through including an unknown level and NA; and dead and aggregate rows receiving neither data nor land. Tolerances are measured, not guessed, and stated in the file header with the number each was calibrated against. Fixture polygons are inset in latitude inside their cells on purpose: under s2 a cell's east-west edge is a great circle rather than a parallel, so a polygon whose edge sits ON a cell boundary loses 5.2e-5 to 2.7e-4 of its area to the clip. Inset polygons reproduce their own area to 1e-14, so the tolerances measure the producer instead of the way a grid cell is represented. The spec was checked to be satisfiable before being committed: a throwaway reference implementation, kept out of the package, makes all 113 expectations pass. Against the real package all 18 blocks fail with "'build_polycell_support' is not an exported object from 'namespace:whep'", which is the intended state until T-A4 lands. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The schema assertion is a subset check, so DA-13's transitional polity_frac and DA-12's second footprint can ride alongside without breaking it. Both are claims about the interim crosswalk geometry, which these polygon fixtures do not exercise, so they belong to T-A5's before-and-after measurement rather than here. Recorded in the header so T-A4 does not read the omission as an oversight. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Add the geodesic polycell producer: intersect the 0.5-degree grid with polity polygons over their validity interval using sf::st_area() on unprojected WGS84 with s2 enabled (matching WHEP's spherical cell-area convention, rejecting the WGS84 ellipsoid and equal-area projections). - Injectable geometries (default get_polity_geometries()); year resolution reads start_year/end_year columns with an exclusive end_year, never the polity_code. - NA-explicit wiki_status/polity_type filtering (keeps NA-status real territories); polygon_status passed through untouched. - Three separately addressable area categories (land + inland_water + ice = polity_area_ha): DA-6 ice subtracted by exact intersection with s2-invalid-feature repair (EA9), GLWD water apportioned pro rata by polity_area_ha. - DA-5 LUH2 reconciliation emitted via attr(result, unassigned); no LUH2 masking, no renormalisation. All 18 T-A3 tests in test_polycell_support.R pass.

Complete T-A4a and T-A4b on top of the core producer. Inputs (R/polycell_layers.R, a new capability: WHEP had no CLM reader). - read_glwd_water() reads the LPJmL CLM binary via its .json sidecar. Measured 67,420 cells, 32,358 with water, 2.4759 Mkm2 -- EA10 exactly. - read_glaciated_areas() reads ne_10m_glaciated_areas and classifies every feature as ok / s2-repaired / still-invalid instead of crashing on it. - read_luh2_terrestrial(vintage) reads (1 - icwtr) * carea from staticData_quarterdeg.nc: 12.9931 Gha, EA1/EA2 exactly. The vintage is recorded in the luh2_vintage output column (DA-9). Producer. - DA-16 interval grain: one row per polycell per interval over which every area is constant, plus expand_polycell_years() for the per-year view. Intervals split at each cell's own breakpoints, because the inland-water apportionment depends on which polities share the cell. Measured 564,304 interval rows against 13,742,974 per-year rows for 1850-2022. - DA-19 water clamp: apportioned GLWD water is capped at the polycell's territory and the excess emitted, so land_area_ha never goes negative. 94 polycells clamp in 2015, almost all of them on the Caspian. - DA-13 shim: polycell_shim_view() is identical() to build_cell_polity(), 68,527 rows, bit-for-bit on polity_frac and cell_area_ha. - DA-12: both crosswalk footprints emitted and reconciled at crosswalk_year. - DA-15 and EA9: 23 polities have no geometry and 7 carry an s2-invalid polygon (4 repairable, 3 not, one of them live Fiji). All carry an explicit coverage_status and are named in the "coverage" attribute, never a silent zero area. - New "overlap" diagnostic: 441 cells in 2015 hold more territory than the cell, because two live polities were handed the same polygon. Emitted, not renormalised: deciding who owns the ground is not the producer's call. - DA-17: polycell_support registered in whep_inputs.csv with an env-var override. Nothing uploaded. Measured on the shipped polities for 2015: land 12.7806 Gha rather than the whole-cell 14.70 Gha, inland water 2.3983 Mkm2, ice 2.2789 Mkm2, water plus ice 3.53% of territory against AM-6's predicted ~3.7%. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The S-A11 diagnostic keyed its rows on the intervals in which a cell was claimed, so a cell held 1900-1950 and unclaimed afterwards had no row covering 2015 and its unclaimed land vanished from that year's slice. Measured on the shipped polities the 2015 figure was halved: 158.05 Mha reported against 315.50 Mha actually unclaimed, the latter confirmed by recomputing the quantity independently from two separate builds. Each cell now contributes rows that partition the whole domain: the intervals it was claimed in, the stretches between them, and the stretch after the last one, with `claimed_land_ha = 0` on the gaps. Cells no polity ever claimed get one row spanning the domain instead of a row with NA years. A year therefore resolves to exactly one row per cell, asserted in the new test. Measured after the fix, 2015: 34,903 cells, 315.50 Mha, 2.428% of LUH2's 12.9931 Gha terrestrial area. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`.pcs_valid_pieces()` discarded every intersection piece that stayed s2-invalid after a planar repair, with no warning, no diagnostic and no change to `coverage_status`. The pieces are not slivers: 21 of them hold 1.4191 Mha across 5 polities, including seven whole Peloponnese and Aegean cells worth 466,032 ha, 10.08% of GRC-1830-1913, which still reported `has_geometry`. The loss broke S-A2 re-aggregation at every pre-1950 year and re-emerged as fake unclaimed land in the S-A11 diagnostic, mixing a producer bug with the genuine unclaimed territory DA-7 exists to keep attributable. The replacement, `.pcs_measure_pieces()`, keeps every piece and measures the unreadable ones with `terra::expanse()`, as `read_glaciated_areas()` already did. A new `area_engine` column records `"s2"` or `"terra"` per row, the `"terra_measured"` attribute lists the affected polycells, and the substitution raises a warning with its magnitude. terra is measured one feature at a time: handed a whole sfc it warns "not all geometries were transferred" on a GEOMETRYCOLLECTION and returns a shorter vector that would recycle against the rows it is assigned to. The ice clip goes through the same helper, so ice can no longer be under-subtracted silently either. Also fixed, all found while re-measuring: - `inland_water_ha` went negative on 56 fully ice-covered Greenland rows, violating T-A3's `expect_col_vals_gte(inland_water_ha, 0)`. Ice and water are independent intersections, so the water headroom is now floored where it is formed rather than only at `land_area_ha`, and the comment no longer claims the clamp bounds both terms. - `split_method` labelled a single-polity cell carrying GLWD water as `polygon_intersection`, hiding that its water was apportioned rather than placed. DA-6 asks the column to record which rule applied, so it now does wherever the pro-rata step ran. - `build_polycell_support()` took 6 arguments against CLAUDE.md's limit. The `example` argument is gone, replaced by `polycell_example_geometries()`, a tiny real fixture the documented examples build from. That also removes the hard-coded toy table, which carried a `year` column although the default grain is interval-keyed. - `inst/scripts/verify_polycell_support.R` re-derives every reported number, so none of them lives only in a scratchpad. Its S-A2 section measures at a year and never sums a polity across intervals, which is how the drop went unnoticed in the first place. Regression tests: GRC-1830-1913 from shipped package data pins the seven recovered cells and the engine they were measured with; a fully ice-covered polycell pins the water floor; a single-polity watered cell pins split_method. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

C0 of T-A5: the characterisation baseline, taken on unmodified consumer code so that every value the polycell migration moves is attributable rather than silent. Nothing here asserts that current behaviour is right; several of the pins record behaviour that is arguably wrong. The one that matters most is deposition. R/n_deposition.R:196 divides the HaNi cell mass by cell_area_ha and :197-199 multiplies it straight back, so the area cancels and today deposition_n_t is value_g_total x polity_frac / 1e6 -- mass is already conserved and the 11% whole-cell defect never reaches it. AM-5's fifth risk is swapping :198 to a land area while leaving :196: that moves the total ~10% DOWN, the direction a reviewer expects a fix to move it, while destroying the conservation property. Its first risk is dividing by each polycell's own land area, which emits a shared cell's mass once per polity and is invisible to every single-polity fixture in the repo. Both are guarded, with multi-polity fixtures throughout. Urban and grass are pinned as pure partitions with NO area term at all, confirming AM-1 against the A2 table: handing their crosswalk both cell_area_ha and land_area_ha must leave the output identical. All five guards were mutation-tested against those exact failure modes in memory (assignInNamespace, no file touched); every mutant died. inst/scripts/characterize_consumer_baseline.R records the numbers that need real inputs: the deployed crosswalk fingerprint DA-12 names as the S-A10 baseline (68,527 / 64,438 / 191, min polity_frac identical to 1/36 per AM-1, not 1/144), and conservation at full crosswalk scale. Two measured corrections to the plan's evidence. Per-cell sum(polity_frac) is NOT exactly 1: two of 64,438 cells fall one ulp short (max |sum - 1| = 1.11e-16), so EA3's "min = max = 1.000000" holds only to six decimals. And the cancellation is exact algebraically but round-trips through IEEE-754, so today's deposition_n_t already differs bitwise from its own identity on 14,775 of 68,527 rows at 3.22e-16 -- no downstream "bit-for-bit" claim can be made about consumer masses, only about the crosswalk columns DA-13 covers. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

BLOCKING REGRESSION. Keeping the s2-invalid clip pieces made their area recoverable, but it also left them in the polycell geometry column, and `.pcs_ice_areas()` then ran `sf::st_intersects()` across that column. s2 aborts on the first such piece, so with the shipped polities and the real ice layer the production call died at ANY year -- `years` is applied only after every polity has been clipped, and GRC-1830-1913 is always in the table. The two populations are now separated before either engine is asked to do anything: s2 clips the pieces it can read, terra clips the rest, so ice is subtracted on both rather than skipped on one. Greece with the real ice layer completes in 14.4 s. The suite missed it because the only test exercising the terra path passed no `ice`, and the two tests that passed `ice` used synthetic geometry with no invalid pieces. The new test combines them. Ice in terra also has to crop before it converts: handing the worldwide ice union to `terra::vect()` took the resident set past 22 GB on a job whose s2 path peaks near 1.2 GB. Cropping to the polycell's bounding box first holds it at 406 MB. DA-21 supersedes the earlier densification escalation, which was a misdiagnosis. The shortfall was never a parallel-vs-great-circle clipping bias: `.pcs_candidate_cells()` enumerated cells from the polygon's COORDINATE box, so the cells a bulging edge reaches into were never offered to the intersection. The window is now that box unioned with `s2::s2_bounds_rect()`, which is exact by construction rather than padded. Measured: SWA-1884-1912 -1.95e-04 -> -1.43e-14 (78 pieces -> 82), KEN-1888-1891 -6.29e-05 -> -4.55e-15 (179 -> 187), ESP-1800-2025 unchanged at -2.92e-14. S-A2 at 1900 now maxes at 6.70e-05 with 3 polities over 1e-6, all three of them carrying terra-measured pieces; at 2015 it maxes at 2.99e-11 with none. Three unreported loss paths closed: - A degenerate or NA-bounded validity interval deleted a polity in full with no warning while still reporting `has_geometry`. It now classifies as `invalid_interval`, warns, and appears in the coverage diagnostic. Latent on the 603-row table, live the moment the source refreshes to overlapping periods. - Cells the water layer and the polycells do not share were silently booked as having no inland water, turning that water into land, or had their water dropped. Both directions now ride in the `"water_unmatched"` attribute, which is what EA10 asked for. - Only the under-claim side of the LUH2 disagreement was emitted, so the over-claim was reconciled away by construction. `over_claimed_land_ha` is now emitted alongside it. Also: `overlap`, `terra_measured` and `water_excess` carry `end_year` as well as `start_year`, so they can be sliced to a year as the roxygen instructs; `@return` documents the `"crosswalk_only"` rows and the NA areas that make an unfiltered `sum(land_area_ha)` return NA; the terra warning pluralises; and the piece counts in the comments and the pinned test are corrected to the measured 1,419,140.84 ha over 21 pieces and 5 polities, with the seven Greek pieces described as the partial cells they are (shares 0.858 down to 1.7e-05) rather than whole ones. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The comment quoted 1,697 cells and 1,488 GLWD cells from an external measurement; the emitted diagnostic reports 1,906 polycell cells holding 0.403 Gha with no water row and 110 WET GLWD cells reaching no polycell. The second side is deliberately restricted to wet cells -- a further 1,378 unmatched GLWD cells carry no water and so lose nothing -- and the comment now says so rather than citing a figure the code does not produce. Measured on the whole-domain production build: shipped geometry table, real ne_10m_glaciated_areas ice, real GLWD water, real LUH2, both crosswalks. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Kills the five mutants the verification sweep found surviving: the over-claim block that passed with the quantity forced to zero, the ice block that admitted a whole-piece, halved or under-cropped result, and the window block that ignored the longitude half of the bounds union. Corrects three figures the code stated but did not produce: the unmatched-water count, the s2-invalid glaciated-feature count, and the clamped-polycell count. INCOMPLETE. An expanded sweep then found four further surviving mutants which are not yet investigated or fixed. Committed green (260 passing) so the work is not lost; it is not finished and has not been verified. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The sweep found four mutants the suite could not detect. Two were missing assertions and two were genuine coverage holes: - M07 "window: drop the s2 longitude UPPER bound" survived because the shipped table exercises only the lower one: F228-1800-1856 extends 10.46407 degrees below its coordinate box while the largest extension above any box is 2.842e-14. The new test asserts the CONTAINMENT PROPERTY exactly -- the window must cover the spherical rectangle on all four sides -- so a margin of 2.8e-14 is enough to fail it. - M10 "water floor removed" survived because no fixture can reach it: built from one geometry, `polity_area - ice` comes out at exactly 0, while on real data ice and territory are two INDEPENDENT intersections and the difference lands at -1e-9. It is now unit-tested on `.pcs_add_water()` directly with that numeric situation, which is what 56 Greenland rows hit. - M18 "terra area: first feature only" survived because every fixture had a single part. Now covered with a GEOMETRYCOLLECTION of two disjoint polygons. - M19 "ice repair skipped on the union" survived because no test ever handed the producer an ice layer whose union is invalid. Measured: all 1,885 kept ne_10m_glaciated_areas features pass `st_is_valid()` and their union does not, so the repair is load-bearing. The smallest reproducing subset is two Greenland features carrying 22,544 vertices and simplification collapses them, so there is no offline fixture: the test uses the real layer and SKIPS without `WHEP_NATURALEARTH_DIR`. That CI gap is stated in the test. Two mutants now survive and both are equivalent, with the measurement recorded next to the code rather than left for the next sweep to re-derive: - M24, skipping `.pcs_polygonal_part()` before terra, changes no area because terra extracts the polygonal parts itself: measured 69,978.2271 ha either way for a two-polygon collection and 0 either way for a line. The helper is kept for its explicit empty return. - M25, dropping the span prefilter, changes no output because the 0.01-degree bulge floor already implies it: the largest bulge any sub-degree span can produce is 0.0011 degrees, and reaching the floor at latitude 45 needs 3.03 degrees. Both equivalences are pinned by tests, so they stop being equivalent the moment the constants move. Also: the reproduction script now reports the S-A2 exception list over EVERY interval at a year inside its own validity, not at one calendar year, and fails loudly if the list grows beyond the five known members -- GRC-1830-1913, DEU-1800-1866, DEU-1866-1871, GBR-1800-1921, FRA-1800-1919 -- or if a member appears without terra-measured pieces. Two of the five are live in neither 1900 nor 2015, which is why a single-year check reports three and misses them. The regenerated Rd picks up the corrected 1,343 clamped polycells. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`.pcs_gaps_before()` was dead code. `transmute()` evaluates sequentially, so `end_year = .data$start_year` read the `start_year` rebound one line earlier rather than the claim's own start; every gap came out zero-length and the `start_year < end_year` filter then deleted all of them. Re-running the broken version over the shipped build returns 0 gap rows where the fixed one returns 7,340, and 6,487 of the 32,248 wet GLWD cells that some polycell reaches had no row at the domain start. The domain-covering diagnostics therefore under-read: the wet unmatched side at 2015 was 110 (its never-reached cells and nothing else) instead of 477. The gap bounds are now built under distinct names and renamed, which makes that class of mistake impossible here. No polycell area moves; only the diagnostics change. `.vps_exception_list()` summed `polity_area_ha` across every interval of a polity before comparing it with the polygon, inflating GRC-1830-1913 1,843x, so the gate could not have detected a real S-A2 regression. It now joins each polity to its own probe year and measures the one interval live there. The verdict is set equality against the expected five, and it prints a machine-checkable counts line plus explicit alerts for growth, shrinkage and any listed interval that carries no terra piece. `inst/scripts/` is under no test, so an unexecuted change there is unverified by construction. The script now takes `polity_codes` (default: the whole table, the production call) and reads WHEP_VPS_POLITY_CODES, so it can be executed in minutes after editing. Six mutants of the gate -- interval summing, relaxed tolerance, and suppressed growth/shrinkage/terra reporting, the last three paired with a positive control that makes the condition genuinely true -- all die against a passing baseline. Also in this commit: - `.pcs_s2_window()` kept the spherical longitude bounds but discarded the latitude bounds when the longitude interval wraps the antimeridian. The latitude bounds stay usable there. Found by widening the domain-partition property test to every polity: it omitted the extreme-latitude cells of F228's nine intervals and of KIR-1800-2025. - Block 46 now reads `min_bulge_deg` from `formals(.pcs_long_edges)` instead of repeating the literal, and block 42 asserts over every polity rather than three. Block 39 covers the whole diagnostic table (42 edges, 30 polities) and pins Egypt's 22nd parallel at 0.113289925 degrees of bulge. - M19 (the ice union invalid while all its features are valid) had no CI protection because it needed the 1.4 GB layer. `tests/testthat/fixtures/ ice_invalid_union.rds` stores the two Greenland features verbatim (22,544 vertices, 106 KB xz) and reproduces the condition offline. Verification: 48 + 6 blocks, 347 + 11 assertions, 0 failures, 0 skips; 29 producer mutants 27 died / 2 survived (both proved equivalent by measurement and pinned by blocks 47 and 48); 6 script mutants all died; `air format .`, `devtools::document()` and `lint_package()` clean (0 lints); `inst/scripts/verify_polycell_support.R` executed end to end. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Verification measured the pre-fix window as differing by 2.8e-14 degrees, which floors to the same cell index, so no candidate cell changed for any of the 580 polities. The comment claimed it had omitted cells from F228 and Kiribati; it had not. The bound is still kept, for the reason that is true: a refreshed geometry source need not be as forgiving, and block 43 makes that enforceable. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Brings the branch up to date with `main`, which is 138 commits ahead of this branch's base `3f479f7a` (AM-17 recorded 50; it has moved since). Merge, not rebase: this branch has published commits and a long review trail. Four conflicts, all resolved by keeping both sides: - `NAMESPACE`: both sides inserted an export in the same alphabetical slot (`read_polycell_support` ours, `read_population` theirs). Kept both, in roxygen's order. `devtools::document()` reproduces the file byte-for-byte, so the manual resolution is what the generator emits. - `R/utils.R`: both sides appended to the same `utils::globalVariables()` block. Kept both, ours first. - `inst/extdata/whep_inputs.csv`: both sides appended rows. Kept all three, main's two first. Main's separate deletion of the retired `luh2_v2h_states` row auto-merged and is preserved -- main ships a test asserting that alias is gone. - `data/whep_inputs.rda`: binary, regenerated from the merged CSV with `Rscript data-raw/whep_inputs.R` rather than resolved by picking a side. 72 aliases = 70 base - 1 retired + 2 main + 1 ours. `DESCRIPTION`, `_pkgdown.yml` and `tests/testthat/test_n_urban.R` auto-merged. No line of `R/polycell_support.R` or `R/polycell_layers.R` was touched. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…requires Three of the four C0 urban blocks stopped being CONSTRUCTIBLE after the merge, and none of them stopped being true. `build_urban_n()` now routes `area_code` through `.manure_territory_to_area_code()` (#463/#512), which aborts on a value it cannot resolve: Could not resolve territory to an area_code. Unrecognised values: "A", "B", and "C". The letters were deliberate: C0 picked codes nothing could look up so no assertion could pass by resolving one. That is now an abort rather than an opaque label, so the fixture takes the numeric codes the rest of this file already carries after main's own ISO3 -> numeric port (203 Spain, 68 France, 231 USA). Only the labels move. Every property, expected value and tolerance is untouched, and each was measured to reproduce EXACTLY under the new codes before the substitution was made, on the merged tree: sum(urban_n_t) 1129.3082821669493 = pop x rate / 1000, rel gap 0 shared-cell split 470.5451176 / 282.3270705 / 188.218047 = 1e6 x rate/1000 x (0.5, 0.3, 0.2), exactly no area term adding cell_area_ha AND land_area_ha to the crosswalk: identical() TRUE; tripling cell_area_ha: identical() TRUE silent drop crosswalk-missing cell still vanishes with no warning; total 941.09023513912 as before Counts before and after the merge are the same file for file: test_n_deposition.R 13 blocks / 34, test_n_urban.R 12 / 26, test_feed_lpjml.R 20 / 47, all green, 0 skips. The full-scale `characterize_consumer_baseline.R` run is byte-identical to the pre-merge run, including the 68,527 / 64,438 / 191 crosswalk fingerprint, min(polity_frac) identical to 1/36, and deposition mass conservation at relative gap 0. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…olvers `polities` uses a half-open validity interval: `start_year` inclusive, `end_year` exclusive. Year 2014 belongs to `RUS-2014-2025`, never to `RUS-1991-2014`. Two implementations read it as inclusive, and the documentation they were written from said "Inclusive". `R/constant_territory.R` filtered `end_year >= yr` with no dedupe. On a boundary year both epochs came back active and `.assign_polity()`'s "keep first match on overlap" handed every cell to whichever sorted first. Measured over the shipped table that is the DISSOLVED predecessor in 315 of 315 cases, because `X-a-b` always sorts before `X-b-c`; the successor received no row at all, and cells lying outside the predecessor's smaller extent were dropped entirely. It is exported, vignetted and runs during `R CMD check`, so this moved published numbers. `R/spatialize_compartments.R` filtered `yr <= end_vals`. Verified dead today: the deployed crosswalk (68,527 rows) and today's producer output (62,780) both carry no validity column, so `.country_grid_is_dynamic()` is FALSE and the resolver returns the grid untouched. It goes live the instant the polycell producer emits `start_year`/`end_year`, which is why it is fixed before rather than after. `R/table_mappings.R` documented `end_year` as "Inclusive", contradicting the convention `polity_end_year` is already documented under twenty lines below and matching what both filters did. Corrected here, with the `country_grid` docs in `spatialize.R` and `spatialize_livestock.R`, so the bug cannot be read back out of the documentation. Year resolution moves to `.active_polities()`, which also keeps one interval per polity using the tie-break `.whep_polity_lookup()` already uses (`R/polities.R:143-153`): exact start first, then latest start. The exclusive bound alone does not settle the 86 genuinely overlapping live polity-years (`PER-1825-1909` beside `PER-1825-1884`, `MNE-1913-1915` beside `MNE-1913-1918`). The dedupe runs AFTER the reported-code filter on the source side, so a source the caller named is never discarded for an interval nobody reported. Measured movement, decomposed so each half stays attributable: exclusive bound alone 665 polity-years / 349 polities 298 reattributed to the successor, 147 polities correctly dissolved on their own end_year, 220 at the 2025 domain end same-polity dedupe 1,263 polity-years / 22 polities both (shipped) 1,912 polity-years / 349 polities, of which 1,103 move from a retired/superseded label onto a live one Named: `RUS-1991-2014` -> `RUS-2014-2025` and `UKR-1991-2014` -> `UKR-2014-2025` at 2014; USA at 1803/1848/1867/1959; 1870-1885 leaves the retired `CAN-1866-1886` for `CAN-1870-1886`. Largest movers by count are BLZ 140, ARG 127, BRA 123, GRC 113, IRQ 94, EGY 84, ROU 84. On a real three-epoch Greek fixture at 1881 the reported total falls from 114.85 to 100 -- the dissolved epoch was attracting donor imputation on top of the double count. The roxygen example and all four pre-existing fixtures are bit-identical (`identical()` TRUE): they were written on non-adjacent epochs and so passed under either convention. Consequence worth knowing: under the exclusive bound 2025 has no active polity, since all 229 open intervals end there. 2024 is the last covered year, and the abort message now says so. Tests assert the property over every polity and every year of the domain instead of hand-picked codes, and add the adjacent-epoch fixture the existing blocks deliberately avoid. 15 mutants, 14 dead; the survivor -- dropping `exact_start` from the tie-break -- is proved equivalent, since `start_year <= yr` makes an exact start the maximum start, checked over all 342 years and 20,000 randomised tables. Related: whep-polities #550 reports the same defect class in `add_polity_code()` (`R/polities.R:139`), which is upstream's and deliberately untouched here. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

C2 made `end_year` exclusive and that was right, but a uniformly exclusive read leaves the current year unrepresentable: all 229 open intervals of the 740-row table end in 2025, so 2024 became the last covered year and `ref_year = 2025` aborted. DA-24 keeps the bound EXCLUSIVE AT A SUCCESSION -- that is what stops the boundary-year double count -- and makes it INCLUSIVE AT THE OPEN END, where nothing succeeds the interval and no double count is possible. "Still open" is read from the data, never from a hardcoded 2025, because PR #551 is already moving the table to 749 rows. An interval is open when it ends on `max(end_year)` of the supplied table AND no later-starting interval of the same polity also reaches that year. The second half is load-bearing rather than belt-and-braces: seven polities carry TWO intervals ending in 2025 -- `AGO-1816-2025` beside `AGO-1975-2025`, and likewise ARG, BLZ, BRA, GRC, IRQ, ROU -- so a bare "end_year is the maximum" test opens 229 intervals where only 222 are open, and the compartment resolver, which has no same-polity dedupe, would return the terminal year twice. Succession is judged only among the intervals that reach the domain end: a sibling that starts later but ends earlier succeeds nothing there, and treating it as a successor would punch a one-year hole in a continuous polity. The rule lives in one predicate, `.covers_year()`, used by both live resolvers. `.active_polities()` passes the polity family; `.filter_country_grid_year()` passes cell + `area_code` and deliberately NOT `polycell_id`, which DA-2 makes a function of the polity code and which therefore changes at every succession, filing an interval and its own successor under different keys. The flag is computed once on the WHOLE table in `build_constant_territory_series()` and indexed for the per-year source subset, because a subset can hide the successor and would reopen a genuinely dissolved epoch. Measured movement against `f7675039`, over every polity and every year of the domain: exactly ONE year changes attribution, 2025, which goes from 0 rows to 222. Nothing else moves -- 1684-2024 is row-for-row identical. The 222 are precisely the polities live in 2024, since no family has an interval starting in 2025. Largest by polygon area: Russia 1681.5 Mha (`RUS-2014-2025`), Canada 977.5, USA 955.2, Brazil 850.4, Australia 770.2, China 755.5, India 315.9, Argentina 278.1, Kazakhstan 271.5, DR Congo 233.7. The compartment resolver moves nothing today: no shipped country grid carries a validity column, so `.country_grid_is_dynamic()` is FALSE, exactly as C2 found. The open-end branch is skipped entirely except on a terminal year, and takes a unique-key fast path that is provably the general case when no compartment repeats. On a 70k-row grid an ordinary year costs 0.007 s and the open end 0.43 s, down from 1.34 s before the fast path. C2's property test is extended, not duplicated: it still asserts over every polity and every year 1684-2025 that no year resolves to two intervals of one polity, and now also that the open end is non-empty (so that assertion is not vacuous exactly where the current polities live), that it resolves to the same set as 2024, that the seven succeeded siblings are enumerated rather than tolerated, that a table ending in 2000 opens 2000 and not 2025, and that `ref_year` on the open end works end to end through the exported function. `tests/testthat/test_datasets_balances.R` re-derived ESP's polity with `polity_end_year >= yr`. Inert -- ESP is one interval and 0 of 10 benchmark years disagree -- but it encoded the wrong convention, so it now states DA-24. `data-raw/balance_coefficients.R` has the same defect and is deliberately left alone: it is issue #565, and fixing it regenerates packaged `.rda` data whose movement needs its own measurement. Every statement of the convention now carries the special case: `R/table_mappings.R` for both `polities` and `polity_area_crosswalk`, `R/constant_territory.R`, `R/spatialize.R`, `R/spatialize_livestock.R`, `R/spatialize_compartments.R`, and `R/polities.R`, where `.area_year_polity_conflicts()` keeps the strict read on purpose and says so with the measurement behind it (190 open intervals, 0 areas resolving to more than one polity on the open end). Sweep baseline-gated before and after: 14 mutants, 12 dead. The two survivors are equivalence claims, both discharged rather than excused. Dropping `.polity_family()` from the resolver's grouping is proved equivalent -- the same-family dedupe downstream already keeps one interval per family on the same latest-start ordering -- at 0 differing years over all 342 years of the shipped table and 0 counterexamples in 20,000 randomised tables. Forcing the slow branch is a deliberate positive control for the fast path and must survive. Full suite 6002 passed / 18 failed / 1 error / 2 skipped / 57 warnings, and a pristine `git archive` of `f7675039` reproduces the identical census -- same four blocks, same 9/4/5 split, same HWSD error. The 18 are AM-25's census failures deferred to PR #551. `lintr::lint_package()` with the CI linter set: 0 lints. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…DA-24) DA-24 settled the convention in `.covers_year()` / `.open_ended_intervals()` (`R/constant_territory.R`, `6fa90e43`): `start_year` inclusive, `end_year` EXCLUSIVE at a succession and INCLUSIVE at the open end. `R/polycell_support.R` kept the uniformly exclusive read at three sites, so the producer stopped one year short of its own data: `expand_polycell_years()` 2025 returned NOTHING -- every polycell of the current world unresolvable `.pcs_add_shim()` a crosswalk describing the domain end attached to no interval, so all 4,420 crosswalk rows fell through to `.pcs_append_crosswalk_only()` and the shim degenerated into a copy of its own input `.pcs_polycell_footprint()` the polycell footprint came back EMPTY while `.pcs_footprints()` still reported a row for it, voiding the reconciliation that measures the migration's movement: 269 disagreeing cells -> 0 The rule is not restated here. All three sites call the existing predicate, via one new pair of private helpers whose only job is to supply the two things `.covers_year()` needs from THIS table. THE DOMAIN. A row with no `polity_code` is `.pcs_append_crosswalk_only()`'s padding, pinned to a synthetic `[crosswalk_year, crosswalk_year + 1)` window. It is not a validity interval: it can never be open, and it must not sit at `max(end_year)` and move the domain end off the year the real intervals reach. THE GROUP is the physical cell plus the polity FAMILY. Wrong is silent in both directions, since the rule opens a group's latest-starting member and closes the rest. TOO FINE and every terminal row is its own group maximum, so all of them open: `polycell_id` and `polity_code` are both a function of the epoch (DA-2), so an interval and its own successor land in different groups, and this grain has no dedupe downstream -- that is the boundary-year double count C2 removed. TOO COARSE and one lineage closes another: `cell_id` alone merges every polity in the cell, and `area_code` is a label rather than an identity -- measured on the shipped table, it is NA for 31 of the 220 live intervals reaching the domain end (and a pasted key stringifies NA, so those collapse into one bucket rather than being skipped), while `SDN-2011-2025` and `SSD-2011-2025` are two lineages under one code, 206. `cell_id` is in the key because a family's successor need not occupy the same cells: on the family alone, a later start in some OTHER cell closes a still-open interval here. `.polity_family()` is exactly what `.active_polities()` groups on, so the package keeps one notion of succession; `.filter_country_grid_year()` reaches the same key from the other end because a country grid need not carry `polity_code` at all. The three sites differ in what a mistake costs, and the grouping was chosen per site rather than assumed. `expand_polycell_years()` is the per-year consumer surface with no dedupe anywhere after it, so a wrongly opened interval lands straight in `polity_area_ha`; its flag is hoisted out of the year loop because being succeeded is a property of the table, not of the year. `.pcs_add_shim()` resolves after the join, since the join can duplicate a row and the flag must be aligned to the rows the mutate sees. `.pcs_polycell_footprint()` ends in `distinct()`, so it cannot double-count -- its risk is the opposite one, a wrongly closed interval dropping a cell and manufacturing the disagreement it exists to measure; dropping the crosswalk-only rows first is what keeps the shim padding out of the footprint, which is a separate guard from keeping it out of the domain. MOVEMENT: exactly one year, measured against a pristine `git archive` of `6fa90e43`. On a bounded build of 151 shipped live polities (19,884 intervals, 6,024 cells, 1684-2025) every year 1684-2024 is row-for-row identical -- the resolved key set matches, the area sums differ by 0, and the WHOLE tibble is `identical()` at 1684, 1800, 1900, 1950, 2000, 2011, 2015, 2023 and 2024. 2025 goes from 0 rows to 4,725, carrying 917.73 Mha, and its polycell set is `identical()` to 2024's. No `polycell_id` resolves twice in any year on either side. With a 4,420-row crosswalk, `crosswalk_year = 2015` (the default, and the only value the shipped pipeline uses) is `identical()` on the table, the shim, both footprints and the 269-row footprint diff; only `crosswalk_year = 2025` moves, from the degenerate 24,304-row / 0-attached / empty-footprint shape to exactly the 2015 one. SWEEP, baseline-gated: 14 mutants, 13 dead. The one survivor, M13, is the deliberate positive control -- dropping the per-table hoist must NOT change an answer -- and the hoist is a performance claim, not a correctness one (342 years of the bounded table expand in 1.77 s). The four candidate grouping keys are separated by a unit test on the helper itself, because `.pcs_split_intervals()` happens to give every interval reaching the domain end in one cell the same start year and would hide three of the four. The baseline reproduced the working tree's 379-passed / 18-failed census before AND after the sweep; two earlier runs were voided by their own gate, once by orphaned concurrent sweeps rewriting the source mid-census, so the source is now checksummed either side of every run. Full suite 1,842 blocks / 5,975 passed / 18 failed / 1 error / 57 warnings / 2 skipped, against 1,832 / 5,925 / 18 / 1 / 57 / 2 for a pristine archive of `6fa90e43` -- the identical failure set, the same three blocks in the same 9/4/5 split and the same HWSD error, AM-25's census failures deferred to PR #551. `lint_package()` with the CI linter set: 0 lints. `rcmdcheck()`: 0 errors, 0 warnings, 0 notes. Not fixed, deliberately: the interval-grain DIAGNOSTIC attributes have the same one-year shortfall one layer out. Filtering them with a plain half-open test at the domain end returns 0 rows where 2024 returns 9 (`overlap`) and 50 (`coverage`), and `inst/scripts/verify_polycell_support.R` holds three such filters (`:190`, `:343`, `:374`). All are inert today because the script pins `year = 2015L`, and the readers of those attributes are consumers, which this commit does not migrate. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

build_n_deposition() splits each cell's HaNi mass across the polities holding that cell. The split key was `polity_frac`: a raw subcell count over Natural Earth 10m polygons at 1/12 degree, quantised to 1/36 of a cell. It is now the polity's share of the cell's territory, `polity_area_ha` as build_polycell_support() measures it geodesically. The key change only. No category split (that is C3b), no new denominator, and `deposition_kgn_ha` is untouched. Either key is a share of the cell, so the source mass is redistributed and never created or destroyed. Measured on the deployed crosswalk over the 63,065 cells both supports carry, with a uniform synthetic field: source mass : 63065000000000 g sum(deposition_n_t)*1e6 OLD : 63065000000000 g | rel gap 0.000e+00 sum(deposition_n_t)*1e6 NEW : 63065000000000 g | rel gap 0.000e+00 NEW total - OLD total : 0 t (rel 0.000e+00) per-cell mass under NEW: min 1000.000000000000 max 1000.000000000000 t cells whose polities disagree on deposition_kgn_ha: 0 An unmigrated consumer is unaffected, proved rather than argued: over the whole 68,527-row crosswalk this commit is identical() to HEAD 0dbc9c4 on all 11 shared output columns. The one difference is the new `method_polity_split`, which records which partition produced a table so the two cannot be confused after the fact; `split =` selects it explicitly and naming a key the support lacks aborts instead of falling back to the coarser one. `deposition_kgn_ha` stays a whole-cell mean shared by every polity of a cell. That is the guard against dividing the allocated mass by each polycell's own territory, which would make rate x area recover the whole cell mass once per polity behind entirely plausible rates. The cell_area_ha round trip is kept rather than simplified away, so substituting a land area on one side alone stays a visible break instead of a silent ~10% fall. build_polycell_support() keys on polity_code, and polity_area_crosswalk folds distinct polities into one area_code (Sudan and South Sudan share 206) or leaves it NA. Deposition rows are keyed on area_code, so a support carrying a duplicated or NA area_code is refused with an instruction to convert at the caller's boundary (DA-23) rather than folded here, where the fold would look like a partition. Movement per S-A10, measured on values and decomposed so the partition change stays attributable from the area_code churn that travels with it. Over the 3,532 shared cells whose polity set is unchanged, per-cell L1 between the old and new share vectors is median 0.0432, p90 0.1190, p99 0.3212, max 1.3396, and 38.6% exceed the 1/36 = 0.0278 quantum AM-8 predicts. The excess is expected: the swap changes the polygon SOURCE (Natural Earth to the periodized whep polities) as well as the estimator, and AM-8 bounded only the estimator. No published balance moves on this commit. Nothing in the package reads `deposition_n_t`; .n_inputs_deposition() reads `deposition_kgn_ha`, which is bit-identical between the two keys. Mutation sweep, baseline-gated at 70 assertions green and byte-restored after each mutant: 19 mutants, the 18 defect mutants all dead and the comment-only positive control alive. Full suite against a pristine git archive of 0dbc9c4: identical 18-failure / 1-error set, the AM-25 census failures deferred to PR #551. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…ger (C3b) build_n_deposition() decomposes each polycell's share of the cell's HaNi mass over the three DA-3 territory categories -- land, inland water and ice -- in DA-19's apportionment shape, the one the producer already uses. The categories are emitted side by side and separately addressable, so a consumer picks a scope per purpose instead of being forced onto one. A support with no category columns cannot be decomposed and carries the single "territory" category, which says the row is undecomposed rather than claiming it is land. DA-14 DECIDED, 2026-08-06, on the measurement this commit checks in: the nitrogen ledger takes the WHOLE territory. That is a scientific correction, not a conservative default. Nitrogen deposited on a lake or a glacier still drives indirect N2O and still reaches the eutrophication pathway, so the impact terms have to account for it; the terrestrial-only scope would have discarded 0.89 Tg N of real flux. Scope is exposed as data$deposition_scope, default "territory", with "land" selectable per the repo's multi-method rule, recorded in a new method_deposition_scope output column, and refused rather than silently downgraded when the support cannot serve it. NO PUBLISHED NUMBER MOVES. Under "territory" the scope factor is a sum divided by itself, exactly 1, so every ledger value is bit-identical -- asserted with expect_identical() over the whole deposition slice of build_n_inputs(), not argued here. The C0 characterisation re-run at full crosswalk scale reproduces AM-14 exactly: 68,527 rows, source 64438000000000 g recovered to a relative gap of 0, the row-wise polity_frac identity bitwise on 53,752 of 68,527 rows with a worst relative gap of 3.22e-16, a 3x cell_area_ha moving 26,122 rows by at most 3.9e-16 with the total bit-identical, and C3a's named guard still at 0 cells whose polities disagree on deposition_kgn_ha. THE DA-14 MEASUREMENT, checked in at inst/scripts/measure_deposition_categories.R and run at 2014 on real HaNi, GLWD and Natural Earth ice input: BEFORE (all territory) 61.6285 Tg N | AFTER (land only) 60.7385 Tg N change -0.8901 Tg N (-1.444%) NEW QUANTITIES: freshwater 0.7248 Tg N (1.176%) | ice 0.1652 Tg N (0.268%) conservation: source 61628543587563.125 g, three categories recover 61628543587563.125 g, relative gap 0 (DA-18 bound 1e-9) per polycell the three categories re-sum to the undecomposed share to 2.78e-13 over 62,725 polycells AM-6's ~3.7% MUST NOT BE QUOTED AS THE BOUND for this question. It is an AREA-weighted bound; measured area-weighted is 3.532% and the DEPOSITION- weighted figure is 1.444%, because deposition is not distributed like territory -- ice is 1.72% of territory and takes 0.27% of deposition. AM-6's named polities do reproduce in rank (Canada #1 -6.93%, USA #3, Russia #4, Sweden #11, Finland #13 -7.50%, Norway #23) but AM-6 misses two classes that dominate the relative ranking: mountain-glacier states (Tajikistan -13.41%, Iceland -10.48%, Switzerland -6.39%, Nepal -6.86%) and African rift-lake states (Malawi -7.58%, Rwanda -5.81%). The Great Lakes box carries 74.88 kt N to freshwater, 10.3% of the world's, best cell ranked 15 of 63,185 -- a real hotspot but not the top one, which is Tibetan and Chinese lakes. THE ICE TERM IS A FLOOR, not an estimate. The deployed crosswalk has no area_code for Greenland at all (0 of its 191 codes, 0 rows at central Greenland), so 4,554 cells holding 177.51 Mha of ice -- 77.9% of the world's -- are unreachable by the ledger. Pre-existing, not introduced here, and the same class as EA9's missing terrestrial cells. PARTITION MISMATCH (AM-29), resolved as far as this commit can. The ledger is self-consistent as shipped because deposition_kgn_ha is a cell constant carrying no partition, so polity_frac via build_ag_land_support() is the only partition inside it. Under a future "land" scope only the ice term is partition-dependent -- DA-19 apportions water pro rata, so every polity of a cell carries the same water fraction whichever partition splits it -- and that is bounded by measurement at 0.074841 Tg of 0.165237 Tg, 0.12% of the total. Consuming the geodesically split land mass directly would NOT be self-consistent and needs build_ag_land_support() to migrate first (C-series). R/n_balance.R FIXES A CLASS, NOT AN INSTANCE, and this is the part to review. Two hardcoded c("method_recycling_n", "method_synthetic") vectors would have swallowed ANY future method stamp, so the new scope column reached build_n_inputs() and was silently dropped before the published balance -- a result that looks like it names its own scope while the reader can never see it. Both sites now route through .nb_input_method_cols(). A DEFECT FOUND BY ITS OWN TEST, recorded because the class keeps recurring. The new one-rate-per-polycell guard first computed rate_spread AFTER deposition_kgn_ha inside the same dplyr::summarise(), so it measured the spread of a scalar and was 0 by construction: an assertion structurally incapable of failing, the third on this branch after .pcs_gaps_before() (AM-20). Fixed by ordering, and pinned as sweep mutant M22. VERIFICATION. Mutation sweep run inside a git ls-files sandbox copy that cannot write the repo, triple-gated (baseline green before each mutant, patch text present exactly once, source md5-restored after each): 25 mutants, 24 defect mutants ALL DEAD, 0 survived, 0 patch-miss, comment-only control ALIVE, all gates TRUE, post-sweep baseline pass 224 fail 0 err 0. The four named risks are M14 (unserviceable scope accepted), M16 (default silently "land"), M17 (scope column mis-recorded), M19 (ledger takes every category) -- all dead. Tests: test_n_deposition.R 113/0/0 (was 70), test_n_balance_inputs.R 111/0/0 (was 85), test_n_balance.R 76/0/0 (was 71), test_n_surplus.R 15/0/0, test_exported_polity_columns.R 29/0/0. air format . clean over the whole repo. Pristine control at d94444e: 1853 blocks, 5979 pass, 18 fail, 0 error, 11 skip, the 18 confined to 3 test_polycell_support.R blocks. NOT YET RUN AT COMMIT TIME, stated rather than implied: the head-side control run (only the pristine side has completed, so "no new failures" is not yet proved), rcmdcheck() and lintr::lint_package() with the CI linter set. Those follow in their own commits. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…dupe PR #577 landed the exclusive `end_year` and DA-24's open-end rule upstream, so this merge drops our `f7675039` where the two overlap and keeps theirs. What #577 does not carry is the same-polity dedupe, and that half is re-landed here because it fixes something the exclusive bound alone does not: on a boundary or overlap year two intervals of one polity both come back active, `.assign_polity()` keeps whichever sorts first, and the dissolved predecessor sorts first in 2,134 of 2,134 co-active pairs on the shipped table because `X-a-b` precedes `X-b-c` lexically. The successor then gets no row at all and every cell outside the predecessor's smaller extent is dropped, so the defect loses mass rather than merely relabelling it. Conflicts and how they were resolved: - `R/constant_territory.R`. Kept our `.active_polities()` / `.covers_year()` / `.open_ended_intervals()` / `.polity_family()`, which are a strict superset of main's inline `.active()`: main's is exclusive everywhere, so it leaves 2025 unresolvable, and it has no dedupe. `R/polycell_support.R` and `R/spatialize_compartments.R` call those helpers directly, so replacing them would put a second copy of the convention back in the producer. Main's measured comment (238 polities carrying a polygon in 1993 under an inclusive read, 453 extra active polity-years over 1850-2024) and main's `as.character()` fix in `.assign_polity()` (#209) are both taken. - `R/table_mappings.R`. Merged rather than chosen. Main's `polities` doc is stale against main's own code: it says an open period "stops one year short" of its `end_year`, while `.polity_join_end_year()` on the same branch widens open periods by one. The merged text keeps main's F51-1947-1993 succession example and their successor-based framing, adds the open-end half, and states the figures re-measured on the 749-row table. - `tests/testthat/test_constant_territory.R`. Both sides' blocks kept. - `NAMESPACE`, `man/*.Rd`. Union, then regenerated with `devtools::document()`. - `inst/extdata/whep_inputs.csv`. Main's newer pin versions for the four LPJmL aliases plus our `polycell_support` row; `data/whep_inputs.rda` regenerated from the merged CSV with `data-raw/whep_inputs.R`. Re-measured on the 749-row table (#551), because every figure moved: - Shared-start-year exceptions: 9, not 8. #551 added `CAN-1886-1949` beside the now-retired `CAN-1886-1948`. Re-pinned as an enumerated list. - Intervals ending on the domain end: 237, of which 229 are open and 8 are succeeded there. `CAN-1948-2025` is the eighth, so a bare "end_year is the maximum" test now opens 237 where 229 are open. Upstream's own `successor` column agrees with the data-derived predicate on every open interval, and a new assertion pins that so the two mechanisms cannot drift. - Self-overlapping polity-years: 2,134 across 23 families on the 703 polygon-carrying rows the resolver actually sees, up from 1,991. On live, non-aggregate rows it is 0, down from 86 across MNE and PER, because #551 retired `MNE-1913-1915` and superseded `PER-1825-1909` rather than deleting them. `get_polity_geometries()` does not filter by `wiki_status`, so the dedupe matters more after #551, not less. FINDING, pinned rather than fixed: the tie-break is status-blind, so on a shared start year a live interval can lose to a dead one. Two cases. `CAN-1886-1949` is harmless, its polygon being `st_equals()` to the retired `CAN-1886-1948`'s. `MNE-1913-1918` is not: the retired `MNE-1913-1915` that wins carries 0.9923 Mha against its 1.5893 Mha, so 1913 and 1914 resolve to a 37.6% smaller Montenegro. Deciding by status would change the function's contract (`wiki_status` is not among the three columns a caller supplies), so it is enumerated in the test instead. Verification. Full suite against a pristine `git archive` of `2287ad22` in the same environment (identical WHEP_* fingerprint on both sides), diffed at block level: pristine 1872 blocks / 6085 pass / 18 fail / 1 error / 2 skip, merge result 1980 / 6464 / 18 / 1 / 2. 110 blocks added, all from main; 2 removed, both deleted by main; 0 shared blocks changed fail/error/skip; the only 2 shared blocks whose pass count moved are the two assertions added here (+3). The failing set is identical -- the same three deferred `test_polycell_support.R` blocks with the same 9/4/5 split, and the same pre-existing HWSD error (#596). The dedupe is mutation-tested in a sandbox that cannot write the repo, baseline re-verified green before every mutant and the source md5-restored after: 9 mutants, 8 dead, 1 (dropping the redundant `exact_start` key) proved equivalent over 342 years on both populations and 228,175 randomised comparisons, with an unmutated control alive. `air format .`, `devtools::document()`, `lint_package()` 0 lints, `rcmdcheck()` 0 errors / 0 warnings / 1 note -- and that note reproduces byte-for-byte on a pristine `origin/main`, so it is inherited (`sovereign_iso3c`, `key_row`, `canonical_area` are undeclared in `globalVariables()` on main). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`.active_polities()` ranked a same-polity overlap on the start year alone, so where two intervals of one polity share a start year nothing decided the winner and a `retired`/`superseded` row could take the year from the live row that replaced it. Measured on the shipped 749-row table: 1,344 polity-years across 10 polities resolved to a dead interval. `wiki_status` is now the LAST key of the ranking, below the start year. A later start is a more specific epoch, so ranking status above it would resurrect a superseded wider period over the narrower one that succeeded it -- the C2 defect in reverse, and measurably 132 further polity-years (`BLX-1921-1999`, `IDN-1889-1945`). The column stays OPTIONAL. Where it is absent every row reads as live, so the key is constant and the ordering is bit-for-bit the one that stood before: verified identical over all 47,741 polity-years against a pristine `git archive` of b778cc1. The difference between the two call styles is documented in `@param polities` rather than left implicit. Movement, measured on values rather than row counts. Exactly one case moves territory: `MNE-1913-1915` (retired, 0.9923 Mha) to `MNE-1913-1918` (draft, 1.5893 Mha), so 1913 and 1914 stop resolving to a Montenegro 37.6% too small; the symmetric difference is 0.5970 Mha. `CAN-1886-1948` to `CAN-1886-1949` (62 yr) and `IDN-1800-1889` to `IDN-1800-1945` (89 yr) are `st_equals()` at 937.96 and 188.51 Mha, so those are relabelling only. The remaining 1,191 are seven reporting-bucket aggregates whose dead side carries no polygon at all, and `build_constant_territory_series()` drops empty geometries before resolving, so its own exported surface moves 153 polity-years, not 1,344. The exception list re-measured with the column supplied still has nine entries, but every one is now a `retired`/`superseded` row the tie-break is supposed to pass over: ZERO live losers, against two before. `.polity_is_live()` states which rows are dead once for the package; `resolve_polity_label()` and `.pcs_prepare_polities()` now read it instead of each carrying its own literal, so the producer's filter and the resolver's tie-break cannot drift. Sweep: 11 mutants, 9 dead, control alive, baseline green before and after every mutant; the one survivor is proved equivalent over 102,988 comparisons. Full suite diffed at block level against a pristine `git archive` of b778cc1 in the same environment: 7 blocks added, 0 removed, 0 shared blocks changed on pass/fail/error/skip, failing set identical. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Two changes that had to land together. The crosswalk: every default reader on the carbon path read `spatialize-country-grid`, whose pin carries only (lon, lat, area_code), so `.normalize_country_grid()` defaulted `cell_area_frac` to 1 and gave a border cell's whole land to one polity (EA4). They now resolve one polycell support through `.carbon_cell_support()`. The area (DA-26): the LUH2 class fractions are rescaled onto the polycell's own `land_area_ha`, so the carbon path and the nitrogen path share one land definition instead of two. `area_basis = c("polycell_land", "luh2_fraction")` keeps the pre-DA-26 arithmetic selectable and records the choice in `method_land_area`, so the crosswalk change and the area change can be measured apart and two tables built on different bases are not indistinguishable afterwards. `cell_area_frac` is the polycell's share of the cell's LAND, never `land_area_ha / cell_area_ha`: everything it splits is already land-only, so that ratio would remove the water twice while the shares still summed to 1. The share denominator is taken over the cell's whole measured land BEFORE any row is dropped, so an unkeyable polity's hectares are lost visibly instead of being renormalised onto its neighbour. DA-23 is handled at one boundary and refused everywhere else: polycells with no `area_code` are dropped with their land reported, polycells sharing an `area_code` in one cell have their land summed with the fold reported, and a support handed to a consumer that is not already one row per cell and `area_code` aborts. `build_ag_land_support()` is pinned to the transitional basis: it is the nitrogen path's land support, whose cropland half is still split by `polity_frac`, and it must not end up with its two halves on two land definitions. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

spatialize_country_n_to_crops() distributes a national synthetic-fertiliser total across grid cells. Each cell's weight was its crop-pattern (or cropland) hectares times `polity_frac`, the crosswalk's subcell-count share of the cell, quantised to 1/36. It is now the polycell's share of the cell's LAND, `land_area_ha / sum(land_area_ha)` over the cell, as build_polycell_support() measures it geodesically. AM-4 makes C5 the CONTROL CASE: the cell weights are renormalised inside each polity-crop-year, so a change of partition redistributes the national total and cannot change it. The prediction held, exactly. Measured at 2015 on real FAOSTAT synthetic-N totals and real harvested-area crop shares, over the 173 polities both supports can place (106.5352 of 106.6550 Mt N): 173 polities; old 106.535153 Mt N, new 106.535153 Mt N difference 0.000000000 t worst per-polity |difference| 0.000000000 t polities moving more than 1e-6 t 0 And no ~11% signal on the weight. Restricted to the 1,011 polity-crop-years landing on exactly the same cells -- the weight alone, with the footprint held fixed -- the L1 distance between the old and new normalised cell-share vectors is median 0.000251, p90 0.0592, p99 0.2310, max 0.3643, with 40.95% exactly zero; per-cell relative share change is median 0.0000%, p99 26.60% and old-share-weighted mean 1.99%, and 57.29% of cells are bit-identical. Over the 13 polities whose whole cell set is unchanged the L1 is median 0.0368, max 0.1897, which is the same order as AM-29's L1 median 0.043 for the same key swap in the deposition path. Over the full migration the L1 tail reaches 2.000000, and that is the FOOTPRINT, not the weight: the two supports differ by 641 crosswalk-only and 4,574 polycell-only cells, and a polity gaining or losing a cell can switch a crop between the crop-pattern route and the uniform-cropland fallback. It vanishes under the same-cell-set cut (max 2.000000 -> 0.3643), and mass is conserved exactly in every case. Cells with a non-zero allocation: 43,272 -> 43,293, 43,177 shared, 95 old-only, 116 new-only. The share is of the cell's LAND, never `land_area_ha / cell_area_ha` -- what it splits is already land-only, so the whole cell as denominator would remove the inland water and ice a second time. And it is a share, never an absolute area: `crop_pattern_ha * land_area_ha` weights every cell a second time by its own size and is no partition at all. Both are AM-5 risk 3 in its C5 form, and both are pinned as mutants that die. `split = c("auto", "land_area_ha", "polity_frac")` keeps the transitional key selectable and records the resolved key in `method_polity_split`, so two tables built on different partitions are not indistinguishable afterwards. `auto` takes the measured land when the support carries it; naming a key the support lacks aborts rather than falling back to the coarser one. DA-23 is refused rather than folded: a support whose `area_code` is duplicated or NA is rejected with an instruction to convert at the caller's boundary. The refusal is a property of the polycell key only, so the transitional path stays exactly what it is today (DA-13), and a `polity_frac` support is used as supplied and never renormalised. The share denominator is the cell's whole supplied land, taken BEFORE the landless cells are dropped. Dropping first renormalises the survivors onto the whole cell and hands them hectares that are not theirs while the shares still sum to 1 -- C7's own defect, pinned here as a mutant. It also bites at the CALLER's boundary, where the DA-23 conversion happens, so `@param cell_polity` names the correct treatment: keep unkeyable land in the denominator under a bucket no country total can join. Measured, that matters: 3,568 polycells carry no `area_code` (101.938 Mha), of which 42.102 Mha sits in 289 cells a keyed polity also holds. Dropping them instead moves 16% of polity-crop-years (L1 p99 0.6232, max 1.3581) with the totals still conserved. Mutation sweep in a `git archive` copy that cannot write the repository, baseline-gated at 69 assertions green, patch text present exactly once, source restored byte-identically after each mutant: 22 mutants, the 21 defect mutants all dead and the comment-only control alive. Two genuine holes were found and closed rather than excused -- `auto`'s preference order was untested (every fixture carried only one key) and the transitional key's no-renormalisation rule was untested (every fixture already summed to 1 per cell). Full suite against a pristine `git archive` of 0477a4d in the same environment, fingerprints diffed and identical, compared at BLOCK level: 17 blocks added (all C5's), 0 removed, 0 of the 2,004 shared blocks changed pass/fail/error/skip, and the failing set is identical (the 18 deferred test_polycell_support.R failures and 2 pre-existing errors). `lint_package()` 0 lints; `rcmdcheck()` 0 errors, 0 warnings, 1 note -- the note inherited from main (canonical_area, key_row, sovereign_iso3c). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

S-A5. `.normalize_country_grid()` defaulted a missing share to `cell_area_frac = 1`, which is exactly how the deployed `spatialize-country-grid` pin (58,795 rows, `lon`/`lat`/`area_code`, 0 multi-polity cells) gave a border cell wholly to one polity while every conservation check still passed. That default is now an abort, and so are an NA share and a share outside [0, 1] beyond float tolerance -- the last of which catches an absolute area substituted for a normalised weight. `landfrac` is removed from the alias list and named as a refusable quantity: it is a LAND fraction, one value per cell and identical for every polity in it, so reading it as a polity share delivers the cell once per polity (AM-5 risk 7). S-A6. `.compartment_join_cols()` aborts rather than silently degrading to a `(lon, lat)` join, which across a polycell grid is many-to-many. `.build_base_grid_cp()` drops pattern cells no compartment claims instead of giving them a whole-cell share under an NA `area_code`. `.build_proxy_grid()` carries only the compartment key and the weight. The acceptance note at the head of `R/spatialize_compartments.R` states what remains irreducible: every driver layer is per physical cell, so two polycells in one cell share one cropland density, one grass productivity and one GLW3 density. That is resolution loss, not cross-border contamination -- the share denominators stay per polity. AM-5 risk 19: the capacity ceiling is soft, and it is soft at scale. `.warn_capacity_breach()` measures and reports it without changing which invariant yields; on the deployed crosswalk at 2000 and 2015, 22,800 of 78,083 polycell-years (29.2%) hold more harvested area than their own ceiling, 308.7 Mha in total. AM-5 risk 26: `.normalize_to_cropland()` and `.get_area_code_from_grid()` deleted. Both were unreachable and both keyed on the physical cell. AM-5 risks 16 and 17, in `inst/scripts/run_spatialize.R`: the `(lon, lat)` re-attachment of `country_grid` is gone and the hand-copied compartment keys now call `.compartment_id_cols()`. Its call to `build_gridded_landuse()` also passed three arguments the engine has not accepted since `config` was introduced, so the script could not run past its first call; fixed. All three are pinned by static checks with positive controls, because the script is neither linted nor executed. Value movement: none. Against pristine b6b36e9 on the deployed crosswalk (68,527 polycells, 3,925 multi-polity cells) at 2000 and 2015, land use is `identical()` over 3,702,607 rows and livestock over 832,386; per (country, crop/species, year) share-vector L1 is 0 for all 15,322 and 2,686 groups, and no cell moves by more than 1e-12. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

DA-13 had `build_polycell_support()` carry `polity_frac` and pad itself with the deployed crosswalk's unreproducible rows, so every unmigrated consumer was provably unchanged while consumers migrated one at a time. That is the only reason C3a, C3b, C5, C7 and C8 could each be attributed a movement. The transition is over for the consumers on our side of DA-28, so it goes. Removed: the `polity_frac` output column and `.pcs_add_shim()`; the `coverage_status == "crosswalk_only"` padding and `.pcs_append_crosswalk_only()`; the exported `polycell_shim_view()` and the `.pcs_strip_diagnostics()` helper that existed only to serve it; and the four now-dead `crosswalk_only` filters in the water-unmatched, overlap, footprint and unassigned diagnostics. Kept, deliberately: `cell_area_ha`. DA-13 named it a shim column, but it is not one. `.nd_polity_share()` requires it and `deposition_kgn_ha` is the cell mass divided by it (`n_deposition.R:241,491`), and `.carbon_support_to_area_code()` requires it too (`carbon_balance.R:1282`), so dropping it aborts two MIGRATED consumers. `data$crosswalk` and `data$crosswalk_year` also stay: they now feed the DA-12 footprint reconciliation alone, which reports the same disagreement the padding used to absorb. The two DA-13 validators are flipped rather than deleted. `test_polycell_support.R`'s "the shim reproduces the crosswalk bit-for-bit" becomes "the DA-13 shim is gone and a crosswalk cannot revive it", asserted on the same fixture and with `data$crosswalk` supplied, so the removal is proved on the path that used to produce the shim rather than on one where it could not have appeared. `verify_polycell_support.R`'s DA-13 section now ABORTS on a `polity_frac` column, a padding row or an unmeasured row instead of printing `identical: TRUE`. `"polity_frac"` is dropped from `utils::globalVariables()` LAST, per AM-4. What that exposes, measured with `codetools`: exactly 5 call sites, all in the three files this PR does not migrate — `water_balance.R` x3 (`.wb_finalise`, `.wb_drop_polity_cols`, `.wb_aggregate_polity`, DA-28), `feed_lpjml.R` x1 (`aggregate_grass_to_polity`, frozen by AM-2) and `feed_intake_redistribute.R` x1 (`.grass_to_cells`). No migrated consumer is among them. Value movement is EXACTLY ZERO, proved rather than asserted. Built on 61 real polity intervals with the real GLWD water, Natural Earth ice, LUH2 and the deployed 68,527-row crosswalk, pre-C9 against post-C9: 74,131 rows -> 7,456, the difference being 66,675 padding rows and nothing else; all 20 shared columns `identical()` in the same row order; all five area masses delta exactly 0; and all nine diagnostic attributes `identical()`, which is what proves the four removed filters were dead. `sum(land_area_ha)` over the unfiltered table goes from NA to 953,941,550.84999514 ha. The whole-ledger measurement is in the report, not here, but its headline belongs on the record: at polity grain the migration's ledger movement is entirely in the deposition INPUT term, -678,612.5 t N (-2.410% of the term, -0.504% of `n_input_full_t`), of which -648,491.3 t (95.6%) is nine polities losing their whole deposition to the DA-23/DA-28 identity gap and only -30,121.2 t (0.107% of the term) is geometry. The deposition SPLIT KEY -- the thing C3a changed -- contributes exactly 0 on 27 of 28 ledger quantities and 1.26e-16 relative on the 28th. Synthetic (C5) moves by exactly 0 t. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

S-A8. `inst/scripts/reconcile_polity_areas.R` measures each polity's polycell total against its official national area and books every material discrepancy to one owner. It changes no producer code: a wrong area is a finding. The reference is FAOSTAT Land Use (RL) from the package's own pin, and the choice of quantity is load-bearing. FAO's Country Notes define "Country area" as land + inland waters + COASTAL waters, which puts Canada at 15.64 Mkm2; "Land area + Inland waters" is used instead, and it is immune to France, whose country area never absorbed its coastal waters (residual -5,732,300 ha). EA6 and AM-29 are BOTH right and neither is a correction of the other: the crosswalk census gives 23 of 191 on the LUH2-land weighting and 21 on the whole-cell weighting (at 45%: 25 and 24). EA6's "Bahrain 46%" is 45.9% on land; AM-29's 29.9% is the same country on cells, and Croatia is 63.9% / 46.5%. Both weightings are now printed side by side so the ambiguity cannot recur. The four island states no longer draw a whole cell: 34.3x / 17.5x / 15.3x / 10.3x of their official area under the crosswalk, 1.18 / 1.00 / 1.16 / 0.58 now. Greenland is the case that proves the reference must be read rather than assumed: FAO states "Greenland 'Country area' refers to area free from ice", so comparing 213 Mha of territory against it reported +419% until the ice was taken off WHEP's side, and -12.9% after. Separating the polygons two live polities both hold moved 16 further polities out of "polygon quality". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

T-A12 closes the three defects the T-A1 verifier logged against `inst/scripts/diagnose_polycell_support.R`. (a) `.pc_polities()` read `whep::polities` and nothing else, so every finding in section P -- the status vocabularies, the dead-row count, the DA-2 code-versus-column disagreement -- could only ever describe whatever snapshot happened to be embedded. `WHEP_POLITIES_PATH` now points the section at any other polity table (an `sf` saved with `saveRDS()`, or any layer `sf::st_read()` opens), the source is printed, and the DA-2 count carries a note saying it belongs to the vintage rather than to the package. Proved end to end: with a table whose two offending `end_year` values are corrected, the count goes 2 -> 0 and the source line names the injected file. Note for the record: the symptom the verifier saw is already gone for an unrelated reason. `whep::polities` is now the re-synced 749-row table, not the 603-row one, so the checked-in artifact prints "2 disagreeing codes" (`NNG-1949-1963`, `TAN-1922-1964`) and agrees with DA-2 rather than contradicting it. That was an upstream re-sync, not a fix -- the missing injection point is what actually gets fixed here. (b) The LUH2 time origin was assumed (`year - 850 + 1`). It is now read from `nc$dim$time$units` and the year vector built from the axis values, with an abort naming the covered span. DA-9 makes vintage switching a supported operation, so a tree with another origin would have been sliced at the wrong year confidently. The file in hand declares a 850 origin over 1173 steps, 850-2022, which the script now prints instead of deriving. (c) Section H hardcoded 2014 and its own 1850 origin while every other section ran at 2015. It now takes the script's evidence year, prints that year next to the mass, and reads its origin from the file too (1850-2020). The plan's EA9 and DA-10 HaNi figures are 2014 measurements, so `WHEP_DIAGNOSE_YEAR=2014` reproduces them exactly: 34.77 Tg NHx, coastal median 0.9937 (n = 53,624) against an interior control of 1.0000 (n = 2,192,108). At 2015 the mass is 35.25 Tg and the coastal median 0.9935. EA1, EA2, EA3, EA6 and Z are byte-identical to the pre-fix run. T-A8 puts the load-bearing consequences in user-facing documentation rather than only in commit messages. `build_polycell_support()` gains two sections and a `@return` paragraph. The four-way land ladder is stated with its numbers -- whole cells 14.3195 Gha, HaNi's mask 13.5977, the polity polygons 13.2795, LUH2 terrestrial 12.9931 -- with the polygon row flagged as vintage-dependent rather than a constant, and the GLWD/CRU mask named as a fifth that is reconciled instead of joined away. `ice_area_ha` is documented as a present-day snapshot that does not vary historically, acceptable only while ice is a reporting category and not a driver. And identity is stated to be `polity_code` alone: measured on the shipped crosswalk, 505 polity codes fold into 201 reporting buckets, 113 of which hold more than one polity and one of which (206) holds Sudan and South Sudan at once, so the conversion belongs at the consumer's boundary where the fold is visible. `build_n_deposition()` gains "Which land definition governs what": WHEP's territory governs PLACEMENT, HaNi's land mask governs the TOTAL. Both rejected alternatives are recorded with their cost (a whole-cell rate times `land_area_ha` sheds ~9% of the source mass; re-referencing to HaNi's own mask moves the total ~4.5%), and the identity that is NOT available is named so nobody looks for it. `read_n_deposition()` gains the HaNi `@source`, verified against Crossref rather than remembered. NEWS.md gains the migration notes: what a consumer changes, the measured movement (all of it in the deposition input term, -0.504% of `n_input_full_t`, 95.6% of it nine polities losing their deposition to the DA-23/DA-28 identity gap and only 0.107% geometry), the island-state fix, the Greenland reference mismatch, DA-14's scope default and why the terrestrial-only scope would have been wrong, and the five surviving `polity_frac` call sites -- `water_balance.R` x3, `feed_lpjml.R`, `feed_intake_redistribute.R` -- stated as deliberate, out of scope, and inside a NOTE the package already had. Verified rather than inherited: 13.5977 Gha and 12.9931 Gha re-measured here, 14.3195 Gha reproduced by the diagnostic, and the crosswalk fold measured directly. Gate, all against a `git archive` of 7d1858d in the same environment: `air format --check .` 0; `devtools::document()` touches only the three expected `.Rd`; `lintr::lint_package()` 0 lints in 20.7 min; `rcmdcheck()` 0 errors / 0 warnings / 1 NOTE with the note text BYTE-IDENTICAL to pristine; pkgdown reference index builds clean with 0 man topics missing from `_pkgdown.yml`. Test suite at block level: 2,039 blocks both sides, 0 added, 0 removed, 0 shared blocks changed on failed/skipped/error/warning/passed/nb, failing and erroring sets identical -- 18 expectations across the same 3 `test_polycell_support.R` blocks and the same 1 HWSD error (#596). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Second reconciliation with `main`, which moved 53 commits since the first one's merge base. Two conflicts, both insertion-versus-insertion, both resolved as a lossless union rather than a choice: - `NAMESPACE`: each side added an export at the same alphabetical slot. Both kept. `devtools::document()` then regenerates the file with no change, which is what proves the ordering rather than my asserting it. - `R/n_balance_spatialize.R`: #579 added the `.cell_polity_*` area-key helpers and C3a/C5 added the `.n_*` split helpers, both immediately before `.n_check_totals_shares()`. Both blocks kept, with main's left exactly where main put it so the next merge does not re-conflict here. The merged file differs from `1c7edf86` by precisely main's 149/6 and from `origin/main` by precisely our 187/20, so no line from either side was dropped. The other fourteen shared files auto-merged; `R/polities.R`, `R/utils.R`, `tests/testthat/test_n_balance_spatialize.R`, `NEWS.md` and `_pkgdown.yml` were each checked to be the same lossless union. #473 was already in the merge base, so T-A11 stays superseded: `read_luh2_landuse()` carries main's `states_source` vintage selector untouched, and this branch's `area_basis` land-definition selector beside it. There is no second vintage mechanism to reconcile. #618 has not merged, so there is nothing to absorb into `R/soil_carbon_inputs.R` yet. The merge leaves that file, `R/feed_lpjml.R`, `R/water_balance.R`, `R/n_urban.R` and `data-raw/balance_coefficients.R` untouched, and main changed none of them either. #579's "the `grid` default reproduces the parquet bit-for-bit" claim is verified, not trusted: `build_cell_polity()` read from the deployed parquet is `identical()` across the merge, 68,527 x 5, every column at max |diff| 0. `grid_area_code` is added only under `area_key = "polity_area"`, which this branch does not opt into. The default does now emit a diagnostic warning naming 8 off-bucket grid codes over 1,069 cells; no test block's warning count moved because of it. `whep::polities` did not move: `data/polities.rda` is the same blob at the merge base, at `1c7edf86` and on `origin/main`, and measures 749 rows / 20 columns / 749 distinct codes on both sides of the merge. AM-44's vintage-dependent area figures therefore stand unchanged. Block-level control against a pristine `git archive 1c7edf8` run in the same environment with identical printed env fingerprints: 47 blocks added and 1 removed, every one of them in the twelve test files main changed (the removed block is a rename main made in `test_energy_co2_extension.R`); 0 shared blocks changed fail, error or skip; the failing set is identical, still the 18 deferred `test_polycell_support.R` expectations plus the one pre-existing HWSD error. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The 18 assertions AM-25 deferred were failing on every CI platform. They are census and identity pins against the old polity table, not behaviour: `R CMD check` treats a failing test as a hard ERROR, and the local gate had been running `rcmdcheck(args = "--no-tests")`, which skips tests entirely. The producer is untouched. Blocks 1 and 2 used `GRC-1830-1913`, whose `wiki_status` the #551 re-sync turned to `superseded`. DA-7's live filter drops it, the build returns 0 rows with no error and no warning, and every assertion then ran against an absent frame. Re-pinned onto its live successor `GRC-1881-1913`, the same `cshapes-europe` feature 350 at a later vintage: it reproduces the defect on all seven originally pinned cells plus 401255 in the Ionian islands the 1886 polygon adds. 8 terra pieces / 530,378.73 ha / 8.34% of the polity, corroborated at 530,417.74 ha by a planar clip measured in an equal-area projection, which shares neither engine with the producer. Both blocks now assert the liveness precondition, so a future supersession fails by name instead of cascading. The two assertions AM-25 flagged as passing vacuously on empty results are guarded rather than rewritten: the census (63 rows) and a positive total are pinned first, so the setequal and the sum have something to compare. Same treatment for the two zero-sums over `outside`. Block 3's long-edge census moves 42 -> 43 edges, 30 polities either way. The physical defect is unchanged: the 49th parallel is the same segment at the same span and bulge, now asserted as one identical triple shared by all three live Canadian intervals, which a relabelling cannot move. `CAN-1866-1886`/`CAN-1886-1948`/`CAN-1948-2025` are retired in favour of `CAN-1870-1886`/`CAN-1886-1949`/`CAN-1949-2025`; that they are retired and absent from the census is now asserted from the table itself. Verification: `devtools::test(stop_on_failure = TRUE)` in the offline job's own configuration FAIL 0; `rcmdcheck()` WITH tests 0/0/1 with `WHEP_*` unset as on CI, and 1 ERROR / 0 / 1 with `WHEP_HWSD_DIR` set, that error being the pre-existing #596; `lint_package()` 0 lints; `air` and `document()` clean. Block-level control against a pristine `git archive` of 79d1ec3 in the same environment: 2085 blocks both sides, 0 added, 0 removed, fail 18 -> 0, and the only three blocks that move at all are the three re-pinned ones. Mutation sweep: control alive, 6 mutants dead, including a 1e-9 relative perturbation of one Canadian bulge that only the new identical-triple assertion catches. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`area_engine` is decided by the platform's s2 build, not by WHEP: s2 validity turns on ULP-level degeneracy, so the clipped pieces the Windows and Linux runners refuse are read back on macOS ARM64. Eight expectations pinned the Windows classification and failed there (run 31154665994) -- five in "an unreadable clip piece is measured, never dropped", three in "an ice layer over an unreadable piece does not abort". Both blocks now assert the property and never the engine: the piece census (63 polycells, all 8 recovered cells present, every area finite and positive), the recovered area against T-A13's engine-independent reference (planar clip, Lambert azimuthal equal-area, 530,417.74 ha), the partition re-aggregating to the polity, and the ice confined to its rectangle. Tolerances are derived, not rounded: from the s2/terra spread measured on this polity's own 55 readable pieces (+2.23e-04 to +9.88e-04 at latitudes 36.25-39.75 -- the global 0.45-0.86% band does not apply, the ellipsoid-over-sphere ratio crosses 1 near latitude 35.3) and from the great-circle bulge of the ice rectangle's edges (271.31 / 92.76 / 476.27 ha of Greek land). Each is stated at its assertion. The `expect_warning("could not measure")` calls become a link -- the producer warns exactly when it substituted -- so a platform needing no substitution is legal while a silent substitution still fails. A new `pcs_s2_refuses()` helper reads the hazard from `.s2_repair()` rather than from `area_engine`, so neither block can pass vacuously once no piece is unreadable. Tests only; no producer change and no published value changes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…support # Conflicts: # NEWS.md # R/spatialize_compartments.R # tests/testthat/test_carbon_balance.R

eduaguilera · 2026-08-07T22:30:33Z

Final post-main acceptance (head 202e5d6, 2026-08-08):

Merged origin/main at f8e55e5 and re-pinned PR Refresh the polities snapshot: geometry now loads under s2 #662's polity snapshot: 751 snapshot rows, 692 live polities, 414,479 measured/retained pieces, 21 terra pieces across 9 polities, and zero DA-22 long edges. The exact full clip took 3,843.01 s.
The census verifier now gates the checked blob, requested/output code-set identity, coverage/status completeness, engine census, S-A2 exceptions, and the all-runtime terra/s2 residual envelope. Its advertised GRC subset reaches Done with 63 pieces, 4 terra pieces, and 227,311.6 ha.
The macOS sf/GEOS failure is closed: missing or unusable intersection idx metadata now falls back to exact per-source intersections. The focused polycell file is green locally with 424 expectations.
Final CI is green: format 31221505737, lint 31221506959, offline tests 31221507000, coverage 31221506978, pkgdown 31221506957 (attempt 2), and all five R-CMD-check platforms including macOS 31221506958.

…ation) Reconciles 51 upstream commits (merge base f8e55e5). Four conflicts, each resolved as a lossless union and verified by diffstat against both parents: * NAMESPACE - insertion vs insertion; both exports kept. devtools::document() regenerated it with no change, which is what proves the ordering. * NEWS.md - insertion vs insertion under one heading; both blocks kept, main's left where main put it so the next merge does not re-conflict on the same seam. 0 deletions either side. * R/polities.R - upstream extracted .area_year_span_conflicts(), so `spans` became that helper's ARGUMENT and is read after the Map(). Our stale `spans <- Map(` would have shadowed it and left `long` undefined, so upstream's code is taken verbatim. Our DA-24 note is re-homed onto .area_year_polity_conflicts(), the declared-span caller where it is true, and re-measured: on polities 751 / 0e52f1f the two readings now DISAGREE (declared 0 ambiguous pairs, open-end-widened 1 - area 7 at 1975, AGO-1975-2025 vs ANG-1905-1975). That is upstream #683. The old note claimed they agreed; it was measured on a superseded snapshot. * tests/testthat/test_run_spatialize.R - append vs append at one seam; both blocks kept. data/polities.rda and data/polity_area_crosswalk.rda are byte-identical across this merge (0e52f1f), so the polycell census pins already keyed to "polities 751 / 0e52f1f" are not invalidated and were not re-pinned. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

… green The merge was textually clean on these files and semantically red: 4 of the 70 test blocks main added fail against this branch's producer. None is a regression -- the block control shows 0 shared blocks changing fail/error/ skip -- and none is fixed by weakening this branch's guards. * test_spatialize.R (x2) and test_spatialize_livestock.R (x1): main's new `.warn_grid_missing_reporters()` fixtures build a `country_grid` with no share column, which is exactly what C8/S-A5 forbids -- on main a share-less grid defaults to `cell_area_frac = 1` and hands a border cell wholly to one polity, and `.abort_missing_polity_share()` refuses rather than defaults. Each fixture cell is owned outright by the single polity keyed on it, so an explicit `cell_area_frac = 1` is the honest value; the missing-reporter behaviour under test is unchanged and the guard is untouched. * test_join_audit.R: main's new whep#669 gate requires every year-free territorial join to be classified, and this branch adds one -- `.carbon_warn_fold()` semi-joins on (lon, lat, area_code). Classified as `diagnostic` rather than dodged: both sides of that join are the SAME carbon support, already filtered to `.carbon_support_year()`, so there is no second year for the key to disagree about, and the join selects which polity codes to name in the DA-23 fold warning rather than reaching a value. The enumerated cap moves 58 -> 59 with the reason in the row, which is the one direction that comment permits. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Four textual conflicts, all additive: NAMESPACE and NEWS.md concatenate, R/polities.R takes main's `.area_year_span_conflicts()` extraction with this branch's DA-24 note moved into the function header, and test_run_spatialize.R keeps both appended test sections. Two integration failures neither branch could see on its own: * main's grid-vintage and re-keyed-grid fixtures carry no share column, which C8's `.normalize_country_grid()` now refuses. Every cell in them is held whole by one reporting area, so the explicit share is 1. * main's whep#669 join gate caught `.carbon_warn_fold()`'s `semi_join` on (lon, lat, area_code). It is classified `diagnostic`: it names the folded polity codes for the warning text alone and cannot reach a number, and both sides are the one snapshot `.carbon_support_year()` has already resolved. `devtools::test()` clean, `lint_package()` 0, `roxygenise()` no drift. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…tal inflation

Merges origin/main (753-row polity snapshot) and closes the three things that kept the default paths from running at all. THE PIN. `polycell_support` is published at 20260810T165201Z-7caf6 and frozen in `inst/extdata/whep_inputs.csv`, so `read_polycell_support()` resolves and the five carbon readers stop aborting. Built from all four layers on the 753-row snapshot; `polity_area_ha == land + inland_water + ice` holds to 1.8e-9 ha over 480,866 rows with no negative area, land at 2015 is 12.9253 Gha against LUH2's independent 12.9931, and ice is 231.4 Mha, the non-Antarctic glaciated area. THE GRID DEFAULT. `.load_country_grid()` takes polycell/centroid/fraction and defaults to the first, read through `.carbon_cell_support()` so the spatialization and carbon paths cannot land on different reference years or folds. `"centroid"` stays selectable to reproduce a published run: it carries no share, which C8 refuses. THE CENSUS. Re-measured on the new snapshot rather than adjusted: live rows 692 -> 694, clipped 664 -> 666, polycells 414,479 -> 414,485, all from Aruba and the Holy See. The terra figures are IDENTICAL to the hectare -- 21 polycells over 9 polities, 1,429,276.70 ha -- which is worth recording because they were first measured on Windows and re-measured here on Linux. Publishing the pin then ran four code paths that had never executed: * `read_luh2_landuse(area_basis = "polycell_land")` -- its real-data test had always taken the centroid fallback. `fraction` is a share of the COMPARTMENT and sums to 1 per (lon, lat, area_code); the test summed over (lon, lat) alone, adding one whole partition per polity of a border cell and reading up to 4.0 across 4,603 cells. Equivalent while every cell had one row. * `.load_country_grid(NULL)` -- pinned by asserting an error while the pin was unpublished, so it stopped testing anything once it was. Now mocks the support reader and asserts the default reads it. * main's "unknown country_grid source" used "polycell" as its bogus value, which is now the default and would have passed vacuously. * `data/whep_inputs.rda` regenerated for the frozen version. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…ared defect Edu pushed the same four semantic-conflict fixes independently while this branch was being merged onto the newer main. Both passes reached the same conclusions, so where they differ this keeps the upstream wording: * the three spatialize fixtures take his comment, which names `.abort_missing_polity_share()` and says the guard is not weakened to accommodate the fixture. * `.carbon_warn_fold`'s baseline row keeps his entry, alphabetically placed. BOTH sides added one and they auto-merged as separate adjacent lines, so the ratchet read 61 and the gate caught the duplicate. * the cap stays 60, not his 59: this branch merges a newer main where whep#698/#709 had already moved it to 59 on their own, so `.carbon_warn_fold` is the sixtieth. His reasoning for the row is kept verbatim. NEWS.md keeps one copy of the polycell entry; his branch re-added it lower in the file after this branch had already placed it at the top. devtools::test() clean, lint_package() 0, roxygenise() no drift. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

github-actions · 2026-08-10T17:25:31Z

+  if (length(v2) > 0L) {
+    return(list(path = v2[[1L]], version = "v2", dir = v2_dir))
+  }
+  v1 <- c(file.path(glwd_dir, "glwd_3", "hdr.adf"), file.path(glwd_dir, "glwd_3.tif"))


[air] _{reported by reviewdog 🐶}

Suggested change

v1 <- c(file.path(glwd_dir, "glwd_3", "hdr.adf"), file.path(glwd_dir, "glwd_3.tif"))

v1 <- c(

file.path(glwd_dir, "glwd_3", "hdr.adf"),

file.path(glwd_dir, "glwd_3.tif")

)

github-actions · 2026-08-10T17:25:31Z

+      1L, 1L, 0L, 0L,
+      1L, 1L, 0L, 0L,


[air] _{reported by reviewdog 🐶}

Suggested change

1L, 1L, 0L, 0L,

1L, 1L, 0L, 0L,

1L,

1L,

0L,

0L,

1L,

1L,

0L,

0L,

github-actions · 2026-08-10T17:25:31Z

+      100, 100, 100, 100,
+      NA, NA, 100, 100,


[air] _{reported by reviewdog 🐶}

Suggested change

100, 100, 100, 100,

NA, NA, 100, 100,

100,

100,

100,

100,

NA,

NA,

100,

100,

…ch's one row Two conflicts, both in files this branch has now merged four times. `test_join_audit.R` is the substantive one, and it moves in BOTH directions: main's ratchet fell to 58 because `.nbx_image_region` left -- the deposited Schulte-Uebbing crosswalk carries IMAGE membership on the canonical cell key, so the gridded boundary no longer reaches IMAGE through a year-free country join. This branch's `.carbon_warn_fold` then puts it at 59. Main's reason for the fall is kept and this branch's reason for the rise appended, and the cap is read back off `.territorial_join_baseline()` rather than inferred: it sums to exactly 59. Arithmetic would have agreed with a duplicated row too, which is how the previous merge briefly reached 61. NEWS.md concatenates, as before. devtools::test() clean, lint_package() 0, roxygenise() regenerates only the man pages main's own commits touched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

eduaguilera and others added 30 commits August 4, 2026 09:28

eduaguilera requested a review from lbm364dl August 7, 2026 04:07

lbm364dl mentioned this pull request Aug 7, 2026

fix(spatialize): stop gridded output carrying two disagreeing territorial keys #653

Merged

eduaguilera added 7 commits August 7, 2026 20:15

test(polycell): pin census to snapshot identity

ca7ee33

fix(polycell): normalize polygonal intersections

075ccfd

Merge remote-tracking branch 'origin/main' into edu/polycell-spatial-…

7e8def1

…support # Conflicts: # NEWS.md # R/spatialize_compartments.R # tests/testthat/test_carbon_balance.R

test(polities): repin refreshed snapshot assertions

90e5907

fix(polycell): restore scalar intersections without idx

e6639d6

fix(polycell): recover intersections without source indices

ce36fe0

style(polycell): shorten intersection helper name

202e5d6

eduaguilera and others added 6 commits August 10, 2026 12:40

Move the GLWD water layer onto the official download and fix its coas…

e1dce5e

…tal inflation

github-actions Bot reviewed Aug 10, 2026

View reviewed changes

Navigation Menu

Replace the two cell-polity crosswalks with the polycell, and move the area denominators onto it - #619

Replace the two cell-polity crosswalks with the polycell, and move the area denominators onto it#619
eduaguilera wants to merge 46 commits into
mainfrom
edu/polycell-spatial-support

eduaguilera commented Aug 7, 2026 •

edited

Loading

Uh oh!

eduaguilera commented Aug 7, 2026

Uh oh!

github-actions Bot Aug 10, 2026

Uh oh!

github-actions Bot Aug 10, 2026

Uh oh!

github-actions Bot Aug 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

eduaguilera commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

The unit

What was wrong

Measured movement, per consumer

Two decisions worth disagreeing with, if you do

Evidence

Known, and deliberately not fixed here

Also in here

Uh oh!

eduaguilera commented Aug 7, 2026

Uh oh!

github-actions Bot Aug 10, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot Aug 10, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot Aug 10, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

eduaguilera commented Aug 7, 2026 •

edited

Loading