fix(#364): coarse-overlay example weights by land_area_km2, not the removed nland

[cboettig]

Fixes #364.

h3-guide.md's coarse-resolution overlay example read nland from the …-hex-weights-res8 / -res9
assets. data-workflows#524 removed that column (2026-08-07, an hour after #357), so the example was a
hard SQL error on any res-8/res-9 question:

SQL Error: Binder Error: Table "p" does not have a column named "nland"

Source fix, not a method change: the weight now comes from the region's land-grid rollup
(ca30x30-ecoregion, assets ecoregion-hex-res9 / -res8), which publishes an exact per-cell
land_area_km2 — the summed area of the actual res-10 land children rather than a count times the
parent's area. That is also verbatim the form the -hex-weights-res9 / -res8 asset descriptions
document, so the injected guide and the per-dataset text stay on the same form (the point of #357).

Verified against prod

Statewide GAP 1+2 at res 8, with the new block:

SELECT 100 * SUM(COALESCE(p.w1 + p.w2, 0) * e.land_area_km2) / SUM(e.land_area_km2)
FROM read_parquet('s3://public-ca30x30/ecoregion/hex-res8/h0=*/data_0.parquet') e
LEFT JOIN read_parquet('s3://public-ca30x30/conserved-areas-terrestrial-2025/hex-weights-res8/h0=*/data_0.parquet') p USING (h8, h0);
-- 26.1348

matching the res-10 ground truth (26.135%) and the nland × h3_cell_area product the guide used to
name. Also ran the filtered three-table shape as written (Mojave Desert as a res-8 feature): 44.87%
at res 8 against 44.70% at res 10 — the residual is the coarse feature's own footprint spill, which
is inherent to asking at res 8, not a weighting error.

The two res-8 assets are 1:1 over the same footprint (525,034 rows each, zero land cells without a
weights row), so the LEFT JOIN + COALESCE is defensive here and load-bearing at res 10.

One deviation from the issue's suggested diff

#364 proposes a three-line block with a bare WHERE <feature filter> and no feature table. That
would not bind — the section's premise is a coarse feature overlaid on a finer layer, so the
feature has to be in the query. I kept the feat CTE from the current block and joined the land grid
to it:

WITH feat AS (
  SELECT DISTINCT h8, h0
  FROM read_parquet('<res-8 feature hex>')
  WHERE <feature filter>
)
SELECT 100 * SUM(COALESCE(p.w1 + p.w2, 0) * e.land_area_km2)
           / SUM(e.land_area_km2) AS pct_conserved
FROM feat f
JOIN read_parquet('<ecoregion hex-res8>') e USING (h8, h0)
LEFT JOIN read_parquet('<conserved-areas hex-weights-res8>') p USING (h8, h0);

Everything else follows the issue: line 318's rule names land_area_km2 and the land-grid assets,
line 331 becomes "the land grid is both the weight and the denominator", and the res-9 / frac
variants at line 336 drop to prose. Net: one fewer inline formula, same number of SQL blocks.

Relation to #362

#362 (prod v0.8.12 → v0.8.13) would ship the broken section, so it should take this commit and
re-gate. Its green gate exercised the res-10 path (CWHR13 percent-conserved), which is unaffected —
it never touched the broken branch. A gate cell asking a res-8 or res-9 feature question would.

[cboettig]

Verified against live data before merge — this is the check that would have caught #364, and it is not something the regression tier can currently do (geo-agent-benchmark#21).

Running this PR's recipe statewide, against the res-10 ground truth for the same scope:

-- this PR's recipe at res 8
WITH feat AS (SELECT DISTINCT h8, h0 FROM read_parquet('s3://public-ca30x30/ecoregion/hex-res8/h0=*/data_0.parquet'))
SELECT 100 * SUM(COALESCE(p.w1 + p.w2, 0) * e.land_area_km2) / SUM(e.land_area_km2)
FROM feat f
JOIN read_parquet('s3://public-ca30x30/ecoregion/hex-res8/h0=*/data_0.parquet') e USING (h8, h0)
LEFT JOIN read_parquet('s3://public-ca30x30/conserved-areas-terrestrial-2025/hex-weights-res8/h0=*/data_0.parquet') p USING (h8, h0);
-- 26.135

this PR, res 8	26.135
res-10 ground truth	26.135
pre-#522 nland count-weighting	25.684
the code on main today	binder error

Exact to three decimals. It also drops the h3_cell_area() call the old form required, so the corrected block is shorter than the one it replaces.

Ready to merge from my side. Next: tag v0.8.14 and promote by digest — #362 (promote v0.8.13) is closed as superseded, since v0.8.13 carries the #364 break.

Coordination note: further edits to this block are sequenced under #367, behind the gate-coverage work, so this should be the last patch here before that design pass.