Open Building Map as a feature, and as the residential split for Microsoft v8

[MFStark]

Adds Open Building Map (OBM) as a model feature, and a microsoft_v8_obm
product that uses it as Microsoft v8's residential split in place of GHSL.

Why

Footprint datasets record where buildings are, not what they are for — a
warehouse and an apartment block look identical from above. Every current
product multiplies total building volume by a proportion_residential that
comes from GHSL, a satellite classification. Nothing in the stack observes
building use.

OBM does: 2.69 billion buildings, each carrying an occupancy type from the GEM
Building Taxonomy.

What this adds

OBM features (obm_20250404_*) — six measures per block, collapsed from the
covariate's eight occupancy classes:

measure
density	sum of the eight parent classes, rescaled where they overlap
height	GHSL ANBH, one storey (2.5 m) imputed where OBM sees a footprint and GHSL sees none
volume, residential_volume	derived
proportion_residential	1 − nonres/total, with unknown counted as residential
p_observed	binary: did OBM observe this fraction, or was it imputed?

microsoft_v8_obm_20250404_residential_volume — Microsoft's own density and
height, untouched; only the residential fraction is swapped:

p = obm_proportion_residential   where OBM sees a building
    ghsl_proportion_residential  where only GHSL does
    1.0                          where neither does

Results

Feeding GHSL alone through the same chain reproduces the shipped
microsoft_v8_residential_volume to 5e-07. That exactness is what makes the
swap attributable — a difference cannot be our arithmetic.

Globally, across 4,797 blocks and 2.11B built pixels:

• +60.4 km³ of non-residential building volume — 1.54× GHSL's total
  (111.6 → 172.1 km³, moving 10.6% → 16.4% of all v8 building)
• OBM supplies the fraction on 48.6% of v8's built pixels but 62.5% of its
  volume — its pixels carry 1.29× the average building density
• 17.8% of pixels reach neither source and take p = 1, but those carry only
  5.7% of the volume

An earlier two-city pilot predicted 1.6×–3.4×. The global answer landing just
below that range is what you would expect: those cities are where OBM's
labelling is unusually complete.

Two things reviewers should know

Every non-residential figure is a floor. 61% of OBM's buildings carry no
occupancy label and our convention counts them as residential. Each one tagged
later can only move the number one way.

Height is not independent of GHSL. Both sides use GHSL's ANBH, so a volume
comparison reduces to area × use. The residential split is independent — it
comes from OSM tags.

Notes

• p_observed exists because proportion_residential cannot carry this: OBM's
  own p = 1 and the no-source fallback write the same number. On one test
  block 98% of OBM's pixels sit at p = 1 already.
• Only residential_volume is written for microsoft_v8_obm. The spliced
  fraction is recoverable as residential_volume / microsoft_v8_volume.
• Provider renamed obm_2025q2 → obm_20250404, after the GFZ upload the data
  came from rather than the quarter we filed it under.