CHELSA present-day baseline (#448) — the reference the future projections are measured against

[cboettig]

Adds the CHELSA v2.1 1981–2010 observed baseline at H3 resolution 8, which is what makes the #564 future projections interpretable. Until now a consumer could not compute warming from this catalog at all — there was nothing to difference against.

Published and verified: verify-stac.py exits 0 against live S3.

What ships

collection	chelsa-2-1/baseline-1981-2010
cells	195,048,994 — identical to all nine future collections, so every cell pairs
columns	11 — seven observed variables plus h8/h5/h4/h0
parents	5,4,0 as scoped on the issue
size	4.05 GB

The schema deliberately differs from the futures, because the source does. This is an observational climatology — one raster per variable, no GCM members — so it carries plain bio1, bio4, … rather than five members plus median/min/max. Plain names so a delta reads directly:

SELECT AVG(f.bio1_median - b.bio1) AS warming_degrees_c
FROM   <future> f JOIN <baseline> b USING (h8)

Validation: Arctic amplification, reproduced

Global land mean for 1981–2010 comes out at 9.409 °C, in line with published land climatology. Joining against the projections at h0 79.2°N (mean cell latitude 76°N), across 456,928 cells:

scenario, 2071–2100 vs baseline	warming
SSP1-2.6	+3.64 °C
SSP3-7.0	+7.05 °C
SSP5-8.5	+8.32 °C (range 4.61–12.10)
cells not warming under SSP5-8.5	0 of 456,928

Against a global land mean of +4.7 °C, that is ~1.8× amplification at high latitude. Nothing in the pipeline can manufacture that — it requires the baseline, the futures, the land mask, the units and the H3 alignment to be simultaneously correct. Structural checks also pass: zero nulls, zero parent mismatches on h3_cell_to_parent.

Reuses the #564 Armada pathway

854 hex jobs (one per h0 × variable) then 122 joins, at the measured 8 Gi / 4 cores. Raw was already staged during #564, so no download. The hex phase finished with zero gaps — consistent with the ~0.02% transient failure rate measured over ~30,000 jobs there.

Two defects caught before publishing

A path mismatch. The data landed at baseline/1981-2010/ while the STAC pointed at baseline-1981-2010/. With a slash the last path segment is 1981-2010, which would have produced the asset key 1981-2010-hex — undescriptive and the same naming defect caught during #564. Moved the data (108/108, old path emptied) rather than degrade the key.

An ephemeral over-request, now recorded in the hex-tuning skill. One join asked for 40 Gi of ephemeral storage to write a 2 KB partition and cycled Leased → Pending → LeaseReturned indefinitely, because no node had that much free — while 121 identical jobs placed fine. It did not run slowly; it never ran, and nothing in the logs says the request is too large.

That makes three dimensions over-requested on this build by inheriting a number instead of measuring it — memory (~6×), cpu (~2.4×), ephemeral (enormous) — each throttling throughput a different way. The skill now covers all three with measurement recipes.

Scope

Seven variables, matching the futures so the pair is symmetric and every column has a defined delta. Extending both to all 19 bioclim variables stays a follow-up.

#448 stays open: Köppen-Geiger (CC-BY-4.0, categorical, mode reducer) and WorldClim (non-redistributable, MinIO-only) are separate ingests with different licences and reducers, not touched here.