k8s(prod): promote v0.8.13 by digest

[cboettig]

On hold — see #364. v0.8.13 carries the #357 coarse-resolution overlay example, which reads
nland from the …-hex-weights-res8/-res9 assets. data-workflows#524 removed that column an hour
after #357 merged, so promoting this digest would ship an example that is a binder error on any
res-8/res-9 question. Fixed in #365; this promotion should be re-cut on a tag that includes it and
re-gated. The gate below stays valid for what it covered — it exercised the res-10 path
(CWHR13 percent-conserved), which is unaffected — but it never touched the broken branch, so a
re-gate wants a cell that asks a res-8 or res-9 feature question.

Promotes prod from v0.8.12 to v0.8.13.

-  image: …:v0.8.12@sha256:332d6df730c1fc66e7b2d317b121828fcb3785339a3a43b49789c3e22e6e0ffd
+  image: …:v0.8.13@sha256:1a3409d56f62bc8c9ba252a8960eb0dad16a240e24c400620eb7afae5233db7f

Digest resolved via the GHCR anonymous-pull manifest HEAD (docker CLI unavailable in this environment). Method validated: resolving v0.8.12 the same way returns sha256:332d6df…, exactly the digest this commit replaces. The v0.8.13 tag and the 19b3c709 git-sha tag resolve to the same digest.

What ships

PR
#357	fractional-overlay examples weight by h3_cell_area, use the published hex-weights assets, bound the denominator with the ecoregion land grid
#360	rebuts the "area terms cancel" rationalization that survived #357
#358	opt-in experimental raster+zarr image — separate tag and endpoint, inert here (EXTRA_DUCKDB_EXTENSIONS unset on the default image)
#351	guidance-validation runbook fix

Gate result (dev @ 19b3c70, regression tier, 3 trials × 2 models)

model	t1	t2	t3
z-ai/glm-5.2	33.3 ✅	33.3 ✅	33.3 ✅
deepseek/deepseek-v4-flash-0731	33.3 ✅	33.3 ✅	33.3 ✅

Zero occurrences of the pre-fix 28.8% form, and zero of the 32.1% marginal form seen in the first gate. Every trial carried the area term and applied the California mask; deepseek stated the ratio explicitly as 100 × SUM(frac × (w1+w2) × cell_area) / SUM(frac × cell_area).

Worth recording: 33.3 is only reachable with both corrections — unweighted+mask gives 32.07, weighted-without-mask 29.89 — so the value corroborates the method independently. But 32.07 rounds inside the cell's ±3pt tolerance, so tolerance alone never separates a correct method from a lucky one; this cell needs a method audit.

Baseline unchanged: Mojave 7,373,400 · GAP 1+2 26,471,500 · hardwood woodland 13.6% · California 26.08–26.1% · BioRank 21.18–21.2 · channelized 22.5% (last two matching the 2026-08-03 run).

Deploy steps after merge

kubectl apply -f k8s/deployment.yaml
kubectl rollout restart deployment/duckdb-mcp -n biodiversity
kubectl -n biodiversity get pods -l app=duckdb-mcp \
  -o custom-columns='NAME:.metadata.name,IMAGE:.status.containerStatuses[0].imageID'   # all 6 must converge

Note

geo-agent-benchmark gold for ca-cwhr13-pct-conserved was corrected in geo-agent-benchmark#13 — it had encoded the same unweighted formula as the guidance, so the suite could not have caught this. data-workflows#523 (res-9/res-8 STAC rollup recipe) is still open and contradicts the shipped guidance at the per-dataset layer; it does not block this promotion but should land soon after. Stale as written: the per-dataset STAC text was corrected by data-workflows#524 on 2026-08-07, before this PR was opened, and #523 is closed as superseded — there is no remaining contradiction at that layer. What #524 also did was drop nland from the weights assets, which is what puts this promotion on hold (#364).

[cboettig]

Hold — this would ship a h3-guide.md example that errors. Filed as #364.

h3-guide.md:326 (the coarse-resolution overlay block from #357) reads nland from the
…-hex-weights-res8 asset. data-workflows#524 removed that column an hour after #357 merged, so the
example is now a binder error, not a wrong number:

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

Same problem in the res-9 formula at line 336 and the prose rule at line 318. The res-10 path this
PR's gate exercised (CWHR13 percent-conserved) is unaffected and correct — which is why the gate is
green: it never touches the broken branch.

The fix is a source change, not a method change: land_area_km2 on ca30x30-ecoregion's
ecoregion-hex-res9/-res8 rollups, which gives 26.135% identically and matches what the
-hex-weights-res9/-res8 asset descriptions already document. Details and a suggested diff in #364.

One stale line in the description above: data-workflows#523 no longer contradicts the shipped
guidance. The per-dataset STAC text was corrected by data-workflows#524 on 2026-08-07, before this PR
was opened, and #523 is now closed as superseded.

[cboettig]

Closing — superseded before it could be merged.

This promotes v0.8.13, which contains the coarse-overlay binder error #364: the example reads nland from …-hex-weights-res8, and data-workflows#522 removed that column (replacing it with land_area_km2 on the land grid). Verified against live data:

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

Merging this would have shipped a res-8 example to production that cannot execute. The gate did not catch it because the regression tier has no res-8 question at all — see geo-agent-benchmark#21.

Replacement path: merge #365 (fixes #364), tag v0.8.14, promote that. Superseded by the promote PR for v0.8.14.

Note prod is still on v0.8.12 and therefore still serving the pre-#357 unweighted overlay guidance — the silent-wrong-answer bug from #356. That is the reason to move quickly on v0.8.14 rather than batching more work into it.