feat(studio): ERA5 climatology — committed product, derived ambient state, stage-1 panel (task 1.1)

[aliakherati]

"I do not see ERA5 data being utilized" — fixed, end to end

The decisions existed (SCIENCE-1, BLOCKING-5); nothing implemented them. This PR is task 1.1
complete: fetch → committed product → derivation → wizard.

The committed product

data/pipelines/era5_zonal_monthly.py fetches ERA5 monthly means — 1991–2020 (WMO normal),
15 levels, 300–5 hPa, 2.5° grid (ASSUMPTION-9) — and reduces 343 MB of raw fields to a 0.1 MB
(month × level × lat) npz, committed at studio/science/data/ with a sha256 manifest. Raw ERA5 is
never redistributed; the derived product is, with attribution. The loader checksum-verifies once per
process and refuses to extrapolate outside 300–5 hPa (ADR-005).

Schema 0.4.0 — the ambient state gets a source

Same structure as the emission system:

	dataset = USER (default)	dataset = ERA5
entered	given_temperature_k, given_h2o_ppmv	(kept, inert)
derived	temperature_k = given, h2o_ppmv = given	interpolated at (latitude, month, pressure)

Pressure stays entered — it places the box, so it is the lookup's coordinate, not its result.
Default USER: every existing config resolves byte-identically. Provenance records the dataset
id + checksum automatically for ERA5 runs (ProvenanceRecord.datasets, carried empty since
Phase 0, now used).

The physics tests paid for themselves before the first commit

The acceptance suite asserts the stratosphere looks like the stratosphere: tropical cold point
185–205 K, summer pole warmer than winter pole, H₂O single-digit ppmv. The first cut of the pipeline
shipped CDS's descending pressure levels into an interpolation that assumed ascending — every
lookup silently read the 300 hPa row, and T(55 hPa) came back 240 K with 414 ppmv of water.
Plausible numbers at a glance; impossible as a stratosphere.

Fixed in the pipeline (both axes sorted ascending) and refused in the loader — a non-monotonic
axis raises rather than being reordered, because the checksum pins what was read.

The headline

ERA5 zonal-mean, 30°N, June, 55 hPa:   T = 209.42 K     (the ensemble typed 210 K)
                                       H₂O = 4.15 ppmv  (typed 6.91; reanalysis dry bias → #94/MLS)
Tropical cold point (Feb, 100 hPa):    191.4 K
December at 30°N:                      207.12 K         (real seasonal cycle, derived live)

The paper's site temperature is confirmed to half a kelvin by 30 years of reanalysis.

In the wizard

Stage 1 gains the ERA5 temperature-profile panel (pressure on a log axis, the box marked on it) —
under USER it is context; under ERA5 the marker is the run's temperature. Flip "Ambient state
from" to era5 and T/H₂O derive from latitude, month and pressure, go stale correctly when you
move them, and carry the dataset checksum into provenance.

357 Python Tier-A (+11), 55 vitest. The climatology tests run in CI — the product is
committed, so this is the first substantive science whose tests CI actually executes.

Stacked on #95. Same merge-then-rebase dance as before if GitHub flags it.