Add ECCC HRDPS forecast quickstart notebook (review draft)

[aldenks]

Quickstart for the new eccc-hrdps-forecast dataset: ECCC's 2.5 km convection-permitting continental model, 6-hourly inits, hourly to 48 hours. Written against the draft "Writing quickstart notebooks" guidelines, which are held back for their own PR.

Draft, not ready to merge: it carries surplus so there is something to choose between.

Not ready to merge

Six static question candidates and three animations, to be cut to two and one. Two draft notes (after the title, and after the community challenge) mark the surplus and must be deleted along with whatever is cut. Recommended keeps:

• Q1, Okanagan wind — the strongest argument for 2.5 km here, and it carries the speed-and-direction convention.
• Q6, Lake Superior lake breeze — a 12 °C drop in one hour at the shoreline with nothing 80 km inland.
• Q7, Prairie convection — the animation. Two nocturnal systems with a dry afternoon between them, which a parameterized model gets structurally wrong.

Each candidate is independent: delete its heading cell, its code cell, and any note cell that follows. An AST check confirms no cross-candidate name dependencies, so any subset runs.

Dataset notes worth knowing

• First rotated pole grid in the archive. y/x are grid_latitude/grid_longitude in degrees, so there is no 1D lat/lon to .sel() on. Three helpers carry the notebook: a cartopy RotatedPole CRS built from spatial_ref, a nearest-cell search over the 2D coordinate arrays, and a lat/lon box slicer.
• Winds are speed and direction, not u/v. Grid-relative u/v would be up to 54° from true north across this domain — measured, not assumed.
• ingested_forecast_length is all-NaT for every init time, and as a scalar coord it lands in every auto-generated plot title as NaT, which reads as broken software. The notebook drops it in the open cell; worth fixing upstream instead.
• CAPE carries -1 and -1000 source markers over a third of the domain at a given hour, varying with the weather; the structure note says to mask below -0.1.

Events

All verified inside the archive window (2026-07-09 onward) and domain: the Bald Range fire near Summerland (7-8 Aug), Thunder Bay 36 in Wabakimi (12-14 Jul), the Prairie heat ridge and its two nocturnal MCSs (11-13 Jul), the ECCC Maritimes rainfall warning (23-24 Aug), and a Lake Superior lake breeze (13 Jul). Sources are named in prose rather than linked, since the 2026 warnings and fire dashboards are live-state pages.

Open question for review

Q5 (Maritimes rain) is the only section that checked the model against something external — ECCC warned 40-60 mm with locals to 100, the forecast gave a 44 mm median and 83 mm at the 90th percentile. The guidelines' "don't compute a number only so the prose can quote it" removed that comparison, so the section now shows an accumulation without saying whether it agrees. That may be right for a quickstart, but it puts "how good is this model" out of scope by construction, and is worth deciding deliberately.

Also flagged: the mm/h conversion on Q7's colorbar partially reverses a unit-conversion cut made on the HRRR notebook. The guidelines were narrowed to "convert only where the stored unit is one nobody reads in" — kg m-2 s-1 earns mm/h, m s-1 does not. One-line revert if that lands differently.

Runs end to end, 29 cells, 8.2 MB, no errors.

🤖 Generated with Claude Code