[JanStreffing]

Draft. Adds ocp_tool/subgrid_orography.py and tools/validate_subgrid_orography.py. Nothing is wired into a pipeline yet and nothing writes an ICMGG.

Why not paleo_subgrid_oro.py

That path shells out to calnoro, an ECHAM tool implementing a different definition of the four gravity-wave-drag fields. Fed ECMWF's own TCO1279 orography it returns a global mean sdor of 43.3 against ECMWF's 60.4, with a point-by-point RMS difference of 74.8 against a field mean of 60.4. Moving its input from ECHAM5 T511 (26 km, 2007) to TCO1279 (9 km) closed about 9% of the gap, so the problem is the formulation, not the data. paleo_subgrid_oro.py is untouched and stays as it is for ECHAM.

What this does

Lott and Miller (1997) statistics computed directly onto the reduced Gaussian grid, following the procedure in the IFS orographic drag documentation and Elvidge et al. (ECMWF Tech Memo 893):

1. average the source to 2'30", about 5 km, since scales below that belong to TOFD and the effective roughness scheme;
2. subtract the large scales, leaving a residual carrying only 5 km to grid length;
3. accumulate the gradient correlation tensor from that residual.

A global ETOPO1 pass takes about 15 seconds.

Validation

tools/validate_subgrid_orography.py compares against what ECMWF ship beside the model, climate.v020/<res>/{stdgwd,isogwd,slogwd}. Against 95_4 over ice-free land, with ETOPO1:

	ratio	corr	ECMWF's own lag-1 coherence
sdor	0.767	0.847	0.819
slor	0.844	0.913	0.807
isor	0.967	0.576	0.416

The last column matters. ECMWF's isor barely agrees with itself one grid box away, because anisotropy is a ratio of the tensor's eigenvalues and is genuinely noisy. All three fields here track ECMWF more closely than ECMWF tracks its own neighbour, so 0.576 is at the field's noise floor rather than short of a target.

Box-mean orography against 95_4/orog is the control, correlation 0.967, which is what says the grid mapping is right.

Things worth a reviewer's attention

• The high-pass cutoff is derived, not fitted. It is the grid length. Held fixed at 150 km the slor ratio drifts 0.910 at TCO79 to 0.955 at TCO319; scaled as 1.5 times the equatorial spacing it stays inside 0.911 to 0.927.
• The 5 km prefilter is physical, not a free parameter. At TCO319, refining it from 3' to 1' takes the slope ratio from 0.93 to 1.73.
• Polar convergence had to be taken out explicitly. A 3 arcmin cell is 5.5 km square at the equator and 5.5 km by 1 km at 80S, which biased the anisotropy into radial spokes about the pole. The zonal prefilter width and the zonal gradient span are now per-row, and box means are area weighted.
• The fields are zeroed below 0.5 land fraction, as ECMWF ship them. Without it, a cell holding an island carries a full-scale anisotropy on a sub-metre standard deviation: 4446 such sea cells on ETOPO1, half with sdor under a metre.

Open

• sdor sits about 20% low at every resolution and setting tried. Stable enough that a single factor would be defensible, but not understood.
• isor degrades with target resolution, 0.52 at TCO79 and TCO95 down to 0.39 at TCO319, and it is not sample starvation: nine times the fine points per box moves it 0.394, 0.410, 0.375.
• ETOPO1 Bed is what is available here, so Greenland and Antarctica are bedrock and are excluded from validation. Production wants the Ice version.

[JanStreffing]

Superseded by #64, which is the same work on a clean branch.

One commit here staged unrelated untracked files that were sitting in the working tree: test_oasis_output/*.nc, an embedded git repository at .oasis3-mct, the calnoro binaries, four config yamls, two docs, install_oasis.sh, some test scripts, and a modification to configs/TCO95_CORE3_nocav.yaml that was not mine. About 15 MB, none of it intended.

Rather than clean it up in place and leave the blobs in history, the branch has been deleted so the objects are unreferenced. #64 carries the same six reviewed commits plus a rename, and touches exactly two files.