fix: handle empty-feature depth slices in global rank

[johannesparty]

Problem

rank_soils_global crashes for some inputs with:

ValueError: Found array with 0 feature(s) (shape=(11, 0)) while a minimum of 1 is required by SimpleImputer.

(seen in production on tag 2026-06-24.1). The backend degrades gracefully (returns ALGORITHM_FAILURE), but the affected lookups get no ranking.

Cause. Per depth slice, rank_soils_global builds a feature matrix and passes it to gower_distances. When a slice has zero usable feature columns — the bedrock column-filter drops every property because the user's sample_pedon row is all-NaN at that depth — gower_distances feeds a shape=(n, 0) array into SimpleImputer, which raises.

Why now (regression). The crash line is old, but #368 (in 2026-06-24.1) changed user property arrays from a contiguous packed list to depth-indexed arrays that leave NaN at any depth the user didn't record. That makes all-NaN slices a normal outcome for gapped / non-zero-start horizon data (exactly the pedons #368 calls out), turning a latent crash into a reachable one.

Fix

• Guard the per-slice gower call (now _slice_gower_distance): a 0-feature slice returns an all-NaN (n, n) matrix instead of calling gower_distances. The existing masked-average and NaN-infill steps already treat NaN distances as "no information", so the slice is ignored and components rank on the depths that do have data. An all-NaN matrix (rather than skipping) keeps the per-slice list aligned with soil_matrix rows.
• Harden dis_max: max(map(np.nanmax, …)) → np.nanmax(…) so an all-NaN slice can't make the max NaN via max()'s ordering.

Tests

New soil_id/tests/test_global_rank_empty_slice.py (fast, no DB/network):

• 0-feature slice → all-NaN (11, 11) matrix, no raise
• characterization: the raw gower_distances path still raises on 0 features
• non-empty slice passes through to gower_distances unchanged

Also verified the live test_global_integration path still ranks normally.

Scope

This is a crash/robustness fix only — it stops the failure and lets ranking proceed on available depths. It does not make global rank use data below 200 cm; that depth space is still hardcoded to 200 (min(bottom, 200), length-200 arrays) and is a separate, later change.

🤖 Generated with Claude Code

[ssfrancis]

It would be fair to say that I reviewed lightly. Carissa or Garo might want to review as well if familiar with gower distance calculations.