[Alex0420W]

Replaces v1.x single ice criterion (Diviner PRP ice-depth + PSR-proximity bonus) with three-class volatile access inside the same 2 km EVA disc: H₂O cold trap (temp_max < 110 K), CO₂/NH₃ cold trap (temp_max < 66 K), and ultra-cold (temp_max < 60 K). Mirrors NASA's documented volatile-diversity priority (Lawrence 2025 FOM names "volatile diversity" as a first-class FOM category) and Wueller 2026's per-class treatment in the peer-reviewed reference. Cold-class thresholds come from Paige et al. 2010, Hayne et al. 2015, and the Hayne et al. 2020 micro-cold-trap classification.

Equal sub-weights (1/3 each) by default — agnostic on NASA's preference ordering between volatile classes; sub_weights= exposed for downstream sensitivity. Combined-score implementation reuses v2.0's scipy.ndimage.convolve circular-kernel pattern; compute_components returns the three sub-scores plus the combined for diagnostic visibility. Per-class component COGs persist alongside the combined score so the catalog surfaces which sites access multiple volatile classes vs single without rerunning preprocess+score. Weight slot unchanged at 0.17 (same as v1.x ice) so the v2.0 → v2.1 comparison is methodology-only.

20 m headline (default-weights, eight active criteria with multi_volatile): 69 sites (was 69 at v2.0); 54 / 69 = 78.3 % within 5 km of an in-scope Wueller site (was 79.7 %); median matched-pair distance 1.84 km (unchanged). Of 69 v2.1 20 m sites: 30 within 10 m of a v2.0 site (identical re-pick); 24 reranked < 1 km, 15 reranked 1-5 km, 0 reranked > 5 km — methodology swap is broader than v2.0's coupling → eva swap (which had 59 of 69 identical) but keeps every site inside its polygon. Per-region site counts unchanged from v2.0.

240 m baseline (full 240 m run, 70-site catalog): 52 / 70 = 74.3 % within 5 km, median 1.87 km, 42 / 73 = 57.5 % Wueller-side.

Per-class breakdown — substantive new content of v2.1: of 69 sites, 0 access all three thermal classes within 2 km EVA disc, 1 accesses two classes (H₂O + CO₂/NH₃), 5 access H₂O only, 63 have no cold-class neighbour within 2 km. The ISRU-volatile-rich subset is concentrated in Nobile Rim 2 (4 of 5 H₂O+ sites, including the only two-class site). The 0 / 69 all-three-classes count isn't a bug — it's the real finding that selene's HLS-compliant catalog is a habitable rim-band catalog, spatially separate from the deepest-cold-trap floors (1-10 km away). Surfaces a gap the v1.x ice criterion couldn't expose.

Per-launch-year HLS cross-check (20 m, matched sites): 23 / 54 pass HLS for 2028 (Artemis IV); 52 / 54 pass HLS in at least one launch year 2025-2032; per-year pass counts: 43 / 23 / 25 / 23 / 31 / 29 / 20 / 42 across 2025-2032.

Continuous-lit illumination metric and rock-abundance safety filter remain deferred indefinitely (data-blocked: PGDA hourly cube not distributed; all public Diviner rock-abundance products clip at ±60-70°). NASA-priority sub-weights remain a candidate follow-on if quantitative weighting becomes public.

Summary

What this PR does, in 1–3 sentences.

Motivation

Why this change is worth making (link the issue if there is one).

Checklist

• ruff check . passes
• ruff format --check . passes
• pytest passes locally
• New behaviour has a test (or an explicit # noqa reason if not)
• Public-facing changes (CLI flags, criterion math, file layout) are
  reflected in the README or the relevant module docstring
• No real-data fixtures were committed; new tests run in CI without
  downloaded data (or are guarded with pytest.mark.skipif)

Scope notes

• Things this PR explicitly does not address (so reviewers know what
  to expect in follow-ups).