feat(simulation): believability gate learns basin-direction variety (T-964, D-245)

A real D-245 gate-shape strengthening, not just a test: the D-256/T-1174
finding proved the believability golden byte-identical under a total
all-North basin_direction collapse — every scalar contrast field is
structurally blind to the one field that regressed. ContrastMetrics
gains land_districts and basin_directions_distinct (both over ALL
districts, no new derive calls — the evidence-backed pick over the
voxel-transect proxy, which washes out at production sample density),
and evaluate_criteria gains 'basin direction variety': pass when
land_districts == 0 (the drained-body guard — an all-ocean body has no
cells that can cast a D8 vote per the aggregator's own exclusion rule,
so a uniform default is legitimate, mirroring the file's existing
nothing-to-vary idiom) or distinct >= 2. Negative test proves the
criterion catches the land-bearing all-North regression; vacuous-pass
test proves the waterworld guard. Golden regenerated and rerun-stable;
both validation bodies (Arbour, Edict) pass at distinct=2.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

server/tests/golden/believability.json

@@ -29,7 +29,9 @@
       "vegetation_classes": 4,
       "terrain_materials": 3,
       "voxel_relief_m": 28,
-      "micro_habitat_distinct": 3
+      "micro_habitat_distinct": 3,
+      "land_districts": 698,
+      "basin_directions_distinct": 2
     },
     "coherence": {
       "water_districts": 46,
@@ -70,7 +72,9 @@
       "vegetation_classes": 3,
       "terrain_materials": 2,
       "voxel_relief_m": 22,
-      "micro_habitat_distinct": 0
+      "micro_habitat_distinct": 0,
+      "land_districts": 1602,
+      "basin_directions_distinct": 2
     },
     "coherence": {
       "water_districts": 19,