diff --git a/server/src/atlas/cascade.rs b/server/src/atlas/cascade.rs
index 559c9bf74..6436e01a8 100644
--- a/server/src/atlas/cascade.rs
+++ b/server/src/atlas/cascade.rs
@@ -306,8 +306,10 @@ pub fn run_cascade_from_heightmap(
                 // shared via scale::HEIGHTMAP_CELLS_PER_DISTRICT so plugin.rs
                 // converts CityPlacement pixel coords with the same constant.
                 // body_id is required for the D-243 §4 climate edge-fuzz warp
-                // domain separation — derive_all_districts builds the region
-                // cache internally.
+                // domain separation — derive_all_districts pre-builds the
+                // region-baseline cache internally on TRUE region keys (each
+                // survey cell's centre world metres → containing district →
+                // region ±1 ring; D-256(c) + PR #199 review).
                 //
                 // Layer1Output.survey_basin_dirs threads the true D8 thalweg
                 // direction into each DistrictProfile.basin_direction (T-1047)
diff --git a/server/src/atlas/district_profile.rs b/server/src/atlas/district_profile.rs
index 86b807558..6cb3a8932 100644
--- a/server/src/atlas/district_profile.rs
+++ b/server/src/atlas/district_profile.rs
@@ -1,10 +1,14 @@
-//! DistrictProfile carrier — ~1 km scale tier of the D-239 refinement chain (T-1023).
+//! DistrictProfile carrier — the coarsest derivation stage of the D-239 refinement
+//! chain (T-1023).
 //!
-//! A `DistrictProfile` is the coarsest derivation stage: ~1 km² cells (~6 000 per body,
-//! roughly 80 × 75), each a pure deterministic function of
-//! `(seed, body_params, terrain_analysis, pos)`. Stored in `BodyWorldState.districts`
-//! so the Atlas can read zone labels without triggering voxel derivation (D-239 §10,
-//! D-203).
+//! In the batch pass a `DistrictProfile` summarizes one survey cell (D-256: an
+//! 8×8 working-pixel block, 64×32 ≈ 2 048 cells per body), each a pure
+//! deterministic function of `(seed, body_params, terrain_analysis, pos)` sampled
+//! at the cell's centre world metres through the shared
+//! `derive_at_metres_with_riparian` core; the on-demand pass derives the same
+//! struct at any absolute world position. Stored in `BodyWorldState.districts`
+//! so the Atlas can read zone labels without triggering voxel derivation
+//! (D-239 §10, D-203).
 //!
 //! ## D-010 compliance
 //!
@@ -18,7 +22,7 @@
 //! enum from `simulation::generator`. The enum is the canonical freeze; adding,
 //! renaming, or removing a zone requires a D-record amendment.
 
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 
 use serde::{Deserialize, Serialize};
 
@@ -2038,23 +2042,25 @@ fn bilinear_bool(mask: &[bool], w: usize, h: usize, px: f64, py: f64) -> f32 {
 ///
 /// `grid_cells_per_district = 8` means each survey cell is 8×8 heightmap cells.
 ///
-/// ## Region baseline (D-256(c))
+/// ## Region baseline (D-256(c), cache pre-build per PR #199 review)
 ///
-/// No pre-built region cache here — the pre-D-256 pre-build keyed on the
-/// SURVEY grid's own pseudo-coordinates (`district_to_region((rx, ry))`),
-/// which does not correspond to the true region a survey cell's world-metres
-/// centre actually falls in (the survey grid can span up to ~2048 distinct
-/// true regions on a big body — it covers the whole body surface in metres,
-/// not a handful of degenerate cells). Since the shared core now floor-divides
-/// the survey cell's TRUE world metres for its own region lookup
-/// ([`derive_at_metres_with_riparian`]'s `district_pos`), the simplest
-/// deterministic option is to let each cell resolve its region baseline
-/// on-the-fly through [`derive_district_profile`]'s own `region_cache`
-/// parameter — passing an empty map here means every lookup misses and
-/// derives on the fly (pure, cheap, four `derive_region_baseline_c` calls per
-/// miss; D-227-legal). This also auto-fixes the former body-uniform
-/// region-(0,0) climate collapse: every survey cell now reads its OWN
-/// region's baseline instead of the pseudo-grid's degenerate region index.
+/// A region-baseline cache is pre-built here on the TRUE region keys the
+/// shared core looks up: each survey cell's centre world metres floor-divides
+/// to its containing district (the same mapping
+/// [`derive_at_metres_with_riparian`] applies internally), and that
+/// district's region ±1 neighbour ring covers every key
+/// [`region_profile::region_baseline_at_district`]'s edge-fuzz blend can
+/// read (base + signed x/y/xy neighbours). The pre-D-256 pre-build keyed on
+/// the SURVEY grid's own pseudo-coordinates (`district_to_region((rx, ry))`)
+/// — the wrong key space entirely, which is what produced the body-uniform
+/// region-(0,0) climate collapse; the true-key rebuild can span up to ~2048
+/// distinct regions plus ring on a big body (the survey grid covers the
+/// whole body surface in metres). Cache-hit and cache-miss are
+/// byte-identical ([`region_profile::build_region_profile`] and the miss
+/// branch in `region_baseline_at_district` compute `mean_temp_c` with the
+/// same expressions — D-227 purity, guarded by the wrapper≡core agreement
+/// tests), so the cache is purely the cost model: each covering region
+/// derives once per body instead of four misses per survey cell.
 ///
 /// `body_id` is the body's string identifier, required for the climate edge-fuzz
 /// warp domain separation.
@@ -2089,11 +2095,36 @@ pub fn derive_all_districts(
     let survey_cols = ta.w.div_ceil(gcpr) as i32;
     let survey_rows = ta.h.div_ceil(gcpr) as i32;
 
-    // D-256(c): no pre-built region cache (see this function's doc) — every
-    // survey cell derives its region baseline on-the-fly through the empty
-    // map below, keyed on the cell's TRUE world-metres region, not a
-    // pseudo-grid index.
-    let region_cache: BTreeMap<RegionPos, RegionProfile> = BTreeMap::new();
+    // Pre-build the region-baseline cache on the TRUE region keys the shared
+    // core will look up (see this function's doc): each survey cell's centre
+    // world metres → containing district → region ±1 ring. Deduped via
+    // BTreeSet (D-010: deterministic iteration), derived once per body.
+    let mut covering_regions: BTreeSet<RegionPos> = BTreeSet::new();
+    for ry in 0..survey_rows {
+        for rx in 0..survey_cols {
+            let (wx, wy) = survey_cell_centre_world_m(
+                SurveyCellPos(rx, ry),
+                gcpr,
+                ta.w,
+                ta.h,
+                body_params.body_radius_km,
+            );
+            // The SAME floor-divide derive_at_metres_with_riparian applies to
+            // reach its region key — one mapping, never a second one.
+            let district_pos: DistrictPos = (
+                (wx / scale::DISTRICT_M as f64).floor() as i32,
+                (wy / scale::DISTRICT_M as f64).floor() as i32,
+            );
+            let base = scale::district_to_region(district_pos);
+            for ndy in -1..=1i32 {
+                for ndx in -1..=1i32 {
+                    covering_regions.insert((base.0 + ndx, base.1 + ndy));
+                }
+            }
+        }
+    }
+    let region_cache =
+        region_profile::derive_regions_for_body(seed, body_params, &climate, covering_regions);
 
     let mut out = BTreeMap::new();
     for ry in 0..survey_rows {