diff --git a/k8s/deployment-grid-rs.yaml b/k8s/deployment-grid-rs.yaml
index 7bc871c2..ca6aa09b 100644
--- a/k8s/deployment-grid-rs.yaml
+++ b/k8s/deployment-grid-rs.yaml
@@ -82,19 +82,17 @@ spec:
               value: "/data/scores"
             - name: RUST_LOG
               value: "info"
-            # talos5 is 4c/4t (i5-7260U). Each replica claims up to this
-            # many forecast hours in parallel. With 2 replicas × 3 → 6
-            # concurrent tasks, the 18-hour chain completes in ~3 min.
-            # Per-task peak RAM is ~250 Mi post-streaming refactor, so
-            # 3×250 ≈ 750 Mi per pod, well under the 2 Gi limit. The
-            # secondary replica may land on a smaller node — 3 keeps it
-            # honest.
+            # talos5 is 4c/4t (i5-7260U). 2 concurrent tasks per pod ×
+            # 2 replicas = 4 cluster-wide, enough to drain f01..f18 in
+            # ~5 min while keeping analysis (f00) room for its native-
+            # duct wgrib2 peak without OOM. Raising to 3 reintroduced
+            # the Stream-A-cutover OOM on the analysis step.
             - name: PROP_GRID_RS_PARALLELISM
-              value: "3"
+              value: "2"
             # Pool size = parallelism + 2 (NOTIFY + retry slack). Kept in
             # sync with PROP_GRID_RS_PARALLELISM when either changes.
             - name: PROP_GRID_RS_PG_CONNS
-              value: "5"
+              value: "4"
             - name: METRICS_ADDR
               value: "0.0.0.0:9100"
           envFrom:
@@ -126,14 +124,16 @@ spec:
           resources:
             requests:
               cpu: 100m
-              memory: 512Mi
+              memory: 768Mi
             limits:
               cpu: "4"
-              # OOM at 1 Gi during active chain work (92k points × 23
-              # bands ≈ 2.1M scores, plus decoded GRIB2 buffers, plus
-              # wgrib2 subprocess memory). 2 Gi leaves headroom; we're
-              # still an order of magnitude below the 6 Gi Elixir budget.
-              memory: 2Gi
+              # Analysis tasks (f00) run the native-level HRRR decode
+              # (~530 MB GRIB2 + wgrib2 working set) on top of 2 concurrent
+              # forecast tasks. 2 Gi OOM'd reliably in the 13 h after
+              # Stream A cutover. 3 Gi leaves room for the analysis
+              # wgrib2 peak without eliminating the forecast-lane
+              # parallelism headroom.
+              memory: 3Gi
           securityContext:
             allowPrivilegeEscalation: false
             runAsNonRoot: true
diff --git a/rust/prop_grid_rs/Dockerfile b/rust/prop_grid_rs/Dockerfile
index cf5c6c01..ca349dfa 100644
--- a/rust/prop_grid_rs/Dockerfile
+++ b/rust/prop_grid_rs/Dockerfile
@@ -6,6 +6,10 @@
 # the already-published Elixir image.
 #
 # Runtime uid 65534 matches the fsGroup on the NFS scores mount.
+#
+# This image ships two binaries: `prop-grid-rs` (default ENTRYPOINT,
+# drains grid_tasks) and `hrrr-point-worker` (k8s `command:` override,
+# drains hrrr_fetch_tasks).
 ARG WGRIB2_IMAGE=git.mcintire.me/graham/prop:latest
 
 FROM ${WGRIB2_IMAGE} AS wgrib2-src