diff --git a/crates/fparkan-runtime/src/lib.rs b/crates/fparkan-runtime/src/lib.rs
index ff7beda..154d588 100644
--- a/crates/fparkan-runtime/src/lib.rs
+++ b/crates/fparkan-runtime/src/lib.rs
@@ -43,6 +43,8 @@ use std::sync::Arc;
 
 pub use fparkan_assets::MissionAssets;
 
+const MISSION_DECODED_PAYLOAD_CACHE_ENTRIES: usize = 256;
+
 /// Engine mode.
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub enum EngineMode {
@@ -684,7 +686,10 @@ fn load_mission_with_options_and_progress(
         })
         .collect();
     record_load_phase(&mut trace, &mut on_phase, MissionLoadPhase::Graph);
-    let repository = CachedResourceRepository::new(vfs.clone());
+    let repository = CachedResourceRepository::with_payload_cache_budget(
+        vfs.clone(),
+        MISSION_DECODED_PAYLOAD_CACHE_ENTRIES,
+    );
     let graph_roots: Vec<_> = mission
         .objects
         .iter()
diff --git a/docs/tomes/05-render.md b/docs/tomes/05-render.md
index 5899182..9565f22 100644
--- a/docs/tomes/05-render.md
+++ b/docs/tomes/05-render.md
@@ -1101,6 +1101,13 @@ graph traversal, edge order and validation semantics. Repeating the controlled
 WEAR and MAT0 expansion before timeout; its remaining bottleneck is TEXM
 validation/archive work, not Vulkan submission or the base prototype graph.
 
+Mission loading now raises the decoded-payload cache entry budget from 64 to
+256 while retaining its 64 MiB byte budget. This avoids premature entry-count
+eviction during resource-rich loads without making memory unbounded. The same
+180-second Part 2 probe nevertheless remained at `GraphVisualTextures`; this
+change is a safe cache-capacity improvement, not measured evidence of a Part 2
+startup advance.
+
 `Assets` now has four ordered diagnostic sub-checkpoints: `AssetModelMeshes`
 (MSH), `AssetWearTables` (WEAR), `AssetMaterials` (MAT0) and `AssetTextures`
 (diffuse TEXM and lightmaps). The callback is observational: it neither changes