diff --git a/src/renderer/vulkan/shaders.zig b/src/renderer/vulkan/shaders.zig
index 8382f5ee9..3f83e8a08 100644
--- a/src/renderer/vulkan/shaders.zig
+++ b/src/renderer/vulkan/shaders.zig
@@ -30,19 +30,11 @@ const DescriptorPool = vulkan.DescriptorPool;
 const Pipeline = @import("Pipeline.zig");
 const math = @import("../../math.zig");
 
-/// Build-time-precompiled SPIR-V blobs for the 9 built-in
-/// shaders. Generated by `src/build/VulkanSpv.zig`. Each decl
-/// is `[]const u32` backed by a comptime-aligned u8 array
-/// (align(@alignOf(u32))) so the underlying storage is on a
-/// 4-byte boundary — Module.initFromSpirv can consume the
-/// slices directly without alignment surgery.
-///
-/// Skipping the runtime glslang call path for built-ins
-/// eliminates the per-process TPoolAllocator high-water-mark
-/// leak that the first-surface compile otherwise leaves
-/// behind (~5 MB heaptrack delta). Custom (user) shaders still
-/// go through glslang at runtime via shadertoy.spirvFromGlsl.
-const vulkan_spv = @import("vulkan_spv");
+// Build-time-precompiled SPIR-V blobs are generated by
+// `src/build/VulkanSpv.zig` and exposed as the `vulkan_spv`
+// module — currently UNUSED at runtime; see the long
+// explanation in `Shaders.init` for why we went back to
+// runtime glslang compilation for built-ins.
 
 const log = std.log.scoped(.vulkan);
 
@@ -920,16 +912,33 @@ pub const Shaders = struct {
         // runtime via shadertoy.spirvFromGlsl, and the per-frame
         // shadertoy post-pipeline below still allocates via
         // `alloc`, so the parameter is still load-bearing.
+        // All built-ins runtime-compiled via glslang. The build-
+        // time SPV precompile path (commits c921c4b2e / 99d26181e)
+        // saves ~5 MB of TPoolAllocator leak per heaptrack but
+        // somehow corrupts kitty image rendering (images fill the
+        // whole window). The diagnostic in 7c045d694... showed
+        // even runtime-compiling image_vert + image_frag while
+        // leaving the other 7 precompiled DIDN'T fix it — so the
+        // bug isn't local to those shaders' SPV, it's some cross-
+        // shader interaction (descriptor binding cross-talk?
+        // ProcessDeferred state divergence between fresh-process
+        // build-time vs sequential-compile runtime?) we don't
+        // understand yet.
+        //
+        // VulkanSpv.zig + src/vulkan_spvgen.zig + the generated
+        // vulkan_spv import stay in tree so a future investigation
+        // can flip the swap back on. ~5 MB cosmetic leak is the
+        // wrong trade-off vs visibly broken images.
         var modules: Modules = .{
-            .bg_color_frag = try Module.initFromSpirv(device, vulkan_spv.bg_color_frag, .fragment),
-            .bg_image_frag = try Module.initFromSpirv(device, vulkan_spv.bg_image_frag, .fragment),
-            .bg_image_vert = try Module.initFromSpirv(device, vulkan_spv.bg_image_vert, .vertex),
-            .cell_bg_frag = try Module.initFromSpirv(device, vulkan_spv.cell_bg_frag, .fragment),
-            .cell_text_frag = try Module.initFromSpirv(device, vulkan_spv.cell_text_frag, .fragment),
-            .cell_text_vert = try Module.initFromSpirv(device, vulkan_spv.cell_text_vert, .vertex),
-            .full_screen_vert = try Module.initFromSpirv(device, vulkan_spv.full_screen_vert, .vertex),
-            .image_frag = try Module.initFromSpirv(device, vulkan_spv.image_frag, .fragment),
-            .image_vert = try Module.initFromSpirv(device, vulkan_spv.image_vert, .vertex),
+            .bg_color_frag = try Module.init(alloc, device, source.bg_color_frag, .fragment),
+            .bg_image_frag = try Module.init(alloc, device, source.bg_image_frag, .fragment),
+            .bg_image_vert = try Module.init(alloc, device, source.bg_image_vert, .vertex),
+            .cell_bg_frag = try Module.init(alloc, device, source.cell_bg_frag, .fragment),
+            .cell_text_frag = try Module.init(alloc, device, source.cell_text_frag, .fragment),
+            .cell_text_vert = try Module.init(alloc, device, source.cell_text_vert, .vertex),
+            .full_screen_vert = try Module.init(alloc, device, source.full_screen_vert, .vertex),
+            .image_frag = try Module.init(alloc, device, source.image_frag, .fragment),
+            .image_vert = try Module.init(alloc, device, source.image_vert, .vertex),
         };
         errdefer {
             inline for (.{