diff --git a/coverage-exemptions.toml b/coverage-exemptions.toml
index 4b5ab2cd..256fc017 100644
--- a/coverage-exemptions.toml
+++ b/coverage-exemptions.toml
@@ -141,7 +141,7 @@ reason = "The DepthUnsupported refusal for a depth-carrying frame on a depthless
 
 [[exempt]]
 file = "crates/rhi/src/vk/pipeline.rs"
-lines = [1414, 1415, 1416]
+lines = [1460, 1461, 1462]
 reason = "The DepthUnsupported refusal for a depth-state pipeline on a depthless adapter - the pipeline-creation twin of the target-side refusal, unreachable for the same reason: the lanes' adapters offer a depth format and the void format query cannot be faulted."
 
 [[exempt]]
diff --git a/crates/rhi/src/vk/pipeline.rs b/crates/rhi/src/vk/pipeline.rs
index f35e35dd..c5be98f8 100644
--- a/crates/rhi/src/vk/pipeline.rs
+++ b/crates/rhi/src/vk/pipeline.rs
@@ -735,6 +735,56 @@ impl<'a> PipelineDesc<'a> {
         }
     }
 
+    /// A depth-only pipeline whose vertex stage writes its own vertex
+    /// list: no per-vertex stream, no fragment stage, no color
+    /// attachment.
+    ///
+    /// **The shape that lets a generative pipeline cast a shadow, and
+    /// until now there was not one.** [`Self::depth_mesh`] hard-sets
+    /// `vertex_input: Some(layout)`, and the frame contract asserts that
+    /// an item names geometry exactly when its pipeline declares
+    /// per-vertex input — so a pipeline that generates its own vertices
+    /// and draws an instance stream could reach a color target and no
+    /// depth-only one. A renderer of that shape could therefore be seen
+    /// and could not cast, which is not a limitation anybody chose.
+    ///
+    /// `vertex_count` is what the stage generates for one instance, the
+    /// same number [`Shaders`] bundles and for the same reason: passed
+    /// beside the SPIR-V it is a second value that compiles in any
+    /// combination, and too low renders part of the geometry while too
+    /// high indexes past the end of the stage's own constant array.
+    ///
+    /// Combine it with [`Self::instance_input`] for the instanced case.
+    /// Depth state must still be declared, exactly as on
+    /// [`Self::depth_mesh`]: a depth-only pipeline that neither tests nor
+    /// writes depth does nothing at all, and creation asserts it.
+    #[must_use]
+    pub fn depth_only(vertex_spirv: &'a [u8], vertex_count: u32) -> Self {
+        Self {
+            vertex_spirv,
+            // Structurally absent, as on `depth_mesh`: the depth-only
+            // format is what licenses the emptiness, and creation asserts
+            // the pairing in both directions.
+            fragment_spirv: &[],
+            target_format: TargetFormat::DepthOnly,
+            vertex_count,
+            blend: Blend::Opaque,
+            // Both faces, as on `depth_mesh`: a caster's silhouette is
+            // the union of what it covers, and culling a face here would
+            // punch holes in a shadow rather than save work worth having.
+            facing: Facing::Both,
+            sampled_bindings: 0,
+            uniform_block: 0,
+            // The difference from `depth_mesh`, and the whole of this
+            // constructor: no per-vertex stream, so the frame contract
+            // expects an item that names no geometry.
+            vertex_input: None,
+            instance_input: None,
+            depth_state: None,
+            push_constant_size: 0,
+        }
+    }
+
     /// Declare per-instance vertex input, in order. Locations and
     /// offsets are derived from position; the shader's `location(n)`
     /// list and this slice describe the same layout or the draw reads
@@ -1821,6 +1871,47 @@ mod tests {
         assert_eq!(desc.address, AddressMode::ClampToEdge);
     }
 
+    /// **The two depth-only shapes differ in exactly one field, and it is
+    /// the one the frame contract reads.**
+    ///
+    /// `Item`'s rule is `pipeline.vertex_input == item.mesh.is_some()`,
+    /// asserted before any GPU call. So `vertex_input` is not a detail of
+    /// these two constructors, it is the whole difference between them:
+    /// one demands geometry and one demands its absence. Everything else
+    /// they set is identical, and pinning that here is what stops a later
+    /// edit from making them differ somewhere a caller cannot see.
+    #[test]
+    fn the_two_depth_only_shapes_differ_only_in_whether_they_want_geometry() {
+        const LAYOUT: &[VertexAttribute] = &[VertexAttribute::Vec3];
+        let generative = PipelineDesc::depth_only(&[1, 2, 3, 4], 6);
+        let over_a_mesh = PipelineDesc::depth_mesh(&[1, 2, 3, 4], LAYOUT);
+
+        assert!(
+            generative.vertex_input.is_none(),
+            "a generative pipeline declaring per-vertex input would be refused a mesh-less item"
+        );
+        assert!(
+            over_a_mesh.vertex_input.is_some(),
+            "a mesh pipeline that declared none would be handed a mesh and ignore it"
+        );
+        assert_eq!(generative.vertex_count, 6, "the stage's own count is kept");
+        assert_eq!(
+            over_a_mesh.vertex_count, 0,
+            "a mesh pipeline takes its count from the geometry"
+        );
+        for desc in [&generative, &over_a_mesh] {
+            assert_eq!(desc.target_format, TargetFormat::DepthOnly);
+            assert!(
+                desc.fragment_spirv.is_empty(),
+                "a depth-only pipeline carries no fragment bytes"
+            );
+            assert!(
+                desc.depth_state.is_none(),
+                "depth state is the caller's to declare, and creation asserts it was"
+            );
+        }
+    }
+
     /// Every attribute's size and Vulkan spelling, every arm. Here rather
     /// than in the device suite for the reason the filter test above
     /// states: that suite skips wherever the validation layer is absent,
diff --git a/crates/rhi/tests/golden.rs b/crates/rhi/tests/golden.rs
index d73c8636..e562e80c 100644
--- a/crates/rhi/tests/golden.rs
+++ b/crates/rhi/tests/golden.rs
@@ -1185,6 +1185,79 @@ fn left_half_quad(depth: f32) -> Vec<u8> {
 /// is the depth buffer itself: the quad's clip-space z where it
 /// covered, the reversed-Z far clear where it did not — a CPU oracle
 /// over the whole rendered-depth path.
+/// **A pipeline that generates its own vertices can now cast.** Until
+/// `PipelineDesc::depth_only` existed it could not: the only depth-only
+/// constructor hard-set `vertex_input: Some(layout)`, and the frame
+/// contract asserts that an item names geometry exactly when its pipeline
+/// declares per-vertex input. So a generative renderer could be seen and
+/// could not cast a shadow, which is not a limitation anybody chose.
+///
+/// What this proves is the seam rather than the picture: the pipeline
+/// builds, and a depth pass whose item names **no geometry** is accepted
+/// and submitted with no validation error. It deliberately does not read
+/// the depth back — the only stage available here writes `z = 0.0`, which
+/// under reversed-Z is the clear value, so a readback cannot tell a draw
+/// that happened from one that did not. Claiming otherwise would be a
+/// test that looks like it checked something.
+///
+/// Probed by swapping `depth_only` for `depth_mesh`: the contract refuses
+/// the item, naming geometry it has no pipeline for.
+#[test]
+fn a_generative_pipeline_can_write_into_a_depth_target() {
+    const SIZE: u32 = 8;
+    let Some(device) = device_or_skip().expect("device bring-up") else {
+        return;
+    };
+    let image = match device.create_render_image(&RenderImageDesc::new(
+        RenderImageKind::Depth,
+        Extent {
+            width: SIZE,
+            height: SIZE,
+        },
+    )) {
+        Ok(image) => image,
+        Err(error) => {
+            assert!(
+                !strict(),
+                "RENEW_GOLDEN=1 but the depth render image was refused: {error}"
+            );
+            eprintln!("SKIP: depth render image refused: {error}");
+            return;
+        }
+    };
+    let caster = device
+        .create_pipeline(
+            &PipelineDesc::depth_only(builtin::TRIANGLE_VS_SPV, 3)
+                .depth_state(DepthState::read_write()),
+        )
+        .expect("a generative depth-only pipeline");
+    let mut target = device
+        .create_offscreen_target(Extent {
+            width: SIZE,
+            height: SIZE,
+        })
+        .expect("offscreen target");
+
+    // No mesh: the whole point. The stage writes its own three vertices.
+    let items = [Item::new(&caster)];
+    let depth_ops = Attachment::new(LoadOp::Clear(ClearValue::Depth(0.0)), StoreOp::Store);
+    // A frame needs at least one surface pass, so the depth pass is
+    // followed by a bare clear. Nothing reads the depth image back — see
+    // the note above for why it could not tell us anything if it did.
+    let color = clear(Color::new(0.0, 0.0, 0.0, 1.0));
+    let passes = [
+        Pass::render_to(&image, depth_ops, &items),
+        Pass::new(&color, &[]),
+    ];
+    target
+        .render(&RenderDesc::new(&passes))
+        .expect("a generative caster renders");
+
+    drop(target);
+    drop(caster);
+    assert_no_validation_errors(&device);
+}
+
 #[expect(
     clippy::too_many_lines,
     reason = "one narrative: a depth-only pass writes, a sampling pass reads it back, and the oracle checks the whole path -- splitting it would hide which half a failure came from. It crossed the bound when deriving the mesh stride from the layout made the constructor wrap."