diff --git a/.config/nextest.toml b/.config/nextest.toml
new file mode 100644
index 00000000..556442e4
--- /dev/null
+++ b/.config/nextest.toml
@@ -0,0 +1,25 @@
+# Nextest configuration for PromptForge workspace.
+
+[profile.default]
+# Report slow tests after 60s and terminate if hung.
+slow-timeout = { period = "60s", terminate-after = 3 }
+# Cleanly terminate lingering background tasks/threads after test returns.
+leak-timeout = "250ms"
+
+[profile.ci]
+# CI profile settings: immediate failure reporting and timeout controls.
+slow-timeout = { period = "60s", terminate-after = 3 }
+leak-timeout = "250ms"
+failure-output = "immediate-final"
+
+[test-groups]
+# Limit concurrency for tensor/FFI-heavy suites to prevent core thrashing.
+heavy = { max-threads = 2 }
+
+[[profile.default.overrides]]
+filter = 'package(promptforge-tool-picker) | package(gateway-stt-backend-whisper) | package(gateway-stt)'
+test-group = 'heavy'
+
+[[profile.ci.overrides]]
+filter = 'package(promptforge-tool-picker) | package(gateway-stt-backend-whisper) | package(gateway-stt)'
+test-group = 'heavy'
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 875c88f4..2603f73e 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -16,12 +16,23 @@ concurrency:
 # to anonymous, which is what the cache is there to make rare.
 env:
   HF_TOKEN: ${{ secrets.HF_TOKEN }}
+  RUSTUP_TOOLCHAIN: stable
 
 jobs:
-  check:
+  fmt:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: rustfmt
+
+      - name: Format
+        run: cargo fmt --all --check
+
+  clippy:
     runs-on: ubuntu-latest
-    # RUSTUP_TOOLCHAIN outranks the repo's rust-toolchain.toml (pinned to the
-    # MSRV for local builds); this job means to test stable.
     env:
       RUSTUP_TOOLCHAIN: stable
     steps:
@@ -29,9 +40,7 @@ jobs:
 
       - uses: dtolnay/rust-toolchain@stable
         with:
-          components: rustfmt, clippy
-
-      - uses: dtolnay/rust-toolchain@1.89.0
+          components: clippy
 
       - name: Cache cargo
         uses: Swatinem/rust-cache@v2
@@ -42,10 +51,13 @@ jobs:
       - uses: actions/setup-node@v4
         with:
           node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
 
-      - name: Install config UI dependencies
-        working-directory: crates/gateway-config-ui/ui
-        run: npm ci
+      - name: Install UI dependencies
+        run: |
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
 
       - name: Build Gateway without Workshop UI tooling
         run: cargo build --locked -p gateway
@@ -66,26 +78,44 @@ jobs:
       - name: Check product dependency boundaries
         run: cargo test -p gateway-stt --test it architecture
 
-      - name: Format
-        run: cargo fmt --all --check
-
-      # The desktop packages stay in their platform jobs because Linux needs
-      # Tauri system libraries. All gateway and STT features are pure Rust
-      # now that whisper.cpp is loaded from a managed runtime artifact.
       - name: Clippy
         run: cargo clippy --workspace --exclude workshop --exclude workshop-server --all-targets --all-features -- -D warnings
 
-      - name: Test
-        run: cargo test --locked --workspace --exclude workshop --exclude workshop-server --all-features
+  test:
+    runs-on: ubuntu-latest
+    env:
+      RUSTUP_TOOLCHAIN: stable
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+
+      - name: Install cargo-nextest
+        uses: taiki-e/install-action@nextest
+
+      - name: Cache cargo
+        uses: Swatinem/rust-cache@v2
+
+      - name: Cache the embedding model
+        uses: ./.github/actions/hf-model-cache
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
+
+      - name: Install UI dependencies
+        run: |
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
+
+      - name: Test (concurrent via nextest)
+        run: cargo nextest run --locked --workspace --exclude workshop --exclude workshop-server --all-features
 
       - name: Doctests
         run: cargo test --workspace --exclude workshop --exclude workshop-server --all-features --doc
 
-      - name: Docs
-        env:
-          RUSTDOCFLAGS: -D warnings
-        run: cargo doc --workspace --no-deps --all-features --exclude workshop --exclude workshop-server
-
       - name: Check headless gateway
         run: cargo check -p gateway --no-default-features
 
@@ -100,6 +130,36 @@ jobs:
             exit 1
           fi
 
+  docs:
+    runs-on: ubuntu-latest
+    env:
+      RUSTUP_TOOLCHAIN: stable
+      RUSTDOCFLAGS: -D warnings
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+
+      - name: Cache cargo
+        uses: Swatinem/rust-cache@v2
+
+      - name: Cache the embedding model
+        uses: ./.github/actions/hf-model-cache
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
+
+      - name: Install UI dependencies
+        run: |
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
+
+      - name: Docs
+        run: cargo doc --workspace --no-deps --all-features --exclude workshop --exclude workshop-server
+
   check-workshop:
     runs-on: windows-latest
     env:
@@ -111,6 +171,9 @@ jobs:
         with:
           components: clippy
 
+      - name: Install cargo-nextest
+        uses: taiki-e/install-action@nextest
+
       - name: Cache cargo
         uses: Swatinem/rust-cache@v2
 
@@ -120,10 +183,13 @@ jobs:
       - uses: actions/setup-node@v4
         with:
           node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
 
       - name: Install UI dependencies
-        working-directory: crates/workshop-server/ui
-        run: npm ci
+        run: |
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
 
       - name: Build featureless Gateway
         run: cargo build --locked -p gateway --no-default-features
@@ -134,8 +200,8 @@ jobs:
       - name: Clippy (workshop)
         run: cargo clippy -p workshop -p workshop-server --all-targets -- -D warnings
 
-      - name: Test (workshop)
-        run: cargo test --locked -p workshop -p workshop-server
+      - name: Test (workshop, concurrent via nextest)
+        run: cargo nextest run --locked -p workshop -p workshop-server
 
       - name: Test Gateway process ownership races
         run: |
@@ -181,15 +247,18 @@ jobs:
       - name: Install Tauri system packages
         run: |
           sudo apt-get update
-          sudo apt-get install -y libwebkit2gtk-4.1-dev libssl-dev librsvg2-dev
+          sudo apt-get install -y --no-install-recommends libwebkit2gtk-4.1-dev libssl-dev librsvg2-dev
 
       - uses: actions/setup-node@v4
         with:
           node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
 
       - name: Install UI dependencies
-        working-directory: crates/workshop-server/ui
-        run: npm ci
+        run: |
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
 
       - name: Build featureless Gateway
         run: cargo build --locked -p gateway --no-default-features
@@ -232,6 +301,8 @@ jobs:
       - uses: actions/setup-node@v4
         with:
           node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
 
       - name: Install UI dependencies
         working-directory: crates/workshop-server/ui
@@ -273,8 +344,6 @@ jobs:
       - uses: actions/checkout@v4
 
       - uses: dtolnay/rust-toolchain@1.89.0
-        with:
-          components: rustfmt
 
       - name: Cache cargo
         uses: Swatinem/rust-cache@v2
@@ -285,31 +354,25 @@ jobs:
       - uses: actions/setup-node@v4
         with:
           node-version: 22
+          cache: npm
+          cache-dependency-path: crates/*/ui/package-lock.json
 
       - name: Install UI dependencies
-        working-directory: crates/workshop-server/ui
-        run: npm ci
-
-      - name: Install config UI dependencies
-        working-directory: crates/gateway-config-ui/ui
-        run: npm ci
-
-      - name: Build and test on MSRV
         run: |
-          cargo build --locked --workspace --exclude workshop --exclude workshop-server --all-features
-          cargo test --locked --workspace --exclude workshop --exclude workshop-server --all-features
+          npm ci --prefix crates/workshop-server/ui
+          npm ci --prefix crates/gateway-config-ui/ui
+
+      - name: Check on MSRV
+        run: cargo check --locked --workspace --exclude workshop --exclude workshop-server --all-targets --all-features
 
   supply-chain:
     runs-on: ubuntu-latest
-    env:
-      RUSTUP_TOOLCHAIN: stable
     steps:
       - uses: actions/checkout@v4
 
-      - uses: dtolnay/rust-toolchain@stable
-
-      - name: Install cargo-deny and cargo-audit
-        run: cargo install cargo-deny cargo-audit --locked
+      - uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-deny, cargo-audit
 
       - name: cargo deny
         run: cargo deny check
diff --git a/crates/gateway/tests/it/realtime_stt/protocol.rs b/crates/gateway/tests/it/realtime_stt/protocol.rs
index a5cc2df4..c0aabb86 100644
--- a/crates/gateway/tests/it/realtime_stt/protocol.rs
+++ b/crates/gateway/tests/it/realtime_stt/protocol.rs
@@ -187,6 +187,7 @@ async fn mounted_route_drives_scripted_wire_ownership_errors_and_privacy() {
             .await
             .expect("interim completion observer joins")
         );
+        tokio::time::sleep(Duration::from_millis(25)).await;
     }
     send(
         &mut socket,
@@ -332,6 +333,7 @@ async fn standard_interims_emit_only_appendable_agreed_deltas() {
             .await
             .expect("interim completion observer joins")
         );
+        tokio::time::sleep(Duration::from_millis(25)).await;
     }
     send(
         &mut socket,
diff --git a/crates/promptforge-core/Cargo.toml b/crates/promptforge-core/Cargo.toml
index 07862b85..0964556f 100644
--- a/crates/promptforge-core/Cargo.toml
+++ b/crates/promptforge-core/Cargo.toml
@@ -34,6 +34,7 @@ tokio = { workspace = true, features = ["macros", "rt", "sync", "time"] }
 [dev-dependencies]
 axum.workspace = true
 promptforge-parser = { workspace = true, features = ["test-support"] }
+promptforge-tool-picker = { workspace = true, features = ["test-fixtures"] }
 tokio.workspace = true
 
 [lints]
diff --git a/crates/promptforge-core/src/execute/tests/live_infer.rs b/crates/promptforge-core/src/execute/tests/live_infer.rs
index c51131d2..d93c659b 100644
--- a/crates/promptforge-core/src/execute/tests/live_infer.rs
+++ b/crates/promptforge-core/src/execute/tests/live_infer.rs
@@ -15,8 +15,7 @@ async fn live_h1_infer_runs_once() {
         ## Result\n\n\
         ```lua\nreturn var.answer\n```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let out = super::super::run(
         &prompt,
@@ -90,8 +89,7 @@ async fn shared_function_resolves_host_globals_when_called() {
         ## Result\n\n\
         ```lua\nreturn read_args()\n```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let out = super::super::run(
         &prompt,
@@ -111,8 +109,7 @@ async fn shared_library_calls_host_apis_at_load_time() {
     // The shared library replays as each section's first chunk with the full
     // host environment installed, so top-level shared code may use `store`,
     // `log`, and `args` at load.
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let store = StoreRef::memory();
     let source = "---\nname: shared-host-load\ndescription: d\npromptforge: 1\n---\n\n\
@@ -179,8 +176,7 @@ async fn captured_bindings_reach_section_execute_and_fanout_vms() {
          ```lua\nreturn binding_names()\n```\n"
     );
     let prompt = parse(&source);
-    let picker = ToolPicker::build(Catalog::new(vec![descriptor]), PickerConfig::default())
-        .expect("tool picker must build");
+    let picker = build_test_picker(Catalog::new(vec![descriptor]), PickerConfig::default());
     let models = test_model_catalog();
     let tools: [Arc<dyn Tool>; 1] = [echo];
     let catalog = ToolCatalog::new(&tools).expect("the fixture tool is unique");
@@ -216,8 +212,7 @@ async fn live_h1_models_infer_resolves_the_default_model_without_touching_sys()
         ## Result\n\n\
         ```lua\nreturn var.answer .. ':' .. tostring(var.sys_untouched)\n```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let out = super::super::run(
         &prompt,
@@ -265,8 +260,7 @@ async fn nested_lua_infer_emits_a_model_turn_observation() {
         ## Result\n\n\
         ```lua\nreturn var.answer\n```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let recorder = Arc::new(Recorder::default());
 
@@ -319,8 +313,7 @@ async fn cancelled_nested_infer_does_not_report_model_turn_failed() {
         return writer:infer('must cancel')\n\
         ```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let recorder = Arc::new(Recorder::default());
     let cancel = crate::cancel::CancelHandle::new();
@@ -432,8 +425,7 @@ async fn live_h1_prose_preserves_non_final_and_final_semantics_and_captures_var(
          ```\n"
     );
     let prompt = parse(&source);
-    let picker = ToolPicker::build(Catalog::new(vec![descriptor]), PickerConfig::default())
-        .expect("tool picker must build");
+    let picker = build_test_picker(Catalog::new(vec![descriptor]), PickerConfig::default());
     let models = test_model_catalog();
     let tools: [Arc<dyn Tool>; 1] = [echo];
     let catalog = ToolCatalog::new(&tools).expect("the fixture tool is unique");
@@ -468,8 +460,7 @@ async fn h1_and_h2_prose_both_run_through_the_shared_block_loop() {
         return reply\n\
         ```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let out = super::super::run(
         &prompt,
@@ -519,8 +510,7 @@ async fn live_h1_chunk_keeps_sys_id_zero_and_the_first_walked_section_takes_one(
         return 'ok'\n\
         ```\n";
     let prompt = parse(source);
-    let picker = ToolPicker::build(Catalog::default(), PickerConfig::default())
-        .expect("empty tool picker must build");
+    let picker = empty_test_picker();
     let models = test_model_catalog();
     let out = super::super::run(
         &prompt,
diff --git a/crates/promptforge-core/src/execute/tests/mod.rs b/crates/promptforge-core/src/execute/tests/mod.rs
index a0b676b1..61131be3 100644
--- a/crates/promptforge-core/src/execute/tests/mod.rs
+++ b/crates/promptforge-core/src/execute/tests/mod.rs
@@ -278,7 +278,7 @@ async fn run(
         .with_similarity_floor(0.0)
         .and_then(|config| config.with_margin(0.0))
         .expect("test thresholds are in the supported domain");
-    let picker = ToolPicker::build(catalog, config).expect("test picker must build");
+    let picker = build_test_picker(catalog, config);
     let tool_catalog = ToolCatalog::new(tools).expect("fixture tools are unique");
     let mut run_config = RunConfig::new(opts.execution).observer(opts.observer);
     if let Some(client) = opts.client {
@@ -298,6 +298,36 @@ async fn run(
     .map_err(Error::from)
 }
 
+pub(super) fn empty_test_picker() -> ToolPicker {
+    ToolPicker::build_with_model(
+        &promptforge_tool_picker::Model::dummy(),
+        Catalog::default(),
+        PickerConfig::default(),
+        None,
+    )
+    .expect("empty test picker must build")
+}
+
+pub(super) fn build_test_picker(catalog: Catalog, config: PickerConfig) -> ToolPicker {
+    if catalog.is_empty() {
+        ToolPicker::build_with_model(
+            &promptforge_tool_picker::Model::dummy(),
+            catalog,
+            config,
+            None,
+        )
+        .expect("empty test picker must build")
+    } else {
+        ToolPicker::build_with_model(shared_test_model(), catalog, config, None)
+            .expect("test picker must build")
+    }
+}
+
+pub(super) fn shared_test_model() -> &'static promptforge_tool_picker::Model {
+    static MODEL: std::sync::OnceLock<promptforge_tool_picker::Model> = std::sync::OnceLock::new();
+    MODEL.get_or_init(|| promptforge_tool_picker::Model::load().expect("the test model loads"))
+}
+
 /// Runs a fixture offline through the real [`run`](super::run) entry point
 /// with a caller-customized [`RunConfig`], returning the typed [`RunError`]
 /// so a test can assert on its kind (limits, cancellation).
@@ -305,14 +335,7 @@ async fn run_with_config(
     test: &TestPrompt,
     configure: impl FnOnce(RunConfig) -> RunConfig,
 ) -> std::result::Result<String, RunError> {
-    let picker = ToolPicker::build(
-        Catalog::new(Vec::new()),
-        PickerConfig::default()
-            .with_similarity_floor(0.0)
-            .and_then(|config| config.with_margin(0.0))
-            .expect("test thresholds are in the supported domain"),
-    )
-    .expect("test picker must build");
+    let picker = empty_test_picker();
     super::run(
         &test.prompt,
         "",
diff --git a/crates/promptforge-core/src/execute/tests/observations.rs b/crates/promptforge-core/src/execute/tests/observations.rs
index 322110ab..d0e0d928 100644
--- a/crates/promptforge-core/src/execute/tests/observations.rs
+++ b/crates/promptforge-core/src/execute/tests/observations.rs
@@ -303,11 +303,6 @@ async fn one_execution_id_spans_parse_and_the_complete_runtime_lifecycle() {
     let recorder = Arc::new(Recorder::default());
     let prompt = Prompt::parse(&source, EXECUTION, recorder.as_ref())
         .expect("the lifecycle fixture must parse");
-    let _picker = ToolPicker::build(
-        Catalog::new(vec![descriptor.clone()]),
-        PickerConfig::default(),
-    )
-    .expect("the lifecycle picker must build");
     let tools: [Arc<dyn Tool>; 1] = [Arc::clone(&tool) as Arc<dyn Tool>];
     let prompt = TestPrompt {
         prompt,
diff --git a/crates/promptforge-core/src/execute/tests/scheduler.rs b/crates/promptforge-core/src/execute/tests/scheduler.rs
index 6ca2d79e..f14f8006 100644
--- a/crates/promptforge-core/src/execute/tests/scheduler.rs
+++ b/crates/promptforge-core/src/execute/tests/scheduler.rs
@@ -1619,8 +1619,7 @@ impl H1Resolution {
     /// model binds resolve, tool binds report absent.
     fn models_only() -> Self {
         Self {
-            picker: ToolPicker::build(Catalog::default(), PickerConfig::default())
-                .expect("empty tool picker must build"),
+            picker: empty_test_picker(),
             models: test_model_catalog(),
             tools: ToolCatalog::default(),
         }
@@ -1629,8 +1628,7 @@ impl H1Resolution {
     /// Everything empty: model binds report absent.
     fn empty() -> Self {
         Self {
-            picker: ToolPicker::build(Catalog::default(), PickerConfig::default())
-                .expect("empty tool picker must build"),
+            picker: empty_test_picker(),
             models: ModelCatalog::empty(),
             tools: ToolCatalog::default(),
         }
@@ -2145,8 +2143,7 @@ async fn live_h1_prose_preserves_non_final_and_final_semantics_and_captures_var(
     let ctx = h1_context(&prompt);
     let tools: [Arc<dyn Tool>; 1] = [echo];
     let resolution = H1Resolution {
-        picker: ToolPicker::build(Catalog::new(vec![descriptor]), PickerConfig::default())
-            .expect("tool picker must build"),
+        picker: build_test_picker(Catalog::new(vec![descriptor]), PickerConfig::default()),
         models: test_model_catalog(),
         tools: ToolCatalog::new(&tools).expect("the fixture tool is unique"),
     };
diff --git a/crates/promptforge-core/tests/suite/support.rs b/crates/promptforge-core/tests/suite/support.rs
index 53b8c84e..e9148617 100644
--- a/crates/promptforge-core/tests/suite/support.rs
+++ b/crates/promptforge-core/tests/suite/support.rs
@@ -48,8 +48,13 @@ pub(super) async fn run(
     store: &StoreRef,
     opts: RunOptions,
 ) -> Result<String, RunError> {
-    let picker = ToolPicker::build(Catalog::default(), Config::default())
-        .expect("empty fixture picker must build");
+    let picker = ToolPicker::build_with_model(
+        &promptforge_tool_picker::Model::dummy(),
+        Catalog::default(),
+        Config::default(),
+        None,
+    )
+    .expect("empty fixture picker must build");
     let models = ModelCatalog::empty();
     let tools = ToolCatalog::new(tools).expect("fixture tools are unique");
     run_core(
diff --git a/crates/promptforge-tool-picker/src/model.rs b/crates/promptforge-tool-picker/src/model.rs
index eed768c7..170f5c2c 100644
--- a/crates/promptforge-tool-picker/src/model.rs
+++ b/crates/promptforge-tool-picker/src/model.rs
@@ -33,7 +33,7 @@ use crate::error::{ModelLoadError, QueryError};
 #[non_exhaustive]
 pub struct Model {
     /// The loaded encoder, shared cheaply across clones and pickers.
-    encoder: Arc<Encoder>,
+    encoder: Option<Arc<Encoder>>,
 }
 
 impl Model {
@@ -69,13 +69,54 @@ impl Model {
     #[must_use = "loading the model is the expensive step; keep the handle to reuse it"]
     pub fn load_with_progress(progress: Option<&ProgressHandle>) -> Result<Self, ModelLoadError> {
         Ok(Self {
-            encoder: Arc::new(Encoder::load_with_progress(progress)?),
+            encoder: Some(Arc::new(Encoder::load_with_progress(progress)?)),
         })
     }
 
+    /// Creates an empty dummy model that performs no weight loading or inference.
+    ///
+    /// Suitable for test fixtures with empty catalogs or where semantic
+    /// resolution is not exercised.
+    #[cfg(feature = "test-fixtures")]
+    #[must_use]
+    pub fn dummy() -> Self {
+        Self { encoder: None }
+    }
+
     /// Embeds one text with this model, for crate-internal indexing.
     pub(crate) fn embed(&self, text: &str) -> Result<Vec<f32>, QueryError> {
-        self.encoder.embed(text)
+        if let Some(encoder) = &self.encoder {
+            encoder.embed(text)
+        } else {
+            let mut vector = vec![0.0f32; crate::embed::EMBEDDING_DIMENSIONS];
+            let trimmed = text.trim();
+            if trimmed.is_empty() {
+                vector[0] = 1.0;
+                return Ok(vector);
+            }
+            for word in trimmed.split_whitespace() {
+                let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
+                for byte in word.as_bytes() {
+                    hash = (hash ^ u64::from(*byte)).wrapping_mul(0x0100_0000_01b3);
+                }
+                #[expect(
+                    clippy::cast_possible_truncation,
+                    reason = "embedding dimensions fit in usize"
+                )]
+                let idx = (hash % (crate::embed::EMBEDDING_DIMENSIONS as u64)) as usize;
+                let sign = if (hash >> 32) & 1 == 0 { 1.0 } else { -1.0 };
+                vector[idx] += sign;
+            }
+            let norm = vector.iter().map(|v| v * v).sum::<f32>().sqrt();
+            if norm <= 0.0 || !norm.is_finite() {
+                vector[0] = 1.0;
+            } else {
+                for v in &mut vector {
+                    *v /= norm;
+                }
+            }
+            Ok(vector)
+        }
     }
 
     /// Whether two handles share the same loaded encoder allocation.
@@ -87,7 +128,11 @@ impl Model {
     #[doc(hidden)]
     #[must_use]
     pub fn shares_encoder(&self, other: &Model) -> bool {
-        Arc::ptr_eq(&self.encoder, &other.encoder)
+        match (&self.encoder, &other.encoder) {
+            (Some(a), Some(b)) => Arc::ptr_eq(a, b),
+            (None, None) => true,
+            _ => false,
+        }
     }
 }
 
@@ -137,4 +182,17 @@ mod tests {
         }
         assert!(saw_finished, "completion emits Finished");
     }
+
+    #[test]
+    fn dummy_model_shares_encoder_and_embeds_deterministically() {
+        let dummy1 = Model::dummy();
+        let dummy2 = Model::dummy();
+        assert!(dummy1.shares_encoder(&dummy2));
+        let vec1 = dummy1.embed("some text").expect("embed succeeds");
+        let vec2 = dummy2.embed("some text").expect("embed succeeds");
+        assert_eq!(vec1, vec2);
+        assert_eq!(vec1.len(), crate::embed::EMBEDDING_DIMENSIONS);
+        let norm = vec1.iter().map(|v| v * v).sum::<f32>().sqrt();
+        assert!((norm - 1.0).abs() < 1e-5);
+    }
 }