VINDEX3: executable container, and a refused expert that cannot become a token

.github/workflows/larql-boundary.yml

@@ -72,8 +72,8 @@ jobs:
 
       - name: Run examples
         run: |
-          cargo run -p larql-boundary --example encode_decode
-          cargo run -p larql-boundary --example gate_decision
+          cargo run -p larql-demos --example encode_decode
+          cargo run -p larql-demos --example gate_decision
 
   coverage:
     name: coverage · ubuntu

.github/workflows/larql-core.yml

@@ -83,11 +83,11 @@ jobs:
 
       - name: Run examples
         run: |
-          cargo run -p larql-core --example edge_demo
-          cargo run -p larql-core --example graph_demo
-          cargo run -p larql-core --example algorithm_demo
-          cargo run -p larql-core --example filter_demo
-          cargo run -p larql-core --example serialization_demo
+          cargo run -p larql-demos --example edge_demo
+          cargo run -p larql-demos --example graph_demo
+          cargo run -p larql-demos --example algorithm_demo
+          cargo run -p larql-demos --example filter_demo
+          cargo run -p larql-demos --example serialization_demo
 
   coverage:
     name: coverage · ubuntu

.github/workflows/larql-demos.yml

@@ -0,0 +1,79 @@
+# larql-demos cross-platform CI
+#
+# The crate holds no logic — every target is an `examples/*.rs` binary —
+# so there is nothing to unit-test or measure coverage on. What CI must
+# guarantee is that the demos still compile against the crates they
+# demonstrate, and that the weight-free ones still run.
+#
+# Demos needing model weights are compiled here but not executed; they
+# are exercised by hand against a real vindex.
+
+name: larql-demos
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'crates/larql-demos/**'
+      - 'Cargo.toml'
+      - 'Cargo.lock'
+      - '.github/workflows/larql-demos.yml'
+  pull_request:
+    branches: [main]
+    paths:
+      - 'crates/larql-demos/**'
+      - 'Cargo.toml'
+      - 'Cargo.lock'
+      - '.github/workflows/larql-demos.yml'
+  workflow_dispatch: {}
+
+jobs:
+  test:
+    name: test · ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 30
+
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest, macos-14]
+
+    steps:
+      - uses: actions/checkout@v7
+
+      - name: Install stable Rust
+        uses: dtolnay/rust-toolchain@stable
+        with:
+          components: clippy, rustfmt
+
+      - name: Cache cargo registry + build artefacts
+        uses: actions/cache@v6
+        with:
+          path: |
+            ~/.cargo/registry
+            ~/.cargo/git
+            target
+          key: ${{ runner.os }}-cargo-demos-${{ hashFiles('**/Cargo.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-cargo-demos-
+
+      - name: Format check
+        run: cargo fmt -p larql-demos -- --check
+
+      - name: Check (all targets)
+        run: cargo check -p larql-demos --all-targets
+
+      - name: Clippy (warnings as errors)
+        run: cargo clippy -p larql-demos --all-targets --no-deps -- -D warnings
+
+      # The weight-free demos, i.e. exactly the set larql-core and
+      # larql-boundary used to run before these moved here.
+      - name: Run the weight-free demos
+        run: |
+          cargo run -p larql-demos --example edge_demo
+          cargo run -p larql-demos --example graph_demo
+          cargo run -p larql-demos --example algorithm_demo
+          cargo run -p larql-demos --example filter_demo
+          cargo run -p larql-demos --example serialization_demo
+          cargo run -p larql-demos --example encode_decode
+          cargo run -p larql-demos --example gate_decision

.github/workflows/larql-vindex.yml

@@ -106,6 +106,18 @@ jobs:
       - name: Clippy
         run: cargo clippy -p larql-vindex --all-targets -- -D warnings
 
+      # E0 — VINDEX2 preservation. Named separately from the general test
+      # step so a failure reads as "the shipped generation regressed" rather
+      # than as one red test among hundreds. Any regression blocks merge:
+      # finding one now identifies the introducing commit, finding one after
+      # another ten changes produces archaeology.
+      #
+      # Scope caveat: this is the generation-boundary subset. The full matrix
+      # (decode, WALK, slice, publish/pull) needs multi-GB checkpoints and
+      # runs locally against tests/goldens/e0 via scripts/e0-capture-goldens.sh.
+      - name: E0 - VINDEX2 preservation (generation boundary)
+        run: cargo test -p larql-vindex --test e0_generation_boundary
+
       - name: Tests
         run: cargo test -p larql-vindex
 

Cargo.toml

@@ -8,11 +8,13 @@ members = [
     "crates/larql-core",
     "crates/larql-vindex",
     "crates/larql-vindex-spec",
+    "crates/larql-execution",
     "crates/larql-factory",
     "crates/larql-inference",
     "crates/larql-kv",
     "crates/larql-lql",
     "crates/larql-cli",
+    "crates/larql-demos",
     "crates/larql-server",
     "crates/larql-router",
     "crates/larql-router-protocol",
@@ -34,6 +36,7 @@ default-members = [
     "crates/larql-kv",
     "crates/larql-lql",
     "crates/larql-cli",
+    "crates/larql-demos",
     "crates/larql-server",
     "crates/larql-router",
     "crates/larql-router-protocol",

Makefile

@@ -54,11 +54,11 @@ larql-core-bench:
 	cargo bench -p larql-core --bench graph
 
 larql-core-examples:
-	cargo run -p larql-core --example edge_demo
-	cargo run -p larql-core --example graph_demo
-	cargo run -p larql-core --example algorithm_demo
-	cargo run -p larql-core --example filter_demo
-	cargo run -p larql-core --example serialization_demo
+	cargo run -p larql-demos --example edge_demo
+	cargo run -p larql-demos --example graph_demo
+	cargo run -p larql-demos --example algorithm_demo
+	cargo run -p larql-demos --example filter_demo
+	cargo run -p larql-demos --example serialization_demo
 
 larql-core-coverage:
 	@if ! command -v cargo-llvm-cov >/dev/null 2>&1; then \
@@ -536,8 +536,8 @@ larql-boundary-bench-test:
 	cargo test -p larql-boundary --benches
 
 larql-boundary-examples:
-	cargo run -p larql-boundary --example encode_decode
-	cargo run -p larql-boundary --example gate_decision
+	cargo run -p larql-demos --example encode_decode
+	cargo run -p larql-demos --example gate_decision
 	cargo run -p larql-boundary --example accuracy
 
 larql-boundary-coverage:

README.md

@@ -376,6 +376,15 @@ gemma3-4b.vindex/
   feature_labels.json      # Probe-confirmed labels
 ```
 
+**Container generations.** `index.json`'s `version` is the sole discriminator —
+schemas 1–2 are **VINDEX2** (what `extract` writes, and what every published
+vindex is today), schema 3 is **VINDEX3**, the successor container for sparse
+models. One binary reads both; `larql show` and `larql verify` dispatch on the
+version and describe each generation in its own terms rather than flattening
+one into the other. VINDEX3 is still draft — the ABI is not frozen and
+`extract` does not emit it yet — so in practice everything below is VINDEX2.
+See [`crates/larql-vindex/docs/vindex3-format-spec.md`](crates/larql-vindex/docs/vindex3-format-spec.md).
+
 Three extraction levels:
 
 | Level | CLI Flag | LQL Syntax | Size (f16) | Enables |
@@ -464,7 +473,7 @@ delta on Metal, which the per-engine bench numbers confirm.
 | `markov-rs` | residual stream | derivative | exact logits under arch contract | **98.0** |
 | `markov-rs-codec` | compressed residuals | derivative | bounded KL | **98.1** |
 | `boundary-per-layer` | per-layer codec residuals | derivative | bounded KL per-layer | **98.7** |
-| `unlimited-context` | KV (within window) + checkpoints | derivative | exact within window | 94.2 |
+| `windowed-checkpoint` | KV (within window) + checkpoints | derivative | exact within window | 94.2 |
 | `turbo-quant` | quantised K/V | canonical (destructive) | bounded KL | 85.0 |
 | `boundary-kv` | K/V + boundary frames | canonical | exact logits | composes `standard` |
 | `apollo` | boundary retrieval store | n/a (retrieval) | task-level | orthogonal |