fix: Prevent panic by Number::modulo (#773)

In Number::modulo, it calls Number::ints_to_bigint, which could panic. The reason is that calling .to_integer() isn't enough to guarantee that .to_bigint_owned() will return Some, but Number::ints_to_bigint assumes it will and calls unwrap(). In particular, it might be that it's a float corresponding to an integer that's larger than F64_SAFE_INTEGER. The fix is to not call Number::ints_to_bigint (and indeed to delete that entire function, which is only used in this one place), and instead only call unwrap when Some is returned.

.dir-locals.el

@@ -0,0 +1,12 @@
+;;; Directory Local Variables            -*- no-byte-compile: t; -*-
+;;; For more information see (info "(emacs) Directory Variables")
+
+;; Regorus is a cargo-verus project (package.metadata.verus.verify = true), so
+;; verus-mode.el runs `cargo verus verify' rather than the raw `verus' binary.
+;; The cargo-verus path ignores `package.metadata.verus.ide.extra_args' and
+;; instead reads `verus-cargo-verus-arguments'. We set it here so that Verus is
+;; invoked with the `verus' Cargo feature enabled.
+;;
+;; Everything before `--' is passed to cargo-verus; everything after `--' is
+;; forwarded to the Verus binary. The `--' is required by verus-mode.el.
+((verus-mode . ((verus-cargo-verus-arguments . ("--features" "verus" "--")))))

.github/workflows/codeql.yml

@@ -62,7 +62,7 @@ jobs:
 
     steps:
     - name: Checkout repository
-      uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
     # Setup language-specific dependencies BEFORE CodeQL init for proper tracing setup
     - name: Setup Rust
@@ -86,26 +86,26 @@ jobs:
 
     - name: Setup Python
       if: matrix.language == 'python'
-      uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
       with:
         python-version: '3.10'
 
     - name: Setup Java
       if: matrix.language == 'java-kotlin'
-      uses: actions/setup-java@be666c2fcd27ec809703dec50e508c2fdc7f6654 # v5.2.0
+      uses: actions/setup-java@03ad4de0992f5dab5e18fcb136590ce7c4a0ac95 # v5.6.0
       with:
         distribution: 'corretto'
         java-version: '8'
 
     - name: Setup Go
       if: matrix.language == 'go'
-      uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6.4.0
+      uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
       with:
         go-version: '1.21'
 
     - name: Setup .NET
       if: matrix.language == 'csharp'
-      uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
+      uses: actions/setup-dotnet@a98b56852c35b8e3190ac28c8c2271da59106c68 # v6.0.0
       with:
         global-json-file: ./bindings/csharp/global.json
 
@@ -115,12 +115,12 @@ jobs:
 
     - name: Setup Node.js
       if: matrix.language == 'javascript-typescript'
-      uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
       with:
         node-version: '18'
 
     - name: Initialize CodeQL
-      uses: github/codeql-action/init@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
+      uses: github/codeql-action/init@e4fba868fa4b1b91e1fdab776edc8cfbe6e9fb81 # v4.37.3
       with:
         languages: ${{ matrix.language }}
         build-mode: ${{ matrix.build-mode }}
@@ -141,7 +141,7 @@ jobs:
 
     - name: Setup Ruby
       if: matrix.language == 'rust' && contains(matrix.working-directory, 'ruby')
-      uses: ruby/setup-ruby@c4e5b1316158f92e3d49443a9d58b31d25ac0f8f # v1.306.0
+      uses: ruby/setup-ruby@95ef2b042f9d7a56d8268cba8559e2842e2ad01b # v1.321.0
       with:
         ruby-version: '3.4.2'
         bundler-cache: true
@@ -188,6 +188,6 @@ jobs:
       run: cargo xtask build-wasm --release
 
     - name: Perform CodeQL Analysis
-      uses: github/codeql-action/analyze@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
+      uses: github/codeql-action/analyze@e4fba868fa4b1b91e1fdab776edc8cfbe6e9fb81 # v4.37.3
       with:
         category: "/language:${{matrix.language}}"

.github/workflows/dependabot-refresh-cargo-lockfiles.yml

@@ -33,7 +33,7 @@ jobs:
       # ONLY cargo update and cargo metadata (which do NOT execute build
       # scripts) may run against this checkout. Do NOT add cargo build/check/
       # test/run steps.
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v4.2.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v4.2.2
         with:
           repository: ${{ github.event.pull_request.head.repo.full_name }}
           ref: ${{ github.event.pull_request.head.sha }}

.github/workflows/dependency-audit.yml

@@ -27,7 +27,7 @@ jobs:
           - bindings/wasm/Cargo.lock
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Run cargo audit
         uses: rustsec/audit-check@v2
@@ -53,7 +53,7 @@ jobs:
           - xtask/Cargo.toml
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Setup Rust
         uses: ./.github/actions/toolchains/rust

.github/workflows/feature-matrix.yml

@@ -67,7 +67,7 @@ jobs:
             features: arc,opa-no-std
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - name: Setup Rust toolchain
       uses: ./.github/actions/toolchains/rust
     - name: Cache cargo

.github/workflows/miri.yml

@@ -14,7 +14,7 @@ jobs:
       MIRIFLAGS: "-Zmiri-disable-isolation"
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
       - uses: ./.github/actions/toolchains/rust
         with:
           toolchain: nightly

.github/workflows/pr-extensions.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - name: Setup Rust toolchain
       uses: ./.github/actions/toolchains/rust
     - name: Cache cargo

.github/workflows/pr.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - name: Setup Rust toolchain
       uses: ./.github/actions/toolchains/rust
     - name: Cache cargo

.github/workflows/publish-java.yml

@@ -35,10 +35,10 @@ jobs:
             os: windows-latest
             extension: dll
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
-      - uses: actions/setup-java@be666c2fcd27ec809703dec50e508c2fdc7f6654 # v5.2.0
+      - uses: actions/setup-java@03ad4de0992f5dab5e18fcb136590ce7c4a0ac95 # v5.6.0
         with:
           java-version: 8
           distribution: "corretto"
@@ -46,7 +46,7 @@ jobs:
         with:
           targets: ${{ matrix.target }}
       - if: ${{ matrix.build_cmd == 'zigbuild' }}
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: "3.11"
       - if: ${{ matrix.build_cmd == 'zigbuild' }}
@@ -66,10 +66,10 @@ jobs:
     runs-on: ubuntu-latest
     needs: build
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
-      - uses: actions/setup-java@be666c2fcd27ec809703dec50e508c2fdc7f6654 # v5.2.0
+      - uses: actions/setup-java@03ad4de0992f5dab5e18fcb136590ce7c4a0ac95 # v5.6.0
         with:
           java-version: 8
           distribution: "corretto"

.github/workflows/publish-python.yml

@@ -20,8 +20,8 @@ jobs:
       matrix:
         target: [x86_64, x86, aarch64, armv7, s390x, ppc64le]
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: '3.10'
       - uses: ./.github/actions/toolchains/rust
@@ -52,8 +52,8 @@ jobs:
       matrix:
         target: [x64, x86]
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: '3.10'
           architecture: ${{ matrix.target }}
@@ -84,8 +84,8 @@ jobs:
       matrix:
         target: [x86_64, aarch64, universal2-apple-darwin]
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: '3.10'
       - uses: ./.github/actions/toolchains/rust

.github/workflows/publish-wasm.yml

@@ -15,11 +15,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       # Setup .npmrc file to publish to npm
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '20.x'
           registry-url: 'https://registry.npmjs.org'

.github/workflows/release-plz.yml

@@ -17,13 +17,13 @@ jobs:
       contents: write
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - name: Install Rust toolchain
         uses: ./.github/actions/toolchains/rust
       - name: Run release-plz
-        uses: MarcoIeni/release-plz-action@1528104d2ca23787631a1c1f022abb64b34c1e11 # v0.5.128
+        uses: MarcoIeni/release-plz-action@2eb1d8bcb770b4c48ccfaad919734b38b51958c9 # v0.5.131
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           CARGO_REGISTRY_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}

.github/workflows/rust-clippy.yml

@@ -32,7 +32,7 @@ jobs:
       actions: read # only required for a private repository by github/codeql-action/upload-sarif to get the Action run status
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Setup Rust toolchain
         uses: ./.github/actions/toolchains/rust
@@ -52,7 +52,7 @@ jobs:
 
       - name: Upload analysis results to GitHub
         if: ${{ hashFiles('rust-clippy-results.sarif') != '' }}
-        uses: github/codeql-action/upload-sarif@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v3.29.11
+        uses: github/codeql-action/upload-sarif@e4fba868fa4b1b91e1fdab776edc8cfbe6e9fb81 # v3.29.11
         with:
           sarif_file: rust-clippy-results.sarif
           wait-for-processing: true

.github/workflows/test-c-cpp.yml

@@ -16,7 +16,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
 

.github/workflows/test-csharp.yml

@@ -39,7 +39,7 @@ jobs:
               **/release/libregorus_ffi.dylib
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - uses: ./.github/actions/toolchains/rust
@@ -73,11 +73,11 @@ jobs:
     needs: build-ffi
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0  
       - uses: ./.github/actions/toolchains/rust
-      - uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
+      - uses: actions/setup-dotnet@a98b56852c35b8e3190ac28c8c2271da59106c68 # v6.0.0
         with:
           global-json-file: ./bindings/csharp/global.json
 
@@ -131,13 +131,13 @@ jobs:
             target: aarch64-apple-darwin
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
 
       - uses: ./.github/actions/toolchains/rust
 
-      - uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
+      - uses: actions/setup-dotnet@a98b56852c35b8e3190ac28c8c2271da59106c68 # v6.0.0
         with:
           global-json-file: ./bindings/csharp/global.json
 

.github/workflows/test-ffi.yml

@@ -16,7 +16,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - uses: ./.github/actions/toolchains/rust

.github/workflows/test-go.yml

@@ -16,7 +16,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - uses: ./.github/actions/toolchains/rust
@@ -30,7 +30,7 @@ jobs:
       - name: Fetch FFI crate dependencies
         run: cargo fetch --locked --manifest-path bindings/ffi/Cargo.toml
 
-      - uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6.4.0
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
         with:
           architecture: x64
 

.github/workflows/test-java.yml

@@ -16,11 +16,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
 
-      - uses: actions/setup-java@be666c2fcd27ec809703dec50e508c2fdc7f6654 # v5.2.0
+      - uses: actions/setup-java@03ad4de0992f5dab5e18fcb136590ce7c4a0ac95 # v5.6.0
         with:
           java-version: 8
           distribution: "corretto"

.github/workflows/test-musl.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - uses: ./.github/actions/toolchains/rust
       with:
         targets: x86_64-unknown-linux-musl

.github/workflows/test-no-std.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - uses: ./.github/actions/toolchains/rust
       with:
         targets: thumbv7m-none-eabi

.github/workflows/test-python.yml

@@ -23,7 +23,7 @@ jobs:
     runs-on: ${{ matrix.host.name }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - uses: ./.github/actions/toolchains/rust
@@ -39,7 +39,7 @@ jobs:
       - name: Fetch Python crate dependencies
         run: cargo fetch --locked --manifest-path bindings/python/Cargo.toml --target ${{ matrix.host.target }}
 
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: "3.10"
           architecture: x64
@@ -68,7 +68,7 @@ jobs:
     runs-on: ${{ matrix.host.name }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
       - uses: ./.github/actions/toolchains/rust
@@ -82,7 +82,7 @@ jobs:
       - name: Fetch Python crate dependencies
         run: cargo fetch --locked --manifest-path bindings/python/Cargo.toml
 
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: ${{ matrix.python-version }}
           architecture: x64

.github/workflows/test-ruby.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
 

.github/workflows/test-wasm.yml

@@ -16,7 +16,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
 
@@ -33,7 +33,7 @@ jobs:
         run: cargo fetch --locked --manifest-path bindings/wasm/Cargo.toml
 
       - name: Setup Node
-        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: 22
 

.github/workflows/tests-debug.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
     - name: Setup Rust toolchain
       uses: ./.github/actions/toolchains/rust
     - name: Cache cargo

.github/workflows/verus.yml

@@ -0,0 +1,80 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+#
+name: verus
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+
+env:
+  CARGO_TERM_COLOR: always
+
+# This workflow only checks out code, downloads a pinned Verus release asset,
+# and runs verification. It never writes to the repository, so restrict the
+# GITHUB_TOKEN to read-only access to repository contents.
+permissions:
+  contents: read
+
+jobs:
+  verify:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+    - name: Setup Rust toolchain
+      uses: ./.github/actions/toolchains/rust
+      with:
+        components: ""
+    - name: Cache cargo
+      uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
+      with:
+        shared-key: ${{ runner.os }}-regorus-verus
+    - name: Install Verus and run verification
+      shell: bash
+      run: |
+        set -euxo pipefail
+        asset_url=https://github.com/verus-lang/verus/releases/download/release%2F0.2026.07.12.0b42f4c/verus-0.2026.07.12.0b42f4c-x86-linux.zip
+        asset_sha256=f6f4f5d08e07d3e1ad721d775bda5ba96b9dd0c73b48fc17f2e071866fbd01c0
+        test -n "$asset_url"
+        curl -fsSL "$asset_url" -o verus.zip
+
+        # Verify the download integrity before trusting/executing its contents.
+        echo "${asset_sha256}  verus.zip" | sha256sum --check --strict
+
+        unzip -q verus.zip -d verus-dist
+
+        # Search under an absolute path so that `find` yields absolute paths;
+        # this keeps the PATH entries below valid regardless of the working
+        # directory.
+        verus_bin="$(find "$PWD/verus-dist" -type f -name verus -perm -u+x | head -n1)"
+        cargo_verus_bin="$(find "$PWD/verus-dist" -type f -name cargo-verus -perm -u+x | head -n1)"
+        version_json="$(find "$PWD/verus-dist" -type f -name version.json | head -n1)"
+        test -n "$verus_bin"
+        test -n "$cargo_verus_bin"
+        test -n "$version_json"
+
+        # Verus is built against a specific Rust toolchain and refuses to run
+        # against any other version. Read the required toolchain from the
+        # release metadata so we track it automatically instead of hardcoding.
+        required_toolchain="$(sed -n 's/.*"toolchain"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p' "$version_json")"
+        test -n "$required_toolchain"
+        echo "Verus requires Rust toolchain: $required_toolchain"
+
+        # Install the exact toolchain Verus expects, including the extra
+        # components (rustc-dev, llvm-tools) that Verus links against and that
+        # are not part of the default rustup profile.
+        rustup toolchain install "$required_toolchain" \
+          --profile minimal \
+          --component rustc-dev --component llvm-tools --component rustfmt
+
+        # Force cargo/rustc to resolve to the Verus toolchain for the commands
+        # below, overriding any repository/directory toolchain override.
+        export RUSTUP_TOOLCHAIN="$required_toolchain"
+
+        # Put cargo-verus on PATH for the commands below.
+        export PATH="$(dirname "$cargo_verus_bin"):$(dirname "$verus_bin"):$PATH"
+        cargo verus --help
+        cargo fetch --locked
+        cargo verus verify --locked --features verus

CHANGELOG.md

@@ -6,6 +6,35 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.11.0](https://github.com/microsoft/regorus/compare/regorus-v0.10.1...regorus-v0.11.0) - 2026-07-21
+
+### Added
+
+- *(compiler)* support registered host-await builtins for natural function call syntax ([#667](https://github.com/microsoft/regorus/pull/667))
+- *(value)* introduce Set storage abstraction ([#740](https://github.com/microsoft/regorus/pull/740))
+
+### Fixed
+
+- *(rvm)* assert every-quantifier results so failing cases don't pass ([#765](https://github.com/microsoft/regorus/pull/765))
+- `Engine::add_data` now deep-merges nested data documents instead of only merging top-level keys. Adding `{ "a": { "x": 1 } }` followed by `{ "a": { "y": 2 } }` now yields `{ "a": { "x": 1, "y": 2 } }` (matching OPA's data-document merge). Nested sets under a shared key are unioned. Only genuine leaf conflicts (the same path holding two different values) are reported as errors. ([#760](https://github.com/microsoft/regorus/pull/760))
+- A zero-arg function producing two different complete values (e.g. `f() := { "a": 1 }` and `f() := { "b": 2 }`) is now reported as a conflict, matching OPA's complete-rule semantics, instead of silently combining the outputs.
+
+### Security
+
+- `Engine::add_data` now rejects data nested beyond 128 levels instead of risking a stack overflow on adversarially deep input.
+
+### Other
+
+- *(deps)* bump the rust-dependencies group across 5 directories with 11 updates ([#764](https://github.com/microsoft/regorus/pull/764))
+- Expand keyword-in-ref coverage for complex parser edge cases (interpreter + RVM) ([#744](https://github.com/microsoft/regorus/pull/744))
+- *(deps)* bump the rust-dependencies group across 5 directories with 4 updates ([#754](https://github.com/microsoft/regorus/pull/754))
+- *(deps)* bump the rust-dependencies group across 5 directories with 6 updates ([#750](https://github.com/microsoft/regorus/pull/750))
+- *(value)* migrate Value::Object to Object storage abstraction ([#736](https://github.com/microsoft/regorus/pull/736))
+- normalize path separators in folder filter on Windows ([#742](https://github.com/microsoft/regorus/pull/742))
+- Introduce Object storage abstraction ([#735](https://github.com/microsoft/regorus/pull/735))
+- *(rvm)* add debug-mode invariant assertions ([#737](https://github.com/microsoft/regorus/pull/737))
+- *(deps)* bump the rust-dependencies group across 5 directories with 5 updates ([#734](https://github.com/microsoft/regorus/pull/734))
+
 ## [0.10.1](https://github.com/microsoft/regorus/compare/regorus-v0.10.0...regorus-v0.10.1) - 2026-05-22
 
 ### Fixed

Cargo.toml

@@ -8,12 +8,17 @@ members = [
 [package]
 name = "regorus"
 description = "A fast, lightweight Rego (OPA policy language) interpreter"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2021"
 license = "MIT AND Apache-2.0 AND BSD-3-Clause"
 repository = "https://github.com/microsoft/regorus"
 keywords = ["interpreter", "no_std", "opa", "policy-as-code", "rego"]
 
+# Support verification with Verus, a Rust verifier (https://github.com/verus-lang/verus)
+
+[package.metadata.verus]
+verify = true
+
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [lib]
@@ -21,6 +26,7 @@ doctest = false
 
 [features]
 default = ["full-opa", "arc", "rvm"]
+verus = ["dep:vstd"]
 
 arc = []
 ast = []
@@ -43,7 +49,7 @@ cache = ["dep:lru"]
 rvm = ["dep:postcard", "dep:indexmap"]
 semver = ["dep:semver"]
 allocator-memory-limits = ["std", "mimalloc", "mimalloc/allocator-memory-limits"]
-std = ["rand/std", "rand/std_rng", "serde_json/std", "indexmap?/std", "msvc_spectre_libs", "dep:parking_lot" ]
+std = ["rand/std", "rand/std_rng", "serde_json/std", "indexmap?/std", "msvc_spectre_libs", "dep:parking_lot", "vstd?/std" ]
 time = ["dep:chrono", "dep:chrono-tz"]
 uuid = ["dep:uuid"]
 urlquery = ["dep:url"]
@@ -98,23 +104,23 @@ rand = ["dep:rand"]
 [dependencies]
 anyhow = { version = "1.0.102", default-features = false }
 serde = {version = "1.0.150", default-features = false, features = ["derive", "rc", "alloc"] }
-serde_json = { version = "1.0.89", default-features = false, features = ["alloc"] }
+serde_json = { version = "1.0.150", default-features = false, features = ["alloc"] }
 hashbrown = { version = "0.17", default-features = false, features = ["default-hasher"], optional = true }
 lazy_static =  { version = "1.4.0", default-features = false }
 thiserror = { version = "2.0", default-features = false }
 
 data-encoding = { version = "2.8.0", optional = true, default-features=false, features = ["alloc"] }
-num-bigint = { version = "0.4", default-features = false }
+num-bigint = { version = "0.5", default-features = false }
 num-traits = { version = "0.2", default-features = false }
 parking_lot = { version = "0.12", optional = true }
-spin = { version = "0.10.0", default-features = false, features = ["mutex", "spin_mutex"] }
+spin = { version = "0.12.0", default-features = false, features = ["mutex", "spin_mutex"] }
 
 globset = { version = "0.4.16", features = ["simd-accel"], default-features = false, optional = true }
 regex = {version = "1.12.3", optional = true, default-features = false }
 semver = {version = "1.0.28", optional = true, default-features = false }
 url = { version = "2.5.4", optional = true }
 uuid = { version = "1.22.0", default-features = false, features = ["v4", "fast-rng"], optional = true }
-jsonschema = { version = "0.46.5", default-features = false, optional = true }
+jsonschema = { version = "0.48.5", default-features = false, optional = true }
 chrono = { version = "0.4.44", optional = true }
 chrono-tz = { version = "0.10.1", optional = true }
 ipnet = { version = "2.12.0", optional = true, default-features = false }
@@ -134,6 +140,11 @@ mimalloc = { package = "regorus-mimalloc", path = "mimalloc", version = "2.2.7",
 indexmap = { version = "2.13.1", default-features = false, features = ["serde"], optional = true }
 postcard = { version = "1.1.3", default-features = false, features = ["alloc"], optional = true }
 
+# Verus-related dependencies.
+# vstd is enabled via the `verus` feature. In no_std builds only the `alloc` feature is used;
+# the crate's `std` feature additionally enables `vstd/std` (matching vstd's default features).
+vstd = { version = "=0.0.0-2026-07-12-0122", optional = true, default-features = false, features = ["alloc"] }
+
 [dev-dependencies]
 anyhow = "1.0.102"
 cfg-if = "1.0.0"
@@ -214,3 +225,7 @@ doctest=false
 # RUSTDOCFLAGS="--cfg docsrs" cargo +nightly doc --all-features --no-deps
 all-features = true
 rustdoc-args = ["--cfg", "docsrs"]
+
+[lints.rust]
+# Allow `verus_keep_ghost` configuration flag (used by Verus)
+unexpected_cfgs = { level = "warn", check-cfg = ['cfg(verus_keep_ghost)'] }

bindings/csharp/Directory.Packages.props

@@ -1,7 +1,7 @@
 <Project>
   <PropertyGroup>
     <ManagePackageVersionsCentrally>true</ManagePackageVersionsCentrally>
-    <RegorusPackageVersion>0.10.1</RegorusPackageVersion>
+    <RegorusPackageVersion>0.11.0</RegorusPackageVersion>
     <RegorusPackageVersionSuffix Condition="'$(VersionSuffix)' != ''">-$(VersionSuffix)</RegorusPackageVersionSuffix>
   </PropertyGroup>
 

bindings/csharp/Regorus.Tests/MemoryGrowthTests.cs

@@ -115,6 +115,10 @@ public class MemoryGrowthTests
 
                 if (i % LogEvery == 0)
                 {
+                    // Collect transient managed garbage so the working-set delta reflects
+                    // retained (leaked) memory rather than uncollected allocations. A real
+                    // native leak from a missed Dispose() would survive GC and still be caught.
+                    ForceFullGc();
                     process.Refresh();
                     var workingSet = process.WorkingSet64;
                     var managed = GC.GetTotalMemory(false);
@@ -228,6 +232,10 @@ public class MemoryGrowthTests
 
                 if (i % LogEvery == 0)
                 {
+                    // Collect transient managed garbage so the working-set delta reflects
+                    // retained (leaked) memory rather than uncollected allocations. A real
+                    // native leak from a missed Dispose() would survive GC and still be caught.
+                    ForceFullGc();
                     process.Refresh();
                     var workingSet = process.WorkingSet64;
                     var managed = GC.GetTotalMemory(false);

bindings/ffi/Cargo.toml

@@ -2,7 +2,7 @@
 
 [package]
 name = "regorus-ffi"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2021"
 license = "MIT AND Apache-2.0 AND BSD-3-Clause"
 
@@ -13,7 +13,7 @@ crate-type = ["cdylib", "staticlib"]
 [dependencies]
 anyhow = "1.0"
 regorus = { path  = "../..", default-features = false }
-serde_json = "1.0.140"
+serde_json = "1.0.150"
 parking_lot = { version = "0.12", optional = true }
 
 [profile.release]

bindings/java/Cargo.toml

@@ -2,7 +2,7 @@
 
 [package]
 name = "regorus-java"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2021"
 repository = "https://github.com/microsoft/regorus/bindings/java"
 description = "Java bindings for Regorus - a fast, lightweight Rego interpreter written in Rust"
@@ -21,6 +21,6 @@ cache = ["regorus/cache"]
 
 [dependencies]
 anyhow = "1.0"
-serde_json = "1.0.112"
+serde_json = "1.0.150"
 jni = "0.22.4"
 regorus = { path  = "../..", default-features = false, features = ["arc", "rvm"] }

bindings/java/pom.xml

@@ -9,7 +9,7 @@
 
   <groupId>com.microsoft.regorus</groupId>
   <artifactId>regorus-java</artifactId>
-  <version>0.10.1</version>
+  <version>0.11.0</version>
 
   <name>Regorus Java</name>
   <description>Java bindings for Regorus - a fast, lightweight Rego interpreter written in Rust</description>
@@ -97,7 +97,7 @@
 
       <plugin>
         <artifactId>maven-surefire-plugin</artifactId>
-        <version>3.5.5</version>
+        <version>3.5.6</version>
         <configuration>
           <!-- Add debug build to Java path, so it's discoverable by JVM. This is only for tests. -->
           <argLine>-Djava.library.path=${project.basedir}/target/debug:${java.library.path}</argLine>

bindings/python/Cargo.toml

@@ -2,7 +2,7 @@
 
 [package]
 name = "regoruspy"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2021"
 repository = "https://github.com/microsoft/regorus/bindings/python"
 description = "Python bindings for Regorus - a fast, lightweight Rego interpreter written in Rust"
@@ -23,7 +23,7 @@ coverage = ["regorus/coverage"]
 [dependencies]
 anyhow = "1.0"
 ordered-float = "5.3.0"
-pyo3 = { version = "0.28.3", features = ["abi3-py310", "anyhow", "extension-module"] }
+pyo3 = { version = "0.29.0", features = ["abi3-py310", "anyhow", "extension-module"] }
 regorus = { path  = "../..", default-features = false, features = ["arc", "rvm"] }
-serde_json = "1.0.140"
+serde_json = "1.0.150"
 

bindings/python/pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["maturin>=1.4,<2.0"]
+requires = ["maturin>=1.14.1,<2.0"]
 build-backend = "maturin"
 
 [project]

bindings/ruby/Gemfile

@@ -11,6 +11,6 @@ gem "minitest", "~> 6.0"
 gem "rake", "~> 13.4"
 gem "rake-compiler", "~> 1.3"
 gem "rake-compiler-dock", "~> 1.12"
-gem "rubocop", "~> 1.86", require: false
-gem "rubocop-minitest", "~> 0.39.1", require: false
+gem "rubocop", "~> 1.88", require: false
+gem "rubocop-minitest", "~> 0.40.0", require: false
 gem "rubocop-rake", "~> 0.7.1", require: false

bindings/ruby/Gemfile.lock

@@ -9,14 +9,14 @@ GEM
   specs:
     ast (2.4.3)
     drb (2.2.3)
-    json (2.19.4)
-    language_server-protocol (3.17.0.5)
+    json (2.21.1)
+    language_server-protocol (3.17.0.6)
     lint_roller (1.1.0)
-    minitest (6.0.5)
+    minitest (6.0.6)
       drb (~> 2.0)
       prism (~> 1.5)
     parallel (2.1.0)
-    parser (3.3.11.1)
+    parser (3.3.12.0)
       ast (~> 2.4.1)
       racc
     prism (1.9.0)
@@ -26,10 +26,10 @@ GEM
     rake-compiler (1.3.1)
       rake
     rake-compiler-dock (1.12.0)
-    rb_sys (0.9.127)
+    rb_sys (0.9.128)
       rake-compiler-dock (= 1.12.0)
     regexp_parser (2.12.0)
-    rubocop (1.86.1)
+    rubocop (1.88.2)
       json (~> 2.3)
       language_server-protocol (~> 3.17.0.2)
       lint_roller (~> 1.1.0)
@@ -40,10 +40,10 @@ GEM
       rubocop-ast (>= 1.49.0, < 2.0)
       ruby-progressbar (~> 1.7)
       unicode-display_width (>= 2.4.0, < 4.0)
-    rubocop-ast (1.49.1)
+    rubocop-ast (1.50.0)
       parser (>= 3.3.7.2)
       prism (~> 1.7)
-    rubocop-minitest (0.39.1)
+    rubocop-minitest (0.40.0)
       lint_roller (~> 1.1)
       rubocop (>= 1.75.0, < 2.0)
       rubocop-ast (>= 1.38.0, < 2.0)
@@ -65,8 +65,8 @@ DEPENDENCIES
   rake-compiler (~> 1.3)
   rake-compiler-dock (~> 1.12)
   regorusrb!
-  rubocop (~> 1.86)
-  rubocop-minitest (~> 0.39.1)
+  rubocop (~> 1.88)
+  rubocop-minitest (~> 0.40.0)
   rubocop-rake (~> 0.7.1)
 
 BUNDLED WITH

bindings/ruby/ext/regorusrb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "regorusrb"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2024"
 description = "Ruby bindings for Regorus - a fast, lightweight Rego interpreter written in Rust"
 license = "MIT AND Apache-2.0 AND BSD-3-Clause"

bindings/ruby/lib/regorus/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Regorus
-  VERSION = "0.10.1"
+  VERSION = "0.11.0"
 end

bindings/wasm/Cargo.toml

@@ -2,7 +2,7 @@
 
 [package]
 name = "regorusjs"
-version = "0.10.1"
+version = "0.11.0"
 edition = "2021"
 repository = "https://github.com/microsoft/regorus/bindings/wasm"
 description = "WASM bindings for Regorus - a fast, lightweight Rego interpreter written in Rust"
@@ -42,7 +42,7 @@ coverage = ["regorus/coverage"]
 [dependencies]
 regorus = { path  = "../..", default-features = false, features = ["arc", "rvm"] }
 serde = { version = "1.0.219", features = ["derive"] }
-serde_json = "1.0.140"
+serde_json = "1.0.150"
 wasm-bindgen = "0.2.100"
 serde-wasm-bindgen = "0.6"
 # Specify uuid as a mandatory dependency so as to enable `js` feature which is now required
@@ -55,7 +55,7 @@ getrandom03 = { package = "getrandom", version = "0.3.1", features = ["std", "wa
 getrandom = { version = "0.4.2", features = ["wasm_js"] }
 
 [dev-dependencies]
-wasm-bindgen-test = "0.3.71"
+wasm-bindgen-test = "0.3.72"
 
 [lints.rust]
 unexpected_cfgs = { level = "warn", check-cfg = ['cfg(wasm_bindgen_unstable_test_coverage)'] }

docs/rvm/architecture.md

@@ -254,7 +254,13 @@ include formatted state snapshots where possible.
 7. **Host await**: In run-to-completion mode, `HostAwait` consumes a response
    from `host_await_responses`. Suspendable mode yields control with a
    `SuspendReason::HostAwait { dest, argument, identifier }` that the host must
-   service.
+   service. The compiler supports two ways to emit `HostAwait`:
+   - **Explicit**: `__builtin_host_await(payload, identifier)` — raw 2-argument
+     form.
+   - **Registered**: `compile_from_policy_with_host_await` accepts a list of
+     `(name, arg_count)` pairs. Calls to registered names are compiled as
+     `HostAwait` with the function name as the identifier literal. Registered
+     names take precedence over user-defined functions and standard builtins.
 8. **Completion**: `Return` wraps the selected register value into
    `InstructionOutcome::Return`, unwinding frames until the entry frame is
    cleared. `RuleReturn` is a specialised variant used by rule execution

docs/rvm/instruction-set.md

@@ -177,6 +177,75 @@ Parameter tables:
 - Suspendable: emits `InstructionOutcome::Suspend` with `SuspendReason::HostAwait`.
   The host must resume with a value that will be written into `dest`.
 
+### Registered host-await builtins
+
+The compiler can be configured with a list of function names that map directly
+to `HostAwait` instructions. This allows policy authors to write natural
+function calls (e.g. `lookup(input.account_id)`) instead of the raw
+`__builtin_host_await(payload, identifier)` builtin.
+
+Registration is done at compile time via `Compiler::compile_from_policy_with_host_await`:
+
+```rust
+let builtins = [("lookup", 1), ("persist", 1)];
+let program = Compiler::compile_from_policy_with_host_await(
+    &compiled_policy, &entry_points, &builtins,
+)?;
+```
+
+Each registered name is a `(name, arg_count)` pair. When the compiler
+encounters a call to a registered name, it emits a `HostAwait` instruction
+with:
+- `arg` = the first argument register
+- `id` = a register loaded with a string literal containing the function name
+
+Both the explicit `__builtin_host_await(arg, id)` call and a registered
+builtin call produce the **same `HostAwait` bytecode instruction**. The only
+difference is how the `id` register is populated: explicit calls take it from
+the second user-supplied argument, while registered calls auto-generate a
+`Load` instruction for the function name string. The VM cannot distinguish
+between the two at runtime.
+
+**Resolution order** in `determine_call_target()`:
+1. `__builtin_host_await` (magic 2-argument form)
+2. Registered host-await builtins (matched by **bare** function name only)
+3. User-defined functions (matched by package-qualified path)
+4. Standard builtins (matched by bare function name)
+
+Registered names shadow both user-defined functions and standard builtins.
+This means `time.parse_duration_ns` can be overridden to route through the
+host instead of the built-in Rust implementation.
+
+**Only unqualified calls are intercepted.** Registration matches a call by
+the name *as written in the policy*. A bare call — `lookup(x)` — is
+intercepted and compiled to a `HostAwait`. A package-qualified call —
+`data.pkg.lookup(x)` — is **not** intercepted; it is resolved normally, as
+if the name were never registered.
+
+```rego
+# "lookup" is registered as a host-await builtin.
+
+package other
+import rego.v1
+lookup(k) := k            # an ordinary rule that happens to share the name
+
+package demo
+import rego.v1
+a := lookup(input.k)             # intercepted     -> HostAwait
+b := data.other.lookup(input.k)  # NOT intercepted -> calls other.lookup
+```
+
+The qualified form is resolved exactly as it would be without registration:
+if a rule exists at that path it is called, otherwise compilation fails with
+`Unknown function`. (A standard builtin like `count` has no qualified form at
+all, so `data.pkg.count(x)` is always an `Unknown function` error, registered
+or not.)
+
+**Argument handling**: The `HostAwait` instruction carries a single `arg`
+register. Registered builtins must use `arg_count: 1`; the compiler rejects
+`arg_count > 1` at registration time. To pass multiple values, use object
+packing: `lookup({"user": x, "resource": y})`.
+
 ---
 
 ## Halt instruction

docs/value/object.md

@@ -0,0 +1,84 @@
+# Object
+
+Opaque container for `Value::Object`'s key→value storage, enabling
+alternative backends without call-site changes.
+
+## Design
+
+`Object` wraps the storage for a key→value collection of `Value`s and
+provides a curated set of methods (`get`, `insert`, `remove`, `iter`,
+`iter_sorted`, `cursor`, serde). The backing store is private; callers
+never see or pattern-match on it, so the representation can change
+without rippling through call sites.
+
+Multiple backends can coexist at runtime. Because the backing store is
+private, different `Object` instances in the same process can use
+different implementations — e.g., a lazy DB-backed object for `input`,
+inline small-map objects for SARIF location records, and a regular
+sorted map elsewhere — all interoperating through the same opaque
+type. This is stronger than the typical Cargo-feature-selected backend
+seen in precedent crates.
+
+Iteration is split intentionally. `iter()` makes no ordering promise,
+which lets backends that don't keep entries sorted skip any sort work.
+`iter_sorted()` returns entries in `Value` order and is what
+serialization and `Ord` rely on for deterministic output. Cursor types
+add resumable, incremental traversal for the RVM iteration state
+without leaking iterator internals.
+
+`Ord` and `PartialOrd` are defined against `iter_sorted()` rather than
+derived from the storage. Two `Object`s built on different backends —
+or with different insertion histories — compare equal whenever their
+sorted entries match, so changing the backend never changes observable
+comparison results.
+
+## Precedents
+
+Other crates that hide storage behind a stable API so the implementation
+can change without breaking callers:
+
+- **`serde_json::Map`** — opaque newtype allowing cargo-feature based
+  swap between `BTreeMap` (canonical order) and `IndexMap` (insertion
+  order).
+- **`toml::Table`** — opaque newtype allowing cargo-feature based swap
+  between `BTreeMap` and `IndexMap`.
+- **`simdjson` DOM** — opaque tree that lazily materializes nodes on
+  access instead of parsing the whole document up front.
+
+## Use cases
+
+- **SARIF small-object pressure** — SARIF reports contain millions of
+  small objects (location records, rule references, message arguments),
+  most with 2-5 keys. A small-map-optimized backend (inline storage
+  for ≤N entries, heap above) eliminates per-object BTreeMap allocation
+  for the common case.
+
+- **Kubernetes admission policies** — large, deeply-nested resource
+  objects (Pod specs, CRDs) where policies typically touch a handful
+  of paths. A lazy-materializing backend (`LazyObjectProvider` over
+  the incoming JSON) parses only the accessed subtrees.
+
+- **Azure Policy aliases** — ARM exposes the same logical property
+  under multiple aliases (e.g. paths like
+  `Microsoft.Compute/virtualMachines/storageProfile.osDisk.managedDisk.id`).
+  An alias-aware backend resolves lookups across canonical and alias
+  forms without rewriting every policy.
+
+- **Azure Policy case-insensitive compare** — ARM property names are
+  case-preserving but case-insensitive on lookup (`tags.Environment`
+  and `tags.environment` resolve identically). A case-insensitive
+  backend centralizes this once at the storage layer instead of at
+  every comparison site.
+
+- **External data sources** — `input` or `data` backed by a database
+  query, CBOR slice, REST endpoint, or other streaming source via a
+  `LazyObjectProvider`. Entries materialize on demand; the policy
+  only pays for what it touches.
+
+- **Eval-time temporaries** — objects constructed during evaluation
+  (comprehensions, intermediate rule results) on a bumpalo arena.
+  The whole arena drops at query end with zero per-entry free cost.
+
+- **Host-language interop** — Python dicts or JS objects accessed via
+  FFI callbacks from the embedding application, without copying into
+  Rust on every binding boundary.

docs/value/set.md

@@ -0,0 +1,79 @@
+# Set
+
+Opaque container for `Value::Set`'s element storage, enabling alternative
+backends without call-site changes. Pairs with [`Object`](object.md) under
+a shared design philosophy.
+
+## Design
+
+`Set` wraps a `BTreeSet<Value>` today but exposes only a curated method
+surface (`contains`, `insert`, `remove`, `iter`, `iter_sorted`, `cursor`,
+`is_subset`, `intersection`, `union`, `difference`, serde). The inner set is
+private — callers cannot pattern-match it or hand out references to the
+backing store, so the backend can change without churn at the ~400 call
+sites that name `Set`.
+
+Two iteration methods reflect a real distinction: `iter()` makes no
+ordering promise (lets future hash/lazy backends skip sorting work);
+`iter_sorted()` guarantees deterministic order (used by serialization and
+`Ord`). Cursor types support incremental traversal needed by the RVM
+iteration state without exposing iterator internals.
+
+`Ord` is hand-written against `iter_sorted` rather than derived, so two
+backends that store elements differently still compare equal when their
+sorted contents match.
+
+## Scenarios enabled
+
+- **Hash-backed storage** — `FxHashSet`-backed inner turns O(log n)
+  membership checks into O(1); swap in for policies where elements aren't
+  compared ordinally.
+- **Lazy/streaming** — wrap a `LazySetProvider` (DB query, CBOR slice,
+  REST endpoint) and materialize elements on demand.
+- **Arena allocation** — bumpalo-backed inner for eval-time temporaries;
+  drop the whole arena at query end with zero per-element free cost.
+- **FFI-backed** — host-language collections (Python set, JS Set) without
+  copying into Rust.
+- **Bloom-filter pre-check** — front a large backing set with a Bloom
+  filter for fast negative-membership tests on read-mostly allowlists.
+
+## Known use cases
+
+- **Azure Policy allowed-values lists** — large allowlists (allowed
+  regions, allowed SKUs, allowed image publishers) compared against
+  single resource values. Hash-backed Set turns O(log n) membership
+  checks into O(1).
+- **SARIF rule deduplication** — collapsing duplicate rule references
+  across thousands of result records. Set-of-objects with structural
+  hashing avoids the BTreeSet sort cost on every insert.
+- **RBAC role membership** — checking whether a principal belongs to any
+  of dozens of role groups. Hash-backed Set scales to thousands of
+  members with constant-time membership.
+- **Azure Policy denied-resource-type sets** — exclusion lists used by
+  deny-effect policies; same hash-backed pattern as allowed-values.
+
+## Precedents
+
+- **`indexmap::IndexSet`** — opaque newtype that pairs hash lookup with
+  insertion-order iteration; precedent for "Set with alternative
+  ordering semantics behind a stable surface."
+- **`hashbrown::HashSet`** — backs Rust's `std::collections::HashSet`
+  and demonstrates a fully swappable backend behind a stable API.
+- **`roaring::RoaringBitmap`** — bitmap-backed integer set. Not
+  applicable to `Value` keys directly, but a precedent for the broader
+  idea of "Set with alternative storage representations chosen by
+  workload shape."
+- **`serde_json`** — note that `serde_json` has no Set equivalent: its
+  Value enum collapses sets into arrays. Regorus's first-class Set with
+  storage abstraction is therefore unusually well-positioned among JSON
+  value libraries.
+
+## Notes
+
+Cursor types are `pub` (referenced by public `IterationState`) but not
+re-exported at the crate root. The crate-internal `Set`/`Map`/`MapEntry`
+aliases for `BTreeSet`/`BTreeMap` in `lib.rs` were renamed to
+`MapSet`/`Map`/`MapEntry` when this type landed, to free the `Set` name
+for the new public type. Future Array and String abstractions follow the
+same shape — see `docs/value/array.md` and `docs/value/string.md` when
+they land.

src/builtins/azure_policy/helpers.rs

@@ -319,7 +319,7 @@ pub fn resolve_path(root: &Value, path: &str) -> Value {
         match &current {
             Value::Object(map) => {
                 let mut next = None;
-                for (key, value) in map.iter() {
+                for (key, value) in map.iter_sorted() {
                     if let Value::String(ref key_str) = *key {
                         if strings::keys::eq(key_str, &segment) {
                             next = Some(value.clone());

src/builtins/azure_policy/template_functions_collection.rs

@@ -8,10 +8,10 @@
 use crate::ast::{Expr, Ref};
 use crate::builtins;
 use crate::lexer::Span;
+use crate::value::Object;
 use crate::value::Value;
 use crate::Rc;
 
-use alloc::collections::BTreeMap;
 use alloc::vec::Vec;
 use anyhow::Result;
 
@@ -72,7 +72,7 @@ fn fn_intersection(
             // Intersection of objects: keep key-value pairs from the first
             // object only when the key exists in every other object AND
             // the value is equal across all of them.
-            let mut result: BTreeMap<Value, Value> = first.as_ref().clone();
+            let mut result: Object = first.as_ref().clone();
             for arg in rest {
                 let Value::Object(ref other) = *arg else {
                     return Ok(Value::Undefined);
@@ -114,7 +114,7 @@ fn fn_union(_span: &Span, _params: &[Ref<Expr>], args: &[Value], _strict: bool)
         Value::Object(_) => {
             // Union of objects: recursive merge. Nested objects are merged
             // recursively; all other types (including arrays) use last-writer-wins.
-            let mut result = BTreeMap::<Value, Value>::new();
+            let mut result = Object::new();
             for arg in args {
                 let Value::Object(ref obj) = *arg else {
                     return Ok(Value::Undefined);
@@ -264,7 +264,7 @@ fn fn_create_object(
         );
     }
 
-    let mut map = BTreeMap::<Value, Value>::new();
+    let mut map = Object::new();
 
     for pair in args.chunks(2) {
         #[allow(clippy::pattern_type_mismatch)]
@@ -280,9 +280,9 @@ fn fn_create_object(
 
 /// Recursively merge two objects.  Nested objects are merged; everything
 /// else (including arrays) uses the value from `incoming`.
-fn merge_objects(base: &BTreeMap<Value, Value>, overlay: &BTreeMap<Value, Value>) -> Value {
+fn merge_objects(base: &Object, overlay: &Object) -> Value {
     let mut result = base.clone();
-    for (k, v) in overlay {
+    for (k, v) in overlay.iter() {
         #[allow(clippy::needless_borrowed_reference)]
         let merged = match (result.get(k), v) {
             (Some(&Value::Object(ref prev)), &Value::Object(ref next)) => merge_objects(prev, next),

src/builtins/azure_policy/template_functions_misc.rs

@@ -8,10 +8,10 @@
 use crate::ast::{Expr, Ref};
 use crate::builtins;
 use crate::lexer::Span;
+use crate::value::Object;
 use crate::value::Value;
 use crate::Rc;
 
-use alloc::collections::BTreeMap;
 use alloc::string::{String, ToString as _};
 use alloc::vec::Vec;
 use anyhow::Result;
@@ -84,8 +84,8 @@ fn fn_items(_span: &Span, _params: &[Ref<Expr>], args: &[Value], _strict: bool)
         return Ok(Value::Undefined);
     };
     let mut result = Vec::with_capacity(obj.len());
-    for (k, v) in obj.as_ref() {
-        let mut entry = BTreeMap::<Value, Value>::new();
+    for (k, v) in obj.iter_sorted() {
+        let mut entry = Object::new();
         entry.insert(Value::from("key"), k.clone());
         entry.insert(Value::from("value"), v.clone());
         result.push(Value::Object(Rc::new(entry)));

src/builtins/encoding.rs

@@ -308,7 +308,7 @@ fn urlquery_encode_object(
 
     {
         let mut pairs = url.query_pairs_mut();
-        for (key, value) in obj.iter() {
+        for (key, value) in obj.iter_sorted() {
             let key = ensure_string(name, &params[0], key)?;
             match value {
                 Value::String(v) => {

src/builtins/graph.rs

@@ -7,10 +7,11 @@ use crate::ast::{Expr, Ref};
 use crate::builtins;
 use crate::builtins::utils::{enforce_limit, ensure_args_count, ensure_object};
 use crate::lexer::Span;
+use crate::value::Object;
 use crate::value::Value;
 use crate::*;
 
-use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::collections::BTreeSet;
 
 use anyhow::{bail, Result};
 
@@ -80,7 +81,7 @@ fn reachable(span: &Span, params: &[Ref<Expr>], args: &[Value], strict: bool) ->
 }
 
 fn visit(
-    graph: &BTreeMap<Value, Value>,
+    graph: &Object,
     visited: &mut BTreeSet<Value>,
     node: &Value,
     path: &mut Vec<Value>,
@@ -211,7 +212,7 @@ fn walk_visit(path: &mut Vec<Value>, value: &Value, paths: &mut Vec<Value>) -> R
             }
         }
         Value::Object(obj) => {
-            for (key, value) in obj.iter() {
+            for (key, value) in obj.iter_sorted() {
                 path.push(key.clone());
                 // Guard path stack growth while traversing object entries.
                 enforce_limit()?;

src/builtins/objects.rs

@@ -205,7 +205,7 @@ fn merge_filters(
                     let vref = match f {
                         Value::Object(obj) => {
                             let obj = Rc::make_mut(obj);
-                            let entry = obj.entry(p.clone()).or_insert_with(Value::new_object);
+                            let entry = obj.get_or_insert_with(p.clone(), Value::new_object);
                             // Guard filter map growth when creating nested objects.
                             enforce_limit()?;
                             entry

src/builtins/strings.rs

@@ -207,7 +207,7 @@ fn to_string(v: &Value, unescape: bool) -> String {
         }
         Value::Object(o) => {
             "{".to_owned()
-                + &o.iter()
+                + &o.iter_sorted()
                     .map(|(k, v)| to_string(k, true) + ": " + &to_string(v, true))
                     .collect::<Vec<String>>()
                     .join(", ")
@@ -568,7 +568,7 @@ fn replace_n(span: &Span, params: &[Ref<Expr>], args: &[Value], _strict: bool) -
     let mut s = ensure_string(name, &params[1], &args[1])?;
 
     let span = params[0].span();
-    for item in obj.as_ref().iter() {
+    for item in obj.as_ref().iter_sorted() {
         match item {
             (Value::String(k), Value::String(v)) => {
                 s = s.replace(k.as_ref(), v.as_ref()).into();

src/builtins/utils.rs

@@ -5,11 +5,12 @@
 use crate::ast::{Expr, Ref};
 use crate::lexer::Span;
 use crate::number::Number;
+use crate::value::Object;
 use crate::Rc;
 use crate::Value;
 use crate::*;
 
-use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::collections::BTreeSet;
 
 use anyhow::{bail, Result};
 
@@ -168,7 +169,7 @@ pub fn ensure_set(fcn: &str, arg: &Expr, v: Value) -> Result<Rc<BTreeSet<Value>>
     })
 }
 
-pub fn ensure_object(fcn: &str, arg: &Expr, v: Value) -> Result<Rc<BTreeMap<Value, Value>>> {
+pub fn ensure_object(fcn: &str, arg: &Expr, v: Value) -> Result<Rc<Object>> {
     Ok(match v {
         Value::Object(o) => o,
         _ => {

src/compiled_policy.rs

@@ -217,7 +217,7 @@ pub(crate) struct CompiledPolicyData {
     pub(crate) default_rules: Map<String, Vec<DefaultRuleInfo>>,
     pub(crate) imports: BTreeMap<String, Ref<Expr>>,
     pub(crate) functions: FunctionTable,
-    pub(crate) rule_paths: Set<String>,
+    pub(crate) rule_paths: MapSet<String>,
     #[cfg(feature = "azure_policy")]
     pub(crate) target_info: Option<TargetInfo>,
     #[cfg(feature = "azure_policy")]

src/engine.rs

@@ -434,7 +434,13 @@ impl Engine {
 
     /// Add data document.
     ///
-    /// The specified data document is merged into existing data document.
+    /// The specified data document is deep-merged into the existing data document. Nested
+    /// objects are merged recursively (matching OPA's data-document merge), so adding
+    /// `{ "a": { "x": 1 } }` and then `{ "a": { "y": 2 } }` yields `{ "a": { "x": 1, "y": 2 } }`.
+    /// A conflict — the same path holding two different values — is an error.
+    ///
+    /// The merge is atomic: if any conflict is detected (including one deep in a nested
+    /// document), the call fails and the existing data document is left unchanged.
     ///
     /// ```
     /// # use regorus::*;
@@ -453,9 +459,13 @@ impl Engine {
     /// // Merge { "z" : 3 }. Conflict error.
     /// assert!(engine.add_data(Value::from_json_str(r#"{ "z" : 3 }"#)?).is_err());
     ///
+    /// // Nested objects are deep-merged. Merge { "y" : { "a" : 10 } } then { "y" : { "b" : 20 } }.
+    /// assert!(engine.add_data(Value::from_json_str(r#"{ "y" : { "a" : 10 } }"#)?).is_ok());
+    /// assert!(engine.add_data(Value::from_json_str(r#"{ "y" : { "b" : 20 } }"#)?).is_ok());
+    ///
     /// assert_eq!(
     ///   engine.eval_query("data".to_string(), false)?.result[0].expressions[0].value,
-    ///   Value::from_json_str(r#"{ "x": 1, "y": {}, "z": 2}"#)?
+    ///   Value::from_json_str(r#"{ "x": 1, "y": { "a": 10, "b": 20 }, "z": 2}"#)?
     /// );
     /// # Ok(())
     /// # }
@@ -464,8 +474,29 @@ impl Engine {
         if data.as_object().is_err() {
             bail!("data must be object");
         }
-        self.prepared = false;
-        self.interpreter.get_init_data_mut().merge(data)
+
+        // add_data is all-or-nothing; the atomic strategy differs by build because the failure
+        // modes do: a conflict (same path, differing values) is possible everywhere, an
+        // allocator-limit failure mid-merge only under `allocator-memory-limits`.
+        #[cfg(not(feature = "allocator-memory-limits"))]
+        {
+            // Conflict is the only failure mode; `check_mergeable` catches it up front without
+            // allocating, so validate then deep-merge in place (zero-copy fast path).
+            self.interpreter.get_init_data().check_mergeable(&data)?;
+            self.prepared = false;
+            self.interpreter.get_init_data_mut().deep_merge(data)
+        }
+        #[cfg(feature = "allocator-memory-limits")]
+        {
+            // A limit failure can strike mid-merge and can't be predicted, so merge into a
+            // candidate and commit only on success. `Value` is copy-on-write, so only touched
+            // subtrees are cloned.
+            let mut candidate = self.interpreter.get_init_data().clone();
+            candidate.deep_merge(data)?;
+            *self.interpreter.get_init_data_mut() = candidate;
+            self.prepared = false;
+            Ok(())
+        }
     }
 
     /// Get the data document.

src/interpreter.rs

@@ -28,7 +28,6 @@ use crate::{Expression, Extension, Location, QueryResult, QueryResults};
 use crate::query::traversal::traverse;
 
 use crate::Rc;
-use alloc::collections::btree_map::Entry as BTreeMapEntry;
 use alloc::collections::{BTreeMap, BTreeSet};
 use anyhow::{anyhow, bail, Result};
 use core::ops::Bound::*;
@@ -61,6 +60,17 @@ enum FunctionModifier {
     Value(Value),
 }
 
+/// How [`Interpreter::update_data`] merges a rule's value into the data document.
+#[derive(Debug, Clone, Copy)]
+enum RuleValueMerge {
+    /// Shallow-merge keeping disjoint keys, so rules sharing a path prefix scaffold into one
+    /// object (`a.foo` + `a.bar` → one `a`) instead of conflicting.
+    Combine,
+    /// Complete-rule semantics: existing value must be absent or exactly equal, else conflict.
+    /// Used for zero-arg function outputs (`f() := …`), which OPA treats like complete rules.
+    Strict,
+}
+
 type RuleValues = BTreeMap<Vec<Value>, (Value, Ref<Expr>)>;
 
 #[derive(Debug)]
@@ -1248,7 +1258,34 @@ impl Interpreter {
             // Apply with modifiers.
             for wm in &stmt.with_mods {
                 let path = Parser::get_path_ref_components(&wm.refr)?;
-                let path: Vec<&str> = path.iter().map(|s| s.text()).collect();
+                let mut path: Vec<String> = path.iter().map(|s| s.text().to_string()).collect();
+
+                // Matching OPA, a leading import alias is rewritten before
+                // any lookups: functions register as overrides below,
+                // anything else becomes a data override. Only the alias
+                // component is replaced so bracketed keys containing dots
+                // survive the rewrite.
+                let rewritten: Option<Vec<String>> = match path.split_first() {
+                    Some((head, rest)) if head.as_str() != "data" => {
+                        self.lookup_import(head).and_then(|import_expr| {
+                            // Use the import target's parsed components, not
+                            // its dot-joined string, so bracketed keys
+                            // containing dots survive in the import path too.
+                            let comps = Parser::get_path_ref_components(import_expr).ok()?;
+                            Some(
+                                comps
+                                    .iter()
+                                    .map(|s| s.text().to_string())
+                                    .chain(rest.iter().cloned())
+                                    .collect(),
+                            )
+                        })
+                    }
+                    _ => None,
+                };
+                if let Some(new_path) = rewritten {
+                    path = new_path;
+                }
                 let mut target = path.join(".");
 
                 let mut target_is_function = self.lookup_function_by_name(&target).is_some()
@@ -1287,11 +1324,17 @@ impl Interpreter {
                             if self.lookup_function_by_name(&function_path).is_none() {
                                 // Lookup without current module path prefixed.
                                 function_path = get_path_string(&wm.r#as, None)?;
-                                if self.lookup_function_by_name(&function_path).is_none()
-                                    && !Self::is_builtin(wm.r#as.span(), &function_path)
-                                {
-                                    // bail!(wm.r#as.span().error("could not evaluate expression"));
-                                    skip_exec = true;
+                                if self.lookup_function_by_name(&function_path).is_none() {
+                                    // Resolve an aliased replacement before builtins.
+                                    let resolved = self
+                                        .resolve_fcn_path_through_imports(&function_path)
+                                        .filter(|r| self.compiled_policy.functions.contains_key(r));
+                                    if let Some(resolved) = resolved {
+                                        function_path = resolved;
+                                    } else if !Self::is_builtin(wm.r#as.span(), &function_path) {
+                                        // bail!(wm.r#as.span().error("could not evaluate expression"));
+                                        skip_exec = true;
+                                    }
                                 }
                             }
                             self.with_functions
@@ -1312,10 +1355,10 @@ impl Interpreter {
                                 *obj = Value::new_object();
                             }
 
-                            obj = obj
-                                .as_object_mut()?
-                                .entry(Value::String(p.to_string().into()))
-                                .or_insert(Value::new_object());
+                            obj = obj.as_object_mut()?.get_or_insert_with(
+                                Value::String(p.to_string().into()),
+                                Value::new_object,
+                            );
                         }
                         *obj = value;
                         // Mark modified rules as processed.
@@ -1682,8 +1725,7 @@ impl Interpreter {
                     let set = obj
                         .as_object_mut()
                         .map_err(|_| anyhow!(span.error("previous value is not an object")))?
-                        .entry(p)
-                        .or_insert(Value::new_set())
+                        .get_or_insert_with(p, Value::new_set)
                         .as_set_mut()
                         .map_err(|_| anyhow!(span.error("previous value is not a set")))?;
                     set.append(value.as_set_mut()?);
@@ -1691,20 +1733,13 @@ impl Interpreter {
                     let obj = obj
                         .as_object_mut()
                         .map_err(|_| anyhow!(span.error("previous value is not an object")))?;
-                    match obj.entry(p) {
-                        BTreeMapEntry::Vacant(v) => {
-                            if value != Value::Undefined {
-                                v.insert(value);
-                            } else {
-                                // TODO: clean this assumption between Undefined vs Object.
-                                v.insert(Value::new_object());
-                            }
-                        }
-                        BTreeMapEntry::Occupied(o) => {
-                            if o.get() != &value && value != Value::Undefined {
-                                bail!(span
-                                    .error("complete rules should not produce multiple outputs"))
-                            }
+                    if value == Value::Undefined {
+                        // TODO: clean this assumption between Undefined vs Object.
+                        obj.get_or_insert_with(p, Value::new_object);
+                    } else {
+                        let existing = obj.get_or_insert_with(p, || value.clone());
+                        if *existing != value {
+                            bail!(span.error("complete rules should not produce multiple outputs"))
                         }
                     }
                 }
@@ -1713,8 +1748,7 @@ impl Interpreter {
                 obj = obj
                     .as_object_mut()
                     .map_err(|_| anyhow!(span.error("previous value is not an object")))?
-                    .entry(p)
-                    .or_insert(Value::new_object());
+                    .get_or_insert_with(p, Value::new_object);
             }
         }
         Ok(())
@@ -1822,8 +1856,7 @@ impl Interpreter {
                     let set = ctx_mut
                         .rule_value
                         .as_object_mut()?
-                        .entry(Value::from_array(comps))
-                        .or_insert(Value::new_set());
+                        .get_or_insert_with(Value::from_array(comps), Value::new_set);
                     if output != Value::Undefined {
                         set.as_set_mut()?.insert(output);
                         return Ok(true);
@@ -1832,20 +1865,13 @@ impl Interpreter {
                 }
 
                 // Non-set rule.
-                match ctx_mut
-                    .rule_value
-                    .as_object_mut()?
-                    .entry(Value::from_array(comps))
-                {
-                    BTreeMapEntry::Vacant(v) => {
-                        v.insert(output);
-                    }
-                    BTreeMapEntry::Occupied(o) if o.get() != &output => bail!(rule_ref
+                let key = Value::from_array(comps);
+                let obj_mut = ctx_mut.rule_value.as_object_mut()?;
+                let existing = obj_mut.get_or_insert_with(key, || output.clone());
+                if *existing != output {
+                    bail!(rule_ref
                         .span()
-                        .error("rules must not produce multiple outputs")),
-                    _ => {
-                        // Rule produced same value.
-                    }
+                        .error("rules must not produce multiple outputs"));
                 }
 
                 return Ok(true);
@@ -2378,6 +2404,72 @@ impl Interpreter {
         }
     }
 
+    /// Look up the import of the current module with the given alias, e.g.
+    /// the `data.a.b` import expression for `b` after `import data.a.b`.
+    fn lookup_import(&self, alias: &str) -> Option<&Ref<Expr>> {
+        if self.compiled_policy.imports.is_empty() {
+            return None;
+        }
+        let import_key = format!("{}.{}", self.current_module_path, alias);
+        self.compiled_policy.imports.get(&import_key)
+    }
+
+    /// Look up the dot-joined target path of an import of the current module
+    /// with the given alias, e.g. `data.a.b` for `b` after `import data.a.b`.
+    fn lookup_import_alias(&self, alias: &str) -> Option<String> {
+        get_path_string(self.lookup_import(alias)?, None).ok()
+    }
+
+    /// Rewrite a path whose leading component is an import alias of the
+    /// current module to the import's target, e.g. `b.f` to `data.a.b.f`
+    /// after `import data.a.b`.
+    fn rewrite_path_through_imports(&self, path: &str) -> Option<String> {
+        if path.starts_with("data.") {
+            return None;
+        }
+
+        let (alias, rest) = match path.split_once('.') {
+            Some((alias, rest)) => (alias, Some(rest)),
+            None => (path, None),
+        };
+        let target = self.lookup_import_alias(alias)?;
+        Some(match rest {
+            Some(rest) => format!("{target}.{rest}"),
+            None => target,
+        })
+    }
+
+    /// Rewrite an import-aliased call path to its target, e.g. `b.f(1)` to
+    /// `data.a.b.f` after `import data.a.b`. Resolves only when the target is
+    /// a known function or default function, so an alias whose target defines
+    /// the called function shadows a like-named builtin namespace, while other
+    /// spellings keep their prior meaning (e.g. a builtin call). OPA instead
+    /// rewrites aliases unconditionally and rejects calls to a missing target
+    /// at compile time.
+    fn resolve_fcn_path_through_imports(&self, path: &str) -> Option<String> {
+        let candidate = self.rewrite_path_through_imports(path)?;
+        (self.compiled_policy.functions.contains_key(&candidate)
+            || self.is_default_function(&candidate))
+        .then_some(candidate)
+    }
+
+    /// True if `path` is the exact path of a `default` function rule.
+    /// `default_rules` also indexes every prefix of a rule path, so it cannot
+    /// be consulted alone: `rule_paths` holds only exact rule paths, and the
+    /// non-empty argument list distinguishes functions from value rules.
+    fn is_default_function(&self, path: &str) -> bool {
+        self.compiled_policy.rule_paths.contains(path)
+            && self
+                .compiled_policy
+                .default_rules
+                .get(path)
+                .is_some_and(|rules| {
+                    rules.iter().any(|(rule, _)| {
+                        matches!(rule.as_ref(), Rule::Default { args, .. } if !args.is_empty())
+                    })
+                })
+    }
+
     fn eval_builtin_call(
         &mut self,
         span: &Span,
@@ -2471,7 +2563,7 @@ impl Interpreter {
             }
             Value::Object(map) => {
                 s.push('{');
-                for (idx, (k, entry_value)) in map.iter().enumerate() {
+                for (idx, (k, entry_value)) in map.iter_sorted().enumerate() {
                     if idx > 0 {
                         s.push_str(", ");
                     }
@@ -2549,6 +2641,13 @@ impl Interpreter {
             param_values.push(self.eval_expr(p)?);
         }
 
+        // Resolve a leading import alias before the `with` override and builtin
+        // lookups, so an override keyed by the full path reaches aliased calls
+        // and the alias shadows a like-named builtin namespace (matching OPA).
+        let fcn_path = self
+            .resolve_fcn_path_through_imports(&fcn_path)
+            .unwrap_or(fcn_path);
+
         let orig_fcn_path = fcn_path.clone();
 
         let mut with_functions_saved = None;
@@ -2725,7 +2824,12 @@ impl Interpreter {
             let value = match self.eval_rule_bodies(ctx, span, bodies) {
                 Ok(v) => v,
                 Err(e) => {
-                    // If the rule produces an error, save the error.
+                    // If the rule produces an error, save the error. Restore
+                    // the caller's module even so: leaving the callee's module
+                    // in place would make the rest of the caller's body
+                    // resolve paths through the wrong module's imports when
+                    // the error is swallowed below in non-strict mode.
+                    self.set_current_module(prev_module)?;
                     errors.push(e);
                     self.scopes = scopes;
                     continue;
@@ -3426,6 +3530,23 @@ impl Interpreter {
         }
     }
 
+    /// Materialize a complete-rule value: the existing value must be absent or *exactly equal*
+    /// to `new`, else it is a conflict.
+    ///
+    /// Unlike the shallow [`Self::merge_rule_value`], differing outputs conflict instead of
+    /// combining — `f() := {"a": 1}` and `f() := {"b": 2}` conflict — matching OPA's semantics
+    /// for zero-arg functions.
+    fn merge_rule_value_strict(span: &Span, value: &mut Value, new: Value) -> Result<()> {
+        if *value == Value::Undefined {
+            *value = new;
+            Ok(())
+        } else if *value == new {
+            Ok(())
+        } else {
+            Err(span.error("rules should not produce multiple outputs."))
+        }
+    }
+
     pub fn get_path_string(refr: &Expr, document: Option<&str>) -> Result<String> {
         let mut comps = vec![];
         let mut expr_opt = Some(refr);
@@ -3681,6 +3802,7 @@ impl Interpreter {
         _refr: &Expr,
         path: &[&str],
         value: Value,
+        merge: RuleValueMerge,
     ) -> Result<()> {
         if value == Value::Undefined {
             return Ok(());
@@ -3688,7 +3810,10 @@ impl Interpreter {
         // Ensure that path is created.
         let vref = Self::make_or_get_value_mut(&mut self.data, path)?;
         if Self::get_value_chained(self.init_data.clone(), path) == Value::Undefined {
-            Self::merge_rule_value(span, vref, value)
+            match merge {
+                RuleValueMerge::Strict => Self::merge_rule_value_strict(span, vref, value),
+                RuleValueMerge::Combine => Self::merge_rule_value(span, vref, value),
+            }
         } else {
             // Retain specified value.
             Ok(())
@@ -3796,7 +3921,13 @@ impl Interpreter {
                         // `a` is created as an empty object.
                         if let Some((_, prefix)) = path.split_last() {
                             if !prefix.is_empty() {
-                                self.update_data(span, refr, prefix, Value::new_object())?;
+                                self.update_data(
+                                    span,
+                                    refr,
+                                    prefix,
+                                    Value::new_object(),
+                                    RuleValueMerge::Combine,
+                                )?;
                             }
                         }
 
@@ -3808,7 +3939,13 @@ impl Interpreter {
                             };
 
                             let value = self.eval_rule_bodies(ctx, span, rule_body)?;
-                            self.update_data(refr.span(), refr, &path[..], value)?;
+                            self.update_data(
+                                refr.span(),
+                                refr,
+                                &path[..],
+                                value,
+                                RuleValueMerge::Strict,
+                            )?;
                         }
                     }
                 }
@@ -4055,6 +4192,7 @@ impl Interpreter {
                         rule_refr,
                         &prefix_path,
                         Value::new_object(),
+                        RuleValueMerge::Combine,
                     )?;
                 }
             }

src/languages/azure_policy/aliases/denormalizer/mod.rs

@@ -213,10 +213,10 @@ pub fn denormalize_with_aliases(
     // Phase 4: Attach properties to result.
     if !properties.is_empty() {
         if let Some(Value::Object(existing_rc)) = result.get_mut("properties") {
-            // Merge directly into the BTreeMap, avoiding full ObjMap round-trip.
+            // Merge directly into the Object, avoiding full ObjMap round-trip.
             let existing = Rc::make_mut(existing_rc);
             for (k, v) in properties {
-                existing.entry(Value::String(k)).or_insert(v);
+                existing.get_or_insert_with(Value::String(k), || v);
             }
         } else {
             obj_insert(&mut result, "properties", make_value(properties));

src/languages/azure_policy/aliases/denormalizer/sub_resource.rs

@@ -7,6 +7,7 @@ use alloc::collections::{BTreeMap, BTreeSet};
 use alloc::string::String;
 use alloc::vec::Vec;
 
+use crate::value::Object;
 use crate::Value;
 
 use super::super::obj_map::{make_value, new_map, obj_insert, val_str, ObjMap};
@@ -141,7 +142,7 @@ fn rewrap_nested_array(
 /// BTreeMap-native recursion for nested sub-resource array re-wrapping,
 /// avoiding ObjMap round-trips on each array element.
 fn rewrap_nested_array_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
+    btree: &mut Object,
     parent_parts: &[&str],
     array_name: &str,
     envelope_fields: &BTreeSet<String>,
@@ -187,10 +188,7 @@ fn rewrap_nested_array_in_btree(
 }
 
 /// Find a key in a BTreeMap using case-insensitive comparison.
-fn find_key_ci_btree(
-    btree: &alloc::collections::BTreeMap<Value, Value>,
-    key: &str,
-) -> Option<Value> {
+fn find_key_ci_btree(btree: &Object, key: &str) -> Option<Value> {
     btree
         .keys()
         .find(|k| val_str(k).is_some_and(|s| s.eq_ignore_ascii_case(key)))

src/languages/azure_policy/aliases/normalizer/element_remap.rs

@@ -6,11 +6,12 @@
 use alloc::string::String;
 use alloc::vec::Vec;
 
+use crate::value::Object;
 use crate::Value;
 
 use super::super::obj_map::{
-    obj_get, obj_get_mut, obj_insert, set_nested_in_btree, set_nested_lowercased,
-    set_nested_verbatim, ObjMap,
+    obj_get, obj_get_mut, obj_insert, set_nested, set_nested_lowercased, set_nested_verbatim,
+    ObjMap,
 };
 use super::super::types::PrecomputedRemap;
 
@@ -118,7 +119,7 @@ fn apply_remap_at_depth(
 /// BTreeMap-native recursion for element-level remap, avoiding ObjMap
 /// round-trips on each array element.
 fn remap_at_depth_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
+    btree: &mut Object,
     array_chain: &[Vec<String>],
     depth: usize,
     source_field: &str,
@@ -177,12 +178,7 @@ fn remap_at_depth_in_btree(
 }
 
 /// Remap a value between dotted paths directly in a BTreeMap.
-fn remap_deep_field_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
-    source: &str,
-    target: &str,
-    lowercase: bool,
-) {
+fn remap_deep_field_in_btree(btree: &mut Object, source: &str, target: &str, lowercase: bool) {
     let val = match read_dotted_path_btree(btree, source) {
         Some(v) => v,
         None => return,
@@ -198,14 +194,11 @@ fn remap_deep_field_in_btree(
         }
         return;
     }
-    set_nested_in_btree(btree, &segments, val, lowercase);
+    set_nested(btree, &segments, val, lowercase);
 }
 
 /// Read a value at a dotted path from a BTreeMap.
-fn read_dotted_path_btree(
-    btree: &alloc::collections::BTreeMap<Value, Value>,
-    path: &str,
-) -> Option<Value> {
+fn read_dotted_path_btree(btree: &Object, path: &str) -> Option<Value> {
     let segments: Vec<&str> = path.split('.').collect();
     let first = segments.first()?;
     let mut cur: &Value = btree.get(&Value::from(*first))?;

src/languages/azure_policy/aliases/normalizer/mod.rs

@@ -13,6 +13,7 @@ mod flatten;
 // Re-export items used by the denormalizer.
 pub(crate) use element_remap::{apply_element_remap, ElementRemap};
 
+use crate::value::Object;
 use crate::Value;
 
 use super::obj_map::{
@@ -109,7 +110,7 @@ pub fn normalize_with_aliases(
 /// Merge `properties` fields into the result map, skipping keys that already
 /// exist.
 fn merge_properties(
-    obj: &alloc::collections::BTreeMap<Value, Value>,
+    obj: &Object,
     result: &mut ObjMap,
     sub_arrays: Option<&alloc::collections::BTreeSet<alloc::string::String>>,
 ) {

src/languages/azure_policy/aliases/obj_map.rs

@@ -4,7 +4,7 @@
 //! Lightweight string-keyed map used during normalization/denormalization.
 //!
 //! Internally uses `hashbrown::HashMap<Rc<str>, Value>` for O(1) lookups,
-//! then converts to `Value::Object` (a `BTreeMap<Value, Value>`) only at
+//! then converts to `Value::Object` (an `Object`) only at
 //! the output boundary via [`make_value`].
 
 use alloc::string::String;
@@ -12,6 +12,7 @@ use alloc::vec::Vec;
 
 use hashbrown::HashMap;
 
+use crate::value::Object;
 use crate::Rc;
 use crate::Value;
 
@@ -81,14 +82,13 @@ pub fn obj_remove(map: &mut ObjMap, key: &str) -> Option<Value> {
 /// Convert an [`ObjMap`] into a [`Value::Object`].
 ///
 /// Keys are converted from `Rc<str>` to `Value::String` and inserted into
-/// a `BTreeMap` to match the `Value::Object` representation.
+/// an `Object` to match the `Value::Object` representation.
 pub fn make_value(map: ObjMap) -> Value {
-    use alloc::collections::BTreeMap;
-    let mut btree = BTreeMap::new();
-    for (k, v) in map {
-        btree.insert(Value::String(k), v);
-    }
-    Value::Object(Rc::new(btree))
+    let obj: Object = map
+        .into_iter()
+        .map(|(k, v)| (Value::String(k), v))
+        .collect();
+    Value::Object(Rc::new(obj))
 }
 
 /// Convert a `Vec<Value>` into a `Value::Array`.
@@ -115,14 +115,14 @@ pub fn extract_type_field(resource: &Value) -> Option<&str> {
     })
 }
 
-/// Convert a `Value::Object` (BTreeMap<Value, Value>) into an [`ObjMap`].
+/// Convert a `Value::Object` (Object) into an [`ObjMap`].
 ///
 /// Non-string keys are silently skipped.
 #[allow(dead_code)]
 pub fn value_to_obj_map(value: &Value) -> Option<ObjMap> {
-    let btree = value.as_object().ok()?;
-    let mut map = ObjMap::with_capacity(btree.len());
-    for (k, v) in btree.iter() {
+    let obj = value.as_object().ok()?;
+    let mut map = ObjMap::with_capacity(obj.len());
+    for (k, v) in obj.iter() {
         if let Value::String(s) = k {
             map.insert(Rc::clone(s), v.clone());
         }
@@ -194,7 +194,7 @@ fn set_nested_inner(obj: &mut ObjMap, segments: &[&str], value: Value, lowercase
     // Descend directly into the BTreeMap, avoiding ObjMap round-trip.
     if let Some(Value::Object(inner_rc)) = obj.get_mut(&*seg) {
         let inner_btree = Rc::make_mut(inner_rc);
-        set_nested_in_btree(
+        set_nested(
             inner_btree,
             segments.get(1..).unwrap_or_default(),
             value,
@@ -203,17 +203,12 @@ fn set_nested_inner(obj: &mut ObjMap, segments: &[&str], value: Value, lowercase
     }
 }
 
-/// Set a value at a path directly in a `BTreeMap<Value, Value>`, creating
+/// Set a value at a path directly in an `Object`, creating
 /// intermediate `Value::Object` nodes as needed.
 ///
 /// This avoids the `btree_to_obj_map` / `obj_map_to_btree` round-trip that
 /// would clone every sibling entry at each nesting level.
-pub fn set_nested_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
-    segments: &[&str],
-    value: Value,
-    lowercase: bool,
-) {
+pub fn set_nested(obj: &mut Object, segments: &[&str], value: Value, lowercase: bool) {
     let Some(&first) = segments.first() else {
         return;
     };
@@ -226,18 +221,18 @@ pub fn set_nested_in_btree(
     let key_val = Value::String(Rc::clone(&key_rc));
 
     if segments.len() == 1 {
-        btree.insert(key_val, value);
+        obj.insert(key_val, value);
         return;
     }
 
     // Ensure an intermediate object exists.
-    if !btree.contains_key(&key_val) {
-        btree.insert(key_val.clone(), make_value(new_map()));
+    if !obj.contains_key(&key_val) {
+        obj.insert(key_val.clone(), make_value(new_map()));
     }
 
-    if let Some(Value::Object(inner_rc)) = btree.get_mut(&key_val) {
+    if let Some(Value::Object(inner_rc)) = obj.get_mut(&key_val) {
         let inner = Rc::make_mut(inner_rc);
-        set_nested_in_btree(
+        set_nested(
             inner,
             segments.get(1..).unwrap_or_default(),
             value,
@@ -352,20 +347,15 @@ fn remove_field_at_depth(obj: &mut ObjMap, array_chain: &[Vec<String>], depth: u
         for elem in inner.iter_mut() {
             if let Value::Object(obj_rc) = elem {
                 let inner_btree = Rc::make_mut(obj_rc);
-                remove_field_at_depth_in_btree(
-                    inner_btree,
-                    array_chain,
-                    depth.saturating_add(1),
-                    field,
-                );
+                remove_field_at_depth_obj(inner_btree, array_chain, depth.saturating_add(1), field);
             }
         }
     }
 }
 
-/// BTreeMap-native recursion for element-level field removal.
-fn remove_field_at_depth_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
+/// Object-native recursion for element-level field removal.
+fn remove_field_at_depth_obj(
+    obj: &mut Object,
     array_chain: &[Vec<String>],
     depth: usize,
     field: &str,
@@ -374,10 +364,10 @@ fn remove_field_at_depth_in_btree(
         let segments: Vec<&str> = field.split('.').collect();
         if segments.len() == 1 {
             if let Some(&seg) = segments.first() {
-                btree.remove(&Value::from(seg));
+                obj.remove(&Value::from(seg));
             }
         } else if segments.len() > 1 {
-            remove_at_dotted_path_in_btree(btree, &segments);
+            remove_at_dotted_path_obj(obj, &segments);
         }
         return;
     };
@@ -389,12 +379,12 @@ fn remove_field_at_depth_in_btree(
 
     let key_val = Value::from(first);
     let arr_val = if nav.len() == 1 {
-        match btree.get_mut(&key_val) {
+        match obj.get_mut(&key_val) {
             Some(v) => v,
             None => return,
         }
     } else {
-        let mut cur: &mut Value = match btree.get_mut(&key_val) {
+        let mut cur: &mut Value = match obj.get_mut(&key_val) {
             Some(v) => v,
             None => return,
         };
@@ -415,27 +405,19 @@ fn remove_field_at_depth_in_btree(
         for elem in inner.iter_mut() {
             if let Value::Object(obj_rc) = elem {
                 let inner_btree = Rc::make_mut(obj_rc);
-                remove_field_at_depth_in_btree(
-                    inner_btree,
-                    array_chain,
-                    depth.saturating_add(1),
-                    field,
-                );
+                remove_field_at_depth_obj(inner_btree, array_chain, depth.saturating_add(1), field);
             }
         }
     }
 }
 
-/// Remove the leaf segment at a dotted path directly in a BTreeMap.
-fn remove_at_dotted_path_in_btree(
-    btree: &mut alloc::collections::BTreeMap<Value, Value>,
-    segments: &[&str],
-) {
+/// Remove the leaf segment at a dotted path directly in an Object.
+fn remove_at_dotted_path_obj(obj: &mut Object, segments: &[&str]) {
     let Some((&leaf, parent_segs)) = segments.split_last() else {
         return;
     };
     if parent_segs.is_empty() {
-        btree.remove(&Value::from(leaf));
+        obj.remove(&Value::from(leaf));
         return;
     }
 
@@ -443,7 +425,7 @@ fn remove_at_dotted_path_in_btree(
         return;
     };
     let first_key = Value::from(first);
-    let parent_val = match btree.get_mut(&first_key) {
+    let parent_val = match obj.get_mut(&first_key) {
         Some(v) => v,
         None => return,
     };

src/languages/azure_policy/compiler/effects.rs

@@ -11,7 +11,7 @@
 //! to fetch a related resource and an optional `existenceCondition` evaluated
 //! inline.
 
-use alloc::collections::BTreeMap;
+use crate::value::Object;
 use alloc::format;
 use alloc::string::ToString as _;
 use alloc::vec::Vec;
@@ -814,7 +814,7 @@ pub(super) fn build_object_from_keys(
     span: &crate::lexer::Span,
 ) -> Result<u8> {
     // Build template: object with all keys set to Undefined.
-    let mut template = BTreeMap::new();
+    let mut template = Object::new();
     for &(key_idx, _) in &keys {
         // key_idx was returned by `add_literal_u16` in the calling code,
         // so it is always in bounds.  We use `.get()` + `?` instead of

src/languages/azure_policy/compiler/metadata.rs

@@ -272,7 +272,7 @@ impl Compiler {
 fn insert_string_set_annotation(
     annot: &mut alloc::collections::BTreeMap<String, Value>,
     key: &str,
-    observed: &BTreeSet<String>,
+    observed: &alloc::collections::BTreeSet<String>,
 ) {
     if !observed.is_empty() {
         let set: BTreeSet<Value> = observed

src/languages/rego/compiler/expressions/collection_literals.rs

@@ -11,8 +11,9 @@ use crate::ast::{Expr, ExprRef};
 use crate::lexer::Span;
 use crate::rvm::instructions::{ArrayCreateParams, ObjectCreateParams, SetCreateParams};
 use crate::rvm::Instruction;
+use crate::value::Object;
 use crate::{Rc, Value};
-use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::collections::BTreeSet;
 use alloc::vec::Vec;
 
 /// Try to evaluate an expression as a compile-time constant.
@@ -43,7 +44,7 @@ pub(in crate::languages::rego::compiler) fn try_eval_const(expr: &Expr) -> Optio
         Expr::Object { fields, .. } => fields
             .iter()
             .map(|(_, k, v)| Some((try_eval_const(k.as_ref())?, try_eval_const(v.as_ref())?)))
-            .collect::<Option<BTreeMap<_, _>>>()
+            .collect::<Option<Object>>()
             .map(|m| Value::Object(Rc::new(m))),
         _ => None,
     }
@@ -117,7 +118,7 @@ impl<'a> Compiler<'a> {
         fields: &[(crate::lexer::Span, ExprRef, ExprRef)],
         span: &Span,
     ) -> Result<Register> {
-        let all_const: Option<BTreeMap<_, _>> = fields
+        let all_const: Option<Object> = fields
             .iter()
             .map(|(_, k, v)| Some((try_eval_const(k.as_ref())?, try_eval_const(v.as_ref())?)))
             .collect();
@@ -166,7 +167,7 @@ impl<'a> Compiler<'a> {
             let mut template_keys = literal_keys.clone();
             template_keys.sort();
 
-            let mut template_obj = BTreeMap::new();
+            let mut template_obj = Object::new();
             for key in &template_keys {
                 template_obj.insert(key.clone(), Value::Undefined);
             }

src/languages/rego/compiler/function_calls.rs

@@ -17,8 +17,20 @@ use crate::lexer::Span;
 use crate::rvm::instructions::{BuiltinCallParams, FunctionCallParams};
 use crate::rvm::Instruction;
 use crate::utils::get_path_string;
-use alloc::{format, string::ToString, vec::Vec};
+use crate::value::Value;
+use alloc::{
+    format,
+    string::{String, ToString},
+    vec::Vec,
+};
 
+/// Resolved destination of a Rego function-call expression. Produced by
+/// [`Compiler::determine_call_target`] and consumed by
+/// [`Compiler::compile_function_call`] to choose which instruction to emit.
+/// Carrying the discrimination in the type (rather than re-matching on a
+/// magic name at the emit site) keeps the host-await handling honest under
+/// future refactors — the compiler will refuse to build if a new variant is
+/// added without updating every match site.
 enum CallTarget {
     User {
         rule_index: u16,
@@ -28,9 +40,14 @@ enum CallTarget {
         builtin_index: u16,
         expected_args: Option<usize>,
     },
-    HostAwait {
-        expected_args: Option<usize>,
-    },
+    /// Explicit `__builtin_host_await(arg, id)` call form (2 user args).
+    /// The identifier is supplied by the policy author at runtime via the
+    /// second argument register.
+    ExplicitHostAwait,
+    /// A registered host-awaitable builtin invoked by its registered name
+    /// (1 user arg). The identifier is the registered name itself and is
+    /// baked into the bytecode as a string literal at compile time.
+    RegisteredHostAwait { identifier: String },
 }
 
 impl<'a> Compiler<'a> {
@@ -46,6 +63,9 @@ impl<'a> Compiler<'a> {
         let original_fcn_path = fcn_path.clone();
         let full_fcn_path = if self.policy.inner.rules.contains_key(&fcn_path) {
             fcn_path
+        } else if let Some(resolved) = self.resolve_fcn_path_through_imports(&original_fcn_path) {
+            // Resolve a leading import alias before module-prefixing and builtins.
+            resolved
         } else {
             get_path_string(fcn, Some(&self.current_package))
                 .map_err(|_| CompilerError::InvalidFunctionExpressionWithPackage.at(&span))?
@@ -59,7 +79,11 @@ impl<'a> Compiler<'a> {
         let expected_args = match &call_target {
             CallTarget::User { expected_args, .. } => *expected_args,
             CallTarget::Builtin { expected_args, .. } => *expected_args,
-            CallTarget::HostAwait { expected_args } => *expected_args,
+            // Both host-await variants have a known fixed arity; carrying it
+            // in the variant lets the rest of the compiler depend on the type
+            // rather than re-matching on the magic name `__builtin_host_await`.
+            CallTarget::ExplicitHostAwait => Some(2),
+            CallTarget::RegisteredHostAwait { .. } => Some(1),
         };
 
         if let Some(expected) = expected_args {
@@ -126,7 +150,8 @@ impl<'a> Compiler<'a> {
                 });
                 self.emit_instruction(Instruction::BuiltinCall { params_index }, &span);
             }
-            CallTarget::HostAwait { .. } => {
+            CallTarget::ExplicitHostAwait => {
+                // Explicit __builtin_host_await(arg, id) — 2 arguments
                 if arg_regs.len() != 2 {
                     return Err(CompilerError::General {
                         message: format!(
@@ -136,7 +161,6 @@ impl<'a> Compiler<'a> {
                     }
                     .at(&span));
                 }
-
                 self.emit_instruction(
                     Instruction::HostAwait {
                         dest,
@@ -146,6 +170,37 @@ impl<'a> Compiler<'a> {
                     &span,
                 );
             }
+            CallTarget::RegisteredHostAwait { identifier } => {
+                // Registered host-awaitable builtin — the identifier is the
+                // registered name and is baked into the bytecode as a literal.
+                if arg_regs.len() != 1 {
+                    return Err(CompilerError::General {
+                        message: format!(
+                            "host-awaitable builtin '{}' expects exactly 1 argument, got {}",
+                            identifier,
+                            arg_regs.len()
+                        ),
+                    }
+                    .at(&span));
+                }
+                let id_reg = self.alloc_register();
+                let literal_idx = self.add_literal(Value::String(identifier.into()));
+                self.emit_instruction(
+                    Instruction::Load {
+                        dest: id_reg,
+                        literal_idx,
+                    },
+                    &span,
+                );
+                self.emit_instruction(
+                    Instruction::HostAwait {
+                        dest,
+                        arg: arg_regs[0],
+                        id: id_reg,
+                    },
+                    &span,
+                );
+            }
         }
 
         if let Some((plan, plan_span)) = &out_param_plan {
@@ -168,6 +223,37 @@ impl<'a> Compiler<'a> {
         Ok(dest)
     }
 
+    /// Rewrite an import-aliased call path to its target, e.g. `b.f(1)` to
+    /// `data.a.b.f` after `import data.a.b`. Resolves only when the target is
+    /// a known function (the `rules` map cannot be used here: it also indexes
+    /// value rules and every rule-path prefix, which must not become callable
+    /// through an alias), so an alias whose target defines the called
+    /// function shadows a like-named builtin namespace, while other spellings
+    /// keep their prior meaning (e.g. a builtin call). OPA instead rewrites
+    /// aliases unconditionally and rejects calls to a missing target at
+    /// compile time.
+    fn resolve_fcn_path_through_imports(&self, path: &str) -> Option<String> {
+        if self.policy.inner.imports.is_empty() || path.starts_with("data.") {
+            return None;
+        }
+        let (alias, rest) = match path.split_once('.') {
+            Some((alias, rest)) => (alias, Some(rest)),
+            None => (path, None),
+        };
+        let import_key = format!("{}.{}", self.current_package, alias);
+        let import_expr = self.policy.inner.imports.get(&import_key)?;
+        let target = get_path_string(import_expr, None).ok()?;
+        let candidate = match rest {
+            Some(rest) => format!("{target}.{rest}"),
+            None => target,
+        };
+        self.policy
+            .inner
+            .functions
+            .contains_key(&candidate)
+            .then_some(candidate)
+    }
+
     fn lookup_builtin_arity(&self, name: &str) -> Option<usize> {
         if name == "print" {
             Some(2)
@@ -187,8 +273,26 @@ impl<'a> Compiler<'a> {
         span: &Span,
     ) -> Result<CallTarget> {
         if original_fcn_path == "__builtin_host_await" {
-            return Ok(CallTarget::HostAwait {
-                expected_args: Some(2),
+            return Ok(CallTarget::ExplicitHostAwait);
+        }
+
+        // Check registered host-awaitable builtins. Registered builtins are
+        // restricted to arg_count == 1 at registration time (see
+        // `Compiler::register_host_await_builtin`), so the variant doesn't
+        // need to carry an arity — it's fixed at 1.
+        //
+        // We deliberately match against `original_fcn_path` only, not
+        // `full_fcn_path`. Registration intercepts the *unqualified* call
+        // form (e.g. `lookup(x)` inside the policy's own package). A
+        // package-qualified call like `data.other.lookup(x)` is left to
+        // resolve through the normal user-defined / builtin path, so a
+        // registered name does not leak into unrelated packages that
+        // happen to expose a rule with the same identifier. This is
+        // documented on `register_host_await_builtin`; the
+        // `registered_host_await.yaml` suite pins the behavior.
+        if self.host_await_builtins.contains_key(original_fcn_path) {
+            return Ok(CallTarget::RegisteredHostAwait {
+                identifier: original_fcn_path.to_string(),
             });
         }
 

src/languages/rego/compiler/loops.rs

@@ -13,7 +13,7 @@ use crate::ast::{self, ExprRef, LiteralStmt, Query};
 use crate::compiler::destructuring_planner::plans::BindingPlan;
 use crate::compiler::hoist::{HoistedLoop, LoopType};
 use crate::lexer::Span;
-use crate::rvm::instructions::{LoopMode, LoopStartParams};
+use crate::rvm::instructions::{GuardMode, LoopMode, LoopStartParams};
 use crate::rvm::Instruction;
 use crate::Value;
 use alloc::format;
@@ -197,6 +197,19 @@ impl<'a> Compiler<'a> {
             *end = loop_end;
         }
 
+        // The loop writes its overall pass/fail into `result_reg`
+        // (`success_count == total_iterations` for `Every`). The enclosing query
+        // must fail (evaluate to undefined) when the quantifier does not hold, so
+        // guard on `result_reg` here. Without this the `every` result is computed
+        // but discarded, leaving the surrounding rule to always succeed.
+        self.emit_instruction(
+            Instruction::Guard {
+                register: result_reg,
+                mode: GuardMode::Condition,
+            },
+            span,
+        );
+
         Ok(())
     }
 
@@ -312,6 +325,25 @@ impl<'a> Compiler<'a> {
             *end = loop_end;
         }
 
+        // A hoisted index-iteration loop inside an `every` body acts as a
+        // condition on the current iteration: if the indexed reference matches
+        // nothing the iteration must fail. The `every` body emits no context
+        // yield, so the loop result register is otherwise discarded (same
+        // situation as `some ... in`). Guard on it so a non-matching indexed
+        // reference fails the enclosing `every` iteration.
+        if matches!(
+            self.context_stack.last().map(|c| &c.context_type),
+            Some(ContextType::Every)
+        ) {
+            self.emit_instruction(
+                Instruction::Guard {
+                    register: result_reg,
+                    mode: GuardMode::Condition,
+                },
+                collection.span(),
+            );
+        }
+
         Ok(())
     }
 

src/languages/rego/compiler/mod.rs

@@ -26,7 +26,9 @@ use crate::rvm::program::{Program, RuleType, SpanInfo};
 use crate::CompiledPolicy;
 use crate::Value;
 use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::format;
 use alloc::string::String;
+use alloc::string::ToString as _;
 use alloc::vec;
 use alloc::vec::Vec;
 use indexmap::IndexMap;
@@ -139,6 +141,10 @@ pub struct Compiler<'a> {
     current_call_stack: Vec<u16>,
     entry_points: IndexMap<String, usize>,
     soft_assert_mode: bool,
+    /// Registered host-awaitable builtins: name → expected arg count.
+    /// When the compiler encounters a call to one of these names, it emits a
+    /// `HostAwait` instruction instead of a regular function or builtin call.
+    host_await_builtins: BTreeMap<String, usize>,
 }
 
 impl<'a> Compiler<'a> {
@@ -173,9 +179,75 @@ impl<'a> Compiler<'a> {
             current_call_stack: Vec::new(),
             entry_points: IndexMap::new(),
             soft_assert_mode: false,
+            host_await_builtins: BTreeMap::new(),
         }
     }
 
+    /// Register a function name as a host-awaitable builtin.
+    ///
+    /// When the compiler encounters an **unqualified** call to `name(arg)`
+    /// (i.e. `name(arg)` from inside the policy's own package, not
+    /// `data.pkg.name(arg)` or any other package-qualified form), it will
+    /// emit a `HostAwait` instruction with the argument and `name` as the
+    /// identifier, instead of treating it as a user-defined or standard
+    /// builtin function.
+    ///
+    /// Package-qualified calls (e.g. `data.other.name(arg)`) are **not**
+    /// intercepted by registration. Those resolve through the normal
+    /// user-defined / builtin lookup against their fully-qualified path
+    /// (`data.other.name`).
+    ///
+    /// `arg_count` must be exactly 1. The `HostAwait` instruction carries a
+    /// single argument register; use object packing to pass multiple values
+    /// (e.g. `name({"key1": v1, "key2": v2})`).
+    ///
+    /// Returns `Err` when:
+    /// - `name` is the reserved identifier `__builtin_host_await`,
+    /// - `name` is empty, only whitespace, or has leading/trailing
+    ///   whitespace (whitespace-padded names would never match the
+    ///   trimmed identifier produced by the Rego parser, creating dead
+    ///   registrations),
+    /// - `name` is already registered (duplicate registration is rejected
+    ///   rather than silently overwritten),
+    /// - `arg_count` is not exactly 1.
+    pub fn register_host_await_builtin(&mut self, name: &str, arg_count: usize) -> Result<()> {
+        if name == "__builtin_host_await" {
+            return Err(CompilerError::General {
+                message: "__builtin_host_await is a reserved name and cannot be registered as a host-await builtin"
+                    .to_string(),
+            }
+            .into());
+        }
+        if name.is_empty() || name != name.trim() {
+            return Err(CompilerError::General {
+                message: format!(
+                    "host-await builtin name {name:?} must not be empty or contain leading/trailing whitespace"
+                ),
+            }
+            .into());
+        }
+        if self.host_await_builtins.contains_key(name) {
+            return Err(CompilerError::General {
+                message: format!(
+                    "host-await builtin '{name}' is already registered; \
+                     duplicate registration is not allowed"
+                ),
+            }
+            .into());
+        }
+        if arg_count != 1 {
+            return Err(CompilerError::General {
+                message: format!(
+                    "registered host-await builtin '{name}' must have arg_count == 1, got {arg_count}. \
+                     Use object packing to pass multiple values."
+                ),
+            }
+            .into());
+        }
+        self.host_await_builtins.insert(name.to_string(), arg_count);
+        Ok(())
+    }
+
     pub(super) fn with_soft_assert_mode<F, R>(&mut self, enabled: bool, f: F) -> R
     where
         F: FnOnce(&mut Self) -> R,

src/languages/rego/compiler/queries.rs

@@ -70,12 +70,31 @@ impl<'a> Compiler<'a> {
                     ..
                 } = &stmt.literal
                 {
-                    self.compile_some_in_loop_with_remaining_statements(
+                    let some_result_reg = self.compile_some_in_loop_with_remaining_statements(
                         key,
                         value,
                         collection,
                         &stmts[idx..],
                     )?;
+                    // Inside an `every` body a `some ... in` acts as a condition
+                    // on the current iteration: if it matches nothing the
+                    // iteration must fail. Unlike a top-level rule body (where
+                    // per-iteration context yields produce the results), the
+                    // `every` body has no yield, so the loop result register is
+                    // otherwise discarded. Guard on it so a `some` that matches
+                    // nothing fails the enclosing `every` iteration.
+                    if matches!(
+                        self.context_stack.last().map(|c| &c.context_type),
+                        Some(ContextType::Every)
+                    ) {
+                        self.emit_instruction(
+                            Instruction::Guard {
+                                register: some_result_reg,
+                                mode: GuardMode::Condition,
+                            },
+                            &stmt.span,
+                        );
+                    }
                     return Ok(());
                 }
             }

src/languages/rego/compiler/rules.rs

@@ -234,8 +234,20 @@ impl<'a> Compiler<'a> {
     pub fn compile_from_policy(
         policy: &CompiledPolicy,
         entry_points: &[&str],
+    ) -> Result<Arc<Program>> {
+        Self::compile_from_policy_with_host_await(policy, entry_points, &[])
+    }
+
+    /// Compile from a CompiledPolicy to RVM Program with registered host-awaitable builtins.
+    pub fn compile_from_policy_with_host_await(
+        policy: &CompiledPolicy,
+        entry_points: &[&str],
+        host_await_builtins: &[(&str, usize)],
     ) -> Result<Arc<Program>> {
         let mut compiler = Compiler::with_policy(policy);
+        for &(name, arg_count) in host_await_builtins {
+            compiler.register_host_await_builtin(name, arg_count)?;
+        }
         compiler.current_rule_path = "".to_string();
         let rules = policy.get_rules();
 

src/lib.rs

@@ -155,7 +155,7 @@ pub mod target;
 #[cfg(any(test, all(feature = "yaml", feature = "std")))]
 pub mod test_utils;
 pub mod utils;
-mod value;
+pub mod value;
 
 #[cfg(feature = "azure_policy")]
 pub use {
@@ -205,10 +205,10 @@ pub use alloc::sync::Arc as Rc;
 pub use alloc::rc::Rc;
 
 #[cfg(feature = "std")]
-use std::collections::{hash_map::Entry as MapEntry, HashMap as Map, HashSet as Set};
+use std::collections::{hash_map::Entry as MapEntry, HashMap as Map, HashSet as MapSet};
 
 #[cfg(not(feature = "std"))]
-use alloc::collections::{btree_map::Entry as MapEntry, BTreeMap as Map, BTreeSet as Set};
+use alloc::collections::{btree_map::Entry as MapEntry, BTreeMap as Map, BTreeSet as MapSet};
 
 use alloc::{
     borrow::ToOwned as _,

src/number.rs

@@ -27,10 +27,14 @@ use num_traits::{One, Signed, ToPrimitive, Zero};
 use serde::ser::Serializer;
 use serde::Serialize;
 
+#[cfg(feature = "verus")]
+use vstd::prelude::*;
+
 use crate::*;
 
 pub type BigInt = NumBigInt;
 
+#[cfg_attr(feature = "verus", verus_verify)]
 const F64_SAFE_INTEGER: f64 = 9_007_199_254_740_992.0; // 2^53
 
 #[derive(Clone)]
@@ -140,13 +144,9 @@ impl Number {
         }
     }
 
-    fn ints_to_bigint(a: &Number, b: &Number) -> (BigInt, BigInt) {
-        (a.to_bigint_owned().unwrap(), b.to_bigint_owned().unwrap())
-    }
-
     fn normalize_float(value: f64) -> Number {
-        if let Some(int) = Self::float_to_small_bigint(value) {
-            return Self::from_bigint_owned(int);
+        if let Some(i) = Self::float_to_small_bigint(value) {
+            return Self::from_bigint_owned(i);
         }
         Number::Float(value)
     }
@@ -587,7 +587,11 @@ impl Number {
                 }
             }
             (Number::Int(a), Number::Int(b)) => {
-                if *a % *b == 0 {
+                if *a == i64::MIN && *b == -1 {
+                    // Rust panics on i64::MIN % -1i64, so handle it specially
+                    let quotient = BigInt::from(*a) / BigInt::from(*b);
+                    Ok(Number::from_bigint_owned(quotient))
+                } else if *a % *b == 0 {
                     if let Some(q) = a.checked_div(*b) {
                         Ok(Number::Int(q))
                     } else {
@@ -656,15 +660,17 @@ impl Number {
     }
 
     pub fn modulo(self, rhs: &Self) -> Result<Number> {
-        if rhs.is_zero() {
+        // Conversion fails for a non-integral float, and also for an integral
+        // one whose magnitude exceeds 2^53, which cannot be represented exactly.
+        let (a, b) = match (self.to_bigint_owned(), rhs.to_bigint_owned()) {
+            (Some(a), Some(b)) => (a, b),
+            _ => bail!("modulo on floating-point number"),
+        };
+
+        if b.is_zero() {
             bail!("modulo by zero");
         }
 
-        if !self.is_integer() || !rhs.is_integer() {
-            bail!("modulo on floating-point number");
-        }
-
-        let (a, b) = Number::ints_to_bigint(&self, rhs);
         let rem = a % &b;
         Ok(Number::from_bigint_owned(rem))
     }
@@ -685,6 +691,7 @@ impl Number {
         }
     }
 
+    #[allow(clippy::if_then_some_else_none)]
     fn ensure_integers(a: &Number, b: &Number) -> Option<(BigInt, BigInt)> {
         if a.is_integer() && b.is_integer() {
             Some((a.to_bigint_owned()?, b.to_bigint_owned()?))
@@ -776,7 +783,8 @@ impl Number {
         if e >= 0 {
             Ok(two_pow_positive(e as u32))
         } else {
-            let denom = two_pow_positive((-e) as u32);
+            // Must cast to i64 before negating in case it's i32::MIN
+            let denom = two_pow_positive((-(e as i64)) as u32);
             Number::from(1u64).divide(&denom)
         }
     }
@@ -785,7 +793,8 @@ impl Number {
         if e >= 0 {
             Ok(ten_pow_positive(e as u32))
         } else {
-            let denom = ten_pow_positive((-e) as u32);
+            // Must cast to i64 before negating in case it's i32::MIN
+            let denom = ten_pow_positive((-(e as i64)) as u32);
             Number::from(1u64).divide(&denom)
         }
     }
@@ -984,3 +993,67 @@ fn scientific_parts_to_bigint(mantissa: &str, exponent: i32) -> Option<BigInt> {
 
     Some(value)
 }
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::expect_used)] // tests expect() to assert arithmetic results
+
+    use super::*;
+    use alloc::string::ToString;
+
+    /// Regression test: `i64::MIN / -1` overflows `i64` and panics in Rust's
+    /// native integer division/remainder. `divide` must promote the result
+    /// instead of panicking.
+    #[test]
+    fn i64_min_by_negative_one() {
+        let quotient = Number::Int(i64::MIN)
+            .divide(&Number::Int(-1))
+            .expect("division should succeed");
+
+        // 2^63 does not fit in i64, but does fit in u64.
+        assert_eq!(quotient.as_u64(), Some(9_223_372_036_854_775_808));
+        assert_eq!(quotient.as_i64(), None);
+        assert_eq!(quotient.as_i128(), Some(9_223_372_036_854_775_808));
+        assert_eq!(
+            *quotient.to_big().expect("to_big should succeed"),
+            -BigInt::from(i64::MIN)
+        );
+
+        // The same overflow case reached via the mixed `Int`/`BigInt` path.
+        let big_quotient = Number::Int(i64::MIN)
+            .divide(&Number::BigInt(Rc::new(BigInt::from(-1))))
+            .expect("division should succeed");
+        assert_eq!(big_quotient.as_u64(), Some(9_223_372_036_854_775_808));
+
+        // `i64::MIN % -1` also panics natively; the result must be zero.
+        let remainder = Number::Int(i64::MIN)
+            .modulo(&Number::Int(-1))
+            .expect("modulo should succeed");
+        assert_eq!(remainder.as_i64(), Some(0));
+    }
+
+    #[test]
+    fn modulo_handles_floats_that_are_really_integers() {
+        // An integral float is a valid operand.
+        assert!(matches!(
+            Number::Float(4.0).modulo(&Number::Int(3)),
+            Ok(Number::UInt(1))
+        ));
+        // `1e300` has no fractional part, but it is too large to convert to an
+        // integer exactly. This must report an error, not panic.
+        assert_eq!(
+            Number::Float(1e300)
+                .modulo(&Number::Int(3))
+                .err()
+                .map(|e| e.to_string()),
+            Some("modulo on floating-point number".to_string())
+        );
+        assert_eq!(
+            Number::Int(3)
+                .modulo(&Number::Float(1e300))
+                .err()
+                .map(|e| e.to_string()),
+            Some("modulo on floating-point number".to_string())
+        );
+    }
+}

src/parser.rs

@@ -354,6 +354,33 @@ impl<'source> Parser<'source> {
         }
     }
 
+    /// Parse a field name after `.` in a ref expression.
+    ///
+    /// Unlike [`Self::parse_var`] and [`Self::parse_ident`], this method accepts **any**
+    /// `TokenKind::Ident` token, including reserved keywords (e.g. `as`, `default`, `else`,
+    /// `false`, `if`, `import`, `in`, `not`, `null`, `package`, `some`, `true`, `with`).
+    ///
+    /// The position immediately after `.` is unambiguously a field name, so there is no
+    /// syntactic ambiguity with statement-level keywords. This matches OPA's
+    /// `keywords_in_refs` capability, which is enabled by default in standard OPA builds.
+    ///
+    /// # Example
+    /// ```rego
+    /// allow if { input.v0.package.format == "npm" }  # `package` is a keyword but valid here
+    /// ```
+    fn parse_ref_field(&mut self) -> Result<Span> {
+        let span = self.tok.1.clone();
+        match self.tok.0 {
+            TokenKind::Ident => {
+                self.next_token()?;
+                Ok(span)
+            }
+            _ => Err(self
+                .source
+                .error(self.tok.1.line, self.tok.1.col, "expecting identifier")),
+        }
+    }
+
     fn read_number(&mut self, span: Span) -> Result<Expr> {
         match Number::from_str(span.text()) {
             Ok(v) => Ok(Expr::Number {
@@ -743,9 +770,10 @@ impl<'source> Parser<'source> {
                     );
                 }
                 "." => {
-                    // Read identifier.
+                    // Read identifier. Keywords are allowed as field names in
+                    // dot-notation refs (e.g. `input.package.name`).
                     self.next_token()?;
-                    let field = self.parse_var()?;
+                    let field = self.parse_ref_field()?;
                     span.end = self.end;
 
                     // Disallow any whitespace between . and identifier.
@@ -1418,9 +1446,10 @@ impl<'source> Parser<'source> {
                     );
                 }
                 "." => {
-                    // Read identifier.
+                    // Read identifier. Keywords are allowed as field names in
+                    // dot-notation refs (e.g. `import data.my.package`).
                     self.next_token()?;
-                    let field = self.parse_ident()?;
+                    let field = self.parse_ref_field()?;
                     span.end = self.end;
 
                     // Disallow any whitespace between . and identifier.
@@ -1523,7 +1552,8 @@ impl<'source> Parser<'source> {
                 "." => {
                     let sep_pos = self.tok.1.start;
                     self.next_token()?;
-                    let field = self.parse_var()?;
+                    // Keywords are allowed as field names in dot-notation refs.
+                    let field = self.parse_ref_field()?;
                     span.end = self.end;
 
                     // Disallow any whitespace between . and identifier.

src/rvm/program/metadata.rs

@@ -11,9 +11,9 @@
 //! values are converted through [`MetadataValue`] — a postcard/bincode-safe
 //! enum that avoids `deserialize_any`.
 
+use crate::value::Object;
 use crate::Rc;
-use alloc::collections::BTreeMap;
-use alloc::collections::BTreeSet;
+use alloc::collections::{BTreeMap, BTreeSet};
 use alloc::string::String;
 use alloc::vec::Vec;
 use serde::{Deserialize, Serialize};
@@ -52,7 +52,7 @@ impl ProgramMetadata {
     pub fn to_value(&self) -> crate::value::Value {
         use crate::value::Value;
 
-        let mut obj = BTreeMap::new();
+        let mut obj = Object::new();
         obj.insert(
             Value::String("compiler_version".into()),
             Value::String(self.compiler_version.as_str().into()),
@@ -75,7 +75,7 @@ impl ProgramMetadata {
         );
 
         if !self.annotations.is_empty() {
-            let mut annotations_obj = BTreeMap::new();
+            let mut annotations_obj = Object::new();
             for (k, v) in &self.annotations {
                 annotations_obj.insert(Value::String(k.as_str().into()), v.clone());
             }
@@ -198,7 +198,7 @@ impl MetadataValue {
         match *self {
             MetadataValue::String(ref s) => Value::String(s.as_str().into()),
             MetadataValue::StringSet(ref set) => {
-                let mut bset = alloc::collections::BTreeSet::new();
+                let mut bset = BTreeSet::new();
                 for s in set {
                     bset.insert(Value::String(s.as_str().into()));
                 }
@@ -211,7 +211,7 @@ impl MetadataValue {
                 Value::Array(Rc::new(values))
             }
             MetadataValue::Map(ref map) => {
-                let mut obj = BTreeMap::new();
+                let mut obj = Object::new();
                 for (k, v) in map {
                     obj.insert(Value::String(k.as_str().into()), v.to_value());
                 }
@@ -257,7 +257,6 @@ mod metadata_serde {
 mod tests {
     use super::*;
     use crate::value::Value;
-    use alloc::collections::BTreeSet;
 
     /// Round-trip: Value → MetadataValue → Value must be equivalent for
     /// all lossless variants (strings, bools, integers, arrays, objects).

src/rvm/program/serialization/value.rs

@@ -1,6 +1,6 @@
 // Copyright (c) Microsoft Corporation.
 // Licensed under the MIT License.
-use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::collections::BTreeSet;
 use alloc::format;
 use alloc::string::String;
 use alloc::vec::Vec;
@@ -11,6 +11,7 @@ use serde::ser::{SerializeSeq as _, SerializeTuple as _};
 use serde::{Deserialize, Serialize};
 
 use crate::number::Number;
+use crate::value::Object;
 use crate::value::Value;
 
 const VARIANT_NULL: u32 = 0;
@@ -132,7 +133,7 @@ impl<'a> Serialize for BinarySetRef<'a> {
     }
 }
 
-struct BinaryObjectRef<'a>(&'a BTreeMap<Value, Value>);
+struct BinaryObjectRef<'a>(&'a Object);
 
 impl<'a> Serialize for BinaryObjectRef<'a> {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
@@ -140,7 +141,7 @@ impl<'a> Serialize for BinaryObjectRef<'a> {
         S: serde::Serializer,
     {
         let mut seq = serializer.serialize_seq(Some(self.0.len()))?;
-        for (key, value) in self.0.iter() {
+        for (key, value) in self.0.iter_sorted() {
             seq.serialize_element(&BinaryEntryRef(key, value))?;
         }
         seq.end()
@@ -261,11 +262,11 @@ impl<'de> Visitor<'de> for BinaryValueVisitor {
             }
             (BinaryVariant::Object, variant) => {
                 let entries: Vec<(BinaryValue, BinaryValue)> = variant.newtype_variant()?;
-                let mut map = BTreeMap::new();
+                let mut map = Object::new();
                 for (key, value) in entries {
                     map.insert(key.into_value(), value.into_value());
                 }
-                Ok(BinaryValue(Value::from(map)))
+                Ok(BinaryValue(Value::Object(crate::Rc::new(map))))
             }
             (BinaryVariant::Undefined, variant) => {
                 variant.unit_variant()?;

src/rvm/vm/arithmetic.rs

@@ -6,8 +6,6 @@
 // Disable both to keep patterns consistent within this file.
 #![allow(clippy::pattern_type_mismatch, clippy::needless_borrowed_reference)]
 
-use alloc::collections::BTreeSet;
-
 use crate::number::Number;
 use crate::value::Value;
 
@@ -32,8 +30,9 @@ impl RegoVM {
         match (a, b) {
             (&Value::Number(ref x), &Value::Number(ref y)) => Ok(Value::from(x.sub(y)?)),
             (&Value::Set(ref left), &Value::Set(ref right)) => {
-                let diff: BTreeSet<Value> = left.difference(right).cloned().collect();
-                Ok(Value::from_set(diff))
+                let diff: alloc::collections::BTreeSet<Value> =
+                    left.difference(right).cloned().collect();
+                Ok(Value::from(diff))
             }
             _ => Err(VmError::InvalidSubtraction {
                 left: a.clone(),

src/rvm/vm/comprehension.rs

@@ -2,9 +2,9 @@
 // Licensed under the MIT License.
 
 use crate::rvm::instructions::{ComprehensionBeginParams, ComprehensionMode};
+use crate::value::Object;
 use crate::value::Value;
 use crate::Rc;
-use alloc::collections::BTreeMap;
 use alloc::format;
 use alloc::sync::Arc;
 use alloc::vec::Vec;
@@ -34,12 +34,12 @@ impl RegoVM {
         let initial_result = match params.mode {
             ComprehensionMode::Set => Value::new_set(),
             ComprehensionMode::Array => Value::new_array(),
-            ComprehensionMode::Object => Value::Object(Rc::new(BTreeMap::new())),
+            ComprehensionMode::Object => Value::Object(Rc::new(Object::new())),
         };
         self.set_register(params.result_reg, initial_result.clone())?;
 
         let auto_iterate = params.collection_reg != params.result_reg;
-        let iteration_state = if auto_iterate {
+        let mut iteration_state = if auto_iterate {
             let source_value = self.get_register(params.collection_reg)?.clone();
             match source_value {
                 Value::Array(items) => {
@@ -53,11 +53,9 @@ impl RegoVM {
                     if obj.is_empty() {
                         None
                     } else {
-                        Some(IterationState::Object {
-                            obj,
-                            current_key: None,
-                            first_iteration: true,
-                        })
+                        // O(1) cursor over shared Rc<Object>.
+                        let cursor = obj.cursor();
+                        Some(IterationState::Object { obj, cursor })
                     }
                 }
                 Value::Set(set) => {
@@ -79,7 +77,7 @@ impl RegoVM {
             None
         };
 
-        let has_iteration = if let Some(state) = iteration_state.as_ref() {
+        let has_iteration = if let Some(state) = iteration_state.as_mut() {
             self.setup_next_iteration(state, params.key_reg, params.value_reg)?
         } else {
             false
@@ -123,12 +121,12 @@ impl RegoVM {
         let initial_result = match params.mode {
             ComprehensionMode::Set => Value::new_set(),
             ComprehensionMode::Array => Value::new_array(),
-            ComprehensionMode::Object => Value::Object(Rc::new(BTreeMap::new())),
+            ComprehensionMode::Object => Value::Object(Rc::new(Object::new())),
         };
         self.set_register(params.result_reg, initial_result.clone())?;
 
         let auto_iterate = params.collection_reg != params.result_reg;
-        let iteration_state = if auto_iterate {
+        let mut iteration_state = if auto_iterate {
             let source_value = self.get_register(params.collection_reg)?.clone();
             match source_value {
                 Value::Array(items) => {
@@ -142,11 +140,8 @@ impl RegoVM {
                     if obj.is_empty() {
                         None
                     } else {
-                        Some(IterationState::Object {
-                            obj,
-                            current_key: None,
-                            first_iteration: true,
-                        })
+                        let cursor = obj.cursor();
+                        Some(IterationState::Object { obj, cursor })
                     }
                 }
                 Value::Set(set) => {
@@ -168,7 +163,7 @@ impl RegoVM {
             None
         };
 
-        let has_iteration = if let Some(state) = iteration_state.as_ref() {
+        let has_iteration = if let Some(state) = iteration_state.as_mut() {
             self.setup_next_iteration(state, params.key_reg, params.value_reg)?
         } else {
             false
@@ -231,6 +226,7 @@ impl RegoVM {
         }
     }
 
+    #[allow(clippy::if_then_some_else_none)]
     fn execute_comprehension_yield_run_to_completion(
         &mut self,
         value_reg: u8,
@@ -255,6 +251,21 @@ impl RegoVM {
         };
 
         let result_reg = comprehension_context.result_reg;
+        // Snapshot the iteration value register BEFORE taking the result
+        // register: if the comprehension compiler ever allocates
+        // `result_reg == context.value_reg`, the writeback at the bottom
+        // of this function would clobber the value register, and a
+        // post-writeback read here would feed the wrong value into
+        // `IterationState::Set::current_item`. Only Set needs the snapshot
+        // (Object uses a self-advancing cursor; Array advances by index).
+        let set_resume_snapshot = if matches!(
+            comprehension_context.iteration_state,
+            Some(IterationState::Set { .. })
+        ) {
+            Some(self.get_register(comprehension_context.value_reg)?.clone())
+        } else {
+            None
+        };
         // Take ownership of the result register so Rc refcount stays at 1,
         // allowing Rc::make_mut to mutate in-place instead of deep-cloning.
         let mut current_result = self.take_register(result_reg)?;
@@ -292,29 +303,16 @@ impl RegoVM {
         self.set_register(result_reg, current_result)?;
 
         if let Some(iter_state) = comprehension_context.iteration_state.as_mut() {
-            match *iter_state {
-                IterationState::Object {
-                    ref mut current_key,
-                    ..
-                } => {
-                    let tracked_key =
-                        if comprehension_context.key_reg != comprehension_context.value_reg {
-                            self.get_register(comprehension_context.key_reg)?.clone()
-                        } else {
-                            self.get_register(comprehension_context.value_reg)?.clone()
-                        };
-                    *current_key = Some(tracked_key);
-                }
-                IterationState::Set {
-                    ref mut current_item,
-                    ..
-                } => {
-                    *current_item =
-                        Some(self.get_register(comprehension_context.value_reg)?.clone());
-                }
-                IterationState::Array { .. } | IterationState::Single { .. } => {}
+            // Set's `Bound::Excluded(current_item)` resume scheme needs the
+            // pre-mutation snapshot taken at the top of this function.
+            // Object uses a self-advancing cursor and needs no snapshot.
+            if let IterationState::Set {
+                ref mut current_item,
+                ..
+            } = *iter_state
+            {
+                *current_item = set_resume_snapshot;
             }
-
             iter_state.advance();
             let has_next = self.setup_next_iteration(
                 iter_state,
@@ -335,6 +333,7 @@ impl RegoVM {
         Ok(())
     }
 
+    #[allow(clippy::if_then_some_else_none)]
     fn execute_comprehension_yield_suspendable(
         &mut self,
         value_reg: u8,
@@ -359,8 +358,7 @@ impl RegoVM {
             result_reg_idx,
             key_reg_idx,
             value_reg_idx,
-            iteration_key,
-            iteration_value,
+            iter_is_set,
         ) = {
             let frame =
                 self.execution_stack
@@ -382,8 +380,8 @@ impl RegoVM {
 
                 let result_reg_idx = context.result_reg;
                 let mode = context.mode.clone();
-                let iteration_key = self.get_register(context.key_reg)?.clone();
-                let iteration_value = self.get_register(context.value_reg)?.clone();
+                let iter_is_set =
+                    matches!(context.iteration_state, Some(IterationState::Set { .. }));
 
                 (
                     value_to_add,
@@ -392,8 +390,7 @@ impl RegoVM {
                     result_reg_idx,
                     context.key_reg,
                     context.value_reg,
-                    iteration_key,
-                    iteration_value,
+                    iter_is_set,
                 )
             } else {
                 return Err(VmError::InvalidIteration {
@@ -403,6 +400,18 @@ impl RegoVM {
             }
         };
 
+        // Snapshot the iteration value register BEFORE the result writeback:
+        // if the compiler ever allocates `result_reg == value_reg_idx`, a
+        // post-writeback read would feed the result accumulator into
+        // `IterationState::Set::current_item`, breaking the next iteration.
+        // Only Set needs this (Object cursor self-advances; Array advances
+        // by index).
+        let set_resume_snapshot = if iter_is_set {
+            Some(self.get_register(value_reg_idx)?.clone())
+        } else {
+            None
+        };
+
         // Take ownership of the result register so Rc refcount stays at 1,
         // allowing Rc::make_mut to mutate in-place instead of deep-cloning.
         let mut current_result = self.take_register(result_reg_idx)?;
@@ -450,27 +459,13 @@ impl RegoVM {
             } = &mut frame.kind
             {
                 if let Some(iter_state) = context.iteration_state.as_mut() {
-                    match *iter_state {
-                        IterationState::Object {
-                            ref mut current_key,
-                            ..
-                        } => {
-                            let tracked_key = if context.key_reg != context.value_reg {
-                                iteration_key.clone()
-                            } else {
-                                iteration_value.clone()
-                            };
-                            *current_key = Some(tracked_key);
-                        }
-                        IterationState::Set {
-                            ref mut current_item,
-                            ..
-                        } => {
-                            *current_item = Some(iteration_value.clone());
-                        }
-                        IterationState::Array { .. } | IterationState::Single { .. } => {}
+                    if let IterationState::Set {
+                        ref mut current_item,
+                        ..
+                    } = *iter_state
+                    {
+                        *current_item = set_resume_snapshot;
                     }
-
                     iter_state.advance();
                 }
 
@@ -487,8 +482,21 @@ impl RegoVM {
             }
         };
 
-        if let Some(state) = iteration_state_snapshot.as_ref() {
-            let has_next = self.setup_next_iteration(state, key_reg_idx, value_reg_idx)?;
+        if let Some(mut state) = iteration_state_snapshot {
+            let has_next = self.setup_next_iteration(&mut state, key_reg_idx, value_reg_idx)?;
+
+            // `setup_next_iteration` advances Object's internal cursor; the
+            // owning frame holds the iteration_state, so we must write the
+            // updated state back. (The Array/Set variants are also unchanged
+            // by copy, so the writeback is uniform.)
+            if let Some(frame) = self.execution_stack.get_mut(comprehension_index) {
+                if let FrameKind::Comprehension {
+                    ref mut context, ..
+                } = frame.kind
+                {
+                    context.iteration_state = Some(state);
+                }
+            }
 
             if has_next {
                 if let Some(frame) = self.execution_stack.get_mut(comprehension_index) {
@@ -528,7 +536,6 @@ impl RegoVM {
             Ok(false)
         }
     }
-
     pub(super) fn handle_comprehension_condition_failure_suspendable(&mut self) -> Result<bool> {
         if let Some(mut frame) = self.execution_stack.pop() {
             let handled = if let &mut FrameKind::Comprehension {
@@ -554,11 +561,16 @@ impl RegoVM {
         context: &mut ComprehensionContext,
     ) -> Result<()> {
         if let Some(iter_state) = context.iteration_state.as_mut() {
-            self.capture_comprehension_iteration_position(
-                iter_state,
-                context.key_reg,
-                context.value_reg,
-            )?;
+            // Snapshot the current value into Set's `current_item` so the
+            // next iteration can resume from `Bound::Excluded(current)`.
+            // Object uses a self-advancing cursor and needs no snapshot here.
+            if let IterationState::Set {
+                ref mut current_item,
+                ..
+            } = *iter_state
+            {
+                *current_item = Some(self.get_register(context.value_reg)?.clone());
+            }
             iter_state.advance();
             let has_next =
                 self.setup_next_iteration(iter_state, context.key_reg, context.value_reg)?;
@@ -575,37 +587,10 @@ impl RegoVM {
         Ok(())
     }
 
-    fn capture_comprehension_iteration_position(
-        &mut self,
-        iter_state: &mut IterationState,
-        key_reg: u8,
-        value_reg: u8,
-    ) -> Result<()> {
-        match *iter_state {
-            IterationState::Object {
-                ref mut current_key,
-                ..
-            } => {
-                let tracked_key = if key_reg != value_reg {
-                    self.get_register(key_reg)?.clone()
-                } else {
-                    self.get_register(value_reg)?.clone()
-                };
-                *current_key = Some(tracked_key);
-            }
-            IterationState::Set {
-                ref mut current_item,
-                ..
-            } => {
-                *current_item = Some(self.get_register(value_reg)?.clone());
-            }
-            IterationState::Array { .. } | IterationState::Single { .. } => {}
-        }
-
-        Ok(())
-    }
-
     fn execute_comprehension_end_run_to_completion(&mut self) -> Result<()> {
+        // `ComprehensionEnd` is reached from a loaded program; an empty stack
+        // here means malformed user-supplied bytecode, which must still surface
+        // as a typed error rather than a panic — including in debug builds.
         self.comprehension_stack.pop().map_or_else(
             || {
                 Err(VmError::InvalidIteration {

src/rvm/vm/context.rs

@@ -3,8 +3,9 @@
 
 use crate::rvm::instructions::{ComprehensionMode, LoopMode};
 use crate::value::Value;
+use crate::value::{Object, ObjectCursor};
 use crate::Rc;
-use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::collections::BTreeSet;
 use alloc::vec::Vec;
 
 /// Loop execution context for managing iteration state
@@ -24,7 +25,18 @@ pub struct LoopContext {
     pub current_iteration_failed: bool, // Track if current iteration had condition failures
 }
 
-/// Iterator state for different collection types
+/// Iterator state for different collection types.
+///
+/// Snapshot independence for `Object` is provided by the shared
+/// `Rc<Object>` — `Rc::make_mut` on an aliased Rc allocates a new
+/// collection, leaving the iterator's Rc pointing at the original
+/// pre-mutation state. The `ObjectCursor` is opaque and resumes in
+/// O(log n) for the BTree backend.
+///
+/// `Set` continues to use the pre-existing snapshot-by-cloned-key
+/// approach (`current_item` + `first_iteration`); migration of `Set`
+/// to a cursor-based iterator ships with the `Set` storage abstraction
+/// in a follow-up PR.
 #[derive(Debug, Clone)]
 pub enum IterationState {
     Array {
@@ -32,9 +44,8 @@ pub enum IterationState {
         index: usize,
     },
     Object {
-        obj: Rc<BTreeMap<Value, Value>>,
-        current_key: Option<Value>,
-        first_iteration: bool,
+        obj: Rc<Object>,
+        cursor: ObjectCursor,
     },
     Set {
         items: Rc<BTreeSet<Value>>,
@@ -54,13 +65,21 @@ impl IterationState {
     pub(super) const fn advance(&mut self) {
         match *self {
             Self::Array { ref mut index, .. } => {
+                // Array iteration uses `usize` as the cursor and advances via
+                // `saturating_add(1)`. A cursor already at `usize::MAX` here
+                // means a stuck (non-progressing) iteration was emitted by
+                // malformed bytecode; assert in debug to surface it loudly.
+                debug_assert!(
+                    *index < usize::MAX,
+                    "IterationState::Array index already at usize::MAX on advance"
+                );
                 *index = index.saturating_add(1);
             }
-            Self::Object {
-                ref mut first_iteration,
-                ..
-            }
-            | Self::Set {
+            // For Object the cursor advances inside `setup_next_iteration`
+            // when it pulls the next item via `Object::next`, so `advance`
+            // is a no-op for the cursor-backed Object variant.
+            Self::Object { .. } => {}
+            Self::Set {
                 ref mut first_iteration,
                 ..
             } => {
@@ -69,6 +88,12 @@ impl IterationState {
             Self::Single {
                 ref mut consumed, ..
             } => {
+                // `Single` yields exactly once; advancing a consumed Single
+                // means the compiler emitted a redundant LoopNext.
+                debug_assert!(
+                    !*consumed,
+                    "IterationState::Single advanced after consumption"
+                );
                 *consumed = true;
             }
         }
@@ -107,3 +132,71 @@ pub(super) struct ComprehensionContext {
     /// Resume location for the parent frame once this comprehension completes
     pub(super) resume_pc: usize,
 }
+
+#[cfg(test)]
+#[allow(
+    clippy::expect_used,
+    clippy::unwrap_used,
+    clippy::unreachable,
+    clippy::pattern_type_mismatch,
+    clippy::shadow_unrelated,
+    clippy::panic
+)]
+mod tests {
+    use super::*;
+    use crate::value::Object;
+
+    /// IterationState::Object holds an `Rc<Object>` plus an opaque cursor.
+    /// Mutating an aliased Rc via `Rc::make_mut` allocates a new collection
+    /// (CoW) so the in-flight iterator's source is unaffected.
+    #[test]
+    fn iteration_state_object_is_snapshot_independent_of_source() {
+        let mut obj = Object::new();
+        obj.insert(Value::from("a"), Value::from(1));
+        obj.insert(Value::from("b"), Value::from(2));
+        obj.insert(Value::from("c"), Value::from(3));
+
+        let source = Value::Object(Rc::new(obj));
+
+        let snapshot_obj = match &source {
+            Value::Object(o) => Rc::clone(o),
+            _ => unreachable!(),
+        };
+        let state = IterationState::Object {
+            obj: Rc::clone(&snapshot_obj),
+            cursor: snapshot_obj.cursor(),
+        };
+
+        // Mutate a clone of the source mid-iteration.
+        let mut alias = source.clone();
+        let inner = alias.as_object_mut().expect("object");
+        inner.insert(Value::from("a"), Value::from(999));
+        inner.insert(Value::from("d"), Value::from(4));
+        inner.remove(&Value::from("b"));
+
+        // Drain the snapshot via the cursor — must still report the original
+        // 3 entries with original values.
+        let mut collected: Vec<(Value, Value)> = Vec::new();
+        if let IterationState::Object {
+            ref obj,
+            mut cursor,
+        } = state
+        {
+            while let Some((k, v)) = obj.next(&mut cursor) {
+                collected.push((k.clone(), v.clone()));
+            }
+        } else {
+            unreachable!();
+        }
+        assert_eq!(collected.len(), 3);
+        assert!(collected.contains(&(Value::from("a"), Value::from(1))));
+        assert!(collected.contains(&(Value::from("b"), Value::from(2))));
+        assert!(collected.contains(&(Value::from("c"), Value::from(3))));
+        assert!(!collected.iter().any(|kv| kv.0 == Value::from("d")));
+
+        // The original source Value (untouched) is also unchanged.
+        let src_obj = source.as_object().expect("object");
+        assert_eq!(src_obj.len(), 3);
+        assert_eq!(src_obj.get(&Value::from("a")), Some(&Value::from(1)));
+    }
+}

src/rvm/vm/dispatch.rs

@@ -4,7 +4,6 @@
 use crate::rvm::instructions::{GuardMode, Instruction, LiteralOrRegister};
 use crate::rvm::program::Program;
 use crate::value::Value;
-use alloc::collections::BTreeSet;
 use alloc::vec::Vec;
 use core::mem;
 
@@ -670,7 +669,7 @@ impl RegoVM {
                 }
             }
             SetNew { dest } => {
-                let empty_set = Value::Set(crate::Rc::new(BTreeSet::new()));
+                let empty_set = Value::new_set();
                 self.set_register(dest, empty_set)?;
                 Ok(InstructionOutcome::Continue)
             }
@@ -707,7 +706,7 @@ impl RegoVM {
                     if any_undefined {
                         self.set_register(params.dest, Value::Undefined)?;
                     } else {
-                        let mut set = BTreeSet::new();
+                        let mut set = alloc::collections::BTreeSet::new();
                         for &reg in params.element_registers() {
                             set.insert(self.get_register(reg)?.clone());
                         }

src/rvm/vm/errors.rs

@@ -295,6 +295,13 @@ pub enum VmError {
     #[error("Call rule stack underflow during rule finalization (pc={pc})")]
     CallRuleStackUnderflow { pc: usize },
 
+    #[error("Call rule stack mismatch during rule finalization: expected rule_index {expected}, popped {actual} (pc={pc})")]
+    CallRuleStackMismatch {
+        expected: u16,
+        actual: u16,
+        pc: usize,
+    },
+
     #[error("Internal VM error: {message} (pc={pc})")]
     Internal { message: String, pc: usize },
 }

src/rvm/vm/execution.rs

@@ -117,6 +117,10 @@ impl RegoVM {
         let target = self.convert_pc(target, "jump target")?;
         self.pc = target;
         while self.pc < program.instructions.len() {
+            // Per-instruction sanity check: every iteration of the dispatch
+            // loop must re-enter with the VM in a Running/Ready state and the
+            // working data structures coherent.
+            self.assert_vm_invariants();
             self.memory_check()?;
             if self.executed_instructions >= self.max_instructions {
                 return Err(VmError::InstructionLimitExceeded {
@@ -189,6 +193,9 @@ impl RegoVM {
     }
 
     fn execute_suspendable_entry(&mut self, entry_point_pc: usize) -> Result<Value> {
+        // Precondition: callers (execute_entry_point_by_{index,name}) reset the
+        // VM before invoking this method, so the VM must be in a clean state.
+        self.debug_assert_state_is_clean();
         self.execution_state = ExecutionState::Running;
         self.reset_execution_timer_state();
         match self.run_stackless_from(entry_point_pc) {
@@ -201,6 +208,10 @@ impl RegoVM {
     }
 
     pub fn resume(&mut self, resume_value: Option<Value>) -> Result<Value> {
+        // Precondition is enforced below by returning `VmError::InvalidResumeState`
+        // for any non-`Suspended` state. A `debug_assert!` here would diverge
+        // debug vs release behavior and, when invoked via FFI, would trip the
+        // unwind guard and poison the engine on a recoverable misuse.
         let (reason, mut last_result) = match self.execution_state.clone() {
             ExecutionState::Suspended {
                 reason,
@@ -289,6 +300,9 @@ impl RegoVM {
 
     fn run_stackless_loop(&mut self, program: &Program, last_result: &mut Value) -> Result<()> {
         while !self.execution_stack.is_empty() {
+            // Per-instruction sanity check: see `assert_vm_invariants` for the
+            // exact contract. Compiled out in release.
+            self.assert_vm_invariants();
             self.memory_check()?;
             self.frame_pc_overridden = false;
             let should_finalize_rule = self.execution_stack.last().is_some_and(|frame| {

src/rvm/vm/loops.rs

@@ -3,6 +3,7 @@
 
 use crate::rvm::instructions::LoopMode;
 use crate::value::Value;
+use crate::Rc;
 
 use super::context::{IterationState, LoopContext};
 use super::errors::{Result, VmError};
@@ -89,13 +90,13 @@ impl RegoVM {
     ) -> Result<()> {
         self.set_register(params.result_reg, Value::Bool(false))?;
 
-        let iteration_state = match self.resolve_iteration_state(mode, &params)? {
+        let mut iteration_state = match self.resolve_iteration_state(mode, &params)? {
             Some(state) => state,
             None => return Ok(()),
         };
 
         let has_next =
-            self.setup_next_iteration(&iteration_state, params.key_reg, params.value_reg)?;
+            self.setup_next_iteration(&mut iteration_state, params.key_reg, params.value_reg)?;
         if !has_next {
             self.pc = usize::from(params.loop_end);
             return Ok(());
@@ -155,15 +156,10 @@ impl RegoVM {
                 LoopAction::Continue => {}
             }
 
-            if let &mut IterationState::Object {
-                ref mut current_key,
-                ..
-            } = &mut loop_ctx.iteration_state
-            {
-                if loop_ctx.key_reg != loop_ctx.value_reg {
-                    *current_key = Some(self.get_register(loop_ctx.key_reg)?.clone());
-                }
-            } else if let &mut IterationState::Set {
+            // Snapshot the current value for Set so its next iteration can resume
+            // from `Bound::Excluded(current)`. Object uses a cursor and advances
+            // inside `setup_next_iteration` itself.
+            if let &mut IterationState::Set {
                 ref mut current_item,
                 ..
             } = &mut loop_ctx.iteration_state
@@ -173,7 +169,7 @@ impl RegoVM {
 
             loop_ctx.iteration_state.advance();
             let has_next = self.setup_next_iteration(
-                &loop_ctx.iteration_state,
+                &mut loop_ctx.iteration_state,
                 loop_ctx.key_reg,
                 loop_ctx.value_reg,
             )?;
@@ -211,13 +207,13 @@ impl RegoVM {
     ) -> Result<()> {
         self.set_register(params.result_reg, Value::Bool(false))?;
 
-        let iteration_state = match self.resolve_iteration_state(mode, &params)? {
+        let mut iteration_state = match self.resolve_iteration_state(mode, &params)? {
             Some(state) => state,
             None => return Ok(()),
         };
 
         let has_next =
-            self.setup_next_iteration(&iteration_state, params.key_reg, params.value_reg)?;
+            self.setup_next_iteration(&mut iteration_state, params.key_reg, params.value_reg)?;
         if !has_next {
             self.pc = usize::from(params.loop_end);
             return Ok(());
@@ -316,7 +312,14 @@ impl RegoVM {
                 Ok(())
             }
             LoopAction::Continue => {
-                let (mode, success_count, total_iterations, key_reg, value_reg, iteration_state) = {
+                let (
+                    mode,
+                    success_count,
+                    total_iterations,
+                    key_reg,
+                    value_reg,
+                    mut iteration_state,
+                ) = {
                     let (mode, success_count, total_iterations, key_reg, value_reg) = {
                         let frame = self
                             .execution_stack
@@ -334,11 +337,6 @@ impl RegoVM {
                         }
                     };
 
-                    let key_value = if key_reg != value_reg {
-                        Some(self.get_register(key_reg)?.clone())
-                    } else {
-                        None
-                    };
                     let value_value = self.get_register(value_reg)?.clone();
 
                     let frame = self
@@ -349,20 +347,16 @@ impl RegoVM {
                         &mut FrameKind::Loop {
                             ref mut context, ..
                         } => {
-                            if let &mut IterationState::Object {
-                                ref mut current_key,
-                                ..
-                            } = &mut context.iteration_state
-                            {
-                                if context.key_reg != context.value_reg {
-                                    *current_key = key_value;
-                                }
-                            } else if let &mut IterationState::Set {
+                            // Snapshot the current value for Set so its next
+                            // iteration can resume from `Bound::Excluded(current)`.
+                            // Object uses a cursor and advances inside
+                            // `setup_next_iteration` itself.
+                            if let &mut IterationState::Set {
                                 ref mut current_item,
                                 ..
                             } = &mut context.iteration_state
                             {
-                                *current_item = Some(value_value.clone());
+                                *current_item = Some(value_value);
                             }
 
                             context.iteration_state.advance();
@@ -381,7 +375,21 @@ impl RegoVM {
                     }
                 };
 
-                let has_next = self.setup_next_iteration(&iteration_state, key_reg, value_reg)?;
+                let has_next =
+                    self.setup_next_iteration(&mut iteration_state, key_reg, value_reg)?;
+
+                // `setup_next_iteration` advances Object's internal cursor;
+                // the owning frame holds the iteration_state, so we must
+                // write the updated state back. (Array/Set are unchanged by
+                // the call, so the writeback is uniform.)
+                if let Some(frame) = self.execution_stack.last_mut() {
+                    if let FrameKind::Loop {
+                        ref mut context, ..
+                    } = frame.kind
+                    {
+                        context.iteration_state = iteration_state;
+                    }
+                }
 
                 if has_next {
                     if let Some(frame) = self.execution_stack.last_mut() {
@@ -459,10 +467,14 @@ impl RegoVM {
                         self.handle_empty_collection(mode, params.result_reg, params.loop_end)?;
                         return Ok(None);
                     }
+                    // O(1) resumable cursor over the shared Rc<Object>.
+                    // No eager pair snapshot: avoids O(N) setup, O(N) memory
+                    // floor, and O(N) memory-limit checks. Snapshot
+                    // independence is via the shared Rc (CoW).
+                    let cursor = obj.cursor();
                     Ok(Some(IterationState::Object {
-                        obj: obj.clone(),
-                        current_key: None,
-                        first_iteration: true,
+                        obj: Rc::clone(obj),
+                        cursor,
                     }))
                 }
             }
@@ -483,8 +495,15 @@ impl RegoVM {
                     // over a virtual null element.
                     Ok(Some(IterationState::Single { consumed: false }))
                 } else {
-                    // Standard Rego or count/forEach: non-collection → immediate result.
-                    let result = non_collection_result(mode);
+                    // Standard Rego: iterating a non-collection scalar (number,
+                    // string, bool, null, undefined) yields no iterations. For
+                    // `every` this makes the quantifier undefined (it fails) — it
+                    // is NOT vacuously true, which only applies to a genuinely
+                    // empty collection. `any`/`forEach` remain false.
+                    let result = match *mode {
+                        LoopMode::Every => Value::Undefined,
+                        LoopMode::Any | LoopMode::ForEach => Value::Bool(false),
+                    };
                     self.set_register(params.result_reg, result)?;
                     self.pc = usize::from(params.loop_end).saturating_sub(1);
                     Ok(None)
@@ -512,7 +531,7 @@ impl RegoVM {
 
     pub(super) fn setup_next_iteration(
         &mut self,
-        state: &IterationState,
+        state: &mut IterationState,
         key_reg: u8,
         value_reg: u8,
     ) -> Result<bool> {
@@ -538,33 +557,19 @@ impl RegoVM {
             }
             IterationState::Object {
                 ref obj,
-                ref current_key,
-                ref first_iteration,
+                ref mut cursor,
             } => {
-                if *first_iteration {
-                    if let Some((key, value)) = obj.iter().next() {
-                        if key_reg != value_reg {
-                            self.set_register(key_reg, key.clone())?;
-                        }
-                        self.set_register(value_reg, value.clone())?;
-                        Ok(true)
-                    } else {
-                        Ok(false)
-                    }
-                } else if let Some(ref current) = *current_key {
-                    let mut range_iter = obj.range((
-                        core::ops::Bound::Excluded(current),
-                        core::ops::Bound::Unbounded,
-                    ));
-                    if let Some((key, value)) = range_iter.next() {
-                        if key_reg != value_reg {
-                            self.set_register(key_reg, key.clone())?;
-                        }
-                        self.set_register(value_reg, value.clone())?;
-                        Ok(true)
-                    } else {
-                        Ok(false)
+                // Object iterates via a resumable cursor on the shared
+                // `Rc<Object>`; `next` both yields the current entry and
+                // advances the cursor. No explicit `current_key` snapshot is
+                // needed — see the doc on `IterationState`.
+                if let Some((key, value)) = obj.next(cursor) {
+                    let value = value.clone();
+                    if key_reg != value_reg {
+                        self.set_register(key_reg, key.clone())?;
                     }
+                    self.set_register(value_reg, value)?;
+                    Ok(true)
                 } else {
                     Ok(false)
                 }

src/rvm/vm/rules.rs

@@ -277,6 +277,19 @@ impl RegoVM {
             .call_rule_stack
             .pop()
             .ok_or(VmError::CallRuleStackUnderflow { pc: self.pc })?;
+        // Stack discipline: the context we just popped must belong to the
+        // rule we are finalizing. A mismatch indicates a missing push or an
+        // extra pop somewhere in this rule's execution and would otherwise
+        // silently restore the wrong return_pc / rule_type. Surface as a
+        // typed VmError so the contract holds the same in debug and release
+        // builds (avoiding FFI poisoning via a debug-only panic).
+        if rule_index != call_context.rule_index {
+            return Err(VmError::CallRuleStackMismatch {
+                expected: rule_index,
+                actual: call_context.rule_index,
+                pc: self.pc,
+            });
+        }
         self.pc = call_context.return_pc;
 
         let result_from_rule = if !rule_failed_due_to_inconsistency {
@@ -831,6 +844,9 @@ impl RegoVM {
 
         self.registers = parent_registers;
 
+        // Underflow here means malformed/poisoned program state; surface as a
+        // typed error rather than a debug-only panic so the public load_program
+        // contract holds the same in debug and release.
         if self.call_rule_stack.pop().is_none() {
             return Err(VmError::CallRuleStackUnderflow { pc: self.pc });
         }

src/rvm/vm/state.rs

@@ -34,6 +34,139 @@ impl RegoVM {
 
         // Builtin cache entries only live for a single execution
         self.builtins_cache.clear();
+
+        // Postcondition: every stack/cache that `reset_execution_state` touches
+        // must be in its documented "clean" shape. This catches accidental
+        // omissions in future edits to this function.
+        self.debug_assert_state_is_clean();
+    }
+
+    /// Debug-only postcondition for `reset_execution_state`.
+    ///
+    /// Asserts the invariants every caller of `reset_execution_state` relies on
+    /// before starting a fresh execution. The body is fully gated by
+    /// `#[cfg(debug_assertions)]` so this is a zero-cost no-op in release.
+    #[inline]
+    pub(super) fn debug_assert_state_is_clean(&self) {
+        #[cfg(debug_assertions)]
+        {
+            // --- Stacks: every per-execution stack must be drained. ---
+            debug_assert!(
+                self.execution_stack.is_empty(),
+                "reset_execution_state postcondition: execution_stack must be empty"
+            );
+            debug_assert!(
+                self.loop_stack.is_empty(),
+                "reset_execution_state postcondition: loop_stack must be empty"
+            );
+            debug_assert!(
+                self.comprehension_stack.is_empty(),
+                "reset_execution_state postcondition: comprehension_stack must be empty"
+            );
+            debug_assert!(
+                self.call_rule_stack.is_empty(),
+                "reset_execution_state postcondition: call_rule_stack must be empty"
+            );
+            debug_assert!(
+                self.register_stack.is_empty(),
+                "reset_execution_state postcondition: register_stack must be empty"
+            );
+
+            // --- Caches: cleared so a new program/input cannot read stale entries. ---
+            debug_assert!(
+                self.builtins_cache.is_empty(),
+                "reset_execution_state postcondition: builtins_cache must be empty"
+            );
+
+            // --- Registers: window resized to the program's base count and zeroed. ---
+            debug_assert_eq!(
+                self.registers.len(),
+                self.base_register_count,
+                "reset_execution_state postcondition: registers must be sized to base_register_count"
+            );
+            debug_assert!(
+                self.registers.iter().all(|v| matches!(v, Value::Undefined)),
+                "reset_execution_state postcondition: all registers must be Undefined"
+            );
+
+            // --- Rule cache: sized to the current program and marked uncomputed. ---
+            debug_assert_eq!(
+                self.rule_cache.len(),
+                self.program.rule_infos.len(),
+                "reset_execution_state postcondition: rule_cache size must match program rule_infos"
+            );
+            debug_assert!(
+                self.rule_cache.iter().all(|entry| !entry.0),
+                "reset_execution_state postcondition: rule_cache entries must be uncomputed"
+            );
+
+            // --- Counters and execution-state machine: zeroed and back to Ready. ---
+            debug_assert_eq!(
+                self.pc, 0,
+                "reset_execution_state postcondition: pc must be 0"
+            );
+            debug_assert_eq!(
+                self.executed_instructions, 0,
+                "reset_execution_state postcondition: executed_instructions must be 0"
+            );
+            debug_assert!(
+                matches!(self.execution_state, ExecutionState::Ready),
+                "reset_execution_state postcondition: execution_state must be Ready"
+            );
+        }
+    }
+
+    /// Per-opcode VM invariants checked from the inner dispatch loop.
+    ///
+    /// These hold every time control re-enters the dispatch loop with another
+    /// instruction to execute. Only conditions that are *purely VM-internal*
+    /// (i.e. cannot be made false by any host-supplied program or out-of-order
+    /// API call) are asserted here — anything reachable from `load_program`
+    /// input must surface as a typed `VmError` instead, to avoid panicking in
+    /// debug builds and poisoning the engine across FFI.
+    ///
+    /// Fully `#[cfg(debug_assertions)]`-gated so the method body compiles out
+    /// in release.
+    #[inline]
+    pub(super) fn assert_vm_invariants(&self) {
+        #[cfg(debug_assertions)]
+        {
+            // The dispatch loop only runs while execution is live. Once the VM
+            // has transitioned to a terminal state (Suspended/Completed/Error)
+            // the loop must have exited. Note `Ready` is also valid here because
+            // some entry points (e.g. `execute_entry_point_by_index` in
+            // RunToCompletion mode) drive `jump_to` without flipping the state.
+            // `execution_state` is mutated only inside the VM and is not
+            // host-controllable.
+            debug_assert!(
+                matches!(
+                    self.execution_state,
+                    ExecutionState::Ready | ExecutionState::Running
+                ),
+                "vm invariant: execution_state must be Ready or Running inside the dispatch loop, was {:?}",
+                self.execution_state
+            );
+
+            // Rule cache is sized once at reset (against the currently loaded
+            // program) and the VM does not resize it mid-execution. Any
+            // mismatch here would indicate an internal accounting bug rather
+            // than malformed input.
+            debug_assert_eq!(
+                self.rule_cache.len(),
+                self.program.rule_infos.len(),
+                "vm invariant: rule_cache size must equal program.rule_infos size"
+            );
+
+            // NOTE: `!registers.is_empty()` and an `execution_stack` depth
+            // ceiling were intentionally *not* asserted here: both can be
+            // triggered by a host-loaded program (registers via
+            // `RuleInfo::num_registers == 0`; stack depth via deeply nested
+            // rules/loops/comprehensions) and would therefore panic in debug
+            // and poison the engine across FFI. Register access is already
+            // guarded by `VmError::RegisterIndexOutOfBounds`; runaway recursion
+            // is bounded in production by `set_max_instructions` and
+            // `memory_check`.
+        }
     }
 
     /// Return all active objects to their respective pools for reuse

src/scheduler.rs

@@ -569,7 +569,7 @@ impl Analyzer {
                         }
                         Ok(false)
                     }
-                    Array { .. } | Object { .. } => Ok(true),
+                    Expr::Array { .. } | Expr::Object { .. } => Ok(true),
                     _ => Ok(false),
                 })?;
                 Ok(true)
@@ -666,7 +666,7 @@ impl Analyzer {
                 Ok(false)
             }
             // TODO: key vs value for object binding
-            Array { .. } | Object { .. } => Ok(true),
+            Expr::Array { .. } | Expr::Object { .. } => Ok(true),
             _ => Ok(false),
         })?;
         Ok(vars)
@@ -853,7 +853,7 @@ impl Analyzer {
                 Ok(false)
             }
             // TODO: Object key/value
-            Array { .. } | Object { .. } => Ok(true),
+            Expr::Array { .. } | Expr::Object { .. } => Ok(true),
             _ => {
                 non_vars.push(e.clone());
                 Ok(false)

src/schema/tests/suite.rs

@@ -1498,7 +1498,7 @@ fn test_deserialize_object_default_empty_object() {
     let s = Schema::from_serde_json_value(schema).unwrap();
     match s.as_type() {
         Type::Object { default, .. } => {
-            assert_eq!(default, &Some(Value::Object(Rc::new(BTreeMap::new()))));
+            assert_eq!(default, &Some(Value::new_object()));
         }
         _ => panic!("Expected Type::Object"),
     }
@@ -1778,8 +1778,11 @@ fn test_deserialize_enum_values_with_object_non_string_keys() {
     match s.as_type() {
         Type::Enum { values, .. } => match &values[0] {
             Value::Object(obj) => {
-                assert_eq!(obj[&Value::from("1")], Value::from("one"));
-                assert_eq!(obj[&Value::from("true")], Value::from("bool"));
+                assert_eq!(*obj.get(&Value::from("1")).expect("1"), Value::from("one"));
+                assert_eq!(
+                    *obj.get(&Value::from("true")).expect("true"),
+                    Value::from("bool")
+                );
             }
             _ => panic!("Expected object in enum values"),
         },
@@ -1802,18 +1805,21 @@ fn test_deserialize_enum_values_with_deeply_nested_structures() {
     match s.as_type() {
         Type::Enum { values, .. } => match &values[0] {
             Value::Object(obj) => {
-                let a = &obj[&Value::from("a")];
+                let a = obj.get(&Value::from("a")).expect("a");
                 match a {
                     Value::Array(arr) => match &arr[0] {
                         Value::Object(inner) => {
-                            let b = &inner[&Value::from("b")];
+                            let b = inner.get(&Value::from("b")).expect("b");
                             match b {
                                 Value::Array(barr) => {
                                     assert_eq!(barr[0], Value::from(1));
                                     assert_eq!(barr[1], Value::from(2));
                                     match &barr[2] {
                                         Value::Object(cobj) => {
-                                            assert_eq!(cobj[&Value::from("c")], Value::Null);
+                                            assert_eq!(
+                                                *cobj.get(&Value::from("c")).expect("c"),
+                                                Value::Null
+                                            );
                                         }
                                         _ => panic!("Expected object for 'c'"),
                                     }
@@ -1886,8 +1892,11 @@ fn test_deserialize_const_value_object() {
     match s.as_type() {
         Type::Const { value, .. } => match value {
             Value::Object(ref obj) => {
-                assert_eq!(obj[&Value::from("foo")], Value::from("bar"));
-                assert_eq!(obj[&Value::from("baz")], Value::from(1));
+                assert_eq!(
+                    *obj.get(&Value::from("foo")).expect("foo"),
+                    Value::from("bar")
+                );
+                assert_eq!(*obj.get(&Value::from("baz")).expect("baz"), Value::from(1));
             }
             _ => panic!("Expected object for const value"),
         },
@@ -1940,13 +1949,13 @@ fn test_deserialize_const_value_deeply_nested() {
     match s.as_type() {
         Type::Const { value, .. } => match value {
             Value::Object(ref obj) => {
-                let a = &obj[&Value::from("a")];
+                let a = obj.get(&Value::from("a")).expect("a");
                 match a {
                     Value::Array(arr) => {
                         assert_eq!(arr[0], Value::from(1));
                         match &arr[1] {
                             Value::Object(inner) => {
-                                let b = &inner[&Value::from("b")];
+                                let b = inner.get(&Value::from("b")).expect("b");
                                 match b {
                                     Value::Array(barr) => {
                                         assert_eq!(barr[0], Value::Null);

src/schema/validate.rs

@@ -7,6 +7,7 @@
 
 use crate::{
     schema::{error::ValidationError, Schema, Type},
+    value::Object,
     *,
 };
 use alloc::collections::BTreeMap;
@@ -537,7 +538,7 @@ impl SchemaValidator {
     }
 
     fn validate_discriminated_subobject_with_base(
-        object_value: &BTreeMap<Value, Value>,
+        object_value: &Object,
         discriminated_subobject: &crate::schema::DiscriminatedSubobject,
         base_properties: &BTreeMap<String, Schema>,
         base_additional_properties: Option<&Schema>,
@@ -653,7 +654,7 @@ impl SchemaValidator {
     }
 
     fn validate_subobject(
-        object_value: &BTreeMap<Value, Value>,
+        object_value: &Object,
         subobject: &crate::schema::Subobject,
         path: &str,
     ) -> Result<(), ValidationError> {

src/tests/interpreter/mod.rs

@@ -24,6 +24,7 @@ use anyhow::{bail, Result};
 use core::num::NonZeroU32;
 use core::time::Duration;
 use serde::{Deserialize, Serialize};
+use std::collections::{BTreeMap, BTreeSet};
 use test_generator::test_resources;
 use timer_test_support::{
     apply_engine_timer, configure_time_source, reset_time_source, GlobalTimerGuard,
@@ -818,3 +819,446 @@ fn test_get_data() -> Result<()> {
 
     Ok(())
 }
+
+#[test]
+fn test_add_data_deep_merge() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Nested objects under a shared top-level key are deep-merged, not replaced.
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "x" : 1 } }"#)?)?;
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "y" : 2 } }"#)?)?;
+
+    assert_eq!(
+        engine.get_data(),
+        Value::from_json_str(r#"{ "a" : { "x" : 1, "y" : 2 } }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_deep_merge_multi_level() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Merging recurses through multiple levels of nesting.
+    engine.add_data(Value::from_json_str(
+        r#"{ "a" : { "b" : { "x" : 1 } }, "top" : 0 }"#,
+    )?)?;
+    engine.add_data(Value::from_json_str(
+        r#"{ "a" : { "b" : { "y" : 2 }, "c" : 3 } }"#,
+    )?)?;
+
+    assert_eq!(
+        engine.get_data(),
+        Value::from_json_str(r#"{ "a" : { "b" : { "x" : 1, "y" : 2 }, "c" : 3 }, "top" : 0 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_leaf_conflict_errors() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // A genuine leaf conflict (same nested path, different value) is an error.
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "x" : 1 } }"#)?)?;
+    assert!(engine
+        .add_data(Value::from_json_str(r#"{ "a" : { "x" : 2 } }"#)?)
+        .is_err());
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_object_vs_scalar_conflict_errors() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // An object cannot be merged with a scalar at the same path.
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "x" : 1 } }"#)?)?;
+    assert!(engine
+        .add_data(Value::from_json_str(r#"{ "a" : 5 }"#)?)
+        .is_err());
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_equal_leaf_is_noop() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Re-adding identical data (including equal nested leaves) is tolerated as a no-op.
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "x" : 1 } }"#)?)?;
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "x" : 1 }, "b" : 2 }"#)?)?;
+
+    assert_eq!(
+        engine.get_data(),
+        Value::from_json_str(r#"{ "a" : { "x" : 1 }, "b" : 2 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_set_union() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Sets under a shared key are unioned rather than conflicting (consistent with the
+    // rule-evaluation merge, where partial set rules accumulate elements). JSON cannot express
+    // sets, so the data documents are built via the `Value` API.
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+    )])))?;
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([Value::from(2_u64), Value::from(3_u64)])),
+    )])))?;
+
+    let expected = Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([
+            Value::from(1_u64),
+            Value::from(2_u64),
+            Value::from(3_u64),
+        ])),
+    )]));
+    assert_eq!(engine.get_data(), expected);
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_nested_set_union() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // A set nested under an object key exercises the recursive merge: the outer objects are
+    // deep-merged and the inner sets are then unioned.
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("a"),
+        Value::from(BTreeMap::from([(
+            Value::from("s"),
+            Value::from(BTreeSet::from([Value::from(1_u64)])),
+        )])),
+    )])))?;
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("a"),
+        Value::from(BTreeMap::from([(
+            Value::from("s"),
+            Value::from(BTreeSet::from([Value::from(2_u64)])),
+        )])),
+    )])))?;
+
+    let expected = Value::from(BTreeMap::from([(
+        Value::from("a"),
+        Value::from(BTreeMap::from([(
+            Value::from("s"),
+            Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+        )])),
+    )]));
+    assert_eq!(engine.get_data(), expected);
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_equal_set_is_noop() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Re-adding an identical set is tolerated as a no-op (not a conflict).
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+    )])))?;
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+    )])))?;
+
+    let expected = Value::from(BTreeMap::from([(
+        Value::from("s"),
+        Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+    )]));
+    assert_eq!(engine.get_data(), expected);
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_failed_merge_is_atomic() -> Result<()> {
+    let mut engine = Engine::new();
+
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "z" : 1 } }"#)?)?;
+
+    // Mixes a new key `m` with a conflicting leaf `z` (1 vs 3). Because `m` sorts
+    // before `z`, a naive in-place merge would insert `m` and only then hit the `z`
+    // conflict. add_data must be all-or-nothing: the whole call fails AND leaves the
+    // existing data untouched — `m` must not leak in.
+    assert!(engine
+        .add_data(Value::from_json_str(r#"{ "a" : { "m" : 2, "z" : 3 } }"#)?)
+        .is_err());
+
+    assert_eq!(
+        engine.get_data(),
+        Value::from_json_str(r#"{ "a" : { "z" : 1 } }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_failed_set_merge_is_atomic() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Existing data: a set `s` alongside a scalar `z` under `a`.
+    engine.add_data(Value::from(BTreeMap::from([(
+        Value::from("a"),
+        Value::from(BTreeMap::from([
+            (
+                Value::from("s"),
+                Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+            ),
+            (Value::from("z"), Value::from(1_u64)),
+        ])),
+    )])))?;
+
+    // This add would union `s` with {3} but conflicts on `z` (1 vs 2). Since `s`
+    // sorts before `z`, a naive in-place merge would union the set *before* failing
+    // on `z`, leaking {3} into `s`. The atomic add must reject the whole call and
+    // leave `s` as {1, 2}.
+    assert!(engine
+        .add_data(Value::from(BTreeMap::from([(
+            Value::from("a"),
+            Value::from(BTreeMap::from([
+                (
+                    Value::from("s"),
+                    Value::from(BTreeSet::from([Value::from(3_u64)])),
+                ),
+                (Value::from("z"), Value::from(2_u64)),
+            ])),
+        )])))
+        .is_err());
+
+    // `s` must be unchanged ({1, 2}, not {1, 2, 3}) and `z` must still be 1.
+    let expected = Value::from(BTreeMap::from([(
+        Value::from("a"),
+        Value::from(BTreeMap::from([
+            (
+                Value::from("s"),
+                Value::from(BTreeSet::from([Value::from(1_u64), Value::from(2_u64)])),
+            ),
+            (Value::from("z"), Value::from(1_u64)),
+        ])),
+    )]));
+    assert_eq!(engine.get_data(), expected);
+
+    Ok(())
+}
+
+#[test]
+fn test_add_data_failed_array_merge_is_atomic() -> Result<()> {
+    let mut engine = Engine::new();
+
+    engine.add_data(Value::from_json_str(r#"{ "a" : { "arr" : [1, 2] } }"#)?)?;
+
+    // Arrays are atomic leaves (never element-merged), so a differing array at the
+    // same path is a conflict. The new key `aa` sorts before `arr`, so a naive
+    // in-place merge would insert `aa` and only then hit the `arr` conflict. add_data
+    // must reject the whole call and leave the data untouched — `aa` must not leak in.
+    assert!(engine
+        .add_data(Value::from_json_str(
+            r#"{ "a" : { "aa" : 5, "arr" : [3] } }"#
+        )?)
+        .is_err());
+
+    assert_eq!(
+        engine.get_data(),
+        Value::from_json_str(r#"{ "a" : { "arr" : [1, 2] } }"#)?
+    );
+
+    Ok(())
+}
+
+// The `Value::merge` used by `add_data` is shared with the rule-evaluation path
+// (`Interpreter::merge_rule_value`, reached via `with data.* as ...` and rule-value
+// materialization). The tests below pin down that making `merge` recursive changed only the
+// data-document semantics and left rule evaluation — in particular the `with data.* as ...`
+// modifier — behaving exactly as before (an override, never a deep merge).
+
+#[test]
+fn test_with_data_modifier_replaces_nested_object() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Base data provides a nested object with two keys.
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "a" : 1, "b" : 2 } } }"#,
+    )?)?;
+
+    engine.add_policy(
+        "policy.rego".to_string(),
+        r#"
+package test
+
+result := x if {
+    x := data.base.foo with data.base.foo as {"a": 99}
+}
+"#
+        .to_string(),
+    )?;
+
+    // `with data.base.foo as {"a": 99}` REPLACES the whole subtree for the duration of the
+    // rule; it must NOT deep-merge with the base `{ "a": 1, "b": 2 }`. So `b` is gone.
+    assert_eq!(
+        engine
+            .eval_query("data.test.result".to_string(), false)?
+            .result[0]
+            .expressions[0]
+            .value
+            .clone(),
+        Value::from_json_str(r#"{ "a" : 99 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_with_data_modifier_replaces_whole_subtree() -> Result<()> {
+    let mut engine = Engine::new();
+
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : 1, "bar" : 2 } }"#,
+    )?)?;
+
+    engine.add_policy(
+        "policy.rego".to_string(),
+        r#"
+package test
+
+result := x if {
+    x := data.base with data.base as {"only": 3}
+}
+"#
+        .to_string(),
+    )?;
+
+    // `with data.base as {...}` replaces the entire `data.base` object; the original
+    // `foo`/`bar` keys are not merged in.
+    assert_eq!(
+        engine
+            .eval_query("data.test.result".to_string(), false)?
+            .result[0]
+            .expressions[0]
+            .value
+            .clone(),
+        Value::from_json_str(r#"{ "only" : 3 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_with_data_modifier_nested_replace_preserves_siblings() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // `data.base` has a nested `foo` object AND a sibling `bar`.
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "a" : 1, "b" : 2 }, "bar" : 7 } }"#,
+    )?)?;
+
+    engine.add_policy(
+        "policy.rego".to_string(),
+        r#"
+package test
+
+# `with` targets the nested `data.base.foo`, but the rule observes the PARENT `data.base`.
+result := x if {
+    x := data.base with data.base.foo as {"a": 99}
+}
+"#
+        .to_string(),
+    )?;
+
+    // The nested `foo` is deep-replaced (its `b` is gone — `with` never merges), while the
+    // sibling `bar` under the same parent is preserved.
+    assert_eq!(
+        engine
+            .eval_query("data.test.result".to_string(), false)?
+            .result[0]
+            .expressions[0]
+            .value
+            .clone(),
+        Value::from_json_str(r#"{ "foo" : { "a" : 99 }, "bar" : 7 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_rule_reads_deep_merged_base_data() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Two add_data calls deep-merge into a single nested object...
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "a" : 1 } } }"#,
+    )?)?;
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "b" : 2 } } }"#,
+    )?)?;
+
+    engine.add_policy(
+        "policy.rego".to_string(),
+        r#"
+package test
+
+a := data.base.foo.a
+b := data.base.foo.b
+"#
+        .to_string(),
+    )?;
+
+    // ...and both merged leaves are visible to rule evaluation.
+    assert_eq!(
+        engine.eval_query("data.test".to_string(), false)?.result[0].expressions[0]
+            .value
+            .clone(),
+        Value::from_json_str(r#"{ "a" : 1, "b" : 2 }"#)?
+    );
+
+    Ok(())
+}
+
+#[test]
+fn test_rule_values_coexist_with_merged_base_data() -> Result<()> {
+    let mut engine = Engine::new();
+
+    // Deep-merged base data under `base`...
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "a" : 1 } } }"#,
+    )?)?;
+    engine.add_data(Value::from_json_str(
+        r#"{ "base" : { "foo" : { "b" : 2 } } }"#,
+    )?)?;
+
+    engine.add_policy(
+        "policy.rego".to_string(),
+        r#"
+package test
+
+computed := data.base.foo.a + data.base.foo.b
+"#
+        .to_string(),
+    )?;
+
+    let data = engine.eval_query("data".to_string(), false)?.result[0].expressions[0]
+        .value
+        .clone();
+
+    // Base data is preserved and deep-merged...
+    assert_eq!(
+        data["base"],
+        Value::from_json_str(r#"{ "foo" : { "a" : 1, "b" : 2 } }"#)?
+    );
+    // ...and the rule-computed value materializes alongside it without disturbing the merge.
+    assert_eq!(data["test"]["computed"], Value::from(3_u64));
+
+    Ok(())
+}

src/value/mod.rs

@@ -11,6 +11,24 @@
     clippy::as_conversions
 )] // value helpers index paths directly for performance
 
+mod object;
+mod set;
+
+#[cfg(test)]
+mod tests;
+
+#[allow(unused_imports)] // surface for downstream PRs
+pub use object::{IntoIter, Iter, IterMut, Object};
+#[allow(unused_imports)] // surface for downstream PRs
+pub use set::Set;
+
+#[cfg(feature = "rvm")]
+#[allow(unused_imports)] // surface for downstream PRs
+pub use object::ObjectCursor;
+#[cfg(feature = "rvm")]
+#[allow(unused_imports)] // surface for downstream PRs
+pub use set::SetCursor;
+
 use crate::number::Number;
 
 use alloc::collections::{BTreeMap, BTreeSet};
@@ -23,7 +41,7 @@ use core::str::FromStr;
 
 use anyhow::{anyhow, bail, Result};
 use serde::de::{self, Deserializer, Error as DeError, MapAccess, SeqAccess, Visitor};
-use serde::ser::{SerializeMap, Serializer};
+use serde::ser::Serializer;
 use serde::{Deserialize, Serialize};
 
 use crate::*;
@@ -63,7 +81,7 @@ pub enum Value {
 
     /// An object.
     /// Unlike JSON, keys can be any value, not just string.
-    Object(Rc<BTreeMap<Value, Value>>),
+    Object(Rc<Object>),
 
     /// Undefined value.
     /// Used to indicate the absence of a value.
@@ -86,26 +104,15 @@ impl Serialize for Value {
     where
         S: Serializer,
     {
-        use serde::ser::Error;
         match self {
             Value::Null => serializer.serialize_unit(),
             Value::Bool(b) => serializer.serialize_bool(*b),
             Value::String(s) => serializer.serialize_str(s.as_ref()),
             Value::Number(n) => n.serialize(serializer),
             Value::Array(a) => a.serialize(serializer),
-            Value::Object(fields) => {
-                let mut map = serializer.serialize_map(Some(fields.len()))?;
-                for (k, v) in fields.iter() {
-                    match k {
-                        Value::String(_) => map.serialize_entry(k, v)?,
-                        _ => {
-                            let key_str = serde_json::to_string(k).map_err(Error::custom)?;
-                            map.serialize_entry(&key_str, v)?
-                        }
-                    }
-                }
-                map.end()
-            }
+            // Delegate to the Object/Set serializers — single canonical path,
+            // handles non-string-key stringification internally.
+            Value::Object(fields) => fields.serialize(serializer),
 
             // display set as an array
             Value::Set(s) => s.serialize(serializer),
@@ -345,7 +352,7 @@ impl Value {
     /// assert_eq!(array[4], Value::from(12345u64));
     /// let obj = array[5].as_object().expect("not an object");
     /// assert_eq!(obj.len(), 1);
-    /// assert_eq!(obj[&Value::from("name")], Value::from("regorus"));
+    /// assert_eq!(obj.get(&Value::from("name")).expect("missing name"), &Value::from("regorus"));
     /// # Ok(())
     /// # }
     /// ```
@@ -800,7 +807,7 @@ impl From<BTreeMap<Value, Value>> for Value {
     /// # Ok(())
     /// # }
     fn from(s: BTreeMap<Value, Value>) -> Self {
-        Value::Object(Rc::new(s))
+        Value::Object(Rc::new(Object::from(s)))
     }
 }
 
@@ -1279,16 +1286,16 @@ impl Value {
         }
     }
 
-    /// Cast value to [`& BTreeMap<Value, Value>`] if [`Value::Object`].
+    /// Cast value to [`&Object`] if [`Value::Object`].
     /// ```
     /// # use regorus::*;
-    /// # use std::collections::BTreeMap;
+    /// # use regorus::value::Object;
     /// # fn main() -> anyhow::Result<()> {
     /// let v = Value::from(
     ///    [(Value::from("Hello"), Value::from("World"))]
     ///        .iter()
     ///        .cloned()
-    ///        .collect::<BTreeMap<Value, Value>>(),
+    ///        .collect::<Object>(),
     /// );
     /// assert_eq!(
     ///    v.as_object()?.iter().next(),
@@ -1296,28 +1303,28 @@ impl Value {
     /// );
     /// # Ok(())
     /// # }
-    pub fn as_object(&self) -> Result<&BTreeMap<Value, Value>> {
+    pub fn as_object(&self) -> Result<&Object> {
         match self {
             Value::Object(m) => Ok(m),
             _ => Err(anyhow!("not an object")),
         }
     }
 
-    /// Cast value to [`&mut BTreeMap<Value, Value>`] if [`Value::Object`].
+    /// Cast value to [`&mut Object`] if [`Value::Object`].
     /// ```
     /// # use regorus::*;
-    /// # use std::collections::BTreeMap;
+    /// # use regorus::value::Object;
     /// # fn main() -> anyhow::Result<()> {
     /// let mut v = Value::from(
     ///    [(Value::from("Hello"), Value::from("World"))]
     ///        .iter()
     ///        .cloned()
-    ///        .collect::<BTreeMap<Value, Value>>(),
+    ///        .collect::<Object>(),
     /// );
     /// v.as_object_mut()?.insert(Value::from("Good"), Value::from("Bye"));
     /// # Ok(())
     /// # }
-    pub fn as_object_mut(&mut self) -> Result<&mut BTreeMap<Value, Value>> {
+    pub fn as_object_mut(&mut self) -> Result<&mut Object> {
         match self {
             Value::Object(m) => Ok(Rc::make_mut(m)),
             _ => Err(anyhow!("not an object")),
@@ -1325,6 +1332,13 @@ impl Value {
     }
 }
 
+/// Depth cap for `deep_merge`/`check_mergeable`, set at serde_json's default recursion limit.
+///
+/// Prevents a stack overflow from adversarially nested data — an uncatchable abort that poisons
+/// every engine in an FFI process. At serde_json's limit it only backstops `Value`s built without
+/// a parse-time cap: the Python/Ruby native bindings, or programmatic construction.
+const MAX_MERGE_DEPTH: usize = 128;
+
 impl Value {
     pub(crate) fn make_or_get_value_mut<'a>(&'a mut self, paths: &[&str]) -> Result<&'a mut Value> {
         if paths.is_empty() {
@@ -1358,6 +1372,11 @@ impl Value {
         }
     }
 
+    /// Shallow-merge `new` into `self` with strict rule-output semantics.
+    ///
+    /// Objects merge one level deep: a key on both sides must hold the *same* value or it is a
+    /// conflict; sets union; equal values are a no-op. Non-recursive by design — data documents
+    /// use [`Value::deep_merge`] instead.
     pub(crate) fn merge(&mut self, mut new: Value) -> Result<()> {
         if self == &new {
             return Ok(());
@@ -1365,24 +1384,26 @@ impl Value {
         match (self, &mut new) {
             (v @ Value::Undefined, _) => *v = new,
             (Value::Set(ref mut set), Value::Set(new)) => {
-                Rc::make_mut(set).append(Rc::make_mut(new));
-                // Enforce allocator limit after merging set entries.
+                // Union without deep-cloning the RHS set (see `deep_merge`).
+                let dst = Rc::make_mut(set);
+                match Rc::try_unwrap(core::mem::take(new)) {
+                    Ok(owned) => dst.extend(owned),
+                    Err(shared) => dst.extend(shared.iter().cloned()),
+                }
                 enforce_limit_anyhow()?;
             }
             (Value::Object(map), Value::Object(new)) => {
                 for (k, v) in new.iter() {
                     match map.get(k) {
-                        Some(pv) if *pv != *v => {
-                            bail!(
-                                "value for key `{}` generated multiple times: `{}` and `{}`",
-                                serde_json::to_string_pretty(&k).map_err(anyhow::Error::msg)?,
-                                serde_json::to_string_pretty(&pv).map_err(anyhow::Error::msg)?,
-                                serde_json::to_string_pretty(&v).map_err(anyhow::Error::msg)?,
-                            )
-                        }
+                        // Same key, different value: the rule produced two outputs for one path.
+                        Some(pv) if *pv != *v => bail!(
+                            "value for key `{}` generated multiple times: `{}` and `{}`",
+                            serde_json::to_string_pretty(&k).map_err(anyhow::Error::msg)?,
+                            serde_json::to_string_pretty(&pv).map_err(anyhow::Error::msg)?,
+                            serde_json::to_string_pretty(&v).map_err(anyhow::Error::msg)?,
+                        ),
                         _ => {
                             Rc::make_mut(map).insert(k.clone(), v.clone());
-                            // Enforce allocator limit after merging object entries.
                             enforce_limit_anyhow()?;
                         }
                     };
@@ -1392,6 +1413,151 @@ impl Value {
         };
         Ok(())
     }
+
+    /// Recursively deep-merge `new` into `self` — the data-document merge behind [`Engine::add_data`].
+    ///
+    /// Objects recurse per-key, sets union, equal values are a no-op, any other differing pair
+    /// conflicts. Set-union is a regorus extension (OPA data is JSON, which has no sets). Distinct
+    /// from the strict, non-recursive [`Value::merge`] used for rule outputs — use deep-merge ONLY
+    /// for data documents.
+    ///
+    /// [`Engine::add_data`]: crate::Engine::add_data
+    pub(crate) fn deep_merge(&mut self, new: Value) -> Result<()> {
+        self.deep_merge_at(new, 0)
+    }
+
+    /// Depth-tracked worker for [`deep_merge`](Value::deep_merge). See [`MAX_MERGE_DEPTH`].
+    fn deep_merge_at(&mut self, mut new: Value, depth: usize) -> Result<()> {
+        if depth >= MAX_MERGE_DEPTH {
+            bail!("data merge exceeds maximum nesting depth of {MAX_MERGE_DEPTH}");
+        }
+        if self == &new {
+            return Ok(());
+        }
+        match (self, &mut new) {
+            (v @ Value::Undefined, _) => *v = new,
+            (Value::Set(ref mut set), Value::Set(new)) => {
+                // Union without deep-cloning the RHS set: move elements if uniquely owned,
+                // else clone only the element handles (`Rc` bumps), never the whole `BTreeSet`.
+                let dst = Rc::make_mut(set);
+                match Rc::try_unwrap(core::mem::take(new)) {
+                    Ok(owned) => dst.extend(owned),
+                    Err(shared) => dst.extend(shared.iter().cloned()),
+                }
+                enforce_limit_anyhow()?;
+            }
+            (Value::Object(map), Value::Object(new)) => {
+                // What each incoming key requires of the target map. Decided from a read-only
+                // probe so a no-op or a conflict never triggers `Rc::make_mut` (and never clones
+                // a shared map); `make_mut` is taken lazily, only when a key actually mutates.
+                enum Step {
+                    Skip,
+                    Insert,
+                    Recurse,
+                    Conflict,
+                }
+                for (k, v) in new.iter() {
+                    let step = match map.get(k) {
+                        None => Step::Insert,
+                        Some(existing) if existing == v => Step::Skip,
+                        Some(existing)
+                            if matches!(
+                                (existing, v),
+                                (Value::Object(_), Value::Object(_))
+                                    | (Value::Set(_), Value::Set(_))
+                            ) =>
+                        {
+                            Step::Recurse
+                        }
+                        Some(_) => Step::Conflict,
+                    };
+                    match step {
+                        Step::Skip => {}
+                        Step::Insert => {
+                            Rc::make_mut(map).insert(k.clone(), v.clone());
+                            enforce_limit_anyhow()?;
+                        }
+                        // Both sides are containers: recurse so nested objects merge rather than
+                        // the subtree being replaced (OPA data-merge semantics).
+                        Step::Recurse => {
+                            let existing = Rc::make_mut(map).get_mut(k).ok_or_else(|| {
+                                anyhow!("internal error: key vanished during merge")
+                            })?;
+                            existing.deep_merge_at(v.clone(), depth.saturating_add(1))?;
+                        }
+                        Step::Conflict => {
+                            let existing = map.get(k).ok_or_else(|| {
+                                anyhow!("internal error: key vanished during merge")
+                            })?;
+                            bail!(
+                                "value for key `{}` generated multiple times: `{}` and `{}`",
+                                serde_json::to_string_pretty(&k).map_err(anyhow::Error::msg)?,
+                                serde_json::to_string_pretty(&existing)
+                                    .map_err(anyhow::Error::msg)?,
+                                serde_json::to_string_pretty(&v).map_err(anyhow::Error::msg)?,
+                            )
+                        }
+                    }
+                }
+            }
+            _ => bail!("error: could not merge value"),
+        };
+        Ok(())
+    }
+
+    /// Read-only check that [`deep_merge`](Value::deep_merge)-ing `other` into `self` would not
+    /// conflict, without mutating or allocating.
+    ///
+    /// Lets [`Engine::add_data`] validate before merging in place. Since a conflict is the only
+    /// way the default-build merge can fail and it depends only on the inputs, a passing scan
+    /// guarantees the in-place `deep_merge` won't fail — avoiding the alternative of cloning the
+    /// whole document into a candidate just to validate. Only overlapping keys are walked, so
+    /// disjoint additions are near-free.
+    ///
+    /// [`Engine::add_data`]: crate::Engine::add_data
+    #[cfg(not(feature = "allocator-memory-limits"))]
+    pub(crate) fn check_mergeable(&self, other: &Value) -> Result<()> {
+        self.check_mergeable_at(other, 0)
+    }
+
+    /// Depth-tracked worker for [`check_mergeable`](Value::check_mergeable). See [`MAX_MERGE_DEPTH`].
+    #[cfg(not(feature = "allocator-memory-limits"))]
+    fn check_mergeable_at(&self, other: &Value, depth: usize) -> Result<()> {
+        if depth >= MAX_MERGE_DEPTH {
+            bail!("data merge exceeds maximum nesting depth of {MAX_MERGE_DEPTH}");
+        }
+        if self == other {
+            return Ok(());
+        }
+        match (self, other) {
+            (Value::Undefined, _) => Ok(()),
+            // Set union never conflicts.
+            (Value::Set(_), Value::Set(_)) => Ok(()),
+            (Value::Object(dst), Value::Object(src)) => {
+                for (k, sv) in src.iter() {
+                    // Only overlapping keys can conflict.
+                    if let Some(dv) = dst.get(k) {
+                        let both_mergeable = matches!(
+                            (dv, sv),
+                            (Value::Object(_), Value::Object(_)) | (Value::Set(_), Value::Set(_))
+                        );
+                        if both_mergeable {
+                            dv.check_mergeable_at(sv, depth.saturating_add(1))?;
+                        } else if dv != sv {
+                            bail!(
+                                "value for key `{}` generated multiple times: `{}` and `{}`",
+                                serde_json::to_string_pretty(&k).map_err(anyhow::Error::msg)?,
+                                serde_json::to_string_pretty(&dv).map_err(anyhow::Error::msg)?,
+                                serde_json::to_string_pretty(&sv).map_err(anyhow::Error::msg)?,
+                            )
+                        }
+                    }
+                }
+                Ok(())
+            }
+            _ => bail!("error: could not merge value"),
+        }
+    }
 }
 
 impl ops::Index<&Value> for Value {

src/value/object/iter.rs

@@ -0,0 +1,148 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! Opaque iterator types for [`Object`].
+//!
+//! These newtypes wrap the storage backend's iterators so the backend can be
+//! swapped without changing any iterator type signatures observed by callers.
+
+use alloc::collections::btree_map;
+use core::iter::FusedIterator;
+
+use super::Object;
+use crate::value::Value;
+
+/// Owned iterator over `(Value, Value)` entries.
+#[derive(Debug)]
+pub struct IntoIter {
+    pub(super) inner: btree_map::IntoIter<Value, Value>,
+}
+
+impl Iterator for IntoIter {
+    type Item = (Value, Value);
+    #[inline]
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+    #[inline]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+impl DoubleEndedIterator for IntoIter {
+    #[inline]
+    fn next_back(&mut self) -> Option<Self::Item> {
+        self.inner.next_back()
+    }
+}
+
+impl ExactSizeIterator for IntoIter {
+    #[inline]
+    fn len(&self) -> usize {
+        self.inner.len()
+    }
+}
+
+impl FusedIterator for IntoIter {}
+
+/// Borrowed iterator over `(&Value, &Value)` entries.
+#[derive(Debug, Clone)]
+pub struct Iter<'a> {
+    pub(super) inner: btree_map::Iter<'a, Value, Value>,
+}
+
+impl<'a> Iterator for Iter<'a> {
+    type Item = (&'a Value, &'a Value);
+    #[inline]
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+    #[inline]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+impl<'a> DoubleEndedIterator for Iter<'a> {
+    #[inline]
+    fn next_back(&mut self) -> Option<Self::Item> {
+        self.inner.next_back()
+    }
+}
+
+impl<'a> ExactSizeIterator for Iter<'a> {
+    #[inline]
+    fn len(&self) -> usize {
+        self.inner.len()
+    }
+}
+
+impl<'a> FusedIterator for Iter<'a> {}
+
+/// Borrowed iterator over `(&Value, &mut Value)` entries.
+#[derive(Debug)]
+pub struct IterMut<'a> {
+    pub(super) inner: btree_map::IterMut<'a, Value, Value>,
+}
+
+impl<'a> Iterator for IterMut<'a> {
+    type Item = (&'a Value, &'a mut Value);
+    #[inline]
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+    #[inline]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+impl<'a> DoubleEndedIterator for IterMut<'a> {
+    #[inline]
+    fn next_back(&mut self) -> Option<Self::Item> {
+        self.inner.next_back()
+    }
+}
+
+impl<'a> ExactSizeIterator for IterMut<'a> {
+    #[inline]
+    fn len(&self) -> usize {
+        self.inner.len()
+    }
+}
+
+impl<'a> FusedIterator for IterMut<'a> {}
+
+impl IntoIterator for Object {
+    type Item = (Value, Value);
+    type IntoIter = IntoIter;
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        IntoIter {
+            inner: self.inner.into_iter(),
+        }
+    }
+}
+
+impl<'a> IntoIterator for &'a Object {
+    type Item = (&'a Value, &'a Value);
+    type IntoIter = Iter<'a>;
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        Iter {
+            inner: self.inner.iter(),
+        }
+    }
+}
+
+impl<'a> IntoIterator for &'a mut Object {
+    type Item = (&'a Value, &'a mut Value);
+    type IntoIter = IterMut<'a>;
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        IterMut {
+            inner: self.inner.iter_mut(),
+        }
+    }
+}

src/value/object/mod.rs

@@ -0,0 +1,265 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! See [`Object`].
+
+mod iter;
+mod serde;
+
+use alloc::collections::BTreeMap;
+use core::cmp::Ordering;
+use core::fmt;
+use core::ops::Bound;
+
+use crate::value::Value;
+
+pub use iter::{IntoIter, Iter, IterMut};
+
+/// Opaque, ordered key-value map keyed by [`Value`].
+///
+/// The current backing storage is `BTreeMap<Value, Value>`. The inner field
+/// is private so the representation can change (two-tier inline+hash, lazy,
+/// schema-shared) without touching call sites.
+///
+/// # Iteration
+///
+/// - [`Object::iter`] — implementation-defined order; non-resumable.
+/// - [`Object::iter_sorted`] — sorted by `Value::Ord`; non-resumable.
+/// - [`Object::cursor`] / [`Object::next`] — implementation-defined order,
+///   resumable; cheapest per-step cost. Used by interpreter/RVM when iteration
+///   must yield mid-flight.
+#[derive(Default, Clone, Eq, PartialEq)]
+pub struct Object {
+    inner: BTreeMap<Value, Value>,
+}
+
+impl Object {
+    /// Create an empty `Object`.
+    #[inline]
+    pub const fn new() -> Self {
+        Self {
+            inner: BTreeMap::new(),
+        }
+    }
+
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.inner.len()
+    }
+
+    #[inline]
+    pub fn is_empty(&self) -> bool {
+        self.inner.is_empty()
+    }
+
+    #[inline]
+    pub fn get(&self, key: &Value) -> Option<&Value> {
+        self.inner.get(key)
+    }
+
+    #[inline]
+    pub fn contains_key(&self, key: &Value) -> bool {
+        self.inner.contains_key(key)
+    }
+
+    #[inline]
+    pub fn get_mut(&mut self, key: &Value) -> Option<&mut Value> {
+        self.inner.get_mut(key)
+    }
+
+    /// Iteration in implementation-defined order. Non-resumable.
+    ///
+    /// For the current BTree-backed storage this happens to be sorted, but
+    /// callers MUST NOT depend on that. Use [`Object::iter_sorted`] when
+    /// deterministic order is required, or [`Object::cursor`] when iteration
+    /// must yield and resume.
+    #[inline]
+    pub fn iter(&self) -> impl Iterator<Item = (&Value, &Value)> + '_ {
+        self.inner.iter()
+    }
+
+    /// Iteration in sorted key order (by `Value::Ord`). Non-resumable.
+    ///
+    /// Use this for serialization, snapshots, hashing, `Debug`, the
+    /// `object.keys` builtin, etc.
+    #[inline]
+    pub fn iter_sorted(&self) -> Iter<'_> {
+        // BTree backend iterates sorted natively.
+        Iter {
+            inner: self.inner.iter(),
+        }
+    }
+
+    #[inline]
+    pub fn keys(&self) -> impl Iterator<Item = &Value> + '_ {
+        self.inner.keys()
+    }
+
+    /// Keys in sorted order (by `Value::Ord`). Symmetric with
+    /// [`Object::iter_sorted`].
+    #[inline]
+    pub fn keys_sorted(&self) -> impl Iterator<Item = &Value> + '_ {
+        self.iter_sorted().map(|(k, _)| k)
+    }
+
+    #[inline]
+    pub fn values(&self) -> impl Iterator<Item = &Value> + '_ {
+        self.inner.values()
+    }
+
+    #[inline]
+    pub fn iter_mut(&mut self) -> IterMut<'_> {
+        IterMut {
+            inner: self.inner.iter_mut(),
+        }
+    }
+
+    /// Insert a key-value pair. Returns the previous value if any.
+    #[inline]
+    pub fn insert(&mut self, key: Value, value: Value) -> Option<Value> {
+        self.inner.insert(key, value)
+    }
+
+    #[inline]
+    pub fn remove(&mut self, key: &Value) -> Option<Value> {
+        self.inner.remove(key)
+    }
+
+    #[inline]
+    pub fn retain<F>(&mut self, f: F)
+    where
+        F: FnMut(&Value, &mut Value) -> bool,
+    {
+        self.inner.retain(f);
+    }
+
+    #[inline]
+    pub fn clear(&mut self) {
+        self.inner.clear();
+    }
+
+    #[inline]
+    pub fn append(&mut self, other: &mut Object) {
+        self.inner.append(&mut other.inner);
+    }
+
+    /// Gets a mutable reference to the value associated with `key`, inserting
+    /// the result of `default()` if absent. Single O(log n) probe.
+    pub fn get_or_insert_with<F: FnOnce() -> Value>(
+        &mut self,
+        key: Value,
+        default: F,
+    ) -> &mut Value {
+        self.inner.entry(key).or_insert_with(default)
+    }
+
+    /// Wrap into a `Value::Object`.
+    #[inline]
+    pub fn into_value(self) -> Value {
+        Value::Object(crate::Rc::new(self))
+    }
+
+    /// Create a resumable cursor over entries in implementation-defined
+    /// order. Stable for the lifetime of `&self`. O(1).
+    ///
+    /// The cursor is fully self-owned (it stores a clone of the last-seen
+    /// key, not a reference) so it can be stored as a field of a
+    /// long-lived state struct — e.g. an RVM iteration frame that persists
+    /// across instruction dispatches. As a consequence, mutating the
+    /// `Object` between `next()` calls is not rejected by the borrow
+    /// checker; the resulting iteration order in that case is unspecified.
+    #[inline]
+    pub const fn cursor(&self) -> ObjectCursor {
+        ObjectCursor {
+            inner: ObjectCursorInner::BTree(None),
+        }
+    }
+
+    /// Advance `cursor` and yield the next entry. O(log n) for the BTree
+    /// backend (range probe); future hash/inline variants may be O(1).
+    pub fn next<'a>(&'a self, cursor: &mut ObjectCursor) -> Option<(&'a Value, &'a Value)> {
+        let ObjectCursorInner::BTree(ref mut last) = cursor.inner;
+        let next = last.as_ref().map_or_else(
+            || self.inner.iter().next(),
+            |prev| {
+                self.inner
+                    .range((Bound::Excluded(prev.clone()), Bound::Unbounded))
+                    .next()
+            },
+        );
+        let (k, v) = next?;
+        *last = Some(k.clone());
+        Some((k, v))
+    }
+}
+
+/// Opaque resumable cursor over an [`Object`]'s entries in
+/// implementation-defined order.
+///
+/// Self-owned: holds no borrow on the `Object`, so it can be stored as a
+/// field of a long-lived state struct (e.g. an RVM iteration frame).
+#[derive(Debug, Clone)]
+pub struct ObjectCursor {
+    inner: ObjectCursorInner,
+}
+
+#[derive(Debug, Clone)]
+enum ObjectCursorInner {
+    /// BTree backend cursor: tracks last-seen key. `None` means "before start".
+    BTree(Option<Value>),
+}
+
+// ---- Hand-written Ord/PartialOrd ----------------------------------------
+//
+// Implemented in terms of `iter_sorted()` so ordering is consistent with the
+// canonical (sorted) view of the entries and is therefore independent of
+// the storage variant.
+
+impl Ord for Object {
+    fn cmp(&self, other: &Self) -> Ordering {
+        self.iter_sorted().cmp(other.iter_sorted())
+    }
+}
+
+impl PartialOrd for Object {
+    #[inline]
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl fmt::Debug for Object {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // Use sorted iteration so Debug output is stable across storage
+        // variants.
+        f.debug_map().entries(self.iter_sorted()).finish()
+    }
+}
+
+impl Extend<(Value, Value)> for Object {
+    fn extend<I: IntoIterator<Item = (Value, Value)>>(&mut self, iter: I) {
+        self.inner.extend(iter);
+    }
+}
+
+impl FromIterator<(Value, Value)> for Object {
+    fn from_iter<I: IntoIterator<Item = (Value, Value)>>(iter: I) -> Self {
+        Self {
+            inner: BTreeMap::from_iter(iter),
+        }
+    }
+}
+
+impl From<BTreeMap<Value, Value>> for Object {
+    #[inline]
+    fn from(map: BTreeMap<Value, Value>) -> Self {
+        Self { inner: map }
+    }
+}
+
+impl From<Object> for Value {
+    #[inline]
+    fn from(o: Object) -> Self {
+        o.into_value()
+    }
+}

src/value/object/serde.rs

@@ -0,0 +1,59 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! Serde `Serialize`/`Deserialize` impls for [`Object`].
+
+use alloc::string::ToString as _;
+use core::fmt;
+
+use serde::de::{Deserialize, Deserializer, Error as _, MapAccess, Visitor};
+use serde::ser::{Serialize, SerializeMap as _, Serializer};
+
+use super::Object;
+use crate::value::Value;
+
+impl Serialize for Object {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        use serde::ser::Error;
+        let mut map = serializer.serialize_map(Some(self.len()))?;
+        // Sorted iteration: canonical JSON.
+        for (k, v) in self.iter_sorted() {
+            match *k {
+                Value::String(_) => map.serialize_entry(k, v)?,
+                _ => {
+                    // Non-string keys are stringified via serde_json::to_string
+                    // so the resulting JSON has valid string keys.
+                    let key_str = serde_json::to_string(k).map_err(Error::custom)?;
+                    map.serialize_entry(&key_str, v)?;
+                }
+            }
+        }
+        map.end()
+    }
+}
+
+struct ObjectVisitor;
+
+impl<'de> Visitor<'de> for ObjectVisitor {
+    type Value = Object;
+
+    fn expecting(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.write_str("a map of Value to Value")
+    }
+
+    fn visit_map<A: MapAccess<'de>>(self, mut access: A) -> Result<Self::Value, A::Error> {
+        let mut obj = Object::new();
+        while let Some((k, v)) = access.next_entry::<Value, Value>()? {
+            obj.insert(k, v);
+            crate::utils::limits::check_memory_limit_if_needed()
+                .map_err(|err| A::Error::custom(err.to_string()))?;
+        }
+        Ok(obj)
+    }
+}
+
+impl<'de> Deserialize<'de> for Object {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        deserializer.deserialize_map(ObjectVisitor)
+    }
+}

src/value/set/iter.rs

@@ -0,0 +1,103 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! Opaque iterator types for [`Set`].
+//!
+//! These newtypes wrap the storage backend's iterators so the backend can be
+//! swapped without changing any iterator type signatures observed by callers.
+
+use alloc::collections::btree_set;
+use core::iter::FusedIterator;
+
+use super::Set;
+use crate::value::Value;
+
+/// Owned iterator over `Value` elements.
+#[derive(Debug)]
+pub struct IntoIter {
+    pub(super) inner: btree_set::IntoIter<Value>,
+}
+
+impl Iterator for IntoIter {
+    type Item = Value;
+    #[inline]
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+    #[inline]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+impl DoubleEndedIterator for IntoIter {
+    #[inline]
+    fn next_back(&mut self) -> Option<Self::Item> {
+        self.inner.next_back()
+    }
+}
+
+impl ExactSizeIterator for IntoIter {
+    #[inline]
+    fn len(&self) -> usize {
+        self.inner.len()
+    }
+}
+
+impl FusedIterator for IntoIter {}
+
+/// Borrowed iterator over `&Value` elements.
+#[derive(Debug, Clone)]
+pub struct Iter<'a> {
+    pub(super) inner: btree_set::Iter<'a, Value>,
+}
+
+impl<'a> Iterator for Iter<'a> {
+    type Item = &'a Value;
+    #[inline]
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+    #[inline]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+impl<'a> DoubleEndedIterator for Iter<'a> {
+    #[inline]
+    fn next_back(&mut self) -> Option<Self::Item> {
+        self.inner.next_back()
+    }
+}
+
+impl<'a> ExactSizeIterator for Iter<'a> {
+    #[inline]
+    fn len(&self) -> usize {
+        self.inner.len()
+    }
+}
+
+impl<'a> FusedIterator for Iter<'a> {}
+
+impl IntoIterator for Set {
+    type Item = Value;
+    type IntoIter = IntoIter;
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        IntoIter {
+            inner: self.inner.into_iter(),
+        }
+    }
+}
+
+impl<'a> IntoIterator for &'a Set {
+    type Item = &'a Value;
+    type IntoIter = Iter<'a>;
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        Iter {
+            inner: self.inner.iter(),
+        }
+    }
+}

src/value/set/mod.rs

@@ -0,0 +1,269 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! See [`Set`].
+
+mod iter;
+mod serde;
+
+use alloc::collections::BTreeSet;
+use core::cmp::Ordering;
+use core::fmt;
+use core::ops::Bound;
+
+use crate::value::Value;
+
+#[allow(unused_imports)] // surface for downstream PRs
+pub use iter::{IntoIter, Iter};
+
+/// Opaque, ordered set of [`Value`]s.
+///
+/// The current backing storage is `BTreeSet<Value>`. The inner field is
+/// private so the representation can change (hash-backed, lazy, bloom-fronted,
+/// FFI-backed) without touching call sites.
+///
+/// # Iteration
+///
+/// - [`Set::iter`] — implementation-defined order; non-resumable.
+/// - [`Set::iter_sorted`] — sorted by `Value::Ord`; non-resumable.
+/// - [`Set::cursor`] / [`Set::next`] — implementation-defined order,
+///   resumable; cheapest per-step cost. Used by interpreter/RVM when iteration
+///   must yield mid-flight.
+#[derive(Default, Clone, Eq, PartialEq)]
+pub struct Set {
+    inner: BTreeSet<Value>,
+}
+
+impl Set {
+    /// Create an empty `Set`.
+    #[inline]
+    pub const fn new() -> Self {
+        Self {
+            inner: BTreeSet::new(),
+        }
+    }
+
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.inner.len()
+    }
+
+    #[inline]
+    pub fn is_empty(&self) -> bool {
+        self.inner.is_empty()
+    }
+
+    #[inline]
+    pub fn contains(&self, value: &Value) -> bool {
+        self.inner.contains(value)
+    }
+
+    #[inline]
+    pub fn get(&self, value: &Value) -> Option<&Value> {
+        self.inner.get(value)
+    }
+
+    /// First element in sorted order (by `Value::Ord`).
+    #[inline]
+    pub fn first(&self) -> Option<&Value> {
+        self.iter_sorted().next()
+    }
+
+    /// Last element in sorted order (by `Value::Ord`).
+    #[inline]
+    pub fn last(&self) -> Option<&Value> {
+        self.iter_sorted().next_back()
+    }
+
+    /// Iteration in implementation-defined order. Non-resumable.
+    ///
+    /// For the current BTree-backed storage this happens to be sorted, but
+    /// callers MUST NOT depend on that. Use [`Set::iter_sorted`] when
+    /// deterministic order is required, or [`Set::cursor`] when iteration
+    /// must yield and resume.
+    #[inline]
+    pub fn iter(&self) -> impl Iterator<Item = &Value> + '_ {
+        self.inner.iter()
+    }
+
+    /// Iteration in sorted order (by `Value::Ord`). Non-resumable.
+    ///
+    /// Use this for serialization, snapshots, hashing, `Debug`, etc.
+    #[inline]
+    pub fn iter_sorted(&self) -> Iter<'_> {
+        // BTree backend iterates sorted natively.
+        Iter {
+            inner: self.inner.iter(),
+        }
+    }
+
+    /// Insert `value`. Returns `true` if the value was newly inserted.
+    #[inline]
+    pub fn insert(&mut self, value: Value) -> bool {
+        self.inner.insert(value)
+    }
+
+    #[inline]
+    pub fn remove(&mut self, value: &Value) -> bool {
+        self.inner.remove(value)
+    }
+
+    #[inline]
+    pub fn retain<F>(&mut self, f: F)
+    where
+        F: FnMut(&Value) -> bool,
+    {
+        self.inner.retain(f);
+    }
+
+    #[inline]
+    pub fn clear(&mut self) {
+        self.inner.clear();
+    }
+
+    #[inline]
+    pub fn append(&mut self, other: &mut Set) {
+        self.inner.append(&mut other.inner);
+    }
+
+    /// Set intersection. Returns a new `Set` containing the elements
+    /// present in both `self` and `other`.
+    pub fn intersection(&self, other: &Set) -> Set {
+        Set {
+            inner: self.inner.intersection(&other.inner).cloned().collect(),
+        }
+    }
+
+    /// Set union. Returns a new `Set` containing the elements present in
+    /// either `self` or `other`.
+    pub fn union(&self, other: &Set) -> Set {
+        Set {
+            inner: self.inner.union(&other.inner).cloned().collect(),
+        }
+    }
+
+    /// Set difference. Returns a new `Set` containing the elements present
+    /// in `self` but not in `other`.
+    pub fn difference(&self, other: &Set) -> Set {
+        Set {
+            inner: self.inner.difference(&other.inner).cloned().collect(),
+        }
+    }
+
+    #[inline]
+    pub fn is_subset(&self, other: &Set) -> bool {
+        self.inner.is_subset(&other.inner)
+    }
+
+    /// Wrap into a `Value::Set`.
+    #[inline]
+    pub fn into_value(self) -> Value {
+        Value::Set(crate::Rc::new(self.inner))
+    }
+
+    /// Create a resumable cursor over elements in implementation-defined
+    /// order. Stable for the lifetime of `&self`. O(1).
+    ///
+    /// The cursor is fully self-owned (it stores a clone of the last-seen
+    /// element, not a reference) so it can be stored as a field of a
+    /// long-lived state struct — e.g. an RVM iteration frame that persists
+    /// across instruction dispatches. As a consequence, mutating the `Set`
+    /// between `next()` calls is not rejected by the borrow checker; the
+    /// resulting iteration order in that case is unspecified.
+    #[inline]
+    pub const fn cursor(&self) -> SetCursor {
+        SetCursor {
+            inner: SetCursorInner::BTree(None),
+        }
+    }
+
+    /// Advance `cursor` and yield the next element. O(log n) for the BTree
+    /// backend (range probe); future hash/inline variants may be O(1).
+    pub fn next<'a>(&'a self, cursor: &mut SetCursor) -> Option<&'a Value> {
+        let SetCursorInner::BTree(ref mut last) = cursor.inner;
+        let next = last.as_ref().map_or_else(
+            || self.inner.iter().next(),
+            |prev| {
+                // `(Bound<&T>, Bound<&T>)` impls `RangeBounds<T>` — no clone
+                // needed to build the resume bound.
+                self.inner
+                    .range((Bound::Excluded(prev), Bound::Unbounded))
+                    .next()
+            },
+        );
+        let v = next?;
+        *last = Some(v.clone());
+        Some(v)
+    }
+}
+
+/// Opaque resumable cursor over a [`Set`]'s elements in
+/// implementation-defined order.
+///
+/// Self-owned: holds no borrow on the `Set`, so it can be stored as a
+/// field of a long-lived state struct (e.g. an RVM iteration frame).
+#[derive(Debug, Clone)]
+pub struct SetCursor {
+    inner: SetCursorInner,
+}
+
+#[derive(Debug, Clone)]
+enum SetCursorInner {
+    /// BTree backend cursor: tracks last-seen element. `None` means "before start".
+    BTree(Option<Value>),
+}
+
+// ---- Hand-written Ord/PartialOrd ----------------------------------------
+//
+// Implemented in terms of `iter_sorted()` so ordering is consistent with the
+// canonical (sorted) view of the elements and is therefore independent of
+// the storage variant.
+
+impl Ord for Set {
+    fn cmp(&self, other: &Self) -> Ordering {
+        self.iter_sorted().cmp(other.iter_sorted())
+    }
+}
+
+impl PartialOrd for Set {
+    #[inline]
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl fmt::Debug for Set {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // Use sorted iteration so Debug output is stable across storage
+        // variants.
+        f.debug_set().entries(self.iter_sorted()).finish()
+    }
+}
+
+impl Extend<Value> for Set {
+    fn extend<I: IntoIterator<Item = Value>>(&mut self, iter: I) {
+        self.inner.extend(iter);
+    }
+}
+
+impl FromIterator<Value> for Set {
+    fn from_iter<I: IntoIterator<Item = Value>>(iter: I) -> Self {
+        Self {
+            inner: BTreeSet::from_iter(iter),
+        }
+    }
+}
+
+impl From<BTreeSet<Value>> for Set {
+    #[inline]
+    fn from(set: BTreeSet<Value>) -> Self {
+        Self { inner: set }
+    }
+}
+
+impl From<Set> for Value {
+    #[inline]
+    fn from(s: Set) -> Self {
+        s.into_value()
+    }
+}

src/value/set/serde.rs

@@ -0,0 +1,44 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+//! Serde `Serialize`/`Deserialize` impls for [`Set`].
+
+use core::fmt;
+
+use serde::de::{Deserialize, Deserializer, Error as _, SeqAccess, Visitor};
+use serde::ser::{Serialize, Serializer};
+
+use super::Set;
+use crate::value::Value;
+
+impl Serialize for Set {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        // Sets serialize as JSON arrays. Sorted iteration: canonical output.
+        serializer.collect_seq(self.iter_sorted())
+    }
+}
+
+struct SetVisitor;
+
+impl<'de> Visitor<'de> for SetVisitor {
+    type Value = Set;
+
+    fn expecting(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.write_str("a sequence of Values")
+    }
+
+    fn visit_seq<A: SeqAccess<'de>>(self, mut access: A) -> Result<Self::Value, A::Error> {
+        let mut set = Set::new();
+        while let Some(v) = access.next_element::<Value>()? {
+            set.insert(v);
+            crate::utils::limits::check_memory_limit_if_needed().map_err(A::Error::custom)?;
+        }
+        Ok(set)
+    }
+}
+
+impl<'de> Deserialize<'de> for Set {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        deserializer.deserialize_seq(SetVisitor)
+    }
+}

src/value/tests.rs

@@ -0,0 +1,868 @@
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+#![allow(
+    clippy::panic,
+    clippy::expect_used,
+    clippy::unwrap_used,
+    clippy::indexing_slicing,
+    clippy::as_conversions,
+    clippy::arithmetic_side_effects,
+    clippy::unseparated_literal_suffix,
+    clippy::map_unwrap_or,
+    clippy::option_if_let_else,
+    clippy::pattern_type_mismatch
+)]
+
+use alloc::collections::{BTreeMap, BTreeSet};
+use alloc::format;
+use alloc::vec::Vec;
+
+use super::{Object, Set};
+use crate::value::Value;
+
+fn val(i: u64) -> Value {
+    Value::from(i)
+}
+
+fn make_pairs(n: u64) -> Vec<(Value, Value)> {
+    (0..n).map(|i| (val(i), val(i.saturating_mul(2)))).collect()
+}
+
+const SIZES: &[u64] = &[0, 1, 2, 4, 8, 64, 256, 1024];
+
+/// `iter_sorted` must yield entries in the same order as a `BTreeMap` oracle.
+#[test]
+fn object_iter_sorted_matches_btreemap_oracle() {
+    for &n in SIZES {
+        let pairs = make_pairs(n);
+        let oracle: BTreeMap<Value, Value> = pairs.iter().cloned().collect();
+        let obj: Object = pairs.into_iter().collect();
+        let actual: Vec<(&Value, &Value)> = obj.iter_sorted().collect();
+        let expected: Vec<(&Value, &Value)> = oracle.iter().collect();
+        assert_eq!(actual, expected, "size {n}");
+    }
+}
+
+/// `iter` may be in any order, but as a multiset must equal the oracle's entries.
+#[test]
+fn object_iter_multiset_equality_with_oracle() {
+    for &n in SIZES {
+        let pairs = make_pairs(n);
+        let oracle: BTreeMap<Value, Value> = pairs.iter().cloned().collect();
+        let obj: Object = pairs.into_iter().collect();
+        assert_eq!(obj.len(), oracle.len(), "size {n}");
+        let mut a: Vec<(Value, Value)> = obj.iter().map(|(k, v)| (k.clone(), v.clone())).collect();
+        let mut b: Vec<(Value, Value)> =
+            oracle.iter().map(|(k, v)| (k.clone(), v.clone())).collect();
+        a.sort();
+        b.sort();
+        assert_eq!(a, b);
+    }
+}
+
+/// Serialize-then-deserialize must round-trip through JSON without loss.
+#[test]
+fn object_serde_roundtrip() {
+    for &n in &[0_u64, 1, 8, 64] {
+        let pairs: Vec<(Value, Value)> = (0..n)
+            .map(|i| (Value::String(format!("k{i}").into()), val(i)))
+            .collect();
+        let obj: Object = pairs.into_iter().collect();
+        let json = serde_json::to_string(&obj).expect("ser");
+        let back: Object = serde_json::from_str(&json).expect("de");
+        assert_eq!(obj, back, "size {n}");
+    }
+}
+
+/// Equality depends only on contents, not the order keys were inserted.
+#[test]
+fn object_eq_invariant_to_insertion_order() {
+    let mut a = Object::new();
+    let mut b = Object::new();
+    for i in 0..32_u64 {
+        a.insert(val(i), val(i.saturating_add(1)));
+    }
+    for i in (0..32_u64).rev() {
+        b.insert(val(i), val(i.saturating_add(1)));
+    }
+    assert_eq!(a, b);
+}
+
+/// `remove` returns the prior value (or `None`) and `retain` keeps only matching entries.
+#[test]
+fn object_remove_and_retain() {
+    let mut obj: Object = make_pairs(16).into_iter().collect();
+    assert_eq!(obj.remove(&val(0)), Some(val(0)));
+    assert!(obj.remove(&val(100)).is_none());
+    obj.retain(|_, v| {
+        if let Value::Number(ref n) = *v {
+            n.as_u64().is_some_and(|x| x % 4 == 0)
+        } else {
+            false
+        }
+    });
+    for (_, v) in obj.iter_sorted() {
+        if let Value::Number(ref n) = *v {
+            assert_eq!(n.as_u64().expect("u64") % 4, 0);
+        }
+    }
+}
+
+/// `IntoIterator` for `Object` (by value) yields every entry exactly once.
+#[test]
+fn object_into_iterator_owned() {
+    let obj: Object = make_pairs(8).into_iter().collect();
+    let collected: Vec<(Value, Value)> = obj.into_iter().collect();
+    assert_eq!(collected.len(), 8);
+}
+
+// ---- Duplicate-key semantics --------------------------------------------
+
+/// `FromIterator` keeps the last value when the same key appears multiple times.
+#[test]
+fn object_from_iter_last_wins_on_duplicate_keys() {
+    let obj = Object::from_iter([(val(0), val(1)), (val(0), val(2))]);
+    assert_eq!(obj.get(&val(0)), Some(&val(2)));
+    assert_eq!(obj.len(), 1);
+}
+
+/// `From<BTreeMap>` adopts `BTreeMap`'s own last-write-wins semantics for duplicates.
+#[test]
+fn object_from_btreemap_last_wins_on_duplicate_keys() {
+    let mut bm: BTreeMap<Value, Value> = BTreeMap::new();
+    bm.insert(val(0), val(1));
+    bm.insert(val(0), val(2));
+    let obj: Object = bm.into();
+    assert_eq!(obj.get(&val(0)), Some(&val(2)));
+    assert_eq!(obj.len(), 1);
+}
+
+// ---- get_or_insert_with --------------------------------------------------
+
+/// `get_or_insert_with` inserts the default when the key is absent and returns a mutable ref to it.
+#[test]
+fn object_get_or_insert_with_inserts_when_absent() {
+    let mut obj = Object::new();
+    let v = obj.get_or_insert_with(val(7), || val(42));
+    assert_eq!(*v, val(42));
+    *v = val(43);
+    assert_eq!(obj.get(&val(7)), Some(&val(43)));
+}
+
+/// `get_or_insert_with` returns the existing value and never invokes the default closure.
+#[test]
+fn object_get_or_insert_with_returns_existing_when_present() {
+    let mut obj = Object::new();
+    obj.insert(val(7), val(1));
+    let mut closure_called = false;
+    let v = obj.get_or_insert_with(val(7), || {
+        closure_called = true;
+        val(999)
+    });
+    assert_eq!(*v, val(1));
+    assert!(!closure_called, "default closure must not run when present");
+}
+
+// ---- Accessor coverage ---------------------------------------------------
+
+/// Smoke-test every accessor: `contains_key`/`get`/`get_mut`/`keys`/`values`/`iter`/`iter_mut`/`append`/`clear`.
+#[test]
+fn object_accessor_coverage() {
+    let mut obj: Object = make_pairs(4).into_iter().collect();
+
+    assert!(obj.contains_key(&val(0)));
+    assert!(!obj.contains_key(&val(100)));
+
+    assert_eq!(obj.get(&val(2)), Some(&val(4)));
+
+    if let Some(v) = obj.get_mut(&val(1)) {
+        *v = val(999);
+    }
+    assert_eq!(obj.get(&val(1)), Some(&val(999)));
+
+    let keys: Vec<&Value> = obj.keys().collect();
+    assert_eq!(keys.len(), 4);
+    let values: Vec<&Value> = obj.values().collect();
+    assert_eq!(values.len(), 4);
+
+    for (_, v) in obj.iter_mut() {
+        *v = val(0);
+    }
+    for (_, v) in obj.iter() {
+        assert_eq!(*v, val(0));
+    }
+
+    let mut other = Object::new();
+    other.insert(val(100), val(200));
+    obj.append(&mut other);
+    assert!(other.is_empty());
+    assert!(obj.contains_key(&val(100)));
+
+    obj.clear();
+    assert!(obj.is_empty());
+}
+
+// ---- IntoIterator for references -----------------------------------------
+
+/// `IntoIterator` for `&Object` yields shared refs to every entry.
+#[test]
+fn object_into_iterator_ref() {
+    let obj: Object = make_pairs(4).into_iter().collect();
+    let mut count = 0;
+    for (_k, _v) in &obj {
+        count += 1;
+    }
+    assert_eq!(count, 4);
+}
+
+/// `IntoIterator` for `&mut Object` exposes mutable refs to values; mutations persist.
+#[test]
+fn object_into_iterator_ref_mut() {
+    let mut obj: Object = make_pairs(4).into_iter().collect();
+    for (_k, v) in &mut obj {
+        *v = val(0);
+    }
+    for (_, v) in obj.iter() {
+        assert_eq!(*v, val(0));
+    }
+}
+
+// ---- Cursor tests --------------------------------------------------------
+
+/// Driving `cursor`+`next` to completion visits each entry exactly once.
+#[test]
+fn object_cursor_yields_every_entry_once() {
+    for &n in SIZES {
+        let pairs = make_pairs(n);
+        let obj: Object = pairs.clone().into_iter().collect();
+        let mut cursor = obj.cursor();
+        let mut collected: Vec<(Value, Value)> = Vec::new();
+        while let Some((k, v)) = obj.next(&mut cursor) {
+            collected.push((k.clone(), v.clone()));
+        }
+        let mut a = collected;
+        a.sort();
+        let mut b = pairs;
+        b.sort();
+        assert_eq!(a, b, "size {n}");
+    }
+}
+
+/// A freshly-constructed cursor restarts from the beginning, independent of any prior cursor's state.
+#[test]
+fn object_cursor_resumable_fresh_cursor_restarts() {
+    let obj: Object = make_pairs(8).into_iter().collect();
+    let mut c1 = obj.cursor();
+    let _ = obj.next(&mut c1);
+    let _ = obj.next(&mut c1);
+    let mut c2 = obj.cursor();
+    let first_again = obj.next(&mut c2);
+    let first_original = obj.iter().next();
+    assert_eq!(
+        first_again.map(|(k, v)| (k.clone(), v.clone())),
+        first_original.map(|(k, v)| (k.clone(), v.clone()))
+    );
+}
+
+/// When `Object` is shared via `Rc`, `Rc::make_mut` clones — leaving an in-flight cursor on the original snapshot unaffected.
+#[test]
+fn object_cursor_snapshot_independence_via_rc() {
+    use crate::Rc;
+    let mut obj = Object::new();
+    obj.insert(Value::from("a"), Value::from(1));
+    obj.insert(Value::from("b"), Value::from(2));
+    obj.insert(Value::from("c"), Value::from(3));
+    let rc_obj = Rc::new(obj);
+
+    let alias = Rc::clone(&rc_obj);
+    let mut cursor = rc_obj.cursor();
+    let _ = rc_obj.next(&mut cursor);
+
+    let mut alias_for_mut = alias;
+    Rc::make_mut(&mut alias_for_mut).insert(Value::from("d"), Value::from(4));
+    Rc::make_mut(&mut alias_for_mut).remove(&Value::from("a"));
+
+    assert_eq!(rc_obj.len(), 3);
+    let mut remaining = 0;
+    while rc_obj.next(&mut cursor).is_some() {
+        remaining += 1;
+    }
+    assert_eq!(remaining, 2);
+}
+
+/// A cursor over an empty `Object` returns `None` on the first call.
+#[test]
+fn object_cursor_empty_returns_none_immediately() {
+    let obj = Object::new();
+    let mut cursor = obj.cursor();
+    assert!(obj.next(&mut cursor).is_none());
+}
+
+/// Mutating an `Object` between `next()` calls is well-defined: the cursor
+/// must not panic and must terminate. The visit order, and whether
+/// inserted/removed keys appear, is intentionally unspecified — this test
+/// only pins the safety + termination guarantees that callers (e.g. a
+/// future RVM iteration frame) may rely on. It must NOT assert any
+/// particular order or count, or future backend swaps will be forced to
+/// honor an accidental contract.
+#[test]
+fn object_cursor_mutation_between_steps_is_safe_and_terminates() {
+    let mut obj: Object = make_pairs(16).into_iter().collect();
+    let mut cursor = obj.cursor();
+
+    // Yield a few entries before mutating.
+    for _ in 0..3 {
+        let _ = obj.next(&mut cursor);
+    }
+
+    // Interleave mutations and steps. Each yielded entry must, at the
+    // moment of yield, be a real entry in the map.
+    obj.insert(val(100), val(100));
+    if let Some((k, v)) = obj.next(&mut cursor) {
+        assert_eq!(obj.get(k), Some(v));
+    }
+    obj.remove(&val(2));
+    if let Some((k, v)) = obj.next(&mut cursor) {
+        assert_eq!(obj.get(k), Some(v));
+    }
+    obj.clear();
+    // After clear(), draining the cursor must terminate (not panic, not
+    // loop) within a bounded number of calls.
+    let mut terminated = false;
+    for _ in 0..32 {
+        if obj.next(&mut cursor).is_none() {
+            terminated = true;
+            break;
+        }
+    }
+    assert!(terminated, "cursor failed to terminate after clear()");
+}
+
+// ---- Hand-written Ord consistency ---------------------------------------
+
+/// `Ord` (built atop `iter_sorted`) is invariant to insertion order.
+#[test]
+fn object_ord_invariant_to_insertion_order() {
+    let mut a = Object::new();
+    let mut b = Object::new();
+    for i in 0..16_u64 {
+        a.insert(val(i), val(i.saturating_add(1)));
+    }
+    for i in (0..16_u64).rev() {
+        b.insert(val(i), val(i.saturating_add(1)));
+    }
+    use core::cmp::Ordering;
+    assert_eq!(a.cmp(&b), Ordering::Equal);
+}
+
+/// `Ord` agrees with lexicographic comparison of the sorted-entries view.
+#[test]
+fn object_ord_lexicographic_on_sorted_entries() {
+    let a: Object = [(val(0), val(0)), (val(1), val(1))].into_iter().collect();
+    let b: Object = [(val(0), val(0)), (val(2), val(2))].into_iter().collect();
+    assert!(a < b);
+}
+
+/// `empty < non_empty` and a shorter prefix compares less than its extension.
+#[test]
+fn object_ord_empty_and_prefix() {
+    use core::cmp::Ordering;
+    let empty = Object::new();
+    let one: Object = [(val(0), val(0))].into_iter().collect();
+    let two: Object = [(val(0), val(0)), (val(1), val(1))].into_iter().collect();
+    assert_eq!(empty.cmp(&one), Ordering::Less);
+    assert_eq!(one.cmp(&two), Ordering::Less);
+    assert_eq!(two.cmp(&empty), Ordering::Greater);
+}
+
+/// When keys match, `Ord` falls through to comparing values.
+#[test]
+fn object_ord_breaks_ties_on_values() {
+    use core::cmp::Ordering;
+    let a: Object = [(val(0), val(1))].into_iter().collect();
+    let b: Object = [(val(0), val(2))].into_iter().collect();
+    assert_eq!(a.cmp(&b), Ordering::Less);
+}
+
+/// `PartialOrd` must agree with `Ord` for every input pair.
+#[test]
+fn object_partial_cmp_matches_cmp() {
+    let a: Object = [(val(0), val(0)), (val(1), val(1))].into_iter().collect();
+    let b: Object = [(val(0), val(0)), (val(2), val(2))].into_iter().collect();
+    assert_eq!(a.partial_cmp(&b), Some(a.cmp(&b)));
+    assert_eq!(b.partial_cmp(&a), Some(b.cmp(&a)));
+    assert_eq!(a.partial_cmp(&a), Some(core::cmp::Ordering::Equal));
+}
+
+// ---- Debug / keys_sorted determinism ------------------------------------
+
+/// `Debug` output is byte-identical for equal Objects regardless of insertion order.
+#[test]
+fn object_debug_invariant_to_insertion_order() {
+    let mut a = Object::new();
+    let mut b = Object::new();
+    for i in 0..8_u64 {
+        a.insert(val(i), val(i));
+    }
+    for i in (0..8_u64).rev() {
+        b.insert(val(i), val(i));
+    }
+    assert_eq!(format!("{a:?}"), format!("{b:?}"));
+}
+
+/// `keys_sorted` yields exactly `iter_sorted().map(|(k,_)| k)`.
+#[test]
+fn object_keys_sorted_matches_iter_sorted_keys() {
+    let obj: Object = make_pairs(16).into_iter().collect();
+    let from_keys: Vec<&Value> = obj.keys_sorted().collect();
+    let from_iter: Vec<&Value> = obj.iter_sorted().map(|(k, _)| k).collect();
+    assert_eq!(from_keys, from_iter);
+}
+
+// ---- Serde: non-string keys & determinism --------------------------------
+
+/// `Serialize` stringifies non-string keys, and equal Objects produce identical JSON
+/// regardless of insertion order.
+#[test]
+fn object_serialize_non_string_keys_and_deterministic() {
+    let pairs = [
+        (Value::from("alpha"), val(1)),
+        (Value::Bool(true), val(2)),
+        (val(7), val(3)),
+    ];
+    let a: Object = pairs.iter().cloned().collect();
+    let mut b = Object::new();
+    for (k, v) in pairs.iter().rev().cloned() {
+        b.insert(k, v);
+    }
+    let ja = serde_json::to_string(&a).expect("ser a");
+    let jb = serde_json::to_string(&b).expect("ser b");
+    assert_eq!(ja, jb, "serialization must be deterministic");
+
+    // Non-string keys appear as quoted strings in the resulting JSON.
+    let v: serde_json::Value = serde_json::from_str(&ja).expect("parse");
+    let obj = v.as_object().expect("json object");
+    assert!(
+        obj.contains_key("true"),
+        "bool key was not stringified: {ja}"
+    );
+    assert!(
+        obj.contains_key("7"),
+        "number key was not stringified: {ja}"
+    );
+    assert!(obj.contains_key("alpha"));
+}
+
+// ---- Extend / append duplicate-key semantics -----------------------------
+
+/// `extend` overwrites existing entries (last-write-wins) and preserves length when
+/// only existing keys are touched.
+#[test]
+fn object_extend_last_wins_and_empty_noop() {
+    let mut obj: Object = [(val(0), val(0)), (val(1), val(1))].into_iter().collect();
+    obj.extend([(val(0), val(99))]);
+    assert_eq!(obj.get(&val(0)), Some(&val(99)));
+    assert_eq!(obj.len(), 2);
+
+    let before = obj.len();
+    obj.extend(core::iter::empty::<(Value, Value)>());
+    assert_eq!(obj.len(), before, "empty extend is a no-op");
+}
+
+/// `append` drains `other` into `self`, overwriting on overlapping keys.
+#[test]
+fn object_append_overlapping_keys_drain_and_overwrite() {
+    let mut a: Object = [(val(0), val(0)), (val(1), val(1))].into_iter().collect();
+    let mut b: Object = [(val(1), val(99)), (val(2), val(2))].into_iter().collect();
+    a.append(&mut b);
+    assert!(b.is_empty(), "append must drain `other`");
+    assert_eq!(a.len(), 3);
+    assert_eq!(a.get(&val(1)), Some(&val(99)));
+    assert_eq!(a.get(&val(2)), Some(&val(2)));
+}
+
+// ---- Iterator trait surface ---------------------------------------------
+
+/// `DoubleEndedIterator`/`ExactSizeIterator`/`FusedIterator` and `size_hint` all
+/// behave correctly across partial consumption from both ends.
+#[test]
+fn object_iter_sorted_double_ended_and_exact_size() {
+    let obj: Object = make_pairs(4).into_iter().collect();
+    let mut it = obj.iter_sorted();
+    assert_eq!(it.len(), 4);
+    assert_eq!(it.size_hint(), (4, Some(4)));
+
+    let first = it.next().expect("front");
+    let last = it.next_back().expect("back");
+    assert_eq!(it.len(), 2);
+    assert_eq!(it.size_hint(), (2, Some(2)));
+    assert_ne!(first.0, last.0, "front and back must differ for n=4");
+
+    // Drain remaining.
+    while it.next().is_some() {}
+    assert_eq!(it.len(), 0);
+    // FusedIterator: stays None after exhaustion.
+    assert!(it.next().is_none());
+    assert!(it.next().is_none());
+    assert!(it.next_back().is_none());
+}
+
+/// `IntoIter` also honors `DoubleEndedIterator` and `ExactSizeIterator`.
+#[test]
+fn object_into_iter_double_ended_and_exact_size() {
+    let obj: Object = make_pairs(4).into_iter().collect();
+    let mut it = obj.into_iter();
+    assert_eq!(it.len(), 4);
+    let _ = it.next().expect("front");
+    let _ = it.next_back().expect("back");
+    assert_eq!(it.len(), 2);
+    let collected: Vec<_> = it.collect();
+    assert_eq!(collected.len(), 2);
+}
+
+/// `IterMut` decrements its `len()` after consuming from the front.
+#[test]
+fn object_iter_mut_exact_size() {
+    let mut obj: Object = make_pairs(3).into_iter().collect();
+    let mut it = obj.iter_mut();
+    assert_eq!(it.len(), 3);
+    let _ = it.next().expect("front");
+    assert_eq!(it.len(), 2);
+}
+
+/// `Iter` is `Clone`; the clone iterates independently from the same point.
+#[test]
+fn object_iter_sorted_clone_is_independent() {
+    let obj: Object = make_pairs(4).into_iter().collect();
+    let mut a = obj.iter_sorted();
+    let _ = a.next();
+    let b = a.clone();
+    let rest_a: Vec<_> = a.collect();
+    let rest_b: Vec<_> = b.collect();
+    assert_eq!(rest_a, rest_b);
+}
+
+// ---- default / insert ---------------------------------------------------
+
+/// `Object::default()` and `Object::new()` produce equal, empty Objects.
+#[test]
+fn object_default_equals_new_and_is_empty() {
+    let a = Object::default();
+    let b = Object::new();
+    assert_eq!(a, b);
+    assert!(a.is_empty());
+    assert_eq!(a.len(), 0);
+}
+
+/// `insert` returns `None` for a fresh key and `Some(old)` when overwriting.
+#[test]
+fn object_insert_returns_previous_value() {
+    let mut obj = Object::new();
+    assert_eq!(obj.insert(val(0), val(1)), None);
+    assert_eq!(obj.insert(val(0), val(2)), Some(val(1)));
+    assert_eq!(obj.get(&val(0)), Some(&val(2)));
+}
+
+// =========================================================================
+// Set tests
+// =========================================================================
+
+const SET_SIZES: &[u64] = &[0, 1, 2, 4, 8, 64, 256, 1024];
+
+#[test]
+fn set_iter_sorted_matches_btreeset_oracle() {
+    for &n in SET_SIZES {
+        let values: Vec<Value> = (0..n).map(val).collect();
+        let oracle: BTreeSet<Value> = values.iter().cloned().collect();
+        let s: Set = values.into_iter().collect();
+        let actual: Vec<&Value> = s.iter_sorted().collect();
+        let expected: Vec<&Value> = oracle.iter().collect();
+        assert_eq!(actual, expected, "size {n}");
+    }
+}
+
+#[test]
+fn set_iter_multiset_equality_with_oracle() {
+    for &n in SET_SIZES {
+        let values: Vec<Value> = (0..n).map(val).collect();
+        let oracle: BTreeSet<Value> = values.iter().cloned().collect();
+        let s: Set = values.into_iter().collect();
+        let mut a: Vec<Value> = s.iter().cloned().collect();
+        let mut b: Vec<Value> = oracle.iter().cloned().collect();
+        a.sort();
+        b.sort();
+        assert_eq!(a, b);
+    }
+}
+
+#[test]
+fn set_algebra_matches_btreeset() {
+    let a_vals: Vec<Value> = (0..32_u64).map(val).collect();
+    let b_vals: Vec<Value> = (16..48_u64).map(val).collect();
+    let a_btree: BTreeSet<Value> = a_vals.iter().cloned().collect();
+    let b_btree: BTreeSet<Value> = b_vals.iter().cloned().collect();
+    let a: Set = a_vals.into_iter().collect();
+    let b: Set = b_vals.into_iter().collect();
+
+    fn sorted<'a, I: Iterator<Item = &'a Value>>(it: I) -> Vec<&'a Value> {
+        let mut v: Vec<&Value> = it.collect();
+        v.sort();
+        v
+    }
+
+    let inter_set = a.intersection(&b);
+    assert_eq!(
+        sorted(inter_set.iter_sorted()),
+        sorted(a_btree.intersection(&b_btree))
+    );
+
+    let union_set = a.union(&b);
+    assert_eq!(
+        sorted(union_set.iter_sorted()),
+        sorted(a_btree.union(&b_btree))
+    );
+
+    let diff_set = a.difference(&b);
+    assert_eq!(
+        sorted(diff_set.iter_sorted()),
+        sorted(a_btree.difference(&b_btree))
+    );
+
+    // Subset: trivial + non-trivial cases.
+    let proper_subset: Set = (0..16_u64).map(val).collect();
+    let non_subset: Set = (30..50_u64).map(val).collect();
+    assert!(a.is_subset(&a));
+    assert!(proper_subset.is_subset(&a));
+    assert!(!non_subset.is_subset(&a));
+}
+
+#[test]
+fn set_first_last() {
+    let s: Set = (0..16_u64).map(val).collect();
+    assert_eq!(s.first(), Some(&val(0)));
+    assert_eq!(s.last(), Some(&val(15)));
+    assert!(Set::new().first().is_none());
+}
+
+#[test]
+fn set_serde_roundtrip() {
+    for &n in &[0_u64, 1, 8, 64] {
+        let s: Set = (0..n).map(val).collect();
+        let json = serde_json::to_string(&s).expect("ser");
+        let back: Set = serde_json::from_str(&json).expect("de");
+        assert_eq!(s, back, "size {n}");
+    }
+}
+
+#[test]
+fn set_append_drains_other() {
+    let mut a: Set = (0..4_u64).map(val).collect();
+    let mut b: Set = (4..8_u64).map(val).collect();
+    a.append(&mut b);
+    assert_eq!(a.len(), 8);
+    assert!(b.is_empty());
+}
+
+#[test]
+fn set_value_cow_make_mut_isolates_clones() {
+    let a = Value::new_set();
+    let b = a.clone();
+    let mut b_owned = b;
+    b_owned.as_set_mut().expect("set").insert(Value::from("x"));
+    assert_eq!(a.as_set().expect("set").len(), 0);
+    assert_eq!(b_owned.as_set().expect("set").len(), 1);
+}
+
+#[test]
+fn set_from_iter_dedups_duplicates() {
+    let s: Set = [val(1), val(1), val(2), val(2), val(2)]
+        .into_iter()
+        .collect();
+    assert_eq!(s.len(), 2);
+    assert!(s.contains(&val(1)));
+    assert!(s.contains(&val(2)));
+}
+
+#[test]
+fn set_accessor_coverage() {
+    let mut s: Set = (0..4_u64).map(val).collect();
+
+    assert!(s.contains(&val(2)));
+    assert!(!s.contains(&val(100)));
+
+    assert_eq!(s.get(&val(2)), Some(&val(2)));
+    assert!(s.get(&val(100)).is_none());
+
+    assert!(s.remove(&val(2)));
+    assert!(!s.remove(&val(2)));
+    assert_eq!(s.len(), 3);
+
+    s.retain(|v| v != &val(0));
+    assert!(!s.contains(&val(0)));
+    assert_eq!(s.len(), 2);
+
+    s.clear();
+    assert!(s.is_empty());
+    assert!(!s.contains(&val(1)));
+}
+
+#[test]
+fn set_into_iterator_ref() {
+    let s: Set = (0..4_u64).map(val).collect();
+    let mut count = 0;
+    for _v in &s {
+        count += 1;
+    }
+    assert_eq!(count, 4);
+}
+
+#[test]
+fn set_cursor_yields_every_element_once() {
+    for &n in SET_SIZES {
+        let vals: Vec<Value> = (0..n).map(val).collect();
+        let s: Set = vals.clone().into_iter().collect();
+        let mut cursor = s.cursor();
+        let mut collected: Vec<Value> = Vec::new();
+        while let Some(v) = s.next(&mut cursor) {
+            collected.push(v.clone());
+        }
+        let mut a = collected;
+        a.sort();
+        let mut b = vals;
+        b.sort();
+        assert_eq!(a, b, "size {n}");
+    }
+}
+
+#[test]
+fn set_cursor_empty_returns_none_immediately() {
+    let s = Set::new();
+    let mut c = s.cursor();
+    assert!(s.next(&mut c).is_none());
+}
+
+#[test]
+fn set_ord_invariant_to_insertion_order() {
+    let mut a = Set::new();
+    let mut b = Set::new();
+    for i in 0..16_u64 {
+        a.insert(val(i));
+    }
+    for i in (0..16_u64).rev() {
+        b.insert(val(i));
+    }
+    assert_eq!(a.cmp(&b), core::cmp::Ordering::Equal);
+}
+
+fn top_object_rc(v: &Value) -> crate::Rc<Object> {
+    match v {
+        Value::Object(rc) => crate::Rc::clone(rc),
+        other => panic!("expected object, got {other:?}"),
+    }
+}
+
+/// A no-op deep-merge (every incoming key already present with an equal value) must not clone
+/// the target map. `deep_merge` acquires mutable access lazily, so when nothing changes at a
+/// level the shared `Rc` is left untouched.
+#[test]
+fn deep_merge_noop_subset_does_not_clone_object() {
+    let mut a = Value::from_json_str(r#"{"x": {"deep": 1}, "y": 2}"#).unwrap();
+    // Keep a second reference so the map's refcount > 1: eager `make_mut` would clone here.
+    let shared = a.clone();
+    let before = top_object_rc(&a);
+
+    // Strict subset with identical values: no insert, no recurse, no conflict at any level.
+    a.deep_merge(Value::from_json_str(r#"{"y": 2}"#).unwrap())
+        .unwrap();
+
+    let after = top_object_rc(&a);
+    assert!(
+        crate::Rc::ptr_eq(&before, &after),
+        "no-op merge must not clone the shared object map"
+    );
+    assert_eq!(a, shared, "value must be unchanged by a no-op merge");
+}
+
+/// An equal nested object under a shared key is a no-op too — the equality short-circuit runs
+/// before any mutable access, so the map is not cloned.
+#[test]
+fn deep_merge_equal_nested_object_does_not_clone() {
+    let mut a = Value::from_json_str(r#"{"cfg": {"a": 1, "b": 2}, "n": 5}"#).unwrap();
+    let _shared = a.clone();
+    let before = top_object_rc(&a);
+
+    a.deep_merge(Value::from_json_str(r#"{"cfg": {"a": 1, "b": 2}}"#).unwrap())
+        .unwrap();
+
+    let after = top_object_rc(&a);
+    assert!(
+        crate::Rc::ptr_eq(&before, &after),
+        "merging an equal nested object must not clone the map"
+    );
+}
+
+/// A conflict on the first overlapping key is reported without cloning the target map: the
+/// read-only probe detects the conflict before any mutable access is taken.
+#[test]
+fn deep_merge_conflict_does_not_clone_object() {
+    let mut a = Value::from_json_str(r#"{"x": 1, "y": 2}"#).unwrap();
+    let _shared = a.clone();
+    let before = top_object_rc(&a);
+
+    let err = a
+        .deep_merge(Value::from_json_str(r#"{"x": 999}"#).unwrap())
+        .unwrap_err();
+    assert!(format!("{err}").contains("generated multiple times"));
+
+    let after = top_object_rc(&a);
+    assert!(
+        crate::Rc::ptr_eq(&before, &after),
+        "a conflict must not clone the shared object map"
+    );
+}
+
+/// Nest `depth` objects `{"k": {"k": ... leaf}}` iteratively, so building the value can't itself
+/// overflow and there's no parser to cap depth first.
+fn nest(depth: usize, leaf: Value) -> Value {
+    let mut v = leaf;
+    for _ in 0..depth {
+        let mut m = BTreeMap::new();
+        m.insert(Value::from("k"), v);
+        v = Value::from(m);
+    }
+    v
+}
+
+/// Over-deep data must fail with a clean `Err`, not overflow the stack. A `Value` can be built
+/// without serde_json's parse-time cap (the native bindings), so `deep_merge` must guard itself.
+#[test]
+fn deep_merge_rejects_excessive_depth() {
+    let depth = super::MAX_MERGE_DEPTH + 50;
+    // Shared key `k` on both sides forces full-depth recursion; distinct leaves keep the trees
+    // unequal so the equality short-circuit never fires.
+    let mut a = nest(depth, Value::from_json_str(r#"{"a": 1}"#).unwrap());
+    let b = nest(depth, Value::from_json_str(r#"{"b": 2}"#).unwrap());
+
+    let err = a.deep_merge(b).unwrap_err();
+    assert!(
+        format!("{err}").contains("nesting depth"),
+        "expected a depth-limit error, got: {err}"
+    );
+}
+
+/// The pre-scan carries the same guard, so the default build rejects over-deep input up front
+/// (leaving the live document untouched) instead of overflowing during validation.
+#[cfg(not(feature = "allocator-memory-limits"))]
+#[test]
+fn check_mergeable_rejects_excessive_depth() {
+    let depth = super::MAX_MERGE_DEPTH + 50;
+    let a = nest(depth, Value::from_json_str(r#"{"a": 1}"#).unwrap());
+    let b = nest(depth, Value::from_json_str(r#"{"b": 2}"#).unwrap());
+
+    let err = a.check_mergeable(&b).unwrap_err();
+    assert!(
+        format!("{err}").contains("nesting depth"),
+        "expected a depth-limit error, got: {err}"
+    );
+}

tests/azure_policy/mod.rs

@@ -544,16 +544,14 @@ fn make_context(case: &TestCase) -> Result<Value> {
         let map = ctx.as_object_mut()?;
         // Only inject if the caller didn't already provide requestContext
         // in the context object, to avoid clobbering custom test setups.
-        map.entry(Value::from("requestContext")).or_insert(rc_val);
+        map.get_or_insert_with(Value::from("requestContext"), || rc_val);
     } else if let Some(ref api_ver) = case.api_version {
         let map = ctx.as_object_mut()?;
-        if let std::collections::btree_map::Entry::Vacant(e) =
-            map.entry(Value::from("requestContext"))
-        {
+        if !map.contains_key(&Value::from("requestContext")) {
             let mut req_ctx = Value::new_object();
             let rc_map = req_ctx.as_object_mut()?;
             rc_map.insert(Value::from("apiVersion"), Value::from(api_ver.clone()));
-            e.insert(req_ctx);
+            map.insert(Value::from("requestContext"), req_ctx);
         }
     }