Redact host paths and bound structured logging

.github/scripts/download_verified_artifact.ps1

@@ -0,0 +1,48 @@
+param(
+    [Parameter(Mandatory = $true)]
+    [uri] $Url,
+    [Parameter(Mandatory = $true)]
+    [string] $Output,
+    [Parameter(Mandatory = $true)]
+    [ValidatePattern('^[A-Fa-f0-9]{64}$')]
+    [string] $Sha256
+)
+
+$ErrorActionPreference = "Stop"
+if ($Url.Scheme -ne "https" -and -not ($Url.Scheme -eq "http" -and $Url.IsLoopback)) {
+    throw "Artifact URL must use HTTPS."
+}
+
+$maximumAttempts = 3
+$partialOutput = "$Output.$PID.partial"
+$downloaded = $false
+for ($attempt = 1; $attempt -le $maximumAttempts; $attempt++) {
+    try {
+        Invoke-WebRequest -Uri $Url -OutFile $partialOutput -TimeoutSec 120 | Out-Null
+        $downloaded = $true
+        break
+    }
+    catch {
+        Remove-Item -LiteralPath $partialOutput -Force -ErrorAction SilentlyContinue
+        $statusCode = 0
+        $response = $_.Exception.Response
+        if ($null -ne $response -and $null -ne $response.StatusCode) {
+            $statusCode = [int] $response.StatusCode
+        }
+        $transientStatus = $statusCode -eq 0 -or $statusCode -in @(408, 429) -or ($statusCode -ge 500 -and $statusCode -lt 600)
+        if (-not $transientStatus -or $attempt -eq $maximumAttempts) {
+            throw
+        }
+        Start-Sleep -Seconds $attempt
+    }
+}
+
+if (-not $downloaded) {
+    throw "Artifact download failed after $maximumAttempts attempts."
+}
+$actualSha256 = (Get-FileHash -Algorithm SHA256 -Path $partialOutput).Hash.ToLowerInvariant()
+if ($actualSha256 -ne $Sha256.ToLowerInvariant()) {
+    Remove-Item -LiteralPath $partialOutput -Force
+    throw "Artifact SHA-256 mismatch: expected $Sha256, got $actualSha256"
+}
+Move-Item -LiteralPath $partialOutput -Destination $Output -Force

.github/workflows/ci.yml

@@ -109,6 +109,7 @@ jobs:
   unit-tests:
     name: Unit Tests
     runs-on: ubuntu-latest
+    timeout-minutes: 30
     steps:
       - name: Checkout
         uses: actions/checkout@v5
@@ -203,7 +204,9 @@ jobs:
       - name: Run product smoke tests
         run: |
           python -W error -m pytest -v tests/product_smoke \
-            -m "stable" --no-cov --tb=short
+            -m "stable" --no-cov --tb=short \
+            --junitxml="$RUNNER_TEMP/stable-product-smoke.xml"
+          python scripts/check_critical_skips.py "$RUNNER_TEMP/stable-product-smoke.xml"
 
   cross-platform-tests:
     name: Tests (${{ matrix.os }}, Python ${{ matrix.python-version }})
@@ -230,7 +233,7 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y \
-            git build-essential gcc-arm-linux-gnueabihf qemu-user pkg-config wget cmake
+            git build-essential clang gcc-arm-linux-gnueabihf qemu-user pkg-config wget cmake
 
       - name: Install system dependencies (macOS)
         if: runner.os == 'macOS'
@@ -252,11 +255,7 @@ jobs:
           $archive = Join-Path $env:RUNNER_TEMP $archiveName
           $url = "https://github.com/radareorg/radare2/releases/download/$version/$archiveName"
           $expectedSha256 = "913e7d95e7458226a5e783240877f2c1d405398ab38ba176ce9b12e046e31f34"
-          Invoke-WebRequest -Uri $url -OutFile $archive
-          $actualSha256 = (Get-FileHash -Algorithm SHA256 -Path $archive).Hash.ToLowerInvariant()
-          if ($actualSha256 -ne $expectedSha256) {
-            throw "radare2 archive checksum mismatch: expected $expectedSha256, got $actualSha256"
-          }
+          & ".github/scripts/download_verified_artifact.ps1" -Url $url -Output $archive -Sha256 $expectedSha256
           $destination = Join-Path $env:RUNNER_TEMP "radare2-$version"
           Expand-Archive -Path $archive -DestinationPath $destination -Force
           $bin = Join-Path $destination "radare2-$version-w64\bin"
@@ -320,6 +319,7 @@ jobs:
           test_targets = [
               "tests/unit/test_circular_imports.py",
               "tests/unit/test_cli_basic_commands.py::test_cli_version_function",
+              "tests/unit/test_public_surface_contract.py",
               "tests/unit/test_pe_handler_parsing_contract.py",
               "tests/unit/test_pe_handler_struct_parse.py",
               "tests/unit/test_macho_parse_basic.py",
@@ -438,4 +438,5 @@ jobs:
           /tmp/r2morph-wheel-venv/bin/python -W error -m pytest \
             tests/unit/test_circular_imports.py \
             tests/unit/test_cli_basic_commands.py::test_cli_version_function \
+            tests/unit/test_public_surface_contract.py \
             --no-cov --tb=short

.github/workflows/differential-corpus.yml

@@ -11,6 +11,7 @@ concurrency:
       - ".github/workflows/differential-corpus.yml"
       - ".github/workflows/ci.yml"
       - ".github/workflows/adversarial-benchmark.yml"
+      - ".github/scripts/download_verified_artifact.ps1"
       - ".github/scripts/install_radare2.sh"
       - "docs/compatibility-corpus.md"
       - "docs/pass-maturity.md"
@@ -1612,11 +1613,7 @@ jobs:
           $archive = Join-Path $env:RUNNER_TEMP $archiveName
           $url = "https://github.com/radareorg/radare2/releases/download/$version/$archiveName"
           $expectedSha256 = "913e7d95e7458226a5e783240877f2c1d405398ab38ba176ce9b12e046e31f34"
-          Invoke-WebRequest -Uri $url -OutFile $archive
-          $actualSha256 = (Get-FileHash -Algorithm SHA256 -Path $archive).Hash.ToLowerInvariant()
-          if ($actualSha256 -ne $expectedSha256) {
-            throw "radare2 archive checksum mismatch: expected $expectedSha256, got $actualSha256"
-          }
+          & ".github/scripts/download_verified_artifact.ps1" -Url $url -Output $archive -Sha256 $expectedSha256
           $destination = Join-Path $env:RUNNER_TEMP "radare2-$version"
           Expand-Archive -Path $archive -DestinationPath $destination -Force
           $bin = Join-Path $destination "radare2-$version-w64\bin"

.github/workflows/release.yml

@@ -28,6 +28,11 @@ jobs:
         with:
           python-version: "3.12"
 
+      - name: Install release dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r requirements.txt
+
       - name: Extract version from tag
         id: get_version
         run: echo "version=${GITHUB_REF#refs/tags/v}" >> "$GITHUB_OUTPUT"

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## 0.4.0-alpha.3
 
+- Clarifies campaign-specific IDA and Triton availability and limits VM resistance
+  claims to the finite evidence actually measured.
 - Promotes instruction expansion and block reordering into the five-pass Tier 1
   Linux ELF x86-64 contract after native, corpus, composition, and analyzer
   evidence completed.
@@ -13,7 +15,7 @@
   documented wheel gate runs from a normal development installation.
 - Fixes release artifact download in the isolated PyPI publish job.
 - Publishes the same verified ELF x86-64 support contract as alpha.2.
-- Keeps the VM milestone blocked until external human review records signoff.
+- Independent human review approval is recorded for the VM resistance campaign.
 
 ## 0.4.0-alpha.2
 

README.md

@@ -103,8 +103,8 @@ x86-64 or equivalent CodeVirtualization support.
 
 The machine-readable contract is [`docs/support-matrix.json`](docs/support-matrix.json):
 the generated matrix currently records 22/22 evidenced cells for the official
-ELF x86-64 target (100.0% evidence) and 32/242 evidenced cells for non-official
-targets (13.22% evidence). It currently classifies 5 passes as Tier 1 and
+ELF x86-64 target (100.0% evidence) and 35/242 evidenced cells for non-official
+targets (14.46% evidence). It currently classifies 5 passes as Tier 1 and
 17 passes as experimental.
 The scheduled extended maturity smoke now records output-size,
 transform-duration, runtime-duration, and static metric summaries for every
@@ -174,13 +174,12 @@ transform duration, runtime duration, and static analyzer evidence while
 comparing exit code, stdout, stderr, created files, and declared observable
 effects across nine seed-derived inputs.
 
-The latest local adversarial refresh measures all 162 repository fixtures:
+The 2026-09-20 local adversarial campaign measures all 162 repository fixtures:
 `angr` completed 162/162 runs, while radare2, objdump, and the custom analyzer
 also completed all runs. Unicorn completed 155/162 runs; its seven unavailable
-rows are retained with their instruction-support reasons. Binary Ninja, IDA,
-and Triton remain explicit unavailable analyzer slots; Binary Ninja is on hold
-for the current milestone. Ghidra is now provisioned by the scheduled
-campaign. The bounded
+rows are retained with their instruction-support reasons. In that campaign,
+IDA and Triton were unavailable; these rows remain explicit and are not counted
+as completed analysis. Ghidra is provisioned by the scheduled campaign. The bounded
 report is
 [`docs/protection-adversarial-angr-local-2026-09-20.json`](docs/protection-adversarial-angr-local-2026-09-20.json).
 The earlier focused `angr` fixture remains available in
@@ -190,23 +189,38 @@ Binary Ninja is an explicit analyzer slot; the focused local
 fixture records `binary-ninja` as unavailable because `module 'binaryninja' is unavailable`.
 The current installed API is importable, but the headless benchmark reports
 `analyzer license is unavailable`; this remains an explicit evidence gap while
-the slot is on hold.
-The VM review packet still has `human_signoff: not-attested` and
-`release_decision: block-vm-milestone`; memory, direct/indirect calls, returns,
-flags, FP/SIMD, varargs/ABI, unwinding, TLS/signals, SSA, and liveness paths
-remain explicit review scope, and unsupported instructions must fail closed
-instead of producing partial protected functions. Unsupported-instruction
-diagnostics must include instruction address, mnemonic, type, size, bounded
-opcode preview, capability, and reason.
-VM ISA/opcode diversification, dispatcher/handler alternatives,
-superinstructions, anti-tamper, and progressive bytecode protection remain
-explicit human-review scope before the virtualizer milestone can be marked
-complete.
+the slot is on hold. Separately, the 159-sample evidence summarized in
+[`docs/independent-review.json`](docs/independent-review.json) records
+`159 samples, 159 Triton runs`, `159 samples, 318 completed analyses` for IDA,
+and `159 samples, 0 IDA errors`. These are campaign-specific
+results: availability in one campaign does not fill unavailable rows in another,
+and the all-named-analyzers 1.0 criterion remains open.
+The VM review packet records `human_signoff: attested` and
+`release_decision: signoff-recorded`. Human independent review approved the
+reviewed commit `a111f923df43a66826b71ab70db8a484532fbf3b`. The project owner
+confirms Marc Rivero Lopez's independence and specialist experience in binary
+rewriting and virtualization. He reviewed memory, direct/indirect calls,
+returns, flags, FP/SIMD, varargs/ABI,
+unwinding, TLS/signals, SSA, and liveness paths; unsupported instructions must
+fail closed instead of producing partial protected functions. Diagnostics
+include instruction address, mnemonic, type, size, bounded opcode preview,
+capability, and reason. The review covered VM ISA/opcode diversification,
+dispatcher/handler alternatives, superinstructions, anti-tamper, and progressive
+bytecode protection, fuzz properties, failure handling, and bounded report
+data. This approval applies to that commit; changes to the reviewed VM
+implementation require a new independent review.
 Current VM resistance artifacts cover 10 seeds with 255 handlers per seed,
 0 exact normalized cross-seed handler matches, 12 handler stride values, and
 target handler stride diversity recorded in
 [`docs/protection-handler-clustering.json`](docs/protection-handler-clustering.json)
 and [`docs/protection-bytecode-grammar.json`](docs/protection-bytecode-grammar.json).
+These finite seed-diversity measurements describe generated output variation;
+they do not measure generalized attack cost or establish resistance to
+professional reverse engineering. The independent analysis report explicitly
+limits its evidence to one fixture and one pass and states that it does not
+establish general resistance to professional reverse engineering. The report
+does not replace the required three-analyzer release campaign. Published
+resistance results remain an open 1.0 criterion.
 
 ---
 

docs/API.md

@@ -1,5 +1,11 @@
 # r2morph API Documentation
 
+> **Compatibility scope:** This guide includes module-level examples that are
+> useful for development but are not all stable imports. The intended 1.0
+> compatibility surface and deprecation policy are defined in
+> [stability-contract.md](stability-contract.md); currently the package-root
+> exports and documented CLI surface are the contracted entry points.
+
 ## Overview
 
 r2morph is a metamorphic mutation engine for binary transformation. The API is organized into several modules:

docs/compatibility-corpus.md

@@ -38,6 +38,22 @@ compiled case is an additional evidence point, not a support promotion.
 The Windows PE preview job also executes a real NOP-plus-register-substitution
 composition and validates the repaired PE checksum; it remains preview evidence
 and does not promote PE to the official support tier.
+On 2026-11-10, the same PE test module ran locally on macOS 27.2 ARM64 with
+MinGW-w64 GCC 16.2 and Wine 11.0: seven generated-binary mutation tests passed
+and the Windows-only fixture test was skipped. This confirms an additional
+Wine execution path; it is not native Windows evidence or a PE parity claim.
+The PE block-reordering regression now builds a separate control-flow fixture,
+executes the original and mutated images under native Windows or Wine for nine
+inputs, and validates the repaired PE checksum and integrity. A passing local
+run is preview evidence for this pass/target cell only; it does not promote
+PE x86-64 to parity with Linux ELF x86-64.
+On 2026-11-10, a connected licensed static-analysis service inspected the
+original and CodeVirtualization-protected `elf_x86_64` fixture. The bounded
+function and decompilation observations are in
+[`protection-independent-analysis-2026-11-10.json`](protection-independent-analysis-2026-11-10.json).
+The service did not expose its version; runtime equivalence was not measured.
+This one-fixture result does not establish general resistance or replace the
+three-analyzer release campaign.
 The cross-format matrix additionally exercises `basic` and `complex` CFG
 variants for those three ELF targets across the same four individual passes.
 The workflow pins the collected matrix at 73 cases so loss of a preview corpus

docs/human-review-signoff.json

@@ -0,0 +1,11 @@
+{
+  "schema_version": 1,
+  "reviewer": "Marc Rivero Lopez",
+  "review_date_utc": "2026-11-08",
+  "result": "APPROVE",
+  "reviewed_commit": "a111f923df43a66826b71ab70db8a484532fbf3b",
+  "independent": true,
+  "independence_confirmation": "confirmed by the project owner on 2026-11-09",
+  "specialist_expertise_confirmation": "binary rewriting and virtualization experience confirmed by the project owner on 2026-11-09",
+  "findings": []
+}