Bump Vampire/setup-wsl from 3 to 7

Bumps [Vampire/setup-wsl](https://github.com/vampire/setup-wsl) from 3 to 7. - [Release notes](https://github.com/vampire/setup-wsl/releases) - [Commits](https://github.com/vampire/setup-wsl/compare/v3...v7) --- updated-dependencies: - dependency-name: Vampire/setup-wsl dependency-version: '7' dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com>
Bump actions/checkout from 6 to 7 (#1771 )
2026-08-15 16:07:01 +02:00 · 2026-06-29 17:54:42 +00:00 · 2026-06-29 12:47:13 -05:00 · 2026-06-23 14:21:28 -04:00 · 2026-06-16 16:51:14 -04:00 · 2026-06-16 13:16:47 -04:00
119 changed files with 6524 additions and 1798 deletions
@@ -0,0 +1,113 @@
+name: Code-sign Windows binaries
+description: >
+  Sign every .exe under a given path in place via the DefinedNet code-signer
+  Lambda. If `role` or `bucket` is empty, logs a notice and skips signing so
+  forks and dev branches without AWS access still produce usable builds.
+
+inputs:
+  path:
+    description: "Directory whose .exe files should be signed in place"
+    required: true
+  role:
+    description: "IAM role ARN to assume via OIDC; empty disables signing"
+    required: false
+    default: ""
+  bucket:
+    description: "S3 staging bucket the code-signer Lambda reads from; empty disables signing"
+    required: false
+    default: ""
+  region:
+    description: "AWS region for the role and Lambda"
+    required: false
+    default: "us-east-2"
+  function-name:
+    description: "Code-signer Lambda function name"
+    required: false
+    default: "code-signer"
+  key-prefix:
+    description: "S3 key prefix the caller is authorized to write under"
+    required: false
+    default: "code-signing/slackhq/nebula"
+
+runs:
+  using: composite
+  steps:
+    - name: Skip notice
+      if: inputs.role == '' || inputs.bucket == ''
+      shell: sh
+      run: echo "::notice::code-signer role or bucket not set; skipping code signing."
+
+    - name: Configure AWS credentials
+      if: inputs.role != '' && inputs.bucket != ''
+      uses: aws-actions/configure-aws-credentials@v6
+      with:
+        role-to-assume: ${{ inputs.role }}
+        aws-region: ${{ inputs.region }}
+        # Default is 12 retries to ride out IAM trust-policy propagation; once
+        # the role is stable we want a real misconfiguration to fail fast.
+        retry-max-attempts: 5
+
+    - name: Sign .exe files
+      if: inputs.role != '' && inputs.bucket != ''
+      shell: sh
+      env:
+        SIGN_PATH: ${{ inputs.path }}
+        BUCKET: ${{ inputs.bucket }}
+        FUNCTION_NAME: ${{ inputs.function-name }}
+        KEY_PREFIX: ${{ inputs.key-prefix }}
+      run: |
+        set -eu
+        RUN="${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
+
+        find "$SIGN_PATH" -name '*.exe' -print | while read -r path
+        do
+            rel=${path#"$SIGN_PATH"/}
+            file=$(basename "$path")
+            name=${file%.exe}
+            prefix="${KEY_PREFIX}/${RUN}"
+            src="${prefix}/unsigned/${rel}"
+            dst="${prefix}/signed/${rel}"
+
+            echo "::group::Sign ${rel}"
+            echo "Uploading unsigned to s3://${BUCKET}/${src}"
+            aws s3 cp --no-progress "$path" "s3://${BUCKET}/${src}" >/dev/null
+
+            echo "Invoking ${FUNCTION_NAME} Lambda"
+            payload=$(jq -nc \
+                --arg s "$src" \
+                --arg d "$dst" \
+                --arg p "$name" \
+                '{source_key: $s, dest_key: $d, program_name: $p}')
+            meta=$(aws lambda invoke \
+                --function-name "$FUNCTION_NAME" \
+                --cli-binary-format raw-in-base64-out \
+                --payload "$payload" \
+                --output json \
+                /tmp/sign-resp.json)
+            if echo "$meta" | jq -e '.FunctionError != null' >/dev/null
+            then
+                echo "::endgroup::"
+                echo "::error::code-signer Lambda failed for ${rel}"
+                cat /tmp/sign-resp.json >&2
+                exit 1
+            fi
+
+            echo "Downloading signed back to ${path}"
+            aws s3 cp --no-progress "s3://${BUCKET}/${dst}" "$path" >/dev/null
+
+            aws s3 rm "s3://${BUCKET}/${src}" >/dev/null 2>&1 || true
+            aws s3 rm "s3://${BUCKET}/${dst}" >/dev/null 2>&1 || true
+
+            # Sanity-check the bytes we got back actually carry an Authenticode
+            # signature that this machine can validate end to end.
+            status=$(powershell -NoProfile -Command "(Get-AuthenticodeSignature -FilePath '$path').Status" | tr -d '\r')
+            if [ "$status" != "Valid" ]
+            then
+                echo "::endgroup::"
+                echo "::error::${rel} signature status: ${status} (expected Valid)"
+                exit 1
+            fi
+
+            echo "Signed ${rel} (sha256=$(jq -r '.sha256' /tmp/sign-resp.json), status=${status})"
+            echo "::endgroup::"
+        done
@@ -1,34 +0,0 @@
-name: gofmt
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-    paths:
-      - '.github/workflows/gofmt.yml'
-      - '**.go'
-jobs:
-
-  gofmt:
-    name: Run gofmt
-    runs-on: ubuntu-latest
-    steps:
-
-    - uses: actions/checkout@v6
-
-    - uses: actions/setup-go@v6
-      with:
-        go-version: '1.25'
-        check-latest: true
-
-    - name: Install goimports
-      run: |
-        go install golang.org/x/tools/cmd/goimports@latest
-
-    - name: gofmt
-      run: |
-        if [ "$(find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -l)" ]
-        then
-          find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -d
-          exit 1
-        fi
@@ -10,7 +10,7 @@ jobs:
    name: Build Linux/BSD All
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7

      - uses: actions/setup-go@v6
        with:
@@ -24,7 +24,7 @@ jobs:
          mv build/*.tar.gz release

      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: linux-latest
          path: release
@@ -32,8 +32,11 @@ jobs:
  build-windows:
    name: Build Windows
    runs-on: windows-latest
+    permissions:
+      id-token: write
+      contents: read
    steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7

      - uses: actions/setup-go@v6
        with:
@@ -54,8 +57,15 @@ jobs:
          mkdir build\dist\windows
          mv dist\windows\wintun build\dist\windows\

+      - name: Code-sign
+        uses: ./.github/actions/code-sign
+        with:
+          path: build
+          role: ${{ secrets.DEFINED_CODE_SIGNER_ROLE }}
+          bucket: ${{ secrets.DEFINED_CODE_SIGNER_BUCKET }}
+
      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: windows-latest
          path: build
@@ -66,7 +76,7 @@ jobs:
      HAS_SIGNING_CREDS: ${{ secrets.AC_USERNAME != '' }}
    runs-on: macos-latest
    steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7

      - uses: actions/setup-go@v6
        with:
@@ -75,7 +85,7 @@ jobs:

      - name: Import certificates
        if: env.HAS_SIGNING_CREDS == 'true'
-        uses: Apple-Actions/import-codesign-certs@v6
+        uses: Apple-Actions/import-codesign-certs@v7
        with:
          p12-file-base64: ${{ secrets.APPLE_DEVELOPER_CERTIFICATE_P12_BASE64 }}
          p12-password: ${{ secrets.APPLE_DEVELOPER_CERTIFICATE_PASSWORD }}
@@ -104,7 +114,7 @@ jobs:
          fi

      - name: Upload artifacts
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
        with:
          name: darwin-latest
          path: ./release/*
@@ -124,25 +134,25 @@ jobs:
      # be overwritten
      - name: Checkout code
        if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7

      - name: Download artifacts
        if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          name: linux-latest
          path: artifacts

      - name: Login to Docker Hub
        if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          username: ${{ vars.DOCKERHUB_USERNAME }}
          password: ${{ secrets.DOCKERHUB_TOKEN }}

      - name: Set up Docker Buildx
        if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4

      - name: Build and push images
        if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
@@ -153,17 +163,20 @@ jobs:
          mkdir -p build/linux-{amd64,arm64}
          tar -zxvf artifacts/nebula-linux-amd64.tar.gz -C build/linux-amd64/
          tar -zxvf artifacts/nebula-linux-arm64.tar.gz -C build/linux-arm64/
-          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 --tag "${DOCKER_IMAGE_REPO}:${DOCKER_IMAGE_TAG}" --tag "${DOCKER_IMAGE_REPO}:${GITHUB_REF#refs/tags/v}"
+          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 \
+            --build-arg VERSION="${GITHUB_REF#refs/tags/v}" \
+            --build-arg REVISION="${GITHUB_SHA}" \
+            --tag "${DOCKER_IMAGE_REPO}:${DOCKER_IMAGE_TAG}" --tag "${DOCKER_IMAGE_REPO}:${GITHUB_REF#refs/tags/v}"

  release:
    name: Create and Upload Release
    needs: [build-linux, build-darwin, build-windows]
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7

      - name: Download artifacts
-        uses: actions/download-artifact@v7
+        uses: actions/download-artifact@v8
        with:
          path: artifacts

@@ -14,15 +14,23 @@ on:
      - 'go.sum'
 jobs:

-  smoke-extra:
+  smoke-extra-libvirt:
    if: github.ref == 'refs/heads/master' || contains(github.event.pull_request.labels.*.name, 'smoke-test-extra')
-    name: Run extra smoke tests
+    name: ${{ matrix.target }}
    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        target:
+          - freebsd-amd64
+          - openbsd-amd64
+          - netbsd-amd64
+          - linux-amd64-ipv6disable
    env:
      VAGRANT_DEFAULT_PROVIDER: libvirt
    steps:

-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7

    - uses: actions/setup-go@v6
      with:
@@ -40,28 +48,85 @@ jobs:
        sudo chmod 666 /var/run/libvirt/libvirt-sock
        vagrant plugin install vagrant-libvirt

-    - name: freebsd-amd64
-      run: make smoke-vagrant/freebsd-amd64
+    - name: ${{ matrix.target }}
+      run: make smoke-vagrant/${{ matrix.target }}

-    - name: openbsd-amd64
-      run: make smoke-vagrant/openbsd-amd64
+    timeout-minutes: 30

-    - name: netbsd-amd64
-      run: make smoke-vagrant/netbsd-amd64
+  # linux-386 needs VirtualBox, which conflicts with KVM/libvirt -- isolated job.
+  smoke-extra-virtualbox:
+    if: github.ref == 'refs/heads/master' || contains(github.event.pull_request.labels.*.name, 'smoke-test-extra')
+    name: linux-386
+    runs-on: ubuntu-latest
+    env:
+      VAGRANT_DEFAULT_PROVIDER: virtualbox
+    steps:

-    - name: linux-amd64-ipv6disable
-      run: make smoke-vagrant/linux-amd64-ipv6disable
+    - uses: actions/checkout@v7

-    # linux-386 runs last because it requires disabling KVM to use VirtualBox,
-    # which prevents libvirt (used by the other tests) from working after this point.
-    - name: install virtualbox for i386 test
+    - uses: actions/setup-go@v6
+      with:
+        go-version: '1.25'
+        check-latest: true
+
+    - name: add hashicorp source
+      run: wget -O- https://apt.releases.hashicorp.com/gpg | gpg --dearmor | sudo tee /usr/share/keyrings/hashicorp-archive-keyring.gpg && echo "deb [signed-by=/usr/share/keyrings/hashicorp-archive-keyring.gpg] https://apt.releases.hashicorp.com $(lsb_release -cs) main" | sudo tee /etc/apt/sources.list.d/hashicorp.list
+
+    - name: install vagrant and virtualbox
      run: |
-        sudo apt-get install -y virtualbox
+        sudo apt-get update && sudo apt-get install -y vagrant virtualbox
        sudo rmmod kvm_amd kvm_intel kvm 2>/dev/null || true

    - name: linux-386
-      env:
-        VAGRANT_DEFAULT_PROVIDER: virtualbox
      run: make smoke-vagrant/linux-386

    timeout-minutes: 30
+
+  smoke-windows:
+    if: github.ref == 'refs/heads/master' || contains(github.event.pull_request.labels.*.name, 'smoke-test-extra')
+    name: Run windows smoke test
+    runs-on: windows-latest
+    steps:
+
+    - uses: actions/checkout@v7
+
+    - uses: actions/setup-go@v6
+      with:
+        go-version: '1.25'
+        check-latest: true
+
+    # WSL2 + Ubuntu so the smoke can run a real linux peer with its own
+    # netns. iputils-ping is needed for the in-WSL ping check. WSL1 has no
+    # real kernel and would lack /dev/net/tun, so we have to force WSL2.
+    - uses: Vampire/setup-wsl@v7
+      with:
+        distribution: Ubuntu-24.04
+        additional-packages: iputils-ping iproute2
+
+    # Vampire/setup-wsl provisions WSL1 even when the WSL2 platform is present.
+    # Convert the distro to WSL2 explicitly before we try to use /dev/net/tun.
+    - name: convert distro to WSL2
+      shell: pwsh
+      run: |
+        wsl --set-version Ubuntu-24.04 2
+        wsl --shutdown
+        wsl --list --verbose
+
+    - name: build windows nebula
+      run: make bin-windows
+
+    - name: build linux nebula for WSL
+      shell: bash
+      env:
+        GOOS: linux
+        GOARCH: amd64
+      run: |
+        mkdir -p build/linux-amd64
+        go build -o build/linux-amd64/nebula ./cmd/nebula
+
+    - name: run smoke-windows
+      shell: pwsh
+      working-directory: ./.github/workflows/smoke
+      run: ./smoke-windows.ps1
+
+    timeout-minutes: 15
@@ -18,7 +18,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:

-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7

    - uses: actions/setup-go@v6
      with:
@@ -36,6 +36,14 @@ jobs:
      working-directory: ./.github/workflows/smoke
      run: ./smoke.sh

+    - name: setup docker image ipv6
+      working-directory: ./.github/workflows/smoke
+      run: SMOKE_OVERLAY_IPV6=1 ./build.sh
+
+    - name: run smoke ipv6
+      working-directory: ./.github/workflows/smoke
+      run: SMOKE_OVERLAY_IPV6=1 ./smoke.sh
+
    - name: setup relay docker image
      working-directory: ./.github/workflows/smoke
      run: ./build-relay.sh
@@ -5,6 +5,19 @@ set -e -x
 rm -rf ./build
 mkdir ./build

+if [ "$SMOKE_OVERLAY_IPV6" ]
+then
+    LIGHTHOUSE_NIP="fd00:4242:0:0:0:ffff:c0a8:6401"
+    HOST2_NIP="fd00:4242:0:0:0:ffff:c0a8:6402"
+    HOST3_NIP="fd00:4242:0:0:0:ffff:c0a8:6403"
+    HOST4_NIP="fd00:4242:0:0:0:ffff:c0a8:6404"
+else
+    LIGHTHOUSE_NIP="192.168.100.1"
+    HOST2_NIP="192.168.100.2"
+    HOST3_NIP="192.168.100.3"
+    HOST4_NIP="192.168.100.4"
+fi
+
 # Smoke containers run on a dedicated docker network whose subnet is allocated
 # at smoke time, not known at build time. Configs are written with TEST-NET-3
 # placeholder IPs (RFC 5737) and smoke.sh / smoke-vagrant.sh / smoke-relay.sh
@@ -31,24 +44,24 @@ LIGHTHOUSE_IP="203.0.113.2"
        ../genconfig.sh >lighthouse1.yml

    HOST="host2" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
        ../genconfig.sh >host2.yml

    HOST="host3" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
        INBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
        ../genconfig.sh >host3.yml

    HOST="host4" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
        OUTBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
        ../genconfig.sh >host4.yml

    ../../../../nebula-cert ca -curve "${CURVE:-25519}" -name "Smoke Test"
-    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "192.168.100.1/24"
-    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "192.168.100.2/24"
-    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "192.168.100.3/24"
-    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "192.168.100.4/24"
+    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "$LIGHTHOUSE_NIP/24"
+    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "$HOST2_NIP/24"
+    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "$HOST3_NIP/24"
+    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "$HOST4_NIP/24"
 )

 docker build -t "nebula:${NAME:-smoke}" .
@@ -0,0 +1,272 @@
+#!/usr/bin/env pwsh
+# Windows smoke test for the nebula tun + UDP + NLM code paths.
+#
+# Topology:
+#   - lighthouse runs natively on the Windows host (wintun + windows UDP)
+#   - peer runs inside WSL2 (Linux build of nebula, /dev/net/tun)
+#
+# WSL2 gives us a real netns boundary so the loopback fast-path on Windows
+# does not short-circuit the overlay -- when WSL pings the lighthouse VPN IP,
+# Linux has no idea that IP is local to the Windows host, so the packet is
+# forced through nebula. Same in reverse.
+
+$ErrorActionPreference = 'Stop'
+
+# wsl.exe emits UTF-16 LE by default which PowerShell reads as bytes, mangling
+# every captured string. WSL_UTF8 makes wsl.exe emit UTF-8 instead.
+$env:WSL_UTF8 = '1'
+
+$RepoRoot = Resolve-Path "$PSScriptRoot\..\..\.."
+$Nebula = Join-Path $RepoRoot 'nebula.exe'
+$NebulaCert = Join-Path $RepoRoot 'nebula-cert.exe'
+$NebulaLinux = Join-Path $RepoRoot 'build\linux-amd64\nebula'
+
+if (-not (Test-Path $Nebula)) { throw "missing $Nebula; run 'make bin-windows' first" }
+if (-not (Test-Path $NebulaCert)) { throw "missing $NebulaCert; run 'make bin-windows' first" }
+if (-not (Test-Path $NebulaLinux)) { throw "missing $NebulaLinux; build the linux nebula first" }
+
+# Matches the distro installed by Vampire/setup-wsl in smoke-extra.yml.
+$Distro = 'Ubuntu-24.04'
+$listed = (wsl --list --quiet 2>$null) -join "`n"
+if ($listed -notmatch [regex]::Escape($Distro)) {
+    throw "WSL distro $Distro not registered. Got: $listed"
+}
+Write-Host "Using WSL distro: $Distro"
+
+# Windows host as seen from inside WSL: WSL's default-route gateway. We extract
+# it with a regex rather than awk fields so PowerShell does not eat any '$N'
+# tokens, and tabs/double-spaces in `ip route` output do not confuse a cut.
+$ipCmd = 'ip route show default | grep -oE "([0-9]+\.){3}[0-9]+" | head -1'
+$WindowsIp = (wsl -d $Distro -- bash -c $ipCmd).Trim()
+if (-not $WindowsIp) { throw "could not determine Windows host IP from WSL" }
+Write-Host "Windows host IP from WSL: $WindowsIp"
+
+$WorkDir = Join-Path $env:TEMP 'nebula-smoke-windows'
+if (Test-Path $WorkDir) { Remove-Item -Recurse -Force $WorkDir }
+New-Item -ItemType Directory -Path $WorkDir | Out-Null
+
+$WslDir = '/tmp/nebula-smoke'
+wsl -d $Distro -- bash -c "rm -rf $WslDir && mkdir -p $WslDir" | Out-Null
+
+$DevName = 'nebula-smoke'
+$Ip1 = '192.168.241.1'
+$Ip2 = '192.168.241.2'
+$Port = 4242
+
+& $NebulaCert ca -name 'smoke-ca' -out-crt "$WorkDir\ca.crt" -out-key "$WorkDir\ca.key"
+if ($LASTEXITCODE -ne 0) { throw "nebula-cert ca failed (exit $LASTEXITCODE)" }
+
+& $NebulaCert sign -name 'lighthouse' -networks "$Ip1/24" -ca-crt "$WorkDir\ca.crt" -ca-key "$WorkDir\ca.key" -out-crt "$WorkDir\lighthouse.crt" -out-key "$WorkDir\lighthouse.key"
+if ($LASTEXITCODE -ne 0) { throw "nebula-cert sign lighthouse failed (exit $LASTEXITCODE)" }
+
+& $NebulaCert sign -name 'peer' -networks "$Ip2/24" -ca-crt "$WorkDir\ca.crt" -ca-key "$WorkDir\ca.key" -out-crt "$WorkDir\peer.crt" -out-key "$WorkDir\peer.key"
+if ($LASTEXITCODE -ne 0) { throw "nebula-cert sign peer failed (exit $LASTEXITCODE)" }
+
+# Windows lighthouse config.
+@"
+pki:
+  ca: $WorkDir\ca.crt
+  cert: $WorkDir\lighthouse.crt
+  key: $WorkDir\lighthouse.key
+static_host_map: {}
+lighthouse:
+  am_lighthouse: true
+  interval: 60
+  hosts: []
+listen:
+  host: 0.0.0.0
+  port: $Port
+tun:
+  disabled: false
+  dev: $DevName
+  drop_local_broadcast: false
+  drop_multicast: false
+  tx_queue: 500
+  mtu: 1300
+  network_category: private
+logging:
+  level: info
+  format: text
+firewall:
+  outbound_action: drop
+  inbound_action: drop
+  conntrack:
+    tcp_timeout: 12m
+    udp_timeout: 3m
+    default_timeout: 10m
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+"@ | Out-File -FilePath "$WorkDir\lighthouse.yml" -Encoding utf8
+
+# WSL peer config (paths are POSIX, deliberately).
+@"
+pki:
+  ca: $WslDir/ca.crt
+  cert: $WslDir/peer.crt
+  key: $WslDir/peer.key
+static_host_map:
+  "${Ip1}": ["${WindowsIp}:$Port"]
+lighthouse:
+  am_lighthouse: false
+  interval: 60
+  hosts:
+    - "${Ip1}"
+listen:
+  host: 0.0.0.0
+  port: 0
+tun:
+  disabled: false
+  dev: nebula1
+  drop_local_broadcast: false
+  drop_multicast: false
+  tx_queue: 500
+  mtu: 1300
+logging:
+  level: info
+  format: text
+firewall:
+  outbound_action: drop
+  inbound_action: drop
+  conntrack:
+    tcp_timeout: 12m
+    udp_timeout: 3m
+    default_timeout: 10m
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+"@ | Out-File -FilePath "$WorkDir\peer.yml" -Encoding utf8
+
+# Stage WSL artifacts. Convert Windows paths to WSL paths ourselves rather than
+# calling `wslpath`, because PowerShell's argument-passing to external EXEs
+# strips backslashes from path arguments in ways that are hard to escape around.
+function ConvertTo-WslPath {
+    param([string]$WindowsPath)
+    if ($WindowsPath -notmatch '^([A-Za-z]):\\(.*)$') {
+        throw "cannot convert path to WSL: $WindowsPath"
+    }
+    return "/mnt/$($matches[1].ToLower())/$($matches[2].Replace('\','/'))"
+}
+
+$WslWorkDir = ConvertTo-WslPath $WorkDir
+$WslNebulaPath = ConvertTo-WslPath $NebulaLinux
+wsl -d $Distro -- bash -c "cp '$WslWorkDir/ca.crt' '$WslWorkDir/peer.crt' '$WslWorkDir/peer.key' '$WslWorkDir/peer.yml' $WslDir/ && cp '$WslNebulaPath' $WslDir/nebula && chmod +x $WslDir/nebula"
+
+# Make sure WSL has tun support and /dev/net/tun is usable before starting
+# nebula. Diagnostics first so a fail here points at the real problem (e.g.
+# WSL1 distros do not have a real kernel and will not have tun).
+Write-Host '=== WSL diagnostic ==='
+wsl --version 2>&1 | Out-Host
+wsl --list --verbose 2>&1 | Out-Host
+wsl -d $Distro -u root -- uname -a | Out-Host
+wsl -d $Distro -u root -- bash -c "modprobe tun 2>&1 || true; mkdir -p /dev/net; [ -c /dev/net/tun ] || mknod /dev/net/tun c 10 200; chmod 600 /dev/net/tun; ls -l /dev/net/tun"
+if ($LASTEXITCODE -ne 0) { throw "failed to prepare /dev/net/tun in WSL (TUN support missing?)" }
+
+# Deliberately no New-NetFirewallRule calls here -- nebula's windows_bypass_wdf
+# feature is supposed to install WFP permit filters that let inbound traffic
+# through Windows Defender Firewall on its own. If this smoke regresses, that
+# feature regressed.
+
+$lhOut = Join-Path $WorkDir 'lighthouse.out.log'
+$lhErr = Join-Path $WorkDir 'lighthouse.err.log'
+$lhProc = Start-Process -FilePath $Nebula -ArgumentList @('-config', "$WorkDir\lighthouse.yml") `
+    -PassThru -NoNewWindow `
+    -RedirectStandardOutput $lhOut `
+    -RedirectStandardError $lhErr
+
+# Run nebula in WSL as root with no sudo + no shell wrapper. PowerShell's
+# Start-Process arg quoting mangles `bash -c "..."` strings that contain
+# spaces/redirections, so we skip bash entirely and let Start-Process do the
+# stdout/stderr capture itself.
+$peerOut = Join-Path $WorkDir 'peer.out.log'
+$peerErr = Join-Path $WorkDir 'peer.err.log'
+$peerProc = Start-Process -FilePath 'wsl' `
+    -ArgumentList @('-d', $Distro, '-u', 'root', '--', "$WslDir/nebula", '-config', "$WslDir/peer.yml") `
+    -PassThru -NoNewWindow `
+    -RedirectStandardOutput $peerOut `
+    -RedirectStandardError $peerErr
+
+function Wait-Until {
+    param([scriptblock]$Predicate, [int]$TimeoutSec, [string]$What)
+    $deadline = (Get-Date).AddSeconds($TimeoutSec)
+    while ((Get-Date) -lt $deadline) {
+        if (& $Predicate) { return }
+        Start-Sleep -Milliseconds 500
+    }
+    throw "timed out waiting for: $What"
+}
+
+try {
+    Wait-Until -TimeoutSec 30 -What "windows wintun adapter $DevName with NetworkCategory=Private" -Predicate {
+        if ($lhProc.HasExited) { throw "lighthouse exited (code $($lhProc.ExitCode)) before tun was ready" }
+        $p = Get-NetConnectionProfile -InterfaceAlias $DevName -ErrorAction SilentlyContinue
+        $p -and ("$($p.NetworkCategory)" -ieq 'Private')
+    }
+    Write-Host "OK: $DevName NetworkCategory=Private"
+
+    Wait-Until -TimeoutSec 30 -What "WSL nebula1 with $Ip2" -Predicate {
+        if ($peerProc.HasExited) { throw "peer exited (code $($peerProc.ExitCode)) before tun was ready" }
+        $r = wsl -d $Distro -u root -- bash -c "ip -o addr show nebula1 2>/dev/null | grep -q 'inet $Ip2' && echo yes"
+        ("$r").Trim() -eq 'yes'
+    }
+    Write-Host "OK: WSL nebula1 has $Ip2"
+
+    Wait-Until -TimeoutSec 30 -What "ping from WSL peer to windows lighthouse ($Ip1)" -Predicate {
+        if ($peerProc.HasExited) { throw "peer exited (code $($peerProc.ExitCode)) before ping succeeded" }
+        $r = wsl -d $Distro -u root -- bash -c "ping -c1 -W1 $Ip1 >/dev/null 2>&1 && echo OK"
+        ("$r").Trim() -eq 'OK'
+    }
+    Write-Host "OK: WSL peer -> windows lighthouse"
+
+    Wait-Until -TimeoutSec 30 -What "ping from windows lighthouse to WSL peer ($Ip2)" -Predicate {
+        $null = & ping.exe -n 1 -w 1000 $Ip2
+        $LASTEXITCODE -eq 0
+    }
+    Write-Host "OK: windows lighthouse -> WSL peer"
+
+    Write-Host ''
+    Write-Host 'All smoke checks passed.'
+}
+catch {
+    Write-Host ''
+    Write-Host '=== lighthouse stdout ==='
+    Get-Content $lhOut -ErrorAction SilentlyContinue | Out-Host
+    Write-Host '=== lighthouse stderr ==='
+    Get-Content $lhErr -ErrorAction SilentlyContinue | Out-Host
+    Write-Host '=== peer stdout ==='
+    Get-Content $peerOut -ErrorAction SilentlyContinue | Out-Host
+    Write-Host '=== peer stderr ==='
+    Get-Content $peerErr -ErrorAction SilentlyContinue | Out-Host
+    Write-Host '=== nebula WFP filters ==='
+    # Dump nebula-installed filters so we can verify they got registered with
+    # the conditions we expect.
+    $wfpDump = Join-Path $WorkDir 'wfp.xml'
+    netsh wfp show filters file=$wfpDump 2>&1 | Out-Null
+    if (Test-Path $wfpDump) {
+        Select-String -Path $wfpDump -Pattern 'Nebula' -Context 0,80 -ErrorAction SilentlyContinue | Out-Host
+    }
+    throw
+}
+finally {
+    if (-not $lhProc.HasExited) {
+        Stop-Process -Id $lhProc.Id -Force -ErrorAction SilentlyContinue
+        $lhProc.WaitForExit(5000) | Out-Null
+    }
+    wsl -d $Distro -u root -- bash -c "pkill -f $WslDir/nebula 2>/dev/null; true" | Out-Null
+    # pkill returns 1 when no match and wsl propagates that; the smoke is done
+    # so we don't want it to leak into the script's exit code.
+    $global:LASTEXITCODE = 0
+    if ($peerProc -and -not $peerProc.HasExited) {
+        Stop-Process -Id $peerProc.Id -Force -ErrorAction SilentlyContinue
+    }
+}
@@ -47,6 +47,19 @@ HOST2_IP="$PREFIX.3"
 HOST3_IP="$PREFIX.4"
 HOST4_IP="$PREFIX.5"

+if [ "$SMOKE_OVERLAY_IPV6" ]
+then
+    LIGHTHOUSE_NIP="fd00:4242:0:0:0:ffff:c0a8:6401"
+    HOST2_NIP="fd00:4242:0:0:0:ffff:c0a8:6402"
+    HOST3_NIP="fd00:4242:0:0:0:ffff:c0a8:6403"
+    HOST4_NIP="fd00:4242:0:0:0:ffff:c0a8:6404"
+else
+    LIGHTHOUSE_NIP="192.168.100.1"
+    HOST2_NIP="192.168.100.2"
+    HOST3_NIP="192.168.100.3"
+    HOST4_NIP="192.168.100.4"
+fi
+
 # Sed the placeholder TEST-NET-3 IPs in the host configs to the real ones.
 # build/lighthouse1.yml has no IPs to rewrite so it's skipped.
 for f in build/host2.yml build/host3.yml build/host4.yml; do
@@ -80,28 +93,28 @@ docker exec host3 tcpdump -i eth0 -q -w - -U 2>logs/host3.outside.log >logs/host
 docker exec host4 tcpdump -i tun0 -q -w - -U 2>logs/host4.inside.log >logs/host4.inside.pcap &
 docker exec host4 tcpdump -i eth0 -q -w - -U 2>logs/host4.outside.log >logs/host4.outside.pcap &

-docker exec host2 ncat -nklv 0.0.0.0 2000 &
-docker exec host3 ncat -nklv 0.0.0.0 2000 &
-docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -nkluv 0.0.0.0 4000 &
-docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 0.0.0.0 3000 &
-docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 0.0.0.0 3000 &
+docker exec host2 ncat -nklv 2000 &
+docker exec host3 ncat -nklv 2000 &
+docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -nkluv 4000 &
+docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 3000 &
+docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 3000 &

 set +x
 echo
 echo " *** Testing ping from lighthouse1"
 echo
 set -x
-docker exec lighthouse1 ping -c1 192.168.100.2
-docker exec lighthouse1 ping -c1 192.168.100.3
+docker exec lighthouse1 ping -c1 $HOST2_NIP
+docker exec lighthouse1 ping -c1 $HOST3_NIP

 set +x
 echo
 echo " *** Testing ping from host2"
 echo
 set -x
-docker exec host2 ping -c1 192.168.100.1
+docker exec host2 ping -c1 $LIGHTHOUSE_NIP
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ping -c1 192.168.100.3 -w5 || exit 1
+! docker exec host2 ping -c1 $HOST3_NIP -w5 || exit 1

 set +x
 echo
@@ -109,34 +122,34 @@ echo " *** Testing ncat from host2"
 echo
 set -x
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ncat -nzv -w5 192.168.100.3 2000 || exit 1
-! docker exec host2 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
+! docker exec host2 ncat -nzv -w5 $HOST3_NIP 2000 || exit 1
+! docker exec host2 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1

 set +x
 echo
 echo " *** Testing ping from host3"
 echo
 set -x
-docker exec host3 ping -c1 192.168.100.1
-docker exec host3 ping -c1 192.168.100.2
+docker exec host3 ping -c1 $LIGHTHOUSE_NIP
+docker exec host3 ping -c1 $HOST2_NIP

 set +x
 echo
 echo " *** Testing ncat from host3"
 echo
 set -x
-docker exec host3 ncat -nzv -w5 192.168.100.2 2000
-docker exec host3 ncat -nzuv -w5 192.168.100.2 3000 | grep -q host2
+docker exec host3 ncat -nzv -w5 $HOST2_NIP 2000
+docker exec host3 ncat -nzuv -w5 $HOST2_NIP 3000 | grep -q host2

 set +x
 echo
 echo " *** Testing ping from host4"
 echo
 set -x
-docker exec host4 ping -c1 192.168.100.1
+docker exec host4 ping -c1 $LIGHTHOUSE_NIP
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ping -c1 192.168.100.2 -w5 || exit 1
-! docker exec host4 ping -c1 192.168.100.3 -w5 || exit 1
+! docker exec host4 ping -c1 $HOST2_NIP -w5 || exit 1
+! docker exec host4 ping -c1 $HOST3_NIP -w5 || exit 1

 set +x
 echo
@@ -144,10 +157,10 @@ echo " *** Testing ncat from host4"
 echo
 set -x
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ncat -nzv -w5 192.168.100.2 2000 || exit 1
-! docker exec host4 ncat -nzv -w5 192.168.100.3 2000 || exit 1
-! docker exec host4 ncat -nzuv -w5 192.168.100.2 3000 | grep -q host2 || exit 1
-! docker exec host4 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
+! docker exec host4 ncat -nzv -w5 $HOST2_NIP 2000 || exit 1
+! docker exec host4 ncat -nzv -w5 $HOST3_NIP 2000 || exit 1
+! docker exec host4 ncat -nzuv -w5 $HOST2_NIP 3000 | grep -q host2 || exit 1
+! docker exec host4 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1

 set +x
 echo
@@ -159,7 +172,7 @@ set -x
 # cannot initiate UDP to host2. Once host2 initiates a flow to host4:4000,
 # conntrack must let host4's listener reply on that flow. If it doesn't,
 # the echo back from host4 never reaches host2.
-docker exec host2 sh -c "(/usr/bin/echo host2; sleep 2) | ncat -nuv 192.168.100.4 4000" | grep -q helloagainfromhost4
+docker exec host2 sh -c "(/usr/bin/echo host2; sleep 2) | ncat -nuv $HOST4_NIP 4000" | grep -q helloagainfromhost4

 docker exec host4 sh -c 'kill 1'
 docker exec host3 sh -c 'kill 1'
@@ -1,7 +1,7 @@
 # -*- mode: ruby -*-
 # vi: set ft=ruby :
 Vagrant.configure("2") do |config|
-  config.vm.box = "generic/netbsd9"
+  config.vm.box = "DefinedNet/netbsd10"

  config.vm.synced_folder "../build", "/nebula", type: "rsync"
 end
@@ -13,20 +13,28 @@ on:
      - 'go.sum'
 jobs:

-  test-linux:
-    name: Build all and test on ubuntu-linux
+  static:
+    name: Static checks
    runs-on: ubuntu-latest
    steps:

-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7

    - uses: actions/setup-go@v6
      with:
        go-version: '1.25'
        check-latest: true

-    - name: Build
-      run: make all
+    - name: Install goimports
+      run: go install golang.org/x/tools/cmd/goimports@latest
+
+    - name: gofmt
+      run: |
+        if [ "$(find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -l)" ]
+        then
+          find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -d
+          exit 1
+        fi

    - name: Vet
      run: make vet
@@ -36,97 +44,107 @@ jobs:
      with:
        version: v2.5

-    - name: Test
-      run: make test
-
-    - name: End 2 end
-      run: make e2evv
-
-    - name: Build test mobile
-      run: make build-test-mobile
-
-    - uses: actions/upload-artifact@v6
-      with:
-        name: e2e packet flow linux-latest
-        path: e2e/mermaid/linux-latest
-        if-no-files-found: warn
-
-  test-linux-boringcrypto:
-    name: Build and test on linux with boringcrypto
-    runs-on: ubuntu-latest
-    steps:
-
-    - uses: actions/checkout@v6
-
-    - uses: actions/setup-go@v6
-      with:
-        go-version: '1.25'
-        check-latest: true
-
-    - name: Build
-      run: make bin-boringcrypto
-
-    - name: Test
-      run: make test-boringcrypto
-
-    - name: End 2 end
-      run: make e2e GOEXPERIMENT=boringcrypto CGO_ENABLED=1 TEST_ENV="TEST_LOGS=1" TEST_FLAGS="-v -ldflags -checklinkname=0"
-
-  test-linux-pkcs11:
-    name: Build and test on linux with pkcs11
-    runs-on: ubuntu-latest
-    steps:
-
-    - uses: actions/checkout@v6
-
-    - uses: actions/setup-go@v6
-      with:
-        go-version: '1.25'
-        check-latest: true
-
-    - name: Build
-      run: make bin-pkcs11
-
-    - name: Test
-      run: make test-pkcs11
-
  test:
-    name: Build and test on ${{ matrix.os }}
+    name: Test ${{ matrix.name }}
    runs-on: ${{ matrix.os }}
    strategy:
+      fail-fast: false
      matrix:
-        os: [windows-latest, macos-latest]
+        include:
+        - name: linux
+          os: ubuntu-latest
+          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
+          test-cmd: make test
+          e2e-cmd: make e2evv
+        - name: linux-boringcrypto
+          os: ubuntu-latest
+          build-cmd: make bin-boringcrypto
+          test-cmd: make test-boringcrypto
+          e2e-cmd: make e2e GOEXPERIMENT=boringcrypto CGO_ENABLED=1 TEST_ENV="TEST_LOGS=1" TEST_FLAGS="-v -ldflags -checklinkname=0"
+        - name: linux-pkcs11
+          os: ubuntu-latest
+          build-cmd: make bin-pkcs11
+          test-cmd: make test-pkcs11
+          e2e-cmd: ''
+        - name: macos
+          os: macos-latest
+          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
+          test-cmd: make test
+          e2e-cmd: make e2evv
+        - name: windows
+          os: windows-latest
+          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
+          test-cmd: make test
+          e2e-cmd: make e2evv
    steps:

-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7

    - uses: actions/setup-go@v6
      with:
        go-version: '1.25'
        check-latest: true

-    - name: Build nebula
-      run: go build ./cmd/nebula
+    - name: Build
+      run: ${{ matrix.build-cmd }}

-    - name: Build nebula-cert
-      run: go build ./cmd/nebula-cert
-
-    - name: Vet
-      run: make vet
-
-    - name: golangci-lint
-      uses: golangci/golangci-lint-action@v9
-      with:
-        version: v2.5
+    - name: Cross-build darwin-amd64
+      if: matrix.name == 'macos'
+      run: GOARCH=amd64 go build -o /tmp/nebula-amd64 ./cmd/nebula && GOARCH=amd64 go build -o /tmp/nebula-cert-amd64 ./cmd/nebula-cert

    - name: Test
-      run: make test
+      run: ${{ matrix.test-cmd }}

    - name: End 2 end
-      run: make e2evv
+      if: matrix.e2e-cmd != ''
+      run: ${{ matrix.e2e-cmd }}

-    - uses: actions/upload-artifact@v6
+    - uses: actions/upload-artifact@v7
+      if: matrix.e2e-cmd != '' && always()
      with:
-        name: e2e packet flow ${{ matrix.os }}
-        path: e2e/mermaid/${{ matrix.os }}
+        name: e2e packet flow ${{ matrix.name }}
+        path: e2e/mermaid/
        if-no-files-found: warn
+
+  cross-build:
+    name: Cross-build ${{ matrix.name }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+        - {name: linux-arm,   make-target: all-cross-linux-arm}
+        - {name: linux-mips,  make-target: all-cross-linux-mips}
+        - {name: linux-other, make-target: all-cross-linux-other}
+        - {name: freebsd,     make-target: all-freebsd}
+        - {name: openbsd,     make-target: all-openbsd}
+        - {name: netbsd,      make-target: all-netbsd}
+        - {name: windows,     make-target: all-cross-windows}
+        - {name: mobile,      make-target: build-test-mobile}
+    steps:
+
+    - uses: actions/checkout@v7
+
+    - uses: actions/setup-go@v6
+      with:
+        go-version: '1.25'
+        check-latest: true
+
+    - name: Build ${{ matrix.name }}
+      run: make -j"$(nproc)" ${{ matrix.make-target }}
+
+  finish:
+    name: CI status
+    if: always()
+    needs: [static, test, cross-build]
+    runs-on: ubuntu-latest
+    steps:
+
+    - name: Fail if any upstream job failed
+      if: contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled')
+      run: |
+        echo "upstream results: ${{ toJSON(needs) }}"
+        exit 1
+
+    - name: All upstream jobs passed
+      run: echo "ok"
@@ -60,6 +60,18 @@ ALL = $(ALL_LINUX) \
 	windows-amd64 \
 	windows-arm64

+# Cross-build shards used by .github/workflows/test.yml — same as ALL_*
+# but with the arch that has a native CI runner removed, so the cross-build
+# job is not duplicating coverage the native test jobs already give.
+ALL_CROSS_LINUX = $(filter-out linux-amd64,$(ALL_LINUX))
+
+# ALL_CROSS_LINUX further split into family sub-shards so each can run on
+# its own CI runner in parallel. Union of the three must equal
+# ALL_CROSS_LINUX; adding a new linux arch goes into the matching family.
+ALL_CROSS_LINUX_ARM   = linux-arm-5 linux-arm-6 linux-arm-7 linux-arm64
+ALL_CROSS_LINUX_MIPS  = linux-mips linux-mipsle linux-mips64 linux-mips64le linux-mips-softfloat
+ALL_CROSS_LINUX_OTHER = linux-386 linux-ppc64le linux-riscv64 linux-loong64
+
 e2e:
 	$(TEST_ENV) go test -tags=e2e_testing -count=1 $(TEST_FLAGS) ./e2e

@@ -82,6 +94,35 @@ DOCKER_BIN = build/linux-amd64/nebula build/linux-amd64/nebula-cert

 all: $(ALL:%=build/%/nebula) $(ALL:%=build/%/nebula-cert)

+all-linux: $(ALL_LINUX:%=build/%/nebula) $(ALL_LINUX:%=build/%/nebula-cert)
+
+all-freebsd: $(ALL_FREEBSD:%=build/%/nebula) $(ALL_FREEBSD:%=build/%/nebula-cert)
+
+all-openbsd: $(ALL_OPENBSD:%=build/%/nebula) $(ALL_OPENBSD:%=build/%/nebula-cert)
+
+all-netbsd: $(ALL_NETBSD:%=build/%/nebula) $(ALL_NETBSD:%=build/%/nebula-cert)
+
+all-darwin: build/darwin-amd64/nebula build/darwin-amd64/nebula-cert build/darwin-arm64/nebula build/darwin-arm64/nebula-cert
+
+all-windows: build/windows-amd64/nebula.exe build/windows-amd64/nebula-cert.exe build/windows-arm64/nebula.exe build/windows-arm64/nebula-cert.exe
+
+# CI cross-build shards. darwin-arm64 is covered by the native macos-latest
+# job; windows-amd64 is covered by the native windows-latest job; both are
+# omitted here to avoid building them a second time. darwin-amd64 stays in
+# all-cross-darwin because intel mac is only a labeled/master-time native
+# job, so PRs still need cross-build coverage for it.
+all-cross-linux: $(ALL_CROSS_LINUX:%=build/%/nebula) $(ALL_CROSS_LINUX:%=build/%/nebula-cert)
+
+all-cross-linux-arm:   $(ALL_CROSS_LINUX_ARM:%=build/%/nebula)   $(ALL_CROSS_LINUX_ARM:%=build/%/nebula-cert)
+
+all-cross-linux-mips:  $(ALL_CROSS_LINUX_MIPS:%=build/%/nebula)  $(ALL_CROSS_LINUX_MIPS:%=build/%/nebula-cert)
+
+all-cross-linux-other: $(ALL_CROSS_LINUX_OTHER:%=build/%/nebula) $(ALL_CROSS_LINUX_OTHER:%=build/%/nebula-cert)
+
+all-cross-darwin: build/darwin-amd64/nebula build/darwin-amd64/nebula-cert
+
+all-cross-windows: build/windows-arm64/nebula.exe build/windows-arm64/nebula-cert.exe
+
 docker: docker/linux-$(shell go env GOARCH)

 release: $(ALL:%=build/nebula-%.tar.gz)
@@ -227,6 +268,9 @@ smoke-relay-docker: bin-docker
 	cd .github/workflows/smoke/ && ./build-relay.sh
 	cd .github/workflows/smoke/ && ./smoke-relay.sh

+smoke-docker-ipv6: export SMOKE_OVERLAY_IPV6 = 1
+smoke-docker-ipv6: smoke-docker
+
 smoke-docker-race: BUILD_ARGS = -race
 smoke-docker-race: CGO_ENABLED = 1
 smoke-docker-race: smoke-docker
@@ -236,5 +280,5 @@ smoke-vagrant/%: bin-docker build/%/nebula
 	cd .github/workflows/smoke/ && ./smoke-vagrant.sh $*

 .FORCE:
-.PHONY: bench bench-cpu bench-cpu-long bin build-test-mobile e2e e2ev e2evv e2evvv e2evvvv proto release service smoke-docker smoke-docker-race test test-cov-html smoke-vagrant/%
+.PHONY: all all-linux all-freebsd all-openbsd all-netbsd all-darwin all-windows all-cross-linux all-cross-linux-arm all-cross-linux-mips all-cross-linux-other all-cross-darwin all-cross-windows bench bench-cpu bench-cpu-long bin build-test-mobile e2e e2ev e2evv e2evvv e2evvvv proto release service smoke-docker smoke-docker-race test test-cov-html smoke-vagrant/%
 .DEFAULT_GOAL := bin
@@ -2,24 +2,42 @@ package nebula

 import (
 	"context"
+	"fmt"
 	"log/slog"
+	"math"
+	mathbits "math/bits"

 	"github.com/rcrowley/go-metrics"
 )

+const bitsPerWord = 64
+
+// Bits is a sliding-window anti-replay tracker. The window is stored as a
+// circular bitmap packed into uint64 words (8x denser than a []bool), so a
+// length-N window costs N/8 bytes. length must be a power of two.
 type Bits struct {
 	length             uint64
+	lengthMask         uint64
 	current            uint64
-	bits               []bool
+	bits               []uint64
 	lostCounter        metrics.Counter
 	dupeCounter        metrics.Counter
 	outOfWindowCounter metrics.Counter
 }

-func NewBits(bits uint64) *Bits {
+func NewBits(length uint64) *Bits {
+	if length == 0 || length&(length-1) != 0 {
+		panic(fmt.Sprintf("Bits length must be a power of two, got %d", length))
+	}
+
+	nWords := length / bitsPerWord
+	if nWords == 0 {
+		nWords = 1
+	}
 	b := &Bits{
-		length:             bits,
-		bits:               make([]bool, bits, bits),
+		length:             length,
+		lengthMask:         length - 1,
+		bits:               make([]uint64, nWords),
 		current:            0,
 		lostCounter:        metrics.GetOrRegisterCounter("network.packets.lost", nil),
 		dupeCounter:        metrics.GetOrRegisterCounter("network.packets.duplicate", nil),
@@ -27,71 +45,194 @@ func NewBits(bits uint64) *Bits {
 	}

 	// There is no counter value 0, mark it to avoid counting a lost packet later.
-	b.bits[0] = true
-	b.current = 0
+	b.bits[0] = 1
 	return b
 }

+func (b *Bits) get(i uint64) bool {
+	pos := i & b.lengthMask
+	//bit-shifting by 6 because i is a bit index, not a u64 index, and we need to find the u64 without bit in it
+	return b.bits[pos>>6]&(uint64(1)<<(pos&63)) != 0
+}
+
+func (b *Bits) set(i uint64) {
+	pos := i & b.lengthMask
+	b.bits[pos>>6] |= uint64(1) << (pos & 63)
+}
+
+// clearRange clears `count` bits starting at circular position `startPos`
+// (already masked to [0, length)) and returns how many of them were set
+// before the clear. count must be in [1, length].
+func (b *Bits) clearRange(startPos, count uint64) uint64 {
+	wasSet := uint64(0)
+	if count >= b.length {
+		for _, w := range b.bits {
+			wasSet += uint64(mathbits.OnesCount64(w))
+		}
+		clear(b.bits)
+		return wasSet
+	}
+
+	pos := startPos
+	remaining := count
+
+	// handle the potential partial word before pos becomes u64 aligned
+	word := pos >> 6
+	bit := pos & 63
+	take := uint64(64) - bit
+	if take > remaining {
+		take = remaining
+	}
+	if take > b.length-pos {
+		take = b.length - pos
+	}
+	var mask uint64
+	if take == 64 {
+		mask = math.MaxUint64
+	} else {
+		mask = ((uint64(1) << take) - 1) << bit
+	}
+	wasSet += uint64(mathbits.OnesCount64(b.bits[word] & mask))
+	b.bits[word] &^= mask
+	remaining -= take
+	pos = (pos + take) & b.lengthMask
+
+	// Clear whole words, keeping track of the number of set bits
+	for remaining >= 64 {
+		word = pos >> 6
+		wasSet += uint64(mathbits.OnesCount64(b.bits[word]))
+		b.bits[word] = 0
+		remaining -= 64
+		pos = (pos + 64) & b.lengthMask
+	}
+
+	// Clear the remaining partial word
+	if remaining > 0 {
+		word = pos >> 6
+		mask = (uint64(1) << remaining) - 1
+		wasSet += uint64(mathbits.OnesCount64(b.bits[word] & mask))
+		b.bits[word] &^= mask
+	}
+
+	return wasSet
+}
+
+func (b *Bits) strictlyWithinWindow(i uint64) bool {
+	// Handle the case where the window hasn't slid yet. This avoids u64 underflow.
+	inWarmup := b.current < b.length
+	if i < b.length && inWarmup {
+		return true
+	}
+
+	// Next, if the packet is in-window, see if we've seen it before
+	if i > b.current-b.length {
+		return true
+	}
+	return false //not within window!
+}
+
+// Check returns true if i is within (or way out in front of) the window, and not a replay
 func (b *Bits) Check(l *slog.Logger, i uint64) bool {
 	// If i is the next number, return true.
 	if i > b.current {
 		return true
 	}

-	// If i is within the window, check if it's been set already.
-	if i > b.current-b.length || i < b.length && b.current < b.length {
-		return !b.bits[i%b.length]
+	if b.strictlyWithinWindow(i) {
+		return !b.get(i)
 	}

 	// Not within the window
 	if l.Enabled(context.Background(), slog.LevelDebug) {
-		l.Debug("rejected a packet (top)",
-			"current", b.current,
-			"incoming", i,
-		)
+		l.Debug("rejected a packet (top)", "current", b.current, "incoming", i)
 	}
 	return false
 }

+// Update has three branches:
+//   - i == b.current+1: fast path; advance the cursor by one and lose-count
+//     the slot we just stomped (only past warmup; see the i > b.length guard
+//     below).
+//   - i  >  b.current+1: jump path; clear all slots between current and i
+//     (or up to a full window's worth, whichever is smaller) via clearRange,
+//     then mark i. Two arms here: a warmup arm that handles the very first
+//     window before the cursor has slid, and a steady-state arm that treats
+//     every cleared empty slot as a lost packet.
+//   - i  <= b.current: in-window check for duplicates; out-of-window otherwise.
+//
+// NewBits seeds bits[0]=1 so counter 0 looks "received" — Update never
+// clears that marker during warmup (clearRange skips position 0 when
+// startPos=1), and once b.current >= b.length the marker is no longer
+// consulted. The marker prevents a fictitious "lost" hit on the first real
+// counter.
 func (b *Bits) Update(l *slog.Logger, i uint64) bool {
-	// If i is the next number, return true and update current.
+	// Fast path: i is the next expected counter. Split out so the function
+	// stays small and avoids paying for the slow paths' slog argument-build
+	// stack frame on every call. The bit read/test/write is inlined to
+	// touch the backing word once.
 	if i == b.current+1 {
-		// Check if the oldest bit was lost since we are shifting the window by 1 and occupying it with this counter
-		// The very first window can only be tracked as lost once we are on the 2nd window or greater
-		if b.bits[i%b.length] == false && i > b.length {
+		pos := i & b.lengthMask
+		word := pos >> 6
+		mask := uint64(1) << (pos & 63)
+		w := b.bits[word]
+		if i > b.length && w&mask == 0 {
 			b.lostCounter.Inc(1)
 		}
-		b.bits[i%b.length] = true
+		b.bits[word] = w | mask
 		b.current = i
 		return true
 	}
+	return b.updateSlow(l, i)
+}

+// updateSlow handles jumps, in-window backfill, dupes, and out-of-window.
+func (b *Bits) updateSlow(l *slog.Logger, i uint64) bool {
 	// If i is a jump, adjust the window, record lost, update current, and return true
 	if i > b.current {
-		lost := int64(0)
-		// Zero out the bits between the current and the new counter value, limited by the window size,
-		// since the window is shifting
-		for n := b.current + 1; n <= min(i, b.current+b.length); n++ {
-			if b.bits[n%b.length] == false && n > b.length {
-				lost++
+		end := i
+		if end > b.current+b.length {
+			end = b.current + b.length
+		}
+		count := end - b.current
+		startPos := (b.current + 1) & b.lengthMask
+
+		var lost int64
+		if b.current >= b.length {
+			// Steady state: every cleared slot is past warmup, so any unset
+			// bit we evict is a lost packet from the previous cycle.
+			wasSet := b.clearRange(startPos, count)
+			lost = int64(count) - int64(wasSet)
+		} else {
+			// Warmup (the very first window). Some cleared slots represent
+			// packets <= length where eviction is not "lost" in the usual
+			// sense. This branch is taken at most once per connection so we
+			// don't bother optimizing it.
+			for n := b.current + 1; n <= end; n++ {
+				if !b.get(n) && n > b.length {
+					lost++
+				}
 			}
-			b.bits[n%b.length] = false
+			b.clearRange(startPos, count)
 		}

-		// Only record any skipped packets as a result of the window moving further than the window length
-		// Any loss within the new window will be accounted for in future calls
-		lost += max(0, int64(i-b.current-b.length))
+		// Anything past the new window can never be backfilled, so it's lost.
+		if i > b.current+b.length {
+			lost += int64(i - b.current - b.length)
+		}
 		b.lostCounter.Inc(lost)

-		b.bits[i%b.length] = true
+		b.set(i)
 		b.current = i
 		return true
 	}

-	// If i is within the current window but below the current counter,
-	// Check to see if it's a duplicate
-	if i > b.current-b.length || i < b.length && b.current < b.length {
-		if b.current == i || b.bits[i%b.length] == true {
+	// If i is within the current window but below the current counter, check to see if it's a duplicate
+	if b.strictlyWithinWindow(i) {
+		pos := i & b.lengthMask
+		word := pos >> 6
+		mask := uint64(1) << (pos & 63)
+		w := b.bits[word]
+		if b.current == i || w&mask != 0 {
 			if l.Enabled(context.Background(), slog.LevelDebug) {
 				l.Debug("Receive window",
 					"accepted", false,
@@ -104,7 +245,7 @@ func (b *Bits) Update(l *slog.Logger, i uint64) bool {
 			return false
 		}

-		b.bits[i%b.length] = true
+		b.bits[word] = w | mask
 		return true
 	}

@@ -7,61 +7,79 @@ import (
 	"github.com/stretchr/testify/assert"
 )

+// snapshot returns the bitmap as a []bool of length b.length, for readable
+// test assertions against the now-packed []uint64 storage.
+func (b *Bits) snapshot() []bool {
+	out := make([]bool, b.length)
+	for i := uint64(0); i < b.length; i++ {
+		out[i] = b.get(i)
+	}
+	return out
+}
+
+func TestBitsRequiresPowerOfTwo(t *testing.T) {
+	assert.Panics(t, func() { NewBits(10) })
+	assert.Panics(t, func() { NewBits(0) })
+	assert.NotPanics(t, func() { NewBits(1) })
+	assert.NotPanics(t, func() { NewBits(16) })
+	assert.NotPanics(t, func() { NewBits(1024) })
+	assert.NotPanics(t, func() { NewBits(16384) })
+}
+
 func TestBits(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
-
-	// make sure it is the right size
-	assert.Len(t, b.bits, 10)
+	b := NewBits(16)
+	assert.EqualValues(t, 16, b.length)

 	// This is initialized to zero - receive one. This should work.
 	assert.True(t, b.Check(l, 1))
 	assert.True(t, b.Update(l, 1))
 	assert.EqualValues(t, 1, b.current)
-	g := []bool{true, true, false, false, false, false, false, false, false, false}
-	assert.Equal(t, g, b.bits)
+	g := []bool{true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false}
+	assert.Equal(t, g, b.snapshot())

 	// Receive two
 	assert.True(t, b.Check(l, 2))
 	assert.True(t, b.Update(l, 2))
 	assert.EqualValues(t, 2, b.current)
-	g = []bool{true, true, true, false, false, false, false, false, false, false}
-	assert.Equal(t, g, b.bits)
+	g = []bool{true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false}
+	assert.Equal(t, g, b.snapshot())

 	// Receive two again - it will fail
 	assert.False(t, b.Check(l, 2))
 	assert.False(t, b.Update(l, 2))
 	assert.EqualValues(t, 2, b.current)

-	// Jump ahead to 15, which should clear everything and set the 6th element
-	assert.True(t, b.Check(l, 15))
-	assert.True(t, b.Update(l, 15))
-	assert.EqualValues(t, 15, b.current)
-	g = []bool{false, false, false, false, false, true, false, false, false, false}
-	assert.Equal(t, g, b.bits)
+	// Jump ahead to 25, which clears the window and sets slot 25%16 = 9.
+	assert.True(t, b.Check(l, 25))
+	assert.True(t, b.Update(l, 25))
+	assert.EqualValues(t, 25, b.current)
+	g = []bool{false, false, false, false, false, false, false, false, false, true, false, false, false, false, false, false}
+	assert.Equal(t, g, b.snapshot())

-	// Mark 14, which is allowed because it is in the window
-	assert.True(t, b.Check(l, 14))
-	assert.True(t, b.Update(l, 14))
-	assert.EqualValues(t, 15, b.current)
-	g = []bool{false, false, false, false, true, true, false, false, false, false}
-	assert.Equal(t, g, b.bits)
+	// Mark 24, which is in window (current 25, length 16, window covers [10,25]).
+	assert.True(t, b.Check(l, 24))
+	assert.True(t, b.Update(l, 24))
+	assert.EqualValues(t, 25, b.current)
+	g = []bool{false, false, false, false, false, false, false, false, true, true, false, false, false, false, false, false}
+	assert.Equal(t, g, b.snapshot())

-	// Mark 5, which is not allowed because it is not in the window
+	// Mark 5, not allowed because 5 <= current-length (25-16=9).
 	assert.False(t, b.Check(l, 5))
 	assert.False(t, b.Update(l, 5))
-	assert.EqualValues(t, 15, b.current)
-	g = []bool{false, false, false, false, true, true, false, false, false, false}
-	assert.Equal(t, g, b.bits)
+	assert.EqualValues(t, 25, b.current)
+	g = []bool{false, false, false, false, false, false, false, false, true, true, false, false, false, false, false, false}
+	assert.Equal(t, g, b.snapshot())

-	// make sure we handle wrapping around once to the current position
-	b = NewBits(10)
+	// Make sure we handle wrapping around once to the same slot. With
+	// length=16, packets 1 and 17 share slot 1.
+	b = NewBits(16)
 	assert.True(t, b.Update(l, 1))
-	assert.True(t, b.Update(l, 11))
-	assert.Equal(t, []bool{false, true, false, false, false, false, false, false, false, false}, b.bits)
+	assert.True(t, b.Update(l, 17))
+	assert.Equal(t, []bool{false, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false}, b.snapshot())

 	// Walk through a few windows in order
-	b = NewBits(10)
+	b = NewBits(16)
 	for i := uint64(1); i <= 100; i++ {
 		assert.True(t, b.Check(l, i), "Error while checking %v", i)
 		assert.True(t, b.Update(l, i), "Error while updating %v", i)
@@ -72,24 +90,31 @@ func TestBits(t *testing.T) {

 func TestBitsLargeJumps(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
+
+	// length=16. Update(55) from current=0:
+	//   warmup, per-bit loop sees no n>16 with unset bits (slot 0 was set by
+	//   NewBits and gets re-evaluated when n=16; n=16 is not strictly > 16),
+	//   so the loop contributes 0. The jump exceeds the window so we record
+	//   55 - 0 - 16 = 39 packets fell out the back.
+	b := NewBits(16)
 	b.lostCounter.Clear()
+	assert.True(t, b.Update(l, 55))
+	assert.Equal(t, int64(39), b.lostCounter.Count())

-	b = NewBits(10)
-	b.lostCounter.Clear()
-	assert.True(t, b.Update(l, 55)) // We saw packet 55 and can still track 45,46,47,48,49,50,51,52,53,54
-	assert.Equal(t, int64(45), b.lostCounter.Count())
+	// Update(100): clears 16 slots starting at slot 56%16=8. Only slot 7 (for
+	// packet 55) was set, so 16 - 1 = 15 evicted slots had unset bits.
+	// Plus 100 - 55 - 16 = 29 packets fell past the window. Total 44.
+	assert.True(t, b.Update(l, 100))
+	assert.Equal(t, int64(39+44), b.lostCounter.Count())

-	assert.True(t, b.Update(l, 100)) // We saw packet 55 and 100 and can still track 90,91,92,93,94,95,96,97,98,99
-	assert.Equal(t, int64(89), b.lostCounter.Count())
-
-	assert.True(t, b.Update(l, 200)) // We saw packet 55, 100, and 200 and can still track 190,191,192,193,194,195,196,197,198,199
-	assert.Equal(t, int64(188), b.lostCounter.Count())
+	// Update(200): same shape: 16 - 1 = 15 evicted unset, plus 200 - 100 - 16 = 84 past window. Total 99.
+	assert.True(t, b.Update(l, 200))
+	assert.Equal(t, int64(39+44+99), b.lostCounter.Count())
 }

 func TestBitsDupeCounter(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
+	b := NewBits(16)
 	b.lostCounter.Clear()
 	b.dupeCounter.Clear()
 	b.outOfWindowCounter.Clear()
@@ -114,120 +139,117 @@ func TestBitsDupeCounter(t *testing.T) {

 func TestBitsOutOfWindowCounter(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
+	b := NewBits(16)
 	b.lostCounter.Clear()
 	b.dupeCounter.Clear()
 	b.outOfWindowCounter.Clear()

+	// Jump to 20 (warmup branch + 4 past-window packets).
 	assert.True(t, b.Update(l, 20))
 	assert.Equal(t, int64(0), b.outOfWindowCounter.Count())

-	assert.True(t, b.Update(l, 21))
-	assert.True(t, b.Update(l, 22))
-	assert.True(t, b.Update(l, 23))
-	assert.True(t, b.Update(l, 24))
-	assert.True(t, b.Update(l, 25))
-	assert.True(t, b.Update(l, 26))
-	assert.True(t, b.Update(l, 27))
-	assert.True(t, b.Update(l, 28))
-	assert.True(t, b.Update(l, 29))
+	// 9 single-step advances, each evicts a slot whose bit was cleared during
+	// the jump above and whose value was never seen, so each contributes 1
+	// to lostCounter.
+	for n := uint64(21); n <= 29; n++ {
+		assert.True(t, b.Update(l, n))
+	}
 	assert.Equal(t, int64(0), b.outOfWindowCounter.Count())

+	// 0 is below current-length (29-16=13) so it falls outside the window.
 	assert.False(t, b.Update(l, 0))
 	assert.Equal(t, int64(1), b.outOfWindowCounter.Count())

-	assert.Equal(t, int64(19), b.lostCounter.Count()) // packet 0 wasn't lost
+	// 4 from the Update(20) jump + 9 from 21..29.
+	assert.Equal(t, int64(13), b.lostCounter.Count())
 	assert.Equal(t, int64(0), b.dupeCounter.Count())
 	assert.Equal(t, int64(1), b.outOfWindowCounter.Count())
 }

 func TestBitsLostCounter(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
+	b := NewBits(16)
 	b.lostCounter.Clear()
 	b.dupeCounter.Clear()
 	b.outOfWindowCounter.Clear()

-	assert.True(t, b.Update(l, 20))
-	assert.True(t, b.Update(l, 21))
-	assert.True(t, b.Update(l, 22))
-	assert.True(t, b.Update(l, 23))
-	assert.True(t, b.Update(l, 24))
-	assert.True(t, b.Update(l, 25))
-	assert.True(t, b.Update(l, 26))
-	assert.True(t, b.Update(l, 27))
-	assert.True(t, b.Update(l, 28))
-	assert.True(t, b.Update(l, 29))
-	assert.Equal(t, int64(19), b.lostCounter.Count()) // packet 0 wasn't lost
+	// Walk 20..29 like the original, just with a bigger window. Same
+	// reasoning as TestBitsOutOfWindowCounter: 4 past-window from Update(20),
+	// then 9 more from the unit advances.
+	for n := uint64(20); n <= 29; n++ {
+		assert.True(t, b.Update(l, n))
+	}
+	assert.Equal(t, int64(13), b.lostCounter.Count())
 	assert.Equal(t, int64(0), b.dupeCounter.Count())
 	assert.Equal(t, int64(0), b.outOfWindowCounter.Count())

-	b = NewBits(10)
+	b = NewBits(16)
 	b.lostCounter.Clear()
 	b.dupeCounter.Clear()
 	b.outOfWindowCounter.Clear()

-	assert.True(t, b.Update(l, 9))
-	assert.Equal(t, int64(0), b.lostCounter.Count())
-	// 10 will set 0 index, 0 was already set, no lost packets
-	assert.True(t, b.Update(l, 10))
-	assert.Equal(t, int64(0), b.lostCounter.Count())
-	// 11 will set 1 index, 1 was missed, we should see 1 packet lost
-	assert.True(t, b.Update(l, 11))
-	assert.Equal(t, int64(1), b.lostCounter.Count())
-	// Now let's fill in the window, should end up with 8 lost packets
-	assert.True(t, b.Update(l, 12))
-	assert.True(t, b.Update(l, 13))
-	assert.True(t, b.Update(l, 14))
+	// Update(15) clears the warmup window (no lost), sets slot 15.
 	assert.True(t, b.Update(l, 15))
+	assert.Equal(t, int64(0), b.lostCounter.Count())
+
+	// Update(16): slot 0 was already set (NewBits seeded it), and 16 is not
+	// strictly > length, so nothing is recorded as lost.
 	assert.True(t, b.Update(l, 16))
+	assert.Equal(t, int64(0), b.lostCounter.Count())
+
+	// Update(17): we jumped straight from 0 to 15, so slot 1 was cleared
+	// (and never re-set). 17 > 16 is past warmup, so packet 1 is recorded lost.
 	assert.True(t, b.Update(l, 17))
-	assert.True(t, b.Update(l, 18))
-	assert.True(t, b.Update(l, 19))
-	assert.Equal(t, int64(8), b.lostCounter.Count())
+	assert.Equal(t, int64(1), b.lostCounter.Count())

-	// Jump ahead by a window size
-	assert.True(t, b.Update(l, 29))
-	assert.Equal(t, int64(8), b.lostCounter.Count())
-	// Now lets walk ahead normally through the window, the missed packets should fill in
-	assert.True(t, b.Update(l, 30))
-	assert.True(t, b.Update(l, 31))
-	assert.True(t, b.Update(l, 32))
-	assert.True(t, b.Update(l, 33))
-	assert.True(t, b.Update(l, 34))
-	assert.True(t, b.Update(l, 35))
-	assert.True(t, b.Update(l, 36))
-	assert.True(t, b.Update(l, 37))
-	assert.True(t, b.Update(l, 38))
-	// 39 packets tracked, 22 seen, 17 lost
-	assert.Equal(t, int64(17), b.lostCounter.Count())
+	// Fill in 18..30 in single steps. Each i evicts slot i%16. Slots 2..14
+	// were all cleared during Update(15), and we never re-set any of them,
+	// so each i in 18..30 is a fresh lost packet — 13 more.
+	for n := uint64(18); n <= 30; n++ {
+		assert.True(t, b.Update(l, n))
+	}
+	assert.Equal(t, int64(14), b.lostCounter.Count())

-	// Jump ahead by 2 windows, should have recording 1 full window missing
-	assert.True(t, b.Update(l, 58))
-	assert.Equal(t, int64(27), b.lostCounter.Count())
-	// Now lets walk ahead normally through the window, the missed packets should fill in from this window
-	assert.True(t, b.Update(l, 59))
-	assert.True(t, b.Update(l, 60))
-	assert.True(t, b.Update(l, 61))
-	assert.True(t, b.Update(l, 62))
-	assert.True(t, b.Update(l, 63))
-	assert.True(t, b.Update(l, 64))
-	assert.True(t, b.Update(l, 65))
-	assert.True(t, b.Update(l, 66))
-	assert.True(t, b.Update(l, 67))
-	// 68 packets tracked, 32 seen, 36 missed
-	assert.Equal(t, int64(36), b.lostCounter.Count())
+	// Jump ahead by exactly one window size.
+	assert.True(t, b.Update(l, 46))
+	// end = min(46, 30+16) = 46, count = 16, all slots cleared. Before the
+	// jump every slot 0..15 had been set (Update(15), (16), (17), 18..30),
+	// so wasSet=16 and 46 == current+length means no past-window slack:
+	// lost contribution = 0.
+	assert.Equal(t, int64(14), b.lostCounter.Count())
+
+	// Walk 47..55. The Update(46) jump cleared every slot, so only slot 14
+	// (for packet 46) is set when we start. Each subsequent unit step lands
+	// on a slot that was cleared and is past warmup, so it counts as lost.
+	// 9 more = 23.
+	for n := uint64(47); n <= 55; n++ {
+		assert.True(t, b.Update(l, n))
+	}
+	assert.Equal(t, int64(23), b.lostCounter.Count())
+
+	// Jump ahead by two windows: clears the window plus past-window loss.
+	assert.True(t, b.Update(l, 87))
+	// current=55, length=16. end = min(87, 71) = 71. count=16, all slots
+	// cleared. Slots set before the clear are slots 14,15,0..7 (10 total).
+	// Lost from clear = 16 - 10 = 6. Past window: 87 - 55 - 16 = 16. +22.
+	assert.Equal(t, int64(45), b.lostCounter.Count())
 	assert.Equal(t, int64(0), b.dupeCounter.Count())
 	assert.Equal(t, int64(0), b.outOfWindowCounter.Count())
 }

 func TestBitsLostCounterIssue1(t *testing.T) {
 	l := test.NewLogger()
-	b := NewBits(10)
+	b := NewBits(16)
 	b.lostCounter.Clear()
 	b.dupeCounter.Clear()
 	b.outOfWindowCounter.Clear()

+	// Receive 4, backfill 1, then 9, 2, 3, 5, 6, 7 (skip 8), 10, 11, 14.
+	// Then jump to 25 — slot 25%16=9 is being evicted, but it had been set
+	// (we received packet 9), so no spurious lost increment. The original
+	// regression was about double-counting a missing packet when its slot
+	// got cleared on a jump. With the jump path now using clearRange's
+	// word-level wasSet count, the same semantics hold.
 	assert.True(t, b.Update(l, 4))
 	assert.Equal(t, int64(0), b.lostCounter.Count())
 	assert.True(t, b.Update(l, 1))
@@ -244,7 +266,7 @@ func TestBitsLostCounterIssue1(t *testing.T) {
 	assert.Equal(t, int64(0), b.lostCounter.Count())
 	assert.True(t, b.Update(l, 7))
 	assert.Equal(t, int64(0), b.lostCounter.Count())
-	// assert.True(t, b.Update(l, 8))
+	// Skip packet 8.
 	assert.True(t, b.Update(l, 10))
 	assert.Equal(t, int64(0), b.lostCounter.Count())
 	assert.True(t, b.Update(l, 11))
@@ -252,9 +274,23 @@ func TestBitsLostCounterIssue1(t *testing.T) {

 	assert.True(t, b.Update(l, 14))
 	assert.Equal(t, int64(0), b.lostCounter.Count())
-	// Issue seems to be here, we reset missing packet 8 to false here and don't increment the lost counter
-	assert.True(t, b.Update(l, 19))
+
+	// Jump to 25. With length=16, slot 25%16=9 corresponds to packet 9
+	// (which we DID receive), so its bit is set and no lost++ from that
+	// eviction. The trace below shows the only loss is packet 8.
+	assert.True(t, b.Update(l, 25))
+	// current was 14, i=25. end=min(25,30)=25. count=11. startPos=15.
+	// steady? current=14<16, so warmup branch: per-bit n=15..25, count those
+	// with !get(n) AND n>16. n=17..25 are >16. Among slots 17%16=1..25%16=9
+	// did we set slots 1..9 (packets 1..9)? Yes for all but slot 8 (packet 8
+	// was skipped). n=24 maps to slot 8 which is FALSE → lost++. All other
+	// n in 17..25 map to slots that are set. n=16 is not strictly > 16. So
+	// lost = 1.
 	assert.Equal(t, int64(1), b.lostCounter.Count())
+
+	// Fill in 12, 13, 15, 16. Each is below current=25 (in-window). 16 must
+	// recheck slot 0 — it was set by NewBits and then cleared by the
+	// Update(25) jump, so 16 backfills cleanly.
 	assert.True(t, b.Update(l, 12))
 	assert.Equal(t, int64(1), b.lostCounter.Count())
 	assert.True(t, b.Update(l, 13))
@@ -263,29 +299,140 @@ func TestBitsLostCounterIssue1(t *testing.T) {
 	assert.Equal(t, int64(1), b.lostCounter.Count())
 	assert.True(t, b.Update(l, 16))
 	assert.Equal(t, int64(1), b.lostCounter.Count())
-	assert.True(t, b.Update(l, 17))
-	assert.Equal(t, int64(1), b.lostCounter.Count())
-	assert.True(t, b.Update(l, 18))
-	assert.Equal(t, int64(1), b.lostCounter.Count())
-	assert.True(t, b.Update(l, 20))
-	assert.Equal(t, int64(1), b.lostCounter.Count())
-	assert.True(t, b.Update(l, 21))

-	// We missed packet 8 above
+	// We missed packet 8 above and that loss is still recorded once, never
+	// double-counted, never zeroed.
 	assert.Equal(t, int64(1), b.lostCounter.Count())
 	assert.Equal(t, int64(0), b.dupeCounter.Count())
 	assert.Equal(t, int64(0), b.outOfWindowCounter.Count())
 }

-func BenchmarkBits(b *testing.B) {
-	z := NewBits(10)
-	for n := 0; n < b.N; n++ {
-		for i := range z.bits {
-			z.bits[i] = true
-		}
-		for i := range z.bits {
-			z.bits[i] = false
-		}
+// TestBitsWarmupOvershoot exercises the jump path's warmup arm with an
+// overshoot past one full window. NewBits leaves current=0 with only slot 0
+// "set" by the marker. Jumping straight to length+k must (a) clear every
+// slot the jump straddles, (b) count only past-window slack (not the
+// in-window slots, which never had a "lost" tenant during warmup), and
+// (c) leave the cursor at the new counter so subsequent unit advances
+// count from steady state. The marker bit at slot 0 is irrelevant once
+// current >= length.
+func TestBitsWarmupOvershoot(t *testing.T) {
+	l := test.NewLogger()
+	b := NewBits(16)
+	b.lostCounter.Clear()

+	// Jump from current=0 to i=20 (length=16, overshoot=4).
+	// Warmup arm: counts slots in [1..16] where bit unset and n>length.
+	// Only n=16 was unset and >length: but slot 16%16=0 is the marker,
+	// so b.get(16) reads bits[0]=1 and skips. Result: 0 lost from the loop.
+	// Past-window: i - current - length = 20 - 0 - 16 = 4 lost.
+	assert.True(t, b.Update(l, 20))
+	assert.Equal(t, int64(4), b.lostCounter.Count())
+	assert.Equal(t, uint64(20), b.current)
+
+	// Steady state now (current=20 >= length=16). Unit advance to 21
+	// stomps slot 21%16=5, which was cleared by the jump and not reset,
+	// so this is +1 lost.
+	assert.True(t, b.Update(l, 21))
+	assert.Equal(t, int64(5), b.lostCounter.Count())
+}
+
+// TestBitsCheckAcrossWarmupBoundary pins the underflow trick in Check's
+// in-window clause. While in warmup, b.current-b.length underflows uint64
+// to a huge value so the first OR-clause is always false; the second
+// clause (i < length && current < length) carries the in-window check.
+// Once current >= length the regimes flip cleanly.
+func TestBitsCheckAcrossWarmupBoundary(t *testing.T) {
+	l := test.NewLogger()
+	b := NewBits(16)
+
+	// Warmup: current=0. Check(0) must read the marker (set) and return false.
+	assert.False(t, b.Check(l, 0), "marker slot should look already-received")
+	// Warmup: any 0 < i < length is in-window and unset → accepted.
+	for i := uint64(1); i < 16; i++ {
+		assert.True(t, b.Check(l, i), "warmup in-window i=%d should be accepted", i)
+	}
+	// Warmup: i >= length but > current is "next number" so accepted.
+	assert.True(t, b.Check(l, 16))
+	assert.True(t, b.Check(l, 1_000_000))
+
+	// Cross into steady state.
+	assert.True(t, b.Update(l, 100))
+	// Now current=100, length=16. In-window range is [85..100].
+	// 84 is just outside: the underflow clause activates; 84 > 100-16=84 is false.
+	// And the warmup clause is false (current >= length). So out of window.
+	assert.False(t, b.Check(l, 84))
+	// 85 sits at the boundary. 85 > 84 is true → in window, unset → accept.
+	assert.True(t, b.Check(l, 85))
+	// 100 is current itself; not strictly greater, in-window, but already set.
+	assert.False(t, b.Check(l, 100))
+	// Way out: clearly out of window.
+	assert.False(t, b.Check(l, 50))
+}
+
+// TestBitsMarkerInvariant verifies the seeded bits[0]=1 marker behaves
+// correctly across warmup and beyond. Update should never clear the marker
+// during warmup (clearRange skips position 0 when startPos=1), and once
+// current >= length the marker is no longer consulted by Check/Update on
+// the live path — but it must still report counter 0 as a duplicate while
+// we are in warmup.
+func TestBitsMarkerInvariant(t *testing.T) {
+	l := test.NewLogger()
+	b := NewBits(8)
+
+	// Counter 0 is the seeded marker; Check sees it as already received.
+	assert.False(t, b.Check(l, 0))
+	// Update(0) at current=0 hits the duplicate branch.
+	b.dupeCounter.Clear()
+	assert.False(t, b.Update(l, 0))
+	assert.Equal(t, int64(1), b.dupeCounter.Count())
+
+	// Walk forward through warmup; the marker must remain set.
+	for n := uint64(1); n <= 7; n++ {
+		assert.True(t, b.Update(l, n))
+	}
+	// Position 0 (the marker) should still read as set because we never
+	// cleared it; Update(0) still looks like a duplicate.
+	assert.False(t, b.Check(l, 0))
+
+	// Cross into steady state with a unit advance to 8: pos=0, evicts the
+	// marker bit. The lost-counter guard (i > b.length) is false (8 == 8),
+	// so this advance does NOT charge a lost packet — exactly what the
+	// marker is there to prevent.
+	b.lostCounter.Clear()
+	assert.True(t, b.Update(l, 8))
+	assert.Equal(t, int64(0), b.lostCounter.Count())
+	// The slot at pos 0 is now occupied by counter 8.
+	assert.False(t, b.Check(l, 8))
+}
+
+// BenchmarkBitsUpdateInOrder is the steady-state hot path: each call is
+// i == current+1.
+func BenchmarkBitsUpdateInOrder(b *testing.B) {
+	l := test.NewLogger()
+	z := NewBits(16384)
+	for n := 0; n < b.N; n++ {
+		z.Update(l, uint64(n)+1)
+	}
+}
+
+// BenchmarkBitsUpdateReorder simulates light reorder within the window:
+// every other packet arrives one slot behind its predecessor (forces the
+// in-window backfill branch).
+func BenchmarkBitsUpdateReorder(b *testing.B) {
+	l := test.NewLogger()
+	z := NewBits(16384)
+	for n := 0; n < b.N; n++ {
+		base := uint64(n) * 2
+		z.Update(l, base+2)
+		z.Update(l, base+1)
+	}
+}
+
+// BenchmarkBitsUpdateLargeJumps stresses the clearRange word-level path.
+func BenchmarkBitsUpdateLargeJumps(b *testing.B) {
+	l := test.NewLogger()
+	z := NewBits(16384)
+	for n := 0; n < b.N; n++ {
+		z.Update(l, uint64(n+1)*1000)
 	}
 }
@@ -217,6 +217,10 @@ func (ncp *CAPool) verify(c Certificate, now time.Time, certFp string, signerFp
 		return nil, err
 	}

+	if signer.Certificate.Curve() != c.Curve() {
+		return nil, ErrCurveMismatch
+	}
+
 	if signer.Certificate.Expired(now) {
 		return nil, ErrRootExpired
 	}
@@ -654,3 +654,31 @@ func TestCertificateV2_Verify_Subnets(t *testing.T) {
 	_, err = caPool.VerifyCertificate(time.Now(), c)
 	require.NoError(t, err)
 }
+
+func TestCertificateV2_CurveMismatch(t *testing.T) {
+	caIp1 := mustParsePrefixUnmapped("10.0.0.0/16")
+	caIp2 := mustParsePrefixUnmapped("192.168.0.0/24")
+	ca, _, caKey, _ := NewTestCaCert(Version2, Curve_P256, time.Now(), time.Now().Add(10*time.Minute), []netip.Prefix{caIp1, caIp2}, nil, []string{"test"})
+
+	caPem, err := ca.MarshalPEM()
+	require.NoError(t, err)
+
+	caPool := NewCAPool()
+	b, err := caPool.AddCAFromPEM(caPem)
+	require.NoError(t, err)
+	assert.Empty(t, b)
+
+	// ip is outside the network
+	cIp1 := mustParsePrefixUnmapped("10.0.0.1/24")
+	c, _, _, _ := NewTestCert(Version2, Curve_P256, ca, caKey, "test", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{cIp1}, nil, []string{"test"})
+
+	fp, _ := c.Fingerprint()
+	_, err = caPool.verify(c, time.Now(), fp, c.Issuer())
+	require.NoError(t, err)
+	//
+	c2 := c.(*certificateV2)
+	c2.curve = Curve_CURVE25519
+	fp, _ = c.Fingerprint()
+	_, err = caPool.verify(c, time.Now(), fp, c.Issuer())
+	require.Error(t, err)
+}
@@ -112,6 +112,9 @@ func (c *certificateV1) CheckSignature(key []byte) bool {
 	}
 	switch c.details.curve {
 	case Curve_CURVE25519:
+		if len(key) != ed25519.PublicKeySize {
+			return false //avoids a panic internal to ed25519
+		}
 		return ed25519.Verify(key, b, c.signature)
 	case Curve_P256:
 		pubKey, err := ecdsa.ParseUncompressedPublicKey(elliptic.P256(), key)
@@ -151,6 +151,9 @@ func (c *certificateV2) CheckSignature(key []byte) bool {

 	switch c.curve {
 	case Curve_CURVE25519:
+		if len(key) != ed25519.PublicKeySize {
+			return false //avoids a panic internal to ed25519
+		}
 		return ed25519.Verify(key, b, c.signature)
 	case Curve_P256:
 		pubKey, err := ecdsa.ParseUncompressedPublicKey(elliptic.P256(), key)
@@ -22,6 +22,7 @@ var (
 	ErrCaNotFound                 = errors.New("could not find ca for the certificate")
 	ErrUnknownVersion             = errors.New("certificate version unrecognized")
 	ErrCertPubkeyPresent          = errors.New("certificate has unexpected pubkey present")
+	ErrCurveMismatch              = errors.New("certificate curve does not match CA")

 	ErrInvalidPEMBlock                   = errors.New("input did not contain a valid PEM encoded block")
 	ErrInvalidPEMCertificateBanner       = errors.New("bytes did not contain a proper certificate banner")
@@ -13,6 +13,12 @@ import (
 	"golang.org/x/crypto/ed25519"
 )

+// testCertNow is the reference "now" used to derive default before/after times
+// in NewTestCaCert and NewTestCert. Holding it fixed for the lifetime of the
+// test binary keeps CA and leaf defaults aligned at the same second, so a leaf
+// signed with default times can never expire after its CA on a rounding race.
+var testCertNow = time.Now().Round(time.Second)
+
 // NewTestCaCert will create a new ca certificate
 func NewTestCaCert(version Version, curve Curve, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (Certificate, []byte, []byte, []byte) {
 	var err error
@@ -34,10 +40,10 @@ func NewTestCaCert(version Version, curve Curve, before, after time.Time, networ
 	}

 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}

 	t := &TBSCertificate{
@@ -70,11 +76,11 @@ func NewTestCaCert(version Version, curve Curve, before, after time.Time, networ
 // Expiry times are defaulted if you do not pass them in
 func NewTestCert(v Version, curve Curve, ca Certificate, key []byte, name string, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (Certificate, []byte, []byte, []byte) {
 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}

 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}

 	if len(networks) == 0 {
@@ -14,6 +14,12 @@ import (
 	"golang.org/x/crypto/ed25519"
 )

+// testCertNow is the reference "now" used to derive default before/after times
+// in NewTestCaCert and NewTestCert. Holding it fixed for the lifetime of the
+// test binary keeps CA and leaf defaults aligned at the same second, so a leaf
+// signed with default times can never expire after its CA on a rounding race.
+var testCertNow = time.Now().Round(time.Second)
+
 // NewTestCaCert will create a new ca certificate
 func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (cert.Certificate, []byte, []byte, []byte) {
 	var err error
@@ -35,10 +41,10 @@ func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Ti
 	}

 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}

 	t := &cert.TBSCertificate{
@@ -71,11 +77,11 @@ func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Ti
 // Expiry times are defaulted if you do not pass them in
 func NewTestCert(v cert.Version, curve cert.Curve, ca cert.Certificate, key []byte, name string, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (cert.Certificate, []byte, []byte, []byte) {
 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}

 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}

 	var pub, priv []byte
@@ -97,6 +97,19 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		if err = mustFlagString("out-key", cf.outKeyPath); err != nil {
 			return err
 		}
+	} else {
+		// out-key is meaningless under PKCS#11 because the private key never
+		// leaves the HSM; reject it so we never silently accept or claim a
+		// stdout slot for it.
+		outKeySet := false
+		cf.set.Visit(func(f *flag.Flag) {
+			if f.Name == "out-key" {
+				outKeySet = true
+			}
+		})
+		if outKeySet {
+			return newHelpErrorf("cannot set -out-key with -pkcs11")
+		}
 	}
 	if err := mustFlagString("out-crt", cf.outCertPath); err != nil {
 		return err
@@ -171,12 +184,21 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		}
 	}

+	var claims ioClaims
+	if err := reserveOutputs(&claims,
+		"out-key", *cf.outKeyPath,
+		"out-crt", *cf.outCertPath,
+		"out-qr", *cf.outQRPath,
+	); err != nil {
+		return err
+	}
+
 	var passphrase []byte
 	if !isP11 && *cf.encryption {
 		passphrase = []byte(os.Getenv("NEBULA_CA_PASSPHRASE"))
 		if len(passphrase) == 0 {
 			for i := 0; i < 5; i++ {
-				out.Write([]byte("Enter passphrase: "))
+				errOut.Write([]byte("Enter passphrase: "))
 				passphrase, err = pr.ReadPassword()

 				if err == ErrNoTerminal {
@@ -261,14 +283,16 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		Curve:          curve,
 	}

-	if !isP11 {
+	if !isP11 && !isStdio(*cf.outKeyPath) {
 		if _, err := os.Stat(*cf.outKeyPath); err == nil {
 			return fmt.Errorf("refusing to overwrite existing CA key: %s", *cf.outKeyPath)
 		}
 	}

-	if _, err := os.Stat(*cf.outCertPath); err == nil {
-		return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
+	if !isStdio(*cf.outCertPath) {
+		if _, err := os.Stat(*cf.outCertPath); err == nil {
+			return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
+		}
 	}

 	var c cert.Certificate
@@ -294,7 +318,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			b = cert.MarshalSigningPrivateKeyToPEM(curve, rawPriv)
 		}

-		err = os.WriteFile(*cf.outKeyPath, b, 0600)
+		err = writeOutput(*cf.outKeyPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -305,7 +329,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		return fmt.Errorf("error while marshalling certificate: %s", err)
 	}

-	err = os.WriteFile(*cf.outCertPath, b, 0600)
+	err = writeOutput(*cf.outCertPath, b, 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -316,7 +340,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}

-		err = os.WriteFile(*cf.outQRPath, b, 0600)
+		err = writeOutput(*cf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -332,6 +356,7 @@ func caSummary() string {
 func caHelp(out io.Writer) {
 	cf := newCaFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + caSummary() + "\n"))
+	out.Write([]byte(stdioHelpText))
 	cf.set.SetOutput(out)
 	cf.set.PrintDefaults()
 }
@@ -27,6 +27,7 @@ func Test_caHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" ca <flags>: create a self signed certificate authority\n"+
+			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -argon-iterations uint\n"+
 			"    \tOptional: Argon2 iterations parameter used for encrypted private key passphrase (default 1)\n"+
 			"  -argon-memory uint\n"+
@@ -84,7 +85,7 @@ func Test_ca(t *testing.T) {
 		err:      nil,
 	}

-	pwPromptOb := "Enter passphrase: "
+	pwPromptEB := "Enter passphrase: "

 	// required args
 	assertHelpError(t, ca(
@@ -168,8 +169,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.NoError(t, ca(args, ob, eb, testpw))
-	assert.Equal(t, pwPromptOb, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, pwPromptEB, eb.String())

 	// test encrypted key with passphrase environment variable
 	os.Remove(keyF.Name())
@@ -207,8 +208,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.Error(t, ca(args, ob, eb, errpw))
-	assert.Equal(t, pwPromptOb, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, pwPromptEB, eb.String())

 	// test when user fails to enter a password
 	os.Remove(keyF.Name())
@@ -217,8 +218,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.EqualError(t, ca(args, ob, eb, nopw), "no passphrase specified, remove -encrypt flag to write out-key in plaintext")
-	assert.Equal(t, strings.Repeat(pwPromptOb, 5), ob.String()) // prompts 5 times before giving up
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, strings.Repeat(pwPromptEB, 5), eb.String()) // prompts 5 times before giving up

 	// create valid cert/key for overwrite tests
 	os.Remove(keyF.Name())
@@ -247,3 +248,67 @@ func Test_ca(t *testing.T) {
 	os.Remove(keyF.Name())

 }
+
+func Test_ca_stdio(t *testing.T) {
+	nopw := &StubPasswordReader{}
+
+	keyF, err := os.CreateTemp("", "ca.key")
+	require.NoError(t, err)
+	os.Remove(keyF.Name())
+	defer os.Remove(keyF.Name())
+
+	crtF, err := os.CreateTemp("", "ca.crt")
+	require.NoError(t, err)
+	os.Remove(crtF.Name())
+	defer os.Remove(crtF.Name())
+
+	// out-crt on stdout, out-key on disk
+	ob := &bytes.Buffer{}
+	eb := &bytes.Buffer{}
+	require.NoError(t, ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", "-", "-out-key", keyF.Name()}, ob, eb, nopw))
+	assert.Empty(t, eb.String())
+	c, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.True(t, c.IsCA())
+	assert.Equal(t, "test-ca", c.Name())
+
+	// out-key on stdout, out-crt on disk
+	os.Remove(keyF.Name())
+	ob.Reset()
+	eb.Reset()
+	require.NoError(t, ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", crtF.Name(), "-out-key", "-"}, ob, eb, nopw))
+	assert.Empty(t, eb.String())
+	_, _, curve, err := cert.UnmarshalSigningPrivateKeyFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, cert.Curve_CURVE25519, curve)
+
+	// dual stdout is rejected up front
+	os.Remove(crtF.Name())
+	ob.Reset()
+	eb.Reset()
+	require.EqualError(t,
+		ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", "-", "-out-key", "-"}, ob, eb, nopw),
+		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
+	assert.Empty(t, ob.String())
+
+	// an output conflict combined with -encrypt must error BEFORE prompting
+	// for a passphrase; pr would record any read attempt
+	tracker := &trackingPasswordReader{}
+	ob.Reset()
+	eb.Reset()
+	require.EqualError(t,
+		ca([]string{"-name", "test-ca", "-duration", "1h", "-encrypt", "-out-crt", "-", "-out-key", "-"}, ob, eb, tracker),
+		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
+	assert.Empty(t, ob.String())
+	assert.Empty(t, eb.String())
+	assert.Zero(t, tracker.calls, "passphrase prompt should not have been called")
+}
+
+type trackingPasswordReader struct {
+	calls int
+}
+
+func (pr *trackingPasswordReader) ReadPassword() ([]byte, error) {
+	pr.calls++
+	return []byte(""), nil
+}
@@ -42,6 +42,8 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 		if err = mustFlagString("out-key", cf.outKeyPath); err != nil {
 			return err
 		}
+	} else if *cf.outKeyPath != "" {
+		return newHelpErrorf("cannot set -out-key with -pkcs11")
 	}
 	if err = mustFlagString("out-pub", cf.outPubPath); err != nil {
 		return err
@@ -69,6 +71,14 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 		}
 	}

+	var claims ioClaims
+	if err := reserveOutputs(&claims,
+		"out-key", *cf.outKeyPath,
+		"out-pub", *cf.outPubPath,
+	); err != nil {
+		return err
+	}
+
 	if isP11 {
 		p11Client, err := pkclient.FromUrl(*cf.p11url)
 		if err != nil {
@@ -82,12 +92,12 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while getting public key: %w", err)
 		}
 	} else {
-		err = os.WriteFile(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
+		err = writeOutput(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
 	}
-	err = os.WriteFile(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600)
+	err = writeOutput(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-pub: %s", err)
 	}
@@ -102,6 +112,7 @@ func keygenSummary() string {
 func keygenHelp(out io.Writer) {
 	cf := newKeygenFlags()
 	_, _ = out.Write([]byte("Usage of " + os.Args[0] + " " + keygenSummary() + "\n"))
+	_, _ = out.Write([]byte(stdioHelpText))
 	cf.set.SetOutput(out)
 	cf.set.PrintDefaults()
 }
@@ -20,6 +20,7 @@ func Test_keygenHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" keygen <flags>: create a public/private key pair. the public key can be passed to `nebula-cert sign`\n"+
+			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -curve string\n"+
 			"    \tECDH Curve (25519, P256) (default \"25519\")\n"+
 			"  -out-key string\n"+
@@ -93,3 +94,43 @@ func Test_keygen(t *testing.T) {
 	require.NoError(t, err)
 	assert.Len(t, lPub, 32)
 }
+
+func Test_keygen_stdio(t *testing.T) {
+	keyF, err := os.CreateTemp("", "test.key")
+	require.NoError(t, err)
+	os.Remove(keyF.Name())
+	defer os.Remove(keyF.Name())
+
+	pubF, err := os.CreateTemp("", "test.pub")
+	require.NoError(t, err)
+	os.Remove(pubF.Name())
+	defer os.Remove(pubF.Name())
+
+	// out-pub on stdout, out-key on disk
+	ob := &bytes.Buffer{}
+	eb := &bytes.Buffer{}
+	require.NoError(t, keygen([]string{"-out-pub", "-", "-out-key", keyF.Name()}, ob, eb))
+	assert.Empty(t, eb.String())
+	lPub, _, curve, err := cert.UnmarshalPublicKeyFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, cert.Curve_CURVE25519, curve)
+	assert.Len(t, lPub, 32)
+
+	// out-key on stdout, out-pub on disk
+	os.Remove(keyF.Name())
+	ob.Reset()
+	eb.Reset()
+	require.NoError(t, keygen([]string{"-out-pub", pubF.Name(), "-out-key", "-"}, ob, eb))
+	assert.Empty(t, eb.String())
+	lKey, _, curve, err := cert.UnmarshalPrivateKeyFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, cert.Curve_CURVE25519, curve)
+	assert.Len(t, lKey, 32)
+
+	// both on stdout is a conflict caught up front
+	ob.Reset()
+	eb.Reset()
+	require.EqualError(t, keygen([]string{"-out-pub", "-", "-out-key", "-"}, ob, eb),
+		`-out-key and -out-pub both set to "-", only one output may write to stdout`)
+	assert.Empty(t, ob.String())
+}
@@ -22,7 +22,9 @@ func (pr StdinPasswordReader) ReadPassword() ([]byte, error) {
 	}

 	password, err := term.ReadPassword(int(os.Stdin.Fd()))
-	fmt.Println()
+	// Terminal echo is off while reading, so the user's Enter key does not
+	// produce a visible newline. Emit one on stderr to match the prompt.
+	fmt.Fprintln(os.Stderr)

 	return password, err
 }
@@ -40,11 +40,23 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}

-	rawCert, err := os.ReadFile(*pf.path)
+	var claims ioClaims
+	if err := reserveInputs(&claims, "path", *pf.path); err != nil {
+		return err
+	}
+	if err := reserveOutputs(&claims, "out-qr", *pf.outQRPath); err != nil {
+		return err
+	}
+
+	rawCert, err := readInput("path", *pf.path, &claims)
 	if err != nil {
 		return fmt.Errorf("unable to read cert; %s", err)
 	}

+	// When the QR is going to stdout, suppress the human-readable text/json
+	// output so the binary stream is not contaminated.
+	qrToStdout := isStdio(*pf.outQRPath)
+
 	var c cert.Certificate
 	var qrBytes []byte
 	part := 0
@@ -57,11 +69,13 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while unmarshaling cert: %s", err)
 		}

-		if *pf.json {
-			jsonCerts = append(jsonCerts, c)
-		} else {
-			_, _ = out.Write([]byte(c.String()))
-			_, _ = out.Write([]byte("\n"))
+		if !qrToStdout {
+			if *pf.json {
+				jsonCerts = append(jsonCerts, c)
+			} else {
+				_, _ = out.Write([]byte(c.String()))
+				_, _ = out.Write([]byte("\n"))
+			}
 		}

 		if *pf.outQRPath != "" {
@@ -79,7 +93,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		part++
 	}

-	if *pf.json {
+	if *pf.json && !qrToStdout {
 		b, _ := json.Marshal(jsonCerts)
 		_, _ = out.Write(b)
 		_, _ = out.Write([]byte("\n"))
@@ -91,7 +105,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}

-		err = os.WriteFile(*pf.outQRPath, b, 0600)
+		err = writeOutput(*pf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -107,6 +121,7 @@ func printSummary() string {
 func printHelp(out io.Writer) {
 	pf := newPrintFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + printSummary() + "\n"))
+	out.Write([]byte(stdioHelpText))
 	pf.set.SetOutput(out)
 	pf.set.PrintDefaults()
 }
@@ -25,6 +25,7 @@ func Test_printHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" print <flags>: prints details about a certificate\n"+
+			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -json\n"+
 			"    \tOptional: outputs certificates in json format\n"+
 			"  -out-qr string\n"+
@@ -178,6 +179,44 @@ func Test_printCert(t *testing.T) {
 		ob.String(),
 	)
 	assert.Empty(t, eb.String())
+
+	// read cert from stdin
+	ob.Reset()
+	eb.Reset()
+	withStdin(t, bytes.NewReader(p))
+	err = printCert([]string{"-json", "-path", "-"}, ob, eb)
+	require.NoError(t, err)
+	assert.Equal(
+		t,
+		`[{"details":{"curve":"CURVE25519","groups":["hi"],"isCa":false,"issuer":"`+c.Issuer()+`","name":"test","networks":["10.0.0.123/8"],"notAfter":"0001-01-01T00:00:00Z","notBefore":"0001-01-01T00:00:00Z","publicKey":"`+pk+`","unsafeNetworks":[]},"fingerprint":"`+fp+`","signature":"`+sig+`","version":1}]
+`,
+		ob.String(),
+	)
+	assert.Empty(t, eb.String())
+
+	// -out-qr - sends only the PNG to stdout, suppressing the cert dump
+	ob.Reset()
+	eb.Reset()
+	withStdin(t, bytes.NewReader(p))
+	err = printCert([]string{"-path", "-", "-out-qr", "-"}, ob, eb)
+	require.NoError(t, err)
+	assert.Empty(t, eb.String())
+	stdout := ob.Bytes()
+	require.NotEmpty(t, stdout)
+	// PNG magic, no PEM/JSON noise prepended
+	assert.Equal(t, []byte{0x89, 'P', 'N', 'G', 0x0d, 0x0a, 0x1a, 0x0a}, stdout[:8])
+	assert.NotContains(t, string(stdout), "NebulaCertificate")
+	assert.NotContains(t, string(stdout), `"details"`)
+
+	// json + out-qr - still suppresses json
+	ob.Reset()
+	eb.Reset()
+	withStdin(t, bytes.NewReader(p))
+	err = printCert([]string{"-json", "-path", "-", "-out-qr", "-"}, ob, eb)
+	require.NoError(t, err)
+	assert.Empty(t, eb.String())
+	assert.Equal(t, []byte{0x89, 'P', 'N', 'G'}, ob.Bytes()[:4])
+	assert.NotContains(t, ob.String(), `"details"`)
 }

 // NewTestCaCert will generate a CA cert
@@ -85,6 +85,9 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 	if !isP11 && *sf.inPubPath != "" && *sf.outKeyPath != "" {
 		return newHelpErrorf("cannot set both -in-pub and -out-key")
 	}
+	if isP11 && *sf.outKeyPath != "" {
+		return newHelpErrorf("cannot set -out-key with -pkcs11")
+	}

 	var v4Networks []netip.Prefix
 	var v6Networks []netip.Prefix
@@ -102,13 +105,35 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		return newHelpErrorf("-version must be either %v or %v", cert.Version1, cert.Version2)
 	}

+	if *sf.outKeyPath == "" {
+		*sf.outKeyPath = *sf.name + ".key"
+	}
+	if *sf.outCertPath == "" {
+		*sf.outCertPath = *sf.name + ".crt"
+	}
+
+	var claims ioClaims
+	if err := reserveInputs(&claims,
+		"ca-key", *sf.caKeyPath,
+		"ca-crt", *sf.caCertPath,
+		"in-pub", *sf.inPubPath,
+	); err != nil {
+		return err
+	}
+	if err := reserveOutputs(&claims,
+		"out-key", *sf.outKeyPath,
+		"out-crt", *sf.outCertPath,
+		"out-qr", *sf.outQRPath,
+	); err != nil {
+		return err
+	}
+
 	var curve cert.Curve
 	var caKey []byte

 	if !isP11 {
 		var rawCAKey []byte
-		rawCAKey, err := os.ReadFile(*sf.caKeyPath)
-
+		rawCAKey, err = readInput("ca-key", *sf.caKeyPath, &claims)
 		if err != nil {
 			return fmt.Errorf("error while reading ca-key: %s", err)
 		}
@@ -121,7 +146,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 			if len(passphrase) == 0 {
 				// ask for a passphrase until we get one
 				for i := 0; i < 5; i++ {
-					out.Write([]byte("Enter passphrase: "))
+					errOut.Write([]byte("Enter passphrase: "))
 					passphrase, err = pr.ReadPassword()

 					if errors.Is(err, ErrNoTerminal) {
@@ -147,7 +172,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		}
 	}

-	rawCACert, err := os.ReadFile(*sf.caCertPath)
+	rawCACert, err := readInput("ca-crt", *sf.caCertPath, &claims)
 	if err != nil {
 		return fmt.Errorf("error while reading ca-crt: %s", err)
 	}
@@ -245,7 +270,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)

 	if *sf.inPubPath != "" {
 		var pubCurve cert.Curve
-		rawPub, err := os.ReadFile(*sf.inPubPath)
+		rawPub, err := readInput("in-pub", *sf.inPubPath, &claims)
 		if err != nil {
 			return fmt.Errorf("error while reading in-pub: %s", err)
 		}
@@ -266,16 +291,10 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		pub, rawPriv = newKeypair(curve)
 	}

-	if *sf.outKeyPath == "" {
-		*sf.outKeyPath = *sf.name + ".key"
-	}
-
-	if *sf.outCertPath == "" {
-		*sf.outCertPath = *sf.name + ".crt"
-	}
-
-	if _, err := os.Stat(*sf.outCertPath); err == nil {
-		return fmt.Errorf("refusing to overwrite existing cert: %s", *sf.outCertPath)
+	if !isStdio(*sf.outCertPath) {
+		if _, err := os.Stat(*sf.outCertPath); err == nil {
+			return fmt.Errorf("refusing to overwrite existing cert: %s", *sf.outCertPath)
+		}
 	}

 	var crts []cert.Certificate
@@ -360,11 +379,13 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 	}

 	if !isP11 && *sf.inPubPath == "" {
-		if _, err := os.Stat(*sf.outKeyPath); err == nil {
-			return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
+		if !isStdio(*sf.outKeyPath) {
+			if _, err := os.Stat(*sf.outKeyPath); err == nil {
+				return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
+			}
 		}

-		err = os.WriteFile(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
+		err = writeOutput(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -379,7 +400,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		b = append(b, sb...)
 	}

-	err = os.WriteFile(*sf.outCertPath, b, 0600)
+	err = writeOutput(*sf.outCertPath, b, 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -390,7 +411,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}

-		err = os.WriteFile(*sf.outQRPath, b, 0600)
+		err = writeOutput(*sf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -440,6 +461,7 @@ func signSummary() string {
 func signHelp(out io.Writer) {
 	sf := newSignFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + signSummary() + "\n"))
+	out.Write([]byte(stdioHelpText))
 	sf.set.SetOutput(out)
 	sf.set.PrintDefaults()
 }
@@ -27,6 +27,7 @@ func Test_signHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" sign <flags>: create and sign a certificate\n"+
+			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -ca-crt string\n"+
 			"    \tOptional: path to the signing CA cert (default \"ca.crt\")\n"+
 			"  -ca-key string\n"+
@@ -376,15 +377,18 @@ func Test_signCert(t *testing.T) {
 	// test with the proper password
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.NoError(t, signCert(args, ob, eb, testpw))
-	assert.Equal(t, "Enter passphrase: ", ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, "Enter passphrase: ", eb.String())

 	// test with the proper password in the environment
 	os.Remove(crtF.Name())
 	os.Remove(keyF.Name())
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	os.Setenv("NEBULA_CA_PASSPHRASE", string(passphrase))
+	ob.Reset()
+	eb.Reset()
 	require.NoError(t, signCert(args, ob, eb, testpw))
+	assert.Empty(t, ob.String())
 	assert.Empty(t, eb.String())
 	os.Setenv("NEBULA_CA_PASSPHRASE", "")

@@ -395,8 +399,8 @@ func Test_signCert(t *testing.T) {
 	testpw.password = []byte("invalid password")
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, testpw))
-	assert.Equal(t, "Enter passphrase: ", ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, "Enter passphrase: ", eb.String())

 	// test with the wrong password in environment
 	ob.Reset()
@@ -416,8 +420,8 @@ func Test_signCert(t *testing.T) {
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, nopw))
 	// normally the user hitting enter on the prompt would add newlines between these
-	assert.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", eb.String())

 	// test an error condition
 	ob.Reset()
@@ -425,6 +429,106 @@ func Test_signCert(t *testing.T) {

 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, errpw))
-	assert.Equal(t, "Enter passphrase: ", ob.String())
-	assert.Empty(t, eb.String())
+	assert.Empty(t, ob.String())
+	assert.Equal(t, "Enter passphrase: ", eb.String())
+}
+
+func Test_signCert_stdio(t *testing.T) {
+	nopw := &StubPasswordReader{
+		password: []byte(""),
+		err:      nil,
+	}
+
+	caPub, caPriv, _ := ed25519.GenerateKey(rand.Reader)
+	rawCAKey := cert.MarshalSigningPrivateKeyToPEM(cert.Curve_CURVE25519, caPriv)
+
+	ca, _ := NewTestCaCert("ca", caPub, caPriv, time.Now(), time.Now().Add(time.Minute*200), nil, nil, nil)
+	rawCACrt, _ := ca.MarshalPEM()
+
+	caCrtF, err := os.CreateTemp("", "sign-cert.crt")
+	require.NoError(t, err)
+	defer os.Remove(caCrtF.Name())
+	caCrtF.Write(rawCACrt)
+
+	caKeyF, err := os.CreateTemp("", "sign-cert.key")
+	require.NoError(t, err)
+	defer os.Remove(caKeyF.Name())
+	caKeyF.Write(rawCAKey)
+
+	keyF, err := os.CreateTemp("", "sign.key")
+	require.NoError(t, err)
+	os.Remove(keyF.Name())
+	defer os.Remove(keyF.Name())
+
+	// ca-key on stdin, cert to stdout
+	withStdin(t, bytes.NewReader(rawCAKey))
+	ob := &bytes.Buffer{}
+	eb := &bytes.Buffer{}
+	args := []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "-", "-out-key", keyF.Name(), "-duration", "100m"}
+	require.NoError(t, signCert(args, ob, eb, nopw))
+	assert.Empty(t, eb.String())
+
+	lCrt, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, "stdin-test", lCrt.Name())
+	assert.True(t, lCrt.CheckSignature(caPub))
+
+	// two flags reading from stdin should error before any read attempt;
+	// otherwise an interactive shell would hang on io.ReadAll
+	stdinIn := bytes.NewReader(rawCAKey)
+	withStdin(t, stdinIn)
+	ob.Reset()
+	eb.Reset()
+	args = []string{"-version", "1", "-ca-crt", "-", "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "nope", "-out-key", "nope", "-duration", "100m"}
+	require.EqualError(t, signCert(args, ob, eb, nopw),
+		`-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
+	assert.Equal(t, len(rawCAKey), stdinIn.Len(), "stdin should be untouched when conflict is caught up front")
+
+	// two flags writing to stdout should error before any output is written
+	// AND before stdin is consumed
+	stdinR := bytes.NewReader(rawCAKey)
+	withStdin(t, stdinR)
+	ob.Reset()
+	eb.Reset()
+	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "-", "-out-key", "-", "-duration", "100m"}
+	require.EqualError(t, signCert(args, ob, eb, nopw),
+		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
+	assert.Empty(t, ob.String())
+	// stdin should be untouched because the conflict was caught up front
+	assert.Equal(t, len(rawCAKey), stdinR.Len())
+
+	// out-key on stdout, cert on disk
+	keyF2, err := os.CreateTemp("", "sign.key")
+	require.NoError(t, err)
+	os.Remove(keyF2.Name())
+	defer os.Remove(keyF2.Name())
+	crtF, err := os.CreateTemp("", "sign.crt")
+	require.NoError(t, err)
+	os.Remove(crtF.Name())
+	defer os.Remove(crtF.Name())
+
+	ob.Reset()
+	eb.Reset()
+	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", "-", "-duration", "100m"}
+	require.NoError(t, signCert(args, ob, eb, nopw))
+	assert.Empty(t, eb.String())
+	_, _, curve, err := cert.UnmarshalPrivateKeyFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, cert.Curve_CURVE25519, curve)
+
+	// in-pub on stdin (caller already has a keypair, only the cert is generated)
+	inPub, _ := x25519Keypair()
+	rawInPub := cert.MarshalPublicKeyToPEM(cert.Curve_CURVE25519, inPub)
+
+	withStdin(t, bytes.NewReader(rawInPub))
+	os.Remove(crtF.Name())
+	ob.Reset()
+	eb.Reset()
+	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "in-pub-test", "-ip", "1.1.1.1/24", "-in-pub", "-", "-out-crt", "-", "-duration", "100m"}
+	require.NoError(t, signCert(args, ob, eb, nopw))
+	assert.Empty(t, eb.String())
+	stdinCrt, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
+	require.NoError(t, err)
+	assert.Equal(t, "in-pub-test", stdinCrt.Name())
+	assert.Equal(t, inPub, stdinCrt.PublicKey())
 }
@@ -0,0 +1,117 @@
+package main
+
+import (
+	"fmt"
+	"io"
+	"os"
+)
+
+// stdioPath is the special path value that selects stdin (for inputs) or
+// stdout (for outputs) instead of a file on disk.
+const stdioPath = "-"
+
+// stdioHelpText is rendered just under the Usage line of each subcommand
+// help so the - convention is documented once instead of on every flag.
+const stdioHelpText = "  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"
+
+// stdinReader is the source used when an input flag is set to "-".
+// It is a package level var so tests can swap in a deterministic reader.
+// Tests that mutate stdinReader cannot run with t.Parallel().
+var stdinReader io.Reader = os.Stdin
+
+// ioClaims tracks which flags have claimed stdin and stdout during a single
+// command invocation so we can refuse a second flag asking for the same
+// stream.
+type ioClaims struct {
+	in  string
+	out string
+}
+
+func (c *ioClaims) claimIn(flagName string) error {
+	if c.in != "" && c.in != flagName {
+		return fmt.Errorf("-%s and -%s both set to %q, only one input may read from stdin", c.in, flagName, stdioPath)
+	}
+	c.in = flagName
+	return nil
+}
+
+func (c *ioClaims) claimOut(flagName string) error {
+	if c.out != "" && c.out != flagName {
+		return fmt.Errorf("-%s and -%s both set to %q, only one output may write to stdout", c.out, flagName, stdioPath)
+	}
+	c.out = flagName
+	return nil
+}
+
+// reserveInputs walks alternating (flagName, path) pairs and claims stdin
+// for any path equal to stdioPath. It must be called before any input is
+// read so a conflict can be reported immediately instead of blocking on
+// io.ReadAll while waiting for input that will never arrive.
+func reserveInputs(claims *ioClaims, pairs ...string) error {
+	return reserveStdio(claims, "reserveInputs", (*ioClaims).claimIn, pairs)
+}
+
+// reserveOutputs walks alternating (flagName, path) pairs and claims stdout
+// for any path equal to stdioPath. It must be called before any output is
+// written so a conflict cannot leave one stream half written before the
+// second flag fails.
+func reserveOutputs(claims *ioClaims, pairs ...string) error {
+	return reserveStdio(claims, "reserveOutputs", (*ioClaims).claimOut, pairs)
+}
+
+func reserveStdio(claims *ioClaims, who string, claim func(*ioClaims, string) error, pairs []string) error {
+	if len(pairs)%2 != 0 {
+		panic(who + " requires alternating name, path pairs")
+	}
+	for i := 0; i < len(pairs); i += 2 {
+		name, path := pairs[i], pairs[i+1]
+		if path != stdioPath {
+			continue
+		}
+		if err := claim(claims, name); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// readInput returns the bytes referenced by path, reading from stdin when
+// path is stdioPath.
+func readInput(flagName, path string, claims *ioClaims) ([]byte, error) {
+	if path == stdioPath {
+		if err := claims.claimIn(flagName); err != nil {
+			return nil, err
+		}
+		return io.ReadAll(stdinReader)
+	}
+	return os.ReadFile(path)
+}
+
+// openInput returns a reader for path. When path is stdioPath the returned
+// reader wraps stdin and Close is a no-op.
+func openInput(flagName, path string, claims *ioClaims) (io.ReadCloser, error) {
+	if path == stdioPath {
+		if err := claims.claimIn(flagName); err != nil {
+			return nil, err
+		}
+		return io.NopCloser(stdinReader), nil
+	}
+	return os.Open(path)
+}
+
+// writeOutput writes data to path, or to stdout when path is stdioPath. perm
+// is only used for file output. The caller must have already claimed stdout
+// via reserveOutputs before invoking with stdioPath.
+func writeOutput(path string, data []byte, perm os.FileMode, stdout io.Writer) error {
+	if path == stdioPath {
+		_, err := stdout.Write(data)
+		return err
+	}
+	return os.WriteFile(path, data, perm)
+}
+
+// isStdio reports whether path is the stdio sentinel and so should skip
+// existence checks like "refuse to overwrite".
+func isStdio(path string) bool {
+	return path == stdioPath
+}
@@ -0,0 +1,167 @@
+package main
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// withStdin temporarily replaces stdinReader for the duration of t.
+func withStdin(t *testing.T, r io.Reader) {
+	t.Helper()
+	prev := stdinReader
+	stdinReader = r
+	t.Cleanup(func() { stdinReader = prev })
+}
+
+func Test_readInput_stdin(t *testing.T) {
+	withStdin(t, bytes.NewBufferString("hello"))
+	var claims ioClaims
+
+	got, err := readInput("path", "-", &claims)
+	require.NoError(t, err)
+	assert.Equal(t, []byte("hello"), got)
+	assert.Equal(t, "path", claims.in)
+}
+
+func Test_readInput_file(t *testing.T) {
+	dir := t.TempDir()
+	p := filepath.Join(dir, "f")
+	require.NoError(t, os.WriteFile(p, []byte("file"), 0600))
+	var claims ioClaims
+
+	got, err := readInput("path", p, &claims)
+	require.NoError(t, err)
+	assert.Equal(t, []byte("file"), got)
+	assert.Empty(t, claims.in)
+}
+
+func Test_readInput_doubleStdinErrors(t *testing.T) {
+	withStdin(t, bytes.NewBufferString("hello"))
+	var claims ioClaims
+
+	_, err := readInput("ca-key", "-", &claims)
+	require.NoError(t, err)
+
+	_, err = readInput("ca-crt", "-", &claims)
+	require.EqualError(t, err, `-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
+}
+
+func Test_openInput_stdin(t *testing.T) {
+	withStdin(t, bytes.NewBufferString("hi"))
+	var claims ioClaims
+
+	r, err := openInput("ca", "-", &claims)
+	require.NoError(t, err)
+	defer r.Close()
+	b, err := io.ReadAll(r)
+	require.NoError(t, err)
+	assert.Equal(t, []byte("hi"), b)
+}
+
+func Test_openInput_doubleStdinErrors(t *testing.T) {
+	withStdin(t, bytes.NewBufferString("hi"))
+	var claims ioClaims
+
+	r, err := openInput("ca", "-", &claims)
+	require.NoError(t, err)
+	r.Close()
+
+	_, err = openInput("crt", "-", &claims)
+	require.EqualError(t, err, `-ca and -crt both set to "-", only one input may read from stdin`)
+}
+
+func Test_writeOutput_stdout(t *testing.T) {
+	out := &bytes.Buffer{}
+
+	err := writeOutput("-", []byte("payload"), 0600, out)
+	require.NoError(t, err)
+	assert.Equal(t, "payload", out.String())
+}
+
+func Test_writeOutput_file(t *testing.T) {
+	dir := t.TempDir()
+	p := filepath.Join(dir, "f")
+	out := &bytes.Buffer{}
+
+	err := writeOutput(p, []byte("payload"), 0600, out)
+	require.NoError(t, err)
+	assert.Empty(t, out.String())
+	got, err := os.ReadFile(p)
+	require.NoError(t, err)
+	assert.Equal(t, []byte("payload"), got)
+}
+
+func Test_reserveOutputs_noConflict(t *testing.T) {
+	var claims ioClaims
+	require.NoError(t, reserveOutputs(&claims,
+		"out-key", "/tmp/key",
+		"out-crt", "-",
+		"out-qr", "",
+	))
+	assert.Equal(t, "out-crt", claims.out)
+}
+
+func Test_reserveOutputs_conflict(t *testing.T) {
+	var claims ioClaims
+	err := reserveOutputs(&claims,
+		"out-key", "-",
+		"out-crt", "-",
+	)
+	require.EqualError(t, err, `-out-key and -out-crt both set to "-", only one output may write to stdout`)
+}
+
+func Test_reserveOutputs_panicsOnOddPairs(t *testing.T) {
+	defer func() {
+		r := recover()
+		require.NotNil(t, r)
+	}()
+	var claims ioClaims
+	_ = reserveOutputs(&claims, "out-key")
+}
+
+func Test_reserveInputs_noConflict(t *testing.T) {
+	var claims ioClaims
+	require.NoError(t, reserveInputs(&claims,
+		"ca-key", "/tmp/ca.key",
+		"ca-crt", "-",
+		"in-pub", "",
+	))
+	assert.Equal(t, "ca-crt", claims.in)
+}
+
+func Test_reserveInputs_conflict(t *testing.T) {
+	var claims ioClaims
+	err := reserveInputs(&claims,
+		"ca-key", "-",
+		"ca-crt", "-",
+	)
+	require.EqualError(t, err, `-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
+}
+
+func Test_claimIn_idempotent(t *testing.T) {
+	// pre-claim then a lazy re-claim of the same flag should be a no-op
+	var claims ioClaims
+	require.NoError(t, claims.claimIn("ca-key"))
+	require.NoError(t, claims.claimIn("ca-key"))
+	assert.Equal(t, "ca-key", claims.in)
+}
+
+func Test_claimOut_idempotent(t *testing.T) {
+	var claims ioClaims
+	require.NoError(t, claims.claimOut("out-crt"))
+	require.NoError(t, claims.claimOut("out-crt"))
+	assert.Equal(t, "out-crt", claims.out)
+}
+
+func Test_isStdio(t *testing.T) {
+	assert.True(t, isStdio("-"))
+	assert.False(t, isStdio(""))
+	assert.False(t, isStdio("./-"))
+	assert.False(t, isStdio("foo"))
+}
@@ -39,18 +39,26 @@ func verify(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}

-	caFile, err := os.Open(*vf.caPath)
+	var claims ioClaims
+	if err := reserveInputs(&claims,
+		"ca", *vf.caPath,
+		"crt", *vf.certPath,
+	); err != nil {
+		return err
+	}
+
+	caReader, err := openInput("ca", *vf.caPath, &claims)
 	if err != nil {
 		return fmt.Errorf("error while reading ca: %w", err)
 	}
-	defer caFile.Close()
+	defer caReader.Close()

-	caPool, err := cert.NewCAPoolFromPEMReader(caFile)
+	caPool, err := cert.NewCAPoolFromPEMReader(caReader)
 	if err != nil && !errors.Is(err, cert.ErrExpired) {
 		return fmt.Errorf("error while adding ca cert to pool: %w", err)
 	}

-	rawCert, err := os.ReadFile(*vf.certPath)
+	rawCert, err := readInput("crt", *vf.certPath, &claims)
 	if err != nil {
 		return fmt.Errorf("unable to read crt: %w", err)
 	}
@@ -85,6 +93,7 @@ func verifySummary() string {
 func verifyHelp(out io.Writer) {
 	vf := newVerifyFlags()
 	_, _ = out.Write([]byte("Usage of " + os.Args[0] + " " + verifySummary() + "\n"))
+	_, _ = out.Write([]byte(stdioHelpText))
 	vf.set.SetOutput(out)
 	vf.set.PrintDefaults()
 }
@@ -23,6 +23,7 @@ func Test_verifyHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" verify <flags>: verifies a certificate isn't expired and was signed by a trusted authority.\n"+
+			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -ca string\n"+
 			"    \tRequired: path to a file containing one or more ca certificates\n"+
 			"  -crt string\n"+
@@ -122,3 +123,46 @@ func Test_verify(t *testing.T) {
 	assert.Empty(t, eb.String())
 	require.NoError(t, err)
 }
+
+func Test_verify_stdio(t *testing.T) {
+	ob := &bytes.Buffer{}
+	eb := &bytes.Buffer{}
+
+	caPub, caPriv, _ := ed25519.GenerateKey(rand.Reader)
+	ca, _ := NewTestCaCert("test-ca", caPub, caPriv, time.Now().Add(time.Hour*-1), time.Now().Add(time.Hour*2), nil, nil, nil)
+	caPEM, _ := ca.MarshalPEM()
+
+	crt, _ := NewTestCert(ca, caPriv, "test-cert", time.Now().Add(time.Hour*-1), time.Now().Add(time.Hour), nil, nil, nil)
+	crtPEM, _ := crt.MarshalPEM()
+
+	caFile, err := os.CreateTemp("", "verify-ca")
+	require.NoError(t, err)
+	defer os.Remove(caFile.Name())
+	caFile.Write(caPEM)
+
+	// crt on stdin, ca on disk
+	withStdin(t, bytes.NewReader(crtPEM))
+	require.NoError(t, verify([]string{"-ca", caFile.Name(), "-crt", "-"}, ob, eb))
+	assert.Empty(t, ob.String())
+	assert.Empty(t, eb.String())
+
+	// ca on stdin, crt on disk
+	certFile, err := os.CreateTemp("", "verify-cert")
+	require.NoError(t, err)
+	defer os.Remove(certFile.Name())
+	certFile.Write(crtPEM)
+
+	withStdin(t, bytes.NewReader(caPEM))
+	ob.Reset()
+	eb.Reset()
+	require.NoError(t, verify([]string{"-ca", "-", "-crt", certFile.Name()}, ob, eb))
+	assert.Empty(t, ob.String())
+	assert.Empty(t, eb.String())
+
+	// both flags on stdin should error
+	withStdin(t, bytes.NewReader(caPEM))
+	ob.Reset()
+	eb.Reset()
+	require.EqualError(t, verify([]string{"-ca", "-", "-crt", "-"}, ob, eb),
+		`-ca and -crt both set to "-", only one input may read from stdin`)
+}
@@ -61,9 +61,12 @@ func main() {
 	}

 	if *configPath == "" {
-		fmt.Println("-config flag must be set")
-		flag.Usage()
-		os.Exit(1)
+		p, err := config.DefaultPath()
+		if err != nil {
+			fmt.Println(err)
+			os.Exit(1)
+		}
+		*configPath = p
 	}

 	c := config.NewC(l)
@@ -3,8 +3,6 @@ package main
 import (
 	"fmt"
 	"log"
-	"os"
-	"path/filepath"

 	"github.com/kardianos/service"
 	"github.com/slackhq/nebula"
@@ -57,24 +55,13 @@ func (p *program) Stop(s service.Service) error {
 	return nil
 }

-func fileExists(filename string) bool {
-	_, err := os.Stat(filename)
-	if os.IsNotExist(err) {
-		return false
-	}
-	return true
-}
-
 func doService(configPath *string, configTest *bool, build string, serviceFlag *string) error {
 	if *configPath == "" {
-		ex, err := os.Executable()
+		p, err := config.DefaultPath()
 		if err != nil {
 			return err
 		}
-		*configPath = filepath.Dir(ex) + "/config.yaml"
-		if !fileExists(*configPath) {
-			*configPath = filepath.Dir(ex) + "/config.yml"
-		}
+		*configPath = p
 	}

 	svcConfig := &service.Config{
@@ -50,9 +50,12 @@ func main() {
 	}

 	if *configPath == "" {
-		fmt.Println("-config flag must be set")
-		flag.Usage()
-		os.Exit(1)
+		p, err := config.DefaultPath()
+		if err != nil {
+			fmt.Println(err)
+			os.Exit(1)
+		}
+		*configPath = p
 	}

 	l := logging.NewLogger(os.Stdout)
@@ -0,0 +1,29 @@
+package config
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// DefaultPath returns a path to a config file alongside the running executable, preferring config.yaml over config.yml.
+// If neither file exists an error is returned that names both paths checked.
+func DefaultPath() (string, error) {
+	ex, err := os.Executable()
+	if err != nil {
+		return "", err
+	}
+	return defaultPathInDir(filepath.Dir(ex))
+}
+
+func defaultPathInDir(dir string) (string, error) {
+	yamlPath := filepath.Join(dir, "config.yaml")
+	if _, err := os.Stat(yamlPath); err == nil {
+		return yamlPath, nil
+	}
+	ymlPath := filepath.Join(dir, "config.yml")
+	if _, err := os.Stat(ymlPath); err == nil {
+		return ymlPath, nil
+	}
+	return "", fmt.Errorf("no default config found at %s or %s", yamlPath, ymlPath)
+}
@@ -0,0 +1,67 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestDefaultPathInDir(t *testing.T) {
+	t.Run("prefers config.yaml when both exist", func(t *testing.T) {
+		dir := t.TempDir()
+		want := filepath.Join(dir, "config.yaml")
+		other := filepath.Join(dir, "config.yml")
+		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
+		require.NoError(t, os.WriteFile(other, []byte("a: 2"), 0644))
+
+		got, err := defaultPathInDir(dir)
+		require.NoError(t, err)
+		assert.Equal(t, want, got)
+	})
+
+	t.Run("returns config.yaml when only it exists", func(t *testing.T) {
+		dir := t.TempDir()
+		want := filepath.Join(dir, "config.yaml")
+		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
+
+		got, err := defaultPathInDir(dir)
+		require.NoError(t, err)
+		assert.Equal(t, want, got)
+	})
+
+	t.Run("falls back to config.yml when only it exists", func(t *testing.T) {
+		dir := t.TempDir()
+		want := filepath.Join(dir, "config.yml")
+		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
+
+		got, err := defaultPathInDir(dir)
+		require.NoError(t, err)
+		assert.Equal(t, want, got)
+	})
+
+	t.Run("errors when neither exists and names both paths", func(t *testing.T) {
+		dir := t.TempDir()
+		got, err := defaultPathInDir(dir)
+		assert.Empty(t, got)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), filepath.Join(dir, "config.yaml"))
+		assert.Contains(t, err.Error(), filepath.Join(dir, "config.yml"))
+	})
+}
+
+func TestDefaultPath(t *testing.T) {
+	got, err := DefaultPath()
+	if err != nil {
+		ex, exErr := os.Executable()
+		require.NoError(t, exErr)
+		assert.Contains(t, err.Error(), filepath.Dir(ex))
+		return
+	}
+	ex, err := os.Executable()
+	require.NoError(t, err)
+	assert.Equal(t, filepath.Dir(ex), filepath.Dir(got))
+	assert.Contains(t, []string{"config.yaml", "config.yml"}, filepath.Base(got))
+}
@@ -11,7 +11,6 @@ import (
 	"sync/atomic"
 	"time"

-	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/header"
@@ -45,19 +44,16 @@ type connectionManager struct {
 	inactivityTimeout       atomic.Int64
 	dropInactive            atomic.Bool

-	metricsTxPunchy metrics.Counter
-
 	l *slog.Logger
 }

 func newConnectionManagerFromConfig(l *slog.Logger, c *config.C, hm *HostMap, p *Punchy) *connectionManager {
 	cm := &connectionManager{
-		hostMap:         hm,
-		l:               l,
-		punchy:          p,
-		relayUsed:       make(map[uint32]struct{}),
-		relayUsedLock:   &sync.RWMutex{},
-		metricsTxPunchy: metrics.GetOrRegisterCounter("messages.tx.punchy", nil),
+		hostMap:       hm,
+		l:             l,
+		punchy:        p,
+		relayUsed:     make(map[uint32]struct{}),
+		relayUsedLock: &sync.RWMutex{},
 	}

 	cm.reload(c, true)
@@ -140,21 +136,13 @@ func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time)
 	return in, out
 }

-// AddTrafficWatch must be called for every new HostInfo.
-// We will continue to monitor the HostInfo until the tunnel is dropped.
-func (cm *connectionManager) AddTrafficWatch(h *HostInfo) {
-	if h.out.Swap(true) == false {
-		cm.trafficTimer.Add(h.localIndexId, cm.checkInterval)
-	}
-}
-
 func (cm *connectionManager) Start(ctx context.Context) {
 	clockSource := time.NewTicker(cm.trafficTimer.t.tickDuration)
 	defer clockSource.Stop()

 	p := []byte("")
-	// Long-lived buf for the traffic-check goroutine; never released.
-	buf := cm.intf.bufAlloc.Acquire()
+	nb := make([]byte, 12, 12)
+	out := make([]byte, mtu)

 	for {
 		select {
@@ -169,13 +157,13 @@ func (cm *connectionManager) Start(ctx context.Context) {
 					break
 				}

-				cm.doTrafficCheck(localIndex, p, buf, now)
+				cm.doTrafficCheck(localIndex, p, nb, out, now)
 			}
 		}
 	}
 }

-func (cm *connectionManager) doTrafficCheck(localIndex uint32, p []byte, buf *WireBuffer, now time.Time) {
+func (cm *connectionManager) doTrafficCheck(localIndex uint32, p, nb, out []byte, now time.Time) {
 	decision, hostinfo, primary := cm.makeTrafficDecision(localIndex, now)

 	switch decision {
@@ -199,7 +187,7 @@ func (cm *connectionManager) doTrafficCheck(localIndex uint32, p []byte, buf *Wi
 		cm.tryRehandshake(hostinfo)

 	case sendTestPacket:
-		cm.intf.SendMessageToHostInfo(header.Test, header.TestRequest, hostinfo, p, buf)
+		cm.intf.SendMessageToHostInfo(header.Test, header.TestRequest, hostinfo, p, nb, out)
 	}

 	cm.resetRelayTrafficCheck(hostinfo)
@@ -308,12 +296,10 @@ func (cm *connectionManager) migrateRelayUsed(oldhostinfo, newhostinfo *HostInfo
 		if err != nil {
 			cm.l.Error("failed to marshal Control message to migrate relay", "error", err)
 		} else {
-			migBuf := cm.intf.bufAlloc.Acquire()
-			cm.intf.SendMessageToHostInfo(header.Control, 0, newhostinfo, msg, migBuf)
-			cm.intf.bufAlloc.Release(migBuf)
+			cm.intf.SendMessageToHostInfo(header.Control, 0, newhostinfo, msg, make([]byte, 12), make([]byte, mtu))
 			cm.l.Info("send CreateRelayRequest",
-				"relayFrom", req.RelayFromAddr,
-				"relayTo", req.RelayToAddr,
+				"relayFrom", relayFrom,
+				"relayTo", relayTo,
 				"initiatorRelayIndex", req.InitiatorRelayIndex,
 				"responderRelayIndex", req.ResponderRelayIndex,
 				"vpnAddrs", newhostinfo.vpnAddrs,
@@ -371,7 +357,7 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim

 		if !outTraffic {
 			// Send a punch packet to keep the NAT state alive
-			cm.sendPunch(hostinfo)
+			cm.punchy.SendPunch(hostinfo)
 		}

 		return decision, hostinfo, primary
@@ -402,17 +388,16 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim

 			// If we aren't sending or receiving traffic then its an unused tunnel and we don't to test the tunnel.
 			// Just maintain NAT state if configured to do so.
-			cm.sendPunch(hostinfo)
+			cm.punchy.SendPunch(hostinfo)
 			cm.trafficTimer.Add(hostinfo.localIndexId, cm.checkInterval)
 			return doNothing, nil, nil
 		}

-		if cm.punchy.GetTargetEverything() {
-			// This is similar to the old punchy behavior with a slight optimization.
-			// We aren't receiving traffic but we are sending it, punch on all known
-			// ips in case we need to re-prime NAT state
-			cm.sendPunch(hostinfo)
-		}
+		// We aren't receiving traffic but we are sending it. The outbound
+		// traffic itself refreshes the primary remote's NAT state; this
+		// fans out to non-primary remotes, but only if target_all_remotes
+		// is configured.
+		cm.punchy.SendPunchToAll(hostinfo)

 		if cm.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(cm.l).Debug("Tunnel status",
@@ -514,31 +499,6 @@ func (cm *connectionManager) isInvalidCertificate(now time.Time, hostinfo *HostI
 	}
 }

-func (cm *connectionManager) sendPunch(hostinfo *HostInfo) {
-	if !cm.punchy.GetPunch() {
-		// Punching is disabled
-		return
-	}
-
-	if cm.intf.lightHouse.IsAnyLighthouseAddr(hostinfo.vpnAddrs) {
-		// Do not punch to lighthouses, we assume our lighthouse update interval is good enough.
-		// In the event the update interval is not sufficient to maintain NAT state then a publicly available lighthouse
-		// would lose the ability to notify us and punchy.respond would become unreliable.
-		return
-	}
-
-	if cm.punchy.GetTargetEverything() {
-		hostinfo.remotes.ForEach(cm.hostMap.GetPreferredRanges(), func(addr netip.AddrPort, preferred bool) {
-			cm.metricsTxPunchy.Inc(1)
-			cm.intf.outside.WriteTo([]byte{1}, addr)
-		})
-
-	} else if hostinfo.remote.IsValid() {
-		cm.metricsTxPunchy.Inc(1)
-		cm.intf.outside.WriteTo([]byte{1}, hostinfo.remote)
-	}
-}
-
 func (cm *connectionManager) tryRehandshake(hostinfo *HostInfo) {
 	cs := cm.intf.pki.getCertState()
 	curCrt := hostinfo.ConnectionState.myCert
@@ -64,12 +64,12 @@ func Test_NewConnectionManagerTest(t *testing.T) {

 	// Create manager
 	conf := config.NewC(test.NewLogger())
-	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
+	punchy := NewPunchyFromConfig(test.NewLogger(), conf, nil)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
-
 	p := []byte("")
-	buf := NewWireBuffer(mtu, 0)
+	nb := make([]byte, 12, 12)
+	out := make([]byte, mtu)

 	// Add an ip we have established a connection w/ to hostmap
 	hostinfo := &HostInfo{
@@ -92,7 +92,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	assert.True(t, hostinfo.in.Load())

 	// Do a traffic check tick, should not be pending deletion but should not have any in/out packets recorded
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.False(t, hostinfo.pendingDeletion.Load())
 	assert.False(t, hostinfo.out.Load())
 	assert.False(t, hostinfo.in.Load())
@@ -100,7 +100,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	// Do another traffic check tick, this host should be pending deletion now
 	nc.Out(hostinfo)
 	assert.True(t, hostinfo.out.Load())
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.True(t, hostinfo.pendingDeletion.Load())
 	assert.False(t, hostinfo.out.Load())
 	assert.False(t, hostinfo.in.Load())
@@ -108,7 +108,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])

 	// Do a final traffic check tick, the host should now be removed
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.NotContains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs)
 	assert.NotContains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 }
@@ -146,12 +146,12 @@ func Test_NewConnectionManagerTest2(t *testing.T) {

 	// Create manager
 	conf := config.NewC(test.NewLogger())
-	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
+	punchy := NewPunchyFromConfig(test.NewLogger(), conf, nil)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
-
 	p := []byte("")
-	buf := NewWireBuffer(mtu, 0)
+	nb := make([]byte, 12, 12)
+	out := make([]byte, mtu)

 	// Add an ip we have established a connection w/ to hostmap
 	hostinfo := &HostInfo{
@@ -174,14 +174,14 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)

 	// Do a traffic check tick, should not be pending deletion but should not have any in/out packets recorded
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.False(t, hostinfo.pendingDeletion.Load())
 	assert.False(t, hostinfo.out.Load())
 	assert.False(t, hostinfo.in.Load())

 	// Do another traffic check tick, this host should be pending deletion now
 	nc.Out(hostinfo)
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.True(t, hostinfo.pendingDeletion.Load())
 	assert.False(t, hostinfo.out.Load())
 	assert.False(t, hostinfo.in.Load())
@@ -190,7 +190,7 @@ func Test_NewConnectionManagerTest2(t *testing.T) {

 	// We saw traffic, should no longer be pending deletion
 	nc.In(hostinfo)
-	nc.doTrafficCheck(hostinfo.localIndexId, p, buf, time.Now())
+	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
 	assert.False(t, hostinfo.pendingDeletion.Load())
 	assert.False(t, hostinfo.out.Load())
 	assert.False(t, hostinfo.in.Load())
@@ -233,7 +233,7 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	conf.Settings["tunnels"] = map[string]any{
 		"drop_inactive": true,
 	}
-	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
+	punchy := NewPunchyFromConfig(test.NewLogger(), conf, nil)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	assert.True(t, nc.dropInactive.Load())
 	nc.intf = ifce
@@ -358,7 +358,7 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {

 	// Create manager
 	conf := config.NewC(test.NewLogger())
-	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
+	punchy := NewPunchyFromConfig(test.NewLogger(), conf, nil)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
 	ifce.connectionManager = nc
@@ -7,13 +7,14 @@ import (

 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/handshake"
+	"github.com/slackhq/nebula/noiseutil"
 )

 const ReplayWindow = 1024

 type ConnectionState struct {
-	eKey           *NebulaCipherState
-	dKey           *NebulaCipherState
+	eKey           noiseutil.CipherState
+	dKey           noiseutil.CipherState
 	myCert         cert.Certificate
 	peerCert       *cert.CachedCertificate
 	initiator      bool
@@ -31,8 +32,8 @@ func newConnectionStateFromResult(r *handshake.Result) *ConnectionState {
 		myCert:    r.MyCert,
 		initiator: r.Initiator,
 		peerCert:  r.RemoteCert,
-		eKey:      NewNebulaCipherState(r.EKey),
-		dKey:      NewNebulaCipherState(r.DKey),
+		eKey:      noiseutil.NewCipherState(r.EKey, r.Cipher),
+		dKey:      noiseutil.NewCipherState(r.DKey, r.Cipher),
 		window:    NewBits(ReplayWindow),
 	}
 	ci.messageCounter.Add(r.MessageIndex)
@@ -278,9 +278,15 @@ func (c *Control) CloseTunnel(vpnIp netip.Addr, localOnly bool) bool {
 	}

 	if !localOnly {
-		buf := c.f.bufAlloc.Acquire()
-		c.f.send(header.CloseTunnel, 0, hostInfo.ConnectionState, hostInfo, []byte{}, buf)
-		c.f.bufAlloc.Release(buf)
+		c.f.send(
+			header.CloseTunnel,
+			0,
+			hostInfo.ConnectionState,
+			hostInfo,
+			[]byte{},
+			make([]byte, 12, 12),
+			make([]byte, mtu),
+		)
 	}

 	c.f.closeTunnel(hostInfo)
@@ -290,14 +296,11 @@ func (c *Control) CloseTunnel(vpnIp netip.Addr, localOnly bool) bool {
 // CloseAllTunnels is just like CloseTunnel except it goes through and shuts them all down, optionally you can avoid shutting down lighthouse tunnels
 // the int returned is a count of tunnels closed
 func (c *Control) CloseAllTunnels(excludeLighthouses bool) (closed int) {
-	// One WireBuffer for the whole shutdown loop.
-	buf := c.f.bufAlloc.Acquire()
-	defer c.f.bufAlloc.Release(buf)
 	shutdown := func(h *HostInfo) {
 		if excludeLighthouses && c.f.lightHouse.IsAnyLighthouseAddr(h.vpnAddrs) {
 			return
 		}
-		c.f.send(header.CloseTunnel, 0, h.ConnectionState, h, []byte{}, buf)
+		c.f.send(header.CloseTunnel, 0, h.ConnectionState, h, []byte{}, make([]byte, 12, 12), make([]byte, mtu))
 		c.f.closeTunnel(h)

 		c.l.Debug("Sending close tunnel message",
@@ -11,19 +11,21 @@ import (
 	"sync"
 	"sync/atomic"

-	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
 	"github.com/slackhq/nebula/config"
 )

 type dnsServer struct {
 	sync.RWMutex
-	l               *slog.Logger
-	ctx             context.Context
-	dnsMap4         map[string]netip.Addr
-	dnsMap6         map[string]netip.Addr
-	hostMap         *HostMap
-	myVpnAddrsTable *bart.Lite
+	l       *slog.Logger
+	ctx     context.Context
+	dnsMap4 map[string]netip.Addr
+	dnsMap6 map[string]netip.Addr
+	hostMap *HostMap
+	pki     *PKI
+
+	// selfHost is the cached FQDN we last seeded for ourselves
+	selfHost string

 	mux *dns.ServeMux

@@ -55,14 +57,14 @@ type dnsServer struct {
 // they no-op when DNS isn't enabled. Each Start invocation owns a ctx-cancel
 // watcher that tears the listener down on nebula shutdown. The returned
 // pointer is always non-nil, even on error.
-func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, cs *CertState, hostMap *HostMap, c *config.C) (*dnsServer, error) {
+func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, pki *PKI, hostMap *HostMap, c *config.C) (*dnsServer, error) {
 	ds := &dnsServer{
-		l:               l,
-		ctx:             ctx,
-		dnsMap4:         make(map[string]netip.Addr),
-		dnsMap6:         make(map[string]netip.Addr),
-		hostMap:         hostMap,
-		myVpnAddrsTable: cs.myVpnAddrsTable,
+		l:       l,
+		ctx:     ctx,
+		dnsMap4: make(map[string]netip.Addr),
+		dnsMap6: make(map[string]netip.Addr),
+		hostMap: hostMap,
+		pki:     pki,
 	}
 	ds.mux = dns.NewServeMux()
 	ds.mux.HandleFunc(".", ds.handleDnsRequest)
@@ -76,6 +78,7 @@ func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, cs *CertState,
 	if err := ds.reload(c, true); err != nil {
 		return ds, err
 	}
+	ds.seedSelf()
 	return ds, nil
 }

@@ -113,7 +116,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
-		// will repopulate from fresh handshakes.
+		// will repopulate from fresh handshakes and a fresh seedSelf.
 		d.clearRecords()
 		return nil
 	}
@@ -121,17 +124,14 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	if running == nil {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
-		return nil
+	} else if !sameAddr {
+		d.shutdownServer(running, runningStarted, "reload")
+		// Old Start goroutine has now exited; bring up a fresh listener on the new address.
+		go d.Start()
 	}

-	if sameAddr {
-		return nil
-	}
-
-	d.shutdownServer(running, runningStarted, "reload")
-	// Old Start goroutine has now exited; bring up a fresh listener on the
-	// new address.
-	go d.Start()
+	// Refresh the self entry every enabled reload so cert renewals that change our name or VPN addresses are picked up.
+	d.seedSelf()
 	return nil
 }

@@ -249,6 +249,20 @@ func (d *dnsServer) QueryCert(data string) string {
 		return ""
 	}

+	// The hostmap only ever contains peers we have handshaked with, so it never carries an entry for ourselves.
+	// Answer self lookups straight from the local cert state.
+	if cs := d.certState(); cs != nil && cs.myVpnAddrsTable != nil && cs.myVpnAddrsTable.Contains(ip) {
+		c := cs.GetDefaultCertificate()
+		if c == nil {
+			return ""
+		}
+		b, err := c.MarshalJSON()
+		if err != nil {
+			return ""
+		}
+		return string(b)
+	}
+
 	hostinfo := d.hostMap.QueryVpnAddr(ip)
 	if hostinfo == nil {
 		return ""
@@ -266,12 +280,60 @@ func (d *dnsServer) QueryCert(data string) string {
 	return string(b)
 }

-// clearRecords drops all DNS records.
+// clearRecords drops all DNS records, including the self entry.
 func (d *dnsServer) clearRecords() {
 	d.Lock()
 	defer d.Unlock()
 	clear(d.dnsMap4)
 	clear(d.dnsMap6)
+	d.selfHost = ""
+}
+
+// seedSelf inserts (or refreshes) a record for our own cert name pointing at our VPN addresses,
+// so a single-lighthouse network can resolve the lighthouse's own hostname without the two-process workaround.
+func (d *dnsServer) seedSelf() {
+	if !d.enabled.Load() {
+		return
+	}
+	cs := d.certState()
+	if cs == nil {
+		return
+	}
+	c := cs.GetDefaultCertificate()
+	if c == nil {
+		return
+	}
+	newHost := strings.ToLower(c.Name()) + "."
+
+	d.Lock()
+	defer d.Unlock()
+	if d.selfHost != "" && d.selfHost != newHost {
+		delete(d.dnsMap4, d.selfHost)
+		delete(d.dnsMap6, d.selfHost)
+	}
+	d.selfHost = newHost
+	delete(d.dnsMap4, newHost)
+	delete(d.dnsMap6, newHost)
+	haveV4, haveV6 := false, false
+	for _, addr := range cs.myVpnAddrs {
+		if addr.Is4() && !haveV4 {
+			d.dnsMap4[newHost] = addr
+			haveV4 = true
+		} else if addr.Is6() && !haveV6 {
+			d.dnsMap6[newHost] = addr
+			haveV6 = true
+		}
+		if haveV4 && haveV6 {
+			break
+		}
+	}
+}
+
+func (d *dnsServer) certState() *CertState {
+	if d.pki == nil {
+		return nil
+	}
+	return d.pki.getCertState()
 }

 // Add adds the first IPv4 and IPv6 address that appears in `addresses` as the record for `host`
@@ -309,8 +371,12 @@ func (d *dnsServer) isSelfNebulaOrLocalhost(addr string) bool {
 		return true
 	}

+	cs := d.certState()
+	if cs == nil || cs.myVpnAddrsTable == nil {
+		return false
+	}
 	//if we found it in this table, it's good
-	return d.myVpnAddrsTable.Contains(b)
+	return cs.myVpnAddrsTable.Contains(b)
 }

 func (d *dnsServer) parseQuery(m *dns.Msg, w dns.ResponseWriter) {
@@ -9,7 +9,10 @@ import (
 	"testing"
 	"time"

+	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
 	"github.com/slackhq/nebula/config"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -276,6 +279,92 @@ func TestDnsServer_Stop_beforeBind_doesNotHang(t *testing.T) {
 	}
 }

+// newTestPKI builds a minimal *PKI with a single v1 cert whose name and
+// VPN addresses are caller-provided, suitable for exercising seedSelf and
+// QueryCert self handling.
+func newTestPKI(t *testing.T, name string, addrs []netip.Addr) *PKI {
+	t.Helper()
+	networks := make([]netip.Prefix, 0, len(addrs))
+	for _, a := range addrs {
+		bits := 32
+		if a.Is6() {
+			bits = 128
+		}
+		networks = append(networks, netip.PrefixFrom(a, bits))
+	}
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil)
+	c, _, _, _ := cert_test.NewTestCert(cert.Version2, cert.Curve_CURVE25519, ca, caKey, name, time.Time{}, time.Time{}, networks, nil, nil)
+
+	addrsTable := new(bart.Lite)
+	for _, a := range addrs {
+		addrsTable.Insert(netip.PrefixFrom(a, a.BitLen()))
+	}
+
+	cs := &CertState{
+		v2Cert:            c,
+		initiatingVersion: cert.Version2,
+		myVpnAddrs:        addrs,
+		myVpnAddrsTable:   addrsTable,
+	}
+	pki := &PKI{}
+	pki.cs.Store(cs)
+	return pki
+}
+
+func TestDnsServer_seedSelf_addsOwnRecord(t *testing.T) {
+	ds, c := newTestDnsServer(t)
+	myV4 := netip.MustParseAddr("10.0.0.1")
+	myV6 := netip.MustParseAddr("fd00::1")
+	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{myV4, myV6})
+	setDnsConfig(c, "127.0.0.1", "0", true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	ds.seedSelf()
+	got4, exists := ds.Query(dns.TypeA, "lighthouse.")
+	assert.True(t, exists)
+	assert.Equal(t, myV4, got4)
+	got6, exists := ds.Query(dns.TypeAAAA, "lighthouse.")
+	assert.True(t, exists)
+	assert.Equal(t, myV6, got6)
+}
+
+func TestDnsServer_seedSelf_disabled_noOp(t *testing.T) {
+	ds, c := newTestDnsServer(t)
+	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{netip.MustParseAddr("10.0.0.1")})
+	setDnsConfig(c, "127.0.0.1", "0", true, false)
+	require.NoError(t, ds.reload(c, true))
+
+	ds.seedSelf()
+	_, exists := ds.Query(dns.TypeA, "lighthouse.")
+	assert.False(t, exists)
+}
+
+func TestDnsServer_clearRecords_dropsSelfHost(t *testing.T) {
+	ds, c := newTestDnsServer(t)
+	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{netip.MustParseAddr("10.0.0.1")})
+	setDnsConfig(c, "127.0.0.1", "0", true, true)
+	require.NoError(t, ds.reload(c, true))
+	ds.seedSelf()
+	require.NotEmpty(t, ds.selfHost)
+
+	ds.clearRecords()
+	assert.Empty(t, ds.selfHost)
+	_, exists := ds.Query(dns.TypeA, "lighthouse.")
+	assert.False(t, exists)
+}
+
+func TestDnsServer_QueryCert_returnsOwnCert(t *testing.T) {
+	ds, _ := newTestDnsServer(t)
+	myV4 := netip.MustParseAddr("10.0.0.1")
+	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{myV4})
+
+	got := ds.QueryCert(myV4.String() + ".")
+	assert.NotEmpty(t, got, "TXT lookup of our own VPN address should return our cert")
+
+	other := netip.MustParseAddr("10.0.0.99")
+	assert.Empty(t, ds.QueryCert(other.String()+"."), "unknown peer IP should return nothing")
+}
+
 func TestDnsServer_reload_disable_stopsRunningServer(t *testing.T) {
 	port := freeUDPPort(t)
 	ds, c := newTestDnsServer(t)
@@ -1,6 +1,16 @@
 FROM gcr.io/distroless/static:latest

 ARG TARGETOS TARGETARCH
+
+ARG VERSION=dev
+ARG REVISION=unknown
+LABEL org.opencontainers.image.title="nebula" \
+      org.opencontainers.image.description="A scalable overlay networking tool with a focus on performance, simplicity and security" \
+      org.opencontainers.image.vendor="Nebula OSS" \
+      org.opencontainers.image.source="https://github.com/slackhq/nebula" \
+      org.opencontainers.image.version="${VERSION}" \
+      org.opencontainers.image.revision="${REVISION}"
+
 COPY build/$TARGETOS-$TARGETARCH/nebula /nebula
 COPY build/$TARGETOS-$TARGETARCH/nebula-cert /nebula-cert

@@ -1,68 +0,0 @@
-//go:build e2e_testing
-// +build e2e_testing
-
-package e2e
-
-import (
-	"testing"
-	"time"
-
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/cert_test"
-	"github.com/slackhq/nebula/e2e/router"
-)
-
-// BenchmarkHandshake measures end-to-end tunnel establishment time. The two
-// nodes and the router are constructed once before the loop so the timed window
-// is just the handshake itself: trigger packet -> handshake1 -> handshake2 ->
-// cached packet replay -> arrival on the remote TUN. Between iterations we
-// tear down both sides locally (no CloseTunnel notification on the wire) and
-// re-inject the lighthouse address that closeTunnel cleared, so the next
-// iteration runs through a fresh handshake against the same harness.
-func BenchmarkHandshake(b *testing.B) {
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-
-	// Default try_interval is 100ms. The handshake manager schedules handshake1
-	// on its OutboundHandshakeTimer rather than firing immediately on trigger
-	// (the trigger channel only fast-paths static hosts), so a 100ms default
-	// drowns the actual handshake cost. Drop it to 1ms so the bench reflects
-	// the computation, not the wheel cadence.
-	bovr := m{"handshakes": m{"try_interval": "1ms"}}
-	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.1/24", bovr)
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.2/24", bovr)
-
-	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-
-	myControl.Start()
-	theirControl.Start()
-	defer myControl.Stop()
-	defer theirControl.Stop()
-
-	r := router.NewR(b, myControl, theirControl)
-	r.CancelFlowLogs()
-	r.EnableFanIn()
-
-	trigger := BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me"))
-
-	b.ReportAllocs()
-	b.ResetTimer()
-
-	for n := 0; n < b.N; n++ {
-		myControl.InjectTunPacket(trigger)
-		// RouteForAllUntilTxTun returns the moment the cached packet arrives at
-		// the remote TUN, which is also when both sides are fully established.
-		_ = r.RouteForAllUntilTxTun(theirControl)
-
-		b.StopTimer()
-		// Local-only close removes hostmap state on both sides without putting a
-		// CloseTunnel packet on the wire that we'd then have to drain. The
-		// closeTunnel path also clears learned lighthouse state for the peer
-		// when the last hostinfo for that addr goes away, so we re-inject.
-		myControl.CloseTunnel(theirVpnIpNet[0].Addr(), true)
-		theirControl.CloseTunnel(myVpnIpNet[0].Addr(), true)
-		myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-		theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-		b.StartTimer()
-	}
-}
@@ -165,7 +165,7 @@ func TestGoodHandshakeNoOverlap(t *testing.T) {

 	empty := []byte{}
 	t.Log("do something to cause a handshake")
-	myControl.GetF().SendMessageToVpnAddr(header.Test, header.MessageNone, theirVpnIpNet[0].Addr(), empty, nebula.NewWireBuffer(9001, 0))
+	myControl.GetF().SendMessageToVpnAddr(header.Test, header.MessageNone, theirVpnIpNet[0].Addr(), empty, empty, empty)

 	t.Log("Have them consume my stage 0 packet. They have a tunnel now")
 	theirControl.InjectUDPPacket(myControl.GetFromUDP(true))
@@ -18,14 +18,10 @@ import (
 // retry mechanism gives the wg.Wait()-driven goroutines a moment to drain
 // before failing the assertion.
 //
-// IgnoreCurrent is necessary in the parallelized suite: other tests can
-// leave goroutines mid-shutdown when this one runs (Stop is async, the
-// wg.Wait() drain is not blocking on test return). We're checking that
-// *this* test's setup tears down cleanly, not that the whole suite is
-// idle at this moment. Intentionally NOT t.Parallel()'d for the same
-// reason — concurrent test goroutines would always show up.
+// Intentionally NOT t.Parallel()'d: concurrent tests would have their own
+// goroutines running and trip the assertion.
 func TestNoGoroutineLeaks(t *testing.T) {
-	defer goleak.VerifyNone(t, goleak.IgnoreCurrent())
+	defer goleak.VerifyNone(t)

 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
@@ -0,0 +1,125 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"crypto/ed25519"
+	"crypto/rand"
+	"encoding/pem"
+	"net"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/crypto/ssh"
+)
+
+func TestSSHDLifecycle(t *testing.T) {
+	// TestSSHDLifecycle exercises the in-process sshd through several config reloads and a Control.Stop.
+	ca, _, caKey, _ := cert_test.NewTestCaCert(
+		cert.Version1, cert.Curve_CURVE25519,
+		time.Now(), time.Now().Add(10*time.Minute),
+		nil, nil, []string{},
+	)
+
+	hostKeyPEM := generateSSHHostKey(t)
+	clientSigner, clientAuthKey := generateSSHClientKey(t)
+	sshdAddr := allocLoopbackPort(t)
+
+	overrides := m{
+		"sshd": m{
+			"enabled":  true,
+			"listen":   sshdAddr,
+			"host_key": hostKeyPEM,
+			"authorized_users": []m{{
+				"user": "tester",
+				"keys": []string{clientAuthKey},
+			}},
+		},
+	}
+	control, _, _, _ := newSimpleServer(cert.Version1, ca, caKey, "sshd-test", "10.222.0.1/24", overrides)
+	control.Start()
+	t.Cleanup(func() { control.Stop() })
+
+	// sshd binds in a goroutine after Start returns; wait for it.
+	require.Eventually(t, func() bool { return canDial(sshdAddr) }, 2*time.Second, 25*time.Millisecond,
+		"sshd never started listening")
+
+	for i := 1; i <= 3; i++ {
+		out := sshExecReload(t, sshdAddr, clientSigner)
+		assert.Contains(t, out, "Reloading config", "reload cycle %d", i)
+		require.Eventually(t, func() bool { return canDial(sshdAddr) }, 2*time.Second, 25*time.Millisecond,
+			"sshd not listening after reload cycle %d", i)
+	}
+
+	control.Stop()
+	require.Eventually(t, func() bool { return !canDial(sshdAddr) }, 2*time.Second, 25*time.Millisecond,
+		"sshd still listening after Control.Stop")
+}
+
+func canDial(addr string) bool {
+	c, err := net.DialTimeout("tcp", addr, 100*time.Millisecond)
+	if err != nil {
+		return false
+	}
+	_ = c.Close()
+	return true
+}
+
+// allocLoopbackPort grabs an unused TCP port on 127.0.0.1, closes it, and returns the address. There
+// is a small race between releasing the port and the sshd reclaiming it; in practice the OS keeps the
+// port available long enough for the test to bind it.
+func allocLoopbackPort(t *testing.T) string {
+	t.Helper()
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	require.NoError(t, err)
+	addr := l.Addr().String()
+	require.NoError(t, l.Close())
+	return addr
+}
+
+func generateSSHHostKey(t *testing.T) string {
+	t.Helper()
+	_, priv, err := ed25519.GenerateKey(rand.Reader)
+	require.NoError(t, err)
+	block, err := ssh.MarshalPrivateKey(priv, "nebula-e2e-host")
+	require.NoError(t, err)
+	return string(pem.EncodeToMemory(block))
+}
+
+func generateSSHClientKey(t *testing.T) (ssh.Signer, string) {
+	t.Helper()
+	_, priv, err := ed25519.GenerateKey(rand.Reader)
+	require.NoError(t, err)
+	signer, err := ssh.NewSignerFromKey(priv)
+	require.NoError(t, err)
+	auth := strings.TrimSpace(string(ssh.MarshalAuthorizedKey(signer.PublicKey())))
+	return signer, auth
+}
+
+func sshExecReload(t *testing.T, addr string, signer ssh.Signer) string {
+	t.Helper()
+	cfg := &ssh.ClientConfig{
+		User:            "tester",
+		Auth:            []ssh.AuthMethod{ssh.PublicKeys(signer)},
+		HostKeyCallback: ssh.InsecureIgnoreHostKey(),
+		Timeout:         2 * time.Second,
+	}
+	client, err := ssh.Dial("tcp", addr, cfg)
+	require.NoError(t, err)
+	defer client.Close()
+
+	sess, err := client.NewSession()
+	require.NoError(t, err)
+	defer sess.Close()
+
+	// reload tears the channel down before sending exit-status, so Output returns an error on the
+	// channel close. The output buffer still has whatever the reload callback wrote before that.
+	out, _ := sess.Output("reload")
+	return string(out)
+}
@@ -15,6 +15,7 @@ import (
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/udp"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 	"gopkg.in/yaml.v3"
 )

@@ -373,6 +374,100 @@ func TestCrossStackRelaysWork(t *testing.T) {
 	//relayControl.Stop()
 }

+// TestRelayReplayProtection asserts that a relay (forwarding-type) node rejects
+// replayed relay frames. A captured relay frame, re-injected with the same
+// message counter, must be dropped by the replay window rather than re-forwarded
+// to the relay target. Before the fix, handleOutsideRelayPacket authenticated the
+// frame but never advanced the replay window, so every replay was re-forwarded.
+func TestRelayReplayProtection(t *testing.T) {
+	t.Parallel()
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me     ", "10.128.0.1/24,fc00::1/64", m{"relay": m{"use_relays": true}})
+	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "relay  ", "10.128.0.128/24,fc00::128/64", m{"relay": m{"am_relay": true}})
+	theirUdp := netip.MustParseAddrPort("10.0.0.2:4242")
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServerWithUdp(cert.Version2, ca, caKey, "them   ", "fc00::2/64", theirUdp, m{"relay": m{"use_relays": true}})
+
+	myVpnV6 := myVpnIpNet[1]
+	relayVpnV4 := relayVpnIpNet[0]
+	relayVpnV6 := relayVpnIpNet[1]
+	theirVpnV6 := theirVpnIpNet[0]
+
+	// Teach me how to reach the relay and that them is reachable via the relay
+	myControl.InjectLightHouseAddr(relayVpnV4.Addr(), relayUdpAddr)
+	myControl.InjectLightHouseAddr(relayVpnV6.Addr(), relayUdpAddr)
+	myControl.InjectRelays(theirVpnV6.Addr(), []netip.Addr{relayVpnV6.Addr()})
+	relayControl.InjectLightHouseAddr(theirVpnV6.Addr(), theirUdpAddr)
+
+	r := router.NewR(t, myControl, relayControl, theirControl)
+	defer r.RenderFlow()
+
+	myControl.Start()
+	relayControl.Start()
+	theirControl.Start()
+
+	// Establish the relayed tunnel in both directions so all handshakes complete.
+	t.Log("Establish the relayed tunnel")
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnV6.Addr(), 80, myVpnV6.Addr(), 80, []byte("Hi from me")))
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assertUdpPacket(t, []byte("Hi from me"), p, myVpnV6.Addr(), theirVpnV6.Addr(), 80, 80)
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnV6.Addr(), 80, theirVpnV6.Addr(), 80, []byte("Hi from them")))
+	p = r.RouteForAllUntilTxTun(myControl)
+	assertUdpPacket(t, []byte("Hi from them"), p, theirVpnV6.Addr(), myVpnV6.Addr(), 80, 80)
+
+	// Drain anything still queued on me's UDP tx so the next packet we pull is the
+	// relay frame we are about to generate.
+	for myControl.GetFromUDP(false) != nil {
+	}
+
+	// Capture a single legitimate relay frame that me transmits toward the relay.
+	t.Log("Capture a relay frame from me -> relay")
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnV6.Addr(), 80, myVpnV6.Addr(), 80, []byte("replay me")))
+	relayFrame := myControl.GetFromUDP(true)
+	require.Equal(t, relayUdpAddr, relayFrame.To, "captured frame should be addressed to the relay")
+	var fh header.H
+	require.NoError(t, fh.Parse(relayFrame.Data))
+	require.Equal(t, header.Message, fh.Type)
+	require.Equal(t, header.MessageRelay, fh.Subtype)
+
+	// drainForwards counts relay frames the relay forwards toward them within the
+	// settle window. We match on destination + (Message, MessageRelay) so the
+	// relay's own direct traffic to them can't be miscounted.
+	drainForwards := func(settle time.Duration) int {
+		ch := relayControl.GetUDPTxChan()
+		count := 0
+		for {
+			select {
+			case pkt := <-ch:
+				var ph header.H
+				if pkt.To == theirUdpAddr && ph.Parse(pkt.Data) == nil &&
+					ph.Type == header.Message && ph.Subtype == header.MessageRelay {
+					count++
+				}
+				pkt.Release()
+			case <-time.After(settle):
+				return count
+			}
+		}
+	}
+
+	// First delivery of the captured frame: the relay should forward it once.
+	t.Log("Deliver the captured frame once; relay forwards it to them")
+	relayControl.InjectUDPPacket(relayFrame)
+	require.Equal(t, 1, drainForwards(200*time.Millisecond), "relay should forward the first, legitimate copy")
+
+	// Replay the exact same frame several times. A correct replay window rejects
+	// these duplicates so the relay forwards none of them.
+	t.Log("Replay the captured frame; relay must drop the duplicates")
+	const replays = 3
+	for i := 0; i < replays; i++ {
+		relayControl.InjectUDPPacket(relayFrame)
+	}
+	forwarded := drainForwards(200 * time.Millisecond)
+	assert.Equal(t, 0, forwarded, "relay re-forwarded %d/%d replayed relay frames; replay protection is ineffective on relay tunnels", forwarded, replays)
+
+	r.RenderHostmaps("Final hostmaps", myControl, relayControl, theirControl)
+}
+
 func TestCloseTunnelAuthenticated(t *testing.T) {
 	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
@@ -138,6 +138,14 @@ listen:
  # max, net.core.rmem_max and net.core.wmem_max
  #read_buffer: 10485760
  #write_buffer: 10485760
+
+  # On Windows only
+  # When true, Nebula installs a WFP (Windows Filtering Platform) PERMIT filter scoped to UDP at the listener port.
+  # WFP sits below Windows Defender Firewall, so this lets peer handshakes reach Nebula's outside socket regardless
+  # of WDF's inbound rules.
+  # Default true; set to false to leave WDF in charge of inbound decisions on the listener port. Not reloadable.
+  #windows_bypass_wdf: true
+
  # By default, Nebula replies to packets it has no tunnel for with a "recv_error" packet. This packet helps speed up reconnection
  # in the case that Nebula on either side did not shut down cleanly. This response can be abused as a way to discover if Nebula is running
  # on a host though. This option lets you configure if you want to send "recv_error" packets always, never, or only to private network remotes.
@@ -163,17 +171,21 @@ listen:

 punchy:
  # Continues to punch inbound/outbound at a regular interval to avoid expiration of firewall nat mappings
+  # This setting is reloadable.
  punch: true

  # respond means that a node you are trying to reach will connect back out to you if your hole punching fails
  # this is extremely useful if one node is behind a difficult nat, such as a symmetric NAT
  # Default is false
+  # This setting is reloadable.
  #respond: true

  # delays a punch response for misbehaving NATs, default is 1 second.
+  # This setting is reloadable.
  #delay: 1s

  # set the delay before attempting punchy.respond. Default is 5 seconds. respond must be true to take effect.
+  # This setting is reloadable.
  #respond_delay: 5s

 # Cipher allows you to choose between the available ciphers for your network. Options are chachapoly or aes
@@ -282,6 +294,24 @@ tun:
    #  metric: 100
    #  install: true

+  # On Windows only, sets the network category of the nebula interface. Without this, Windows often
+  # leaves the network as "Unidentified" and treats it as Public, which makes the host firewall more
+  # restrictive than you usually want for an overlay between trusted peers. Valid values:
+  #   private             - treat the nebula network as a private/trusted network (default)
+  #   public              - treat it as a public/untrusted network
+  #   domain              - treat it as a domain-authenticated network
+  #   unset               - leave whatever Windows decided alone
+  # Not reloadable.
+  #network_category: private
+
+  # On Windows only
+  # When true, Nebula installs a WFP (Windows Filtering Platform) PERMIT filter scoped to the nebula adapter LUID.
+  # WFP sits below Windows Defender Firewall, so this lets inbound traffic through regardless of WDF rules.
+  # Filters are auto-removed when the adapter goes away.
+  # See listen.windows_bypass_wdf for the matching control over inbound to nebula's outside UDP listener.
+  # Default true; set to false to leave WDF in charge of inbound decisions on the nebula interface. Not reloadable.
+  #windows_bypass_wdf: true
+
  # On linux only, set to true to manage unsafe routes directly on the system route table with gateway routes instead of
  # in nebula configuration files. Default false, not reloadable.
  #use_system_route_table: false
@@ -367,7 +397,7 @@ firewall:
  #   `drop` (default): silently drop the packet.
  #   `reject`: send a reject reply.
  #     - For TCP, this will be a RST "Connection Reset" packet.
-  #     - For other protocols, this will be an ICMP port unreachable packet.
+  #     - For other protocols, this will be an ICMP "Destination unreachable: Communication administratively prohibited" packet.
  outbound_action: drop
  inbound_action: drop

@@ -58,8 +58,9 @@ type Firewall struct {
 	routableNetworks *bart.Lite

 	// assignedNetworks is a list of vpn networks assigned to us in the certificate.
-	assignedNetworks  []netip.Prefix
-	hasUnsafeNetworks bool
+	assignedNetworks []netip.Prefix
+	// unsafeNetworks is the list of unsafe networks issued to us in the certificate
+	unsafeNetworks []netip.Prefix

 	rules        string
 	rulesVersion uint16
@@ -158,10 +159,9 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 		assignedNetworks = append(assignedNetworks, network)
 	}

-	hasUnsafeNetworks := false
-	for _, n := range c.UnsafeNetworks() {
+	unsafeNetworks := c.UnsafeNetworks()
+	for _, n := range unsafeNetworks {
 		routableNetworks.Insert(n)
-		hasUnsafeNetworks = true
 	}

 	return &Firewall{
@@ -169,15 +169,15 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 			Conns:      make(map[firewall.Packet]*conn),
 			TimerWheel: NewTimerWheel[firewall.Packet](tmin, tmax),
 		},
-		InRules:           newFirewallTable(),
-		OutRules:          newFirewallTable(),
-		TCPTimeout:        tcpTimeout,
-		UDPTimeout:        UDPTimeout,
-		DefaultTimeout:    defaultTimeout,
-		routableNetworks:  routableNetworks,
-		assignedNetworks:  assignedNetworks,
-		hasUnsafeNetworks: hasUnsafeNetworks,
-		l:                 l,
+		InRules:          newFirewallTable(),
+		OutRules:         newFirewallTable(),
+		TCPTimeout:       tcpTimeout,
+		UDPTimeout:       UDPTimeout,
+		DefaultTimeout:   defaultTimeout,
+		routableNetworks: routableNetworks,
+		assignedNetworks: assignedNetworks,
+		unsafeNetworks:   unsafeNetworks,
+		l:                l,

 		incomingMetrics: firewallMetrics{
 			droppedLocalAddr:  metrics.GetOrRegisterCounter("firewall.incoming.dropped.local_addr", nil),
@@ -897,7 +897,7 @@ func (flc *firewallLocalCIDR) addRule(f *Firewall, localCidr string) error {
 	}

 	if localCidr == "" {
-		if !f.hasUnsafeNetworks || f.defaultLocalCIDRAny {
+		if len(f.unsafeNetworks) == 0 || f.defaultLocalCIDRAny {
 			flc.Any = true
 			return nil
 		}
@@ -1055,7 +1055,6 @@ func (r *rule) sanity() error {
 }

 func parsePort(s string) (int32, int32, error) {
-	var err error
 	const notAPort int32 = -2
 	if s == "any" {
 		return firewall.PortAny, firewall.PortAny, nil
@@ -1064,11 +1063,11 @@ func parsePort(s string) (int32, int32, error) {
 		return firewall.PortFragment, firewall.PortFragment, nil
 	}
 	if !strings.Contains(s, `-`) {
-		rPort, err := strconv.Atoi(s)
+		rPort, err := parsePortValue("", s)
 		if err != nil {
-			return notAPort, notAPort, fmt.Errorf("was not a number; `%s`", s)
+			return notAPort, notAPort, err
 		}
-		return int32(rPort), int32(rPort), nil
+		return rPort, rPort, nil
 	}

 	sPorts := strings.SplitN(s, `-`, 2)
@@ -1079,22 +1078,40 @@ func parsePort(s string) (int32, int32, error) {
 		return notAPort, notAPort, fmt.Errorf("appears to be a range but could not be parsed; `%s`", s)
 	}

-	rStartPort, err := strconv.Atoi(sPorts[0])
+	startPort, err := parsePortValue("beginning range ", sPorts[0])
 	if err != nil {
-		return notAPort, notAPort, fmt.Errorf("beginning range was not a number; `%s`", sPorts[0])
+		return notAPort, notAPort, err
 	}

-	rEndPort, err := strconv.Atoi(sPorts[1])
+	endPort, err := parsePortValue("ending range ", sPorts[1])
 	if err != nil {
-		return notAPort, notAPort, fmt.Errorf("ending range was not a number; `%s`", sPorts[1])
+		return notAPort, notAPort, err
 	}

-	startPort := int32(rStartPort)
-	endPort := int32(rEndPort)
-
 	if startPort == firewall.PortAny {
 		endPort = firewall.PortAny
 	}

 	return startPort, endPort, nil
 }
+
+// parsePortValue accepts a base-10 decimal in [0, 65535] and returns it
+// widened to int32. Using strconv.ParseUint with bitSize 16 rejects
+// negative input, out-of-range input (>65535), and any non-decimal byte
+// by construction, so the int32 widening that follows is provably safe
+// and cannot collide with firewall.PortAny (0) or firewall.PortFragment
+// (-1) via integer truncation.
+//
+// prefix is prepended to both error messages so callers can disambiguate
+// the single-port path (prefix="") from the range bounds (prefix="beginning
+// range " / "ending range "), preserving the historical error strings.
+func parsePortValue(prefix, s string) (int32, error) {
+	n, err := strconv.ParseUint(s, 10, 16)
+	if err == nil {
+		return int32(n), nil
+	}
+	if errors.Is(err, strconv.ErrRange) {
+		return 0, fmt.Errorf("%sout of range [0,65535]; `%s`", prefix, s)
+	}
+	return 0, fmt.Errorf("%swas not a number; `%s`", prefix, s)
+}
@@ -1029,6 +1029,75 @@ func Test_parsePort(t *testing.T) {
 	require.NoError(t, err)
 }

+// Test_parsePort_invalid covers inputs that must error. The named bug is
+// that int32(strconv.Atoi("4294967296")) truncates to 0 == firewall.PortAny,
+// silently turning a typo into a match-all-ports rule; the rest are
+// representative syntax/range probes.
+func Test_parsePort_invalid(t *testing.T) {
+	tests := []struct {
+		name            string
+		input           string
+		wantErrContains string
+	}{
+		// Numeric overflow (the named bug + boundary).
+		{"named bug: 2^32 truncates to PortAny", "4294967296", "out of range"},
+		{"just above max real port", "65536", "out of range"},
+
+		// Negatives route through the range branch and hit the empty-half
+		// guard; included as defense in depth so a future refactor cannot
+		// accidentally reach the int32 cast.
+		{"negative", "-1", "could not be parsed"},
+
+		// Syntax probes.
+		{"NUL between digits", "4\x002", "was not a number"},
+		{"hex notation", "0x10", "was not a number"},
+		{"scientific notation", "1e3", "was not a number"},
+		{"leading whitespace", " 42", "was not a number"},
+		{"fullwidth digits", "４２", "was not a number"},
+
+		// Range branch.
+		{"range upper out of range", "1-65536", "ending range out of range"},
+		{"range lower out of range", "65536-65537", "beginning range out of range"},
+		{"range with negative upper", "1--1", "ending range was not a number"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, _, err := parsePort(tc.input)
+			require.Error(t, err, "input %q must error", tc.input)
+			require.ErrorContains(t, err, tc.wantErrContains)
+		})
+	}
+}
+
+// Test_parsePort_valid_boundaries locks in success cases at 0, 1, and 65535
+// so a future refactor cannot regress the boundaries.
+func Test_parsePort_valid_boundaries(t *testing.T) {
+	tests := []struct {
+		name      string
+		input     string
+		wantStart int32
+		wantEnd   int32
+	}{
+		{"zero is PortAny", "0", 0, 0},
+		{"min real port", "1", 1, 1},
+		{"max real port", "65535", 65535, 65535},
+		{"range zero to max forces end to zero", "0-65535", 0, 0},
+		{"range max to max", "65535-65535", 65535, 65535},
+		{"range one to max", "1-65535", 1, 65535},
+		{"range with whitespace inside", " 1 - 2    ", 1, 2},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			s, e, err := parsePort(tc.input)
+			require.NoError(t, err)
+			assert.Equal(t, tc.wantStart, s, "start port")
+			assert.Equal(t, tc.wantEnd, e, "end port")
+		})
+	}
+}
+
 func TestNewFirewallFromConfig(t *testing.T) {
 	l := test.NewLogger()
 	// Test a bad rule definition
@@ -9,7 +9,7 @@ require (
 	github.com/armon/go-radix v1.0.0
 	github.com/cyberdelia/go-metrics-graphite v0.0.0-20161219230853-39f87cc3b432
 	github.com/flynn/noise v1.1.0
-	github.com/gaissmai/bart v0.26.0
+	github.com/gaissmai/bart v0.28.0
 	github.com/gogo/protobuf v1.3.2
 	github.com/google/gopacket v1.1.19
 	github.com/kardianos/service v1.2.4
@@ -24,12 +24,12 @@ require (
 	github.com/vishvananda/netlink v1.3.1
 	go.uber.org/goleak v1.3.0
 	go.yaml.in/yaml/v3 v3.0.4
-	golang.org/x/crypto v0.50.0
+	golang.org/x/crypto v0.53.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.52.0
-	golang.org/x/sync v0.20.0
-	golang.org/x/sys v0.43.0
-	golang.org/x/term v0.42.0
+	golang.org/x/net v0.56.0
+	golang.org/x/sync v0.21.0
+	golang.org/x/sys v0.46.0
+	golang.org/x/term v0.44.0
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2
 	golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b
 	golang.zx2c4.com/wireguard/windows v0.6.1
@@ -26,8 +26,8 @@ github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/flynn/noise v1.1.0 h1:KjPQoQCEFdZDiP03phOvGi11+SVVhBG2wOWAorLsstg=
 github.com/flynn/noise v1.1.0/go.mod h1:xbMo+0i6+IGbYdJhF31t2eR1BIU0CYc12+BNAKwUTag=
-github.com/gaissmai/bart v0.26.0 h1:xOZ57E9hJLBiQaSyeZa9wgWhGuzfGACgqp4BE77OkO0=
-github.com/gaissmai/bart v0.26.0/go.mod h1:GREWQfTLRWz/c5FTOsIw+KkscuFkIV5t8Rp7Nd1Td5c=
+github.com/gaissmai/bart v0.28.0 h1:89yZLo8NmyqD0RYgJ3QO9HhqqGGw+oWhf90cZm69Lko=
+github.com/gaissmai/bart v0.28.0/go.mod h1:GREWQfTLRWz/c5FTOsIw+KkscuFkIV5t8Rp7Nd1Td5c=
 github.com/go-kit/kit v0.8.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/kit v0.9.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/log v0.1.0/go.mod h1:zbhenjAZHb184qTLMA9ZjW7ThYL0H2mk7Q6pNt4vbaY=
@@ -162,8 +162,8 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/crypto v0.0.0-20210322153248-0c34fe9e7dc2/go.mod h1:T9bdIzuCu7OtxOm1hfPfRQxPLYneinmdGuTeoZ9dtd4=
-golang.org/x/crypto v0.50.0 h1:zO47/JPrL6vsNkINmLoo/PH1gcxpls50DNogFvB5ZGI=
-golang.org/x/crypto v0.50.0/go.mod h1:3muZ7vA7PBCE6xgPX7nkzzjiUq87kRItoJQM1Yo8S+Q=
+golang.org/x/crypto v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=
+golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090 h1:Di6/M8l0O2lCLc6VVRWhgCiApHV8MnQurBnFSHsQtNY=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090/go.mod h1:FXUEEKJgO7OQYeo8N01OfiKP8RXMtf6e8aTskBGqWdc=
 golang.org/x/lint v0.0.0-20200302205851-738671d3881b/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
@@ -182,8 +182,8 @@ golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLL
 golang.org/x/net v0.0.0-20200625001655-4c5254603344/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
-golang.org/x/net v0.52.0 h1:He/TN1l0e4mmR3QqHMT2Xab3Aj3L9qjbhRm78/6jrW0=
-golang.org/x/net v0.52.0/go.mod h1:R1MAz7uMZxVMualyPXb+VaqGSa3LIaUqk0eEt3w36Sw=
+golang.org/x/net v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=
+golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=
 golang.org/x/oauth2 v0.0.0-20190226205417-e64efc72b421/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181221193216-37e7f081c4d4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -191,8 +191,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201207232520-09787c993a3a/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
-golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
+golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.0.0-20180905080454-ebe1bf3edb33/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20181116152217-5ac8a444bdc5/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
@@ -208,11 +208,11 @@ golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210603081109-ebe580a85c40/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.2.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.43.0 h1:Rlag2XtaFTxp19wS8MXlJwTvoh8ArU6ezoyFsMyCTNI=
-golang.org/x/sys v0.43.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
+golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.42.0 h1:UiKe+zDFmJobeJ5ggPwOshJIVt6/Ft0rcfrXZDLWAWY=
-golang.org/x/term v0.42.0/go.mod h1:Dq/D+snpsbazcBG5+F9Q1n2rXV8Ma+71xEjTRufARgY=
+golang.org/x/term v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
+golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.2/go.mod h1:bEr9sfX3Q8Zfm5fL9x+3itogRgK3+ptLWKqgva+5dAk=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
@@ -13,6 +13,7 @@ var (
 	ErrUnknownSubtype          = errors.New("unknown handshake subtype")
 	ErrMissingContent          = errors.New("expected handshake content but message was empty")
 	ErrUnexpectedContent       = errors.New("received unexpected handshake content")
+	ErrInvalidRemoteIndex      = errors.New("peer sent an invalid index in handshake payload")
 	ErrIndexAllocation         = errors.New("failed to allocate local index")
 	ErrNoCredential            = errors.New("no handshake credential available for cert version")
 	ErrAsymmetricCipherKeys    = errors.New("noise produced only one cipher key")
@@ -31,6 +31,7 @@ type CertVerifier func(cert.Certificate) (*cert.CachedCertificate, error)
 type Result struct {
 	EKey          *noise.CipherState
 	DKey          *noise.CipherState
+	Cipher        noise.CipherFunc // identifies which post-handshake CipherState the data plane should wrap EKey/DKey in
 	MyCert        cert.Certificate
 	RemoteCert    *cert.CachedCertificate
 	RemoteIndex   uint32
@@ -105,6 +106,7 @@ func NewMachine(
 		myVersion:  version,
 		result: &Result{
 			Initiator: initiator,
+			Cipher:    cred.cipherSuite,
 		},
 	}, nil
 }
@@ -310,11 +312,19 @@ func (m *Machine) processPayload(msg []byte, flags msgFlags) error {

 	// Process payload
 	if flags.expectsPayload {
+		var remoteIndex uint32
 		if m.result.Initiator {
-			m.result.RemoteIndex = payload.ResponderIndex
+			remoteIndex = payload.ResponderIndex
 		} else {
-			m.result.RemoteIndex = payload.InitiatorIndex
+			remoteIndex = payload.InitiatorIndex
 		}
+		// The payload presence check above can be satisfied by Time alone, so a payload
+		// could still carry a zero index here. We need to reject it.
+		if remoteIndex == 0 {
+			m.failed = true
+			return ErrInvalidRemoteIndex
+		}
+		m.result.RemoteIndex = remoteIndex
 		m.result.HandshakeTime = payload.Time
 		m.payloadSet = true
 	}
@@ -229,6 +229,24 @@ func TestMachineProcessPayload(t *testing.T) {
 		require.ErrorIs(t, err, ErrUnexpectedContent)
 		assert.True(t, m.Failed())
 	})
+
+	t.Run("zero initiator index on responder is fatal", func(t *testing.T) {
+		m := newTestMachine(t, cs, v, false, 100)
+		bytes := MarshalPayload(nil, Payload{InitiatorIndex: 0, Time: 1})
+		err := m.processPayload(bytes, msgFlags{expectsPayload: true})
+		require.ErrorIs(t, err, ErrInvalidRemoteIndex)
+		assert.True(t, m.Failed())
+		assert.Zero(t, m.result.RemoteIndex)
+	})
+
+	t.Run("zero responder index on initiator is fatal", func(t *testing.T) {
+		m := newTestMachine(t, cs, v, true, 100)
+		bytes := MarshalPayload(nil, Payload{InitiatorIndex: 100, ResponderIndex: 0, Time: 1})
+		err := m.processPayload(bytes, msgFlags{expectsPayload: true})
+		require.ErrorIs(t, err, ErrInvalidRemoteIndex)
+		assert.True(t, m.Failed())
+		assert.Zero(t, m.result.RemoteIndex)
+	})
 }

 // TestMachineRequireComplete checks the fail-on-incomplete-handshake path
@@ -83,6 +83,7 @@ type HandshakeHostInfo struct {
 	initiatingVersionOverride cert.Version     // Should we use a non-default cert version for this handshake?
 	counter                   int64            // How many attempts have we made so far
 	lastRemotes               []netip.AddrPort // Remotes that we sent to during the previous attempt
+	lastRelays                []netip.Addr     // Relays we attempted to use during the previous attempt
 	packetStore               []*cachedPacket  // A set of packets to be transmitted once the handshake completes

 	hostinfo *HostInfo
@@ -217,7 +218,6 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		fields := []any{
 			"udpAddrs", hh.hostinfo.remotes.CopyAddrs(hm.mainHostMap.GetPreferredRanges()),
 			"initiatorIndex", hh.hostinfo.localIndexId,
-			"remoteIndex", hh.hostinfo.remoteIndexId,
 			"durationNs", time.Since(hh.startTime).Nanoseconds(),
 		}
 		// hh.machine can be nil here if buildStage0Packet never succeeded
@@ -323,7 +323,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		)
 	}

-	hm.f.relayManager.StartRelays(hm.f, vpnIp, hostinfo, stage0)
+	hm.f.relayManager.StartRelays(hm.f, vpnIp, hh, stage0)

 	// If a lighthouse triggered this attempt then we are still in the timer wheel and do not need to re-add
 	if !lighthouseTriggered {
@@ -465,7 +465,6 @@ func (hm *HandshakeManager) CheckAndComplete(hostinfo *HostInfo, handshakePacket
 		// We have a collision, but this can happen since we can't control
 		// the remote ID. Just log about the situation as a note.
 		hostinfo.logger(hm.l).Info("New host shadows existing host remoteIndex",
-			"remoteIndex", hostinfo.remoteIndexId,
 			"collision", existingRemoteIndex.vpnAddrs,
 		)
 	}
@@ -488,7 +487,6 @@ func (hm *HandshakeManager) Complete(hostinfo *HostInfo, f *Interface) {
 		// We have a collision, but this can happen since we can't control
 		// the remote ID. Just log about the situation as a note.
 		hostinfo.logger(hm.l).Info("New host shadows existing host remoteIndex",
-			"remoteIndex", hostinfo.remoteIndexId,
 			"collision", existingRemoteIndex.vpnAddrs,
 		)
 	}
@@ -798,7 +796,6 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 	}

 	hm.sendHandshakeResponse(via, response, hostinfo, false)
-	f.connectionManager.AddTrafficWatch(hostinfo)
 	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)

 	// Don't wait for UpdateWorker
@@ -965,17 +962,16 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)

 	hm.Complete(hostinfo, f)
-	f.connectionManager.AddTrafficWatch(hostinfo)

 	if len(hh.packetStore) > 0 {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("Sending stored packets", "count", len(hh.packetStore))
 		}
-		buf := f.bufAlloc.Acquire()
+		nb := make([]byte, 12, 12)
+		out := make([]byte, mtu)
 		for _, cp := range hh.packetStore {
-			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, buf)
+			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, nb, out)
 		}
-		f.bufAlloc.Release(buf)
 		f.cachedPacketMetrics.sent.Inc(int64(len(hh.packetStore)))
 	}

@@ -1085,9 +1081,7 @@ func (hm *HandshakeManager) sendHandshakeResponse(via ViaSender, msg []byte, hos
 		// We received a valid handshake on this relay, so make sure the relay
 		// state reflects that, in case it had been marked Disestablished.
 		via.relayHI.relayState.UpdateRelayForByIdxState(via.remoteIdx, Established)
-		buf := f.bufAlloc.Acquire()
-		f.SendVia(via.relayHI, via.relay, msg, buf)
-		f.bufAlloc.Release(buf)
+		f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
 		f.l.Info("Handshake message sent", append(logFields, "relay", via.relayHI.vpnAddrs[0])...)
 	}
 }
@@ -1104,9 +1098,7 @@ func (hm *HandshakeManager) handleCheckAndCompleteError(err error, existing, hos
 	switch err {
 	case ErrAlreadySeen:
 		if existing.SetRemoteIfPreferred(f.hostMap, via) {
-			buf := f.bufAlloc.Acquire()
-			f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), buf)
-			f.bufAlloc.Release(buf)
+			f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
 		}
 		// Resend the original response. The peer is committed to that response's
 		// ephemeral keys; a freshly-built one would have different keys and break
@@ -1129,9 +1121,7 @@ func (hm *HandshakeManager) handleCheckAndCompleteError(err error, existing, hos
 			"responderIndex", hostinfo.localIndexId,
 			"handshake", hsFields,
 		)
-		buf := f.bufAlloc.Acquire()
-		f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), buf)
-		f.bufAlloc.Release(buf)
+		f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))

 	case ErrLocalIndexCollision:
 		f.l.Error("Failed to add HostInfo due to localIndex collision",
@@ -80,15 +80,15 @@ func testCountTimerWheelEntries(tw *LockingTimerWheel[netip.Addr]) (c int) {
 type mockEncWriter struct {
 }

-func (mw *mockEncWriter) SendMessageToVpnAddr(_ header.MessageType, _ header.MessageSubType, _ netip.Addr, _ []byte, _ *WireBuffer) {
+func (mw *mockEncWriter) SendMessageToVpnAddr(_ header.MessageType, _ header.MessageSubType, _ netip.Addr, _, _, _ []byte) {
 	return
 }

-func (mw *mockEncWriter) SendVia(_ *HostInfo, _ *Relay, _ []byte, _ *WireBuffer) {
+func (mw *mockEncWriter) SendVia(_ *HostInfo, _ *Relay, _, _, _ []byte, _ bool) {
 	return
 }

-func (mw *mockEncWriter) SendMessageToHostInfo(_ header.MessageType, _ header.MessageSubType, _ *HostInfo, _ []byte, _ *WireBuffer) {
+func (mw *mockEncWriter) SendMessageToHostInfo(_ header.MessageType, _ header.MessageSubType, _ *HostInfo, _, _, _ []byte) {
 	return
 }

@@ -174,6 +174,10 @@ func (h *H) SubTypeName() string {
 	return SubTypeName(h.Type, h.Subtype)
 }

+func (h *H) IsValidSubType() bool {
+	return IsValidSubType(h.Type, h.Subtype)
+}
+
 // SubTypeName will transform a nebula message sub type into a human string
 func SubTypeName(t MessageType, s MessageSubType) string {
 	if n, ok := subTypeMap[t]; ok {
@@ -185,6 +189,16 @@ func SubTypeName(t MessageType, s MessageSubType) string {
 	return "unknown"
 }

+func IsValidSubType(t MessageType, s MessageSubType) bool {
+	if n, ok := subTypeMap[t]; ok {
+		if _, ok := (*n)[s]; ok {
+			return true
+		}
+	}
+
+	return false
+}
+
 // NewHeader turns bytes into a header
 func NewHeader(b []byte) (*H, error) {
 	h := new(H)
@@ -138,9 +138,9 @@ func (rs *RelayState) InsertRelayTo(ip netip.Addr) {
 }

 func (rs *RelayState) CopyRelayIps() []netip.Addr {
-	ret := make([]netip.Addr, len(rs.relays))
 	rs.RLock()
 	defer rs.RUnlock()
+	ret := make([]netip.Addr, len(rs.relays))
 	copy(ret, rs.relays)
 	return ret
 }
@@ -308,7 +308,7 @@ type cachedPacket struct {
 	packet         []byte
 }

-type packetCallback func(t header.MessageType, st header.MessageSubType, h *HostInfo, p []byte, buf *WireBuffer)
+type packetCallback func(t header.MessageType, st header.MessageSubType, h *HostInfo, p, nb, out []byte)

 type cachedPacketMetrics struct {
 	sent    metrics.Counter
@@ -623,6 +623,11 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 	hm.Indexes[hostinfo.localIndexId] = hostinfo
 	hm.RemoteIndexes[hostinfo.remoteIndexId] = hostinfo

+	hostinfo.out.Store(true)
+	if f.connectionManager != nil { // f.connectionManager is only nil in some unit tests
+		f.connectionManager.trafficTimer.Add(hostinfo.localIndexId, f.connectionManager.checkInterval)
+	}
+
 	if hm.l.Enabled(context.Background(), slog.LevelDebug) {
 		hm.l.Debug("Hostmap vpnIp added",
 			"hostMap", m{"vpnAddrs": hostinfo.vpnAddrs, "mapTotalSize": len(hm.Hosts),
@@ -691,7 +696,6 @@ func (i *HostInfo) TryPromoteBest(preferredRanges []netip.Prefix, ifce *Interfac
 			}
 		}

-		buf := ifce.bufAlloc.Acquire()
 		i.remotes.ForEach(preferredRanges, func(addr netip.AddrPort, preferred bool) {
 			if remote.IsValid() && (!addr.IsValid() || !preferred) {
 				return
@@ -699,9 +703,8 @@ func (i *HostInfo) TryPromoteBest(preferredRanges []netip.Prefix, ifce *Interfac

 			// Try to send a test packet to that host, this should
 			// cause it to detect a roaming event and switch remotes
-			ifce.sendTo(header.Test, header.TestRequest, i.ConnectionState, i, addr, []byte(""), buf)
+			ifce.sendTo(header.Test, header.TestRequest, i.ConnectionState, i, addr, []byte(""), make([]byte, 12, 12), make([]byte, mtu))
 		})
-		ifce.bufAlloc.Release(buf)
 	}

 	// Re query our lighthouses for new remotes occasionally
@@ -8,13 +8,12 @@ import (
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/iputil"
+	"github.com/slackhq/nebula/noiseutil"
 	"github.com/slackhq/nebula/routing"
 )

-func (f *Interface) consumeInsidePacket(buf *WireBuffer, q int, localCache firewall.ConntrackCache) {
-	packet := buf.IPPacket()
-
-	err := newPacket(packet, false, buf.FwPacket)
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
+	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			f.l.Debug("Error while validating outbound packet",
@@ -27,12 +26,12 @@ func (f *Interface) consumeInsidePacket(buf *WireBuffer, q int, localCache firew

 	// Ignore local broadcast packets
 	if f.dropLocalBroadcast {
-		if f.myBroadcastAddrsTable.Contains(buf.FwPacket.RemoteAddr) {
+		if f.myBroadcastAddrsTable.Contains(fwPacket.RemoteAddr) {
 			return
 		}
 	}

-	if f.myVpnAddrsTable.Contains(buf.FwPacket.RemoteAddr) {
+	if f.myVpnAddrsTable.Contains(fwPacket.RemoteAddr) {
 		// Immediately forward packets from self to self.
 		// This should only happen on Darwin-based and FreeBSD hosts, which
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
@@ -49,20 +48,20 @@ func (f *Interface) consumeInsidePacket(buf *WireBuffer, q int, localCache firew
 	}

 	// Ignore multicast packets
-	if f.dropMulticast && buf.FwPacket.RemoteAddr.IsMulticast() {
+	if f.dropMulticast && fwPacket.RemoteAddr.IsMulticast() {
 		return
 	}

-	hostinfo, ready := f.getOrHandshakeConsiderRouting(buf.FwPacket, func(hh *HandshakeHostInfo) {
+	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
 		hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
 	})

 	if hostinfo == nil {
-		f.rejectInside(packet, buf.Out, q)
+		f.rejectInside(packet, out, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			f.l.Debug("dropping outbound packet, vpnAddr not in our vpn networks or in unsafe networks",
-				"vpnAddr", buf.FwPacket.RemoteAddr,
-				"fwPacket", buf.FwPacket,
+				"vpnAddr", fwPacket.RemoteAddr,
+				"fwPacket", fwPacket,
 			)
 		}
 		return
@@ -72,15 +71,15 @@ func (f *Interface) consumeInsidePacket(buf *WireBuffer, q int, localCache firew
 		return
 	}

-	dropReason := f.firewall.Drop(*buf.FwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
+	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, buf, q)
+		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)

 	} else {
-		f.rejectInside(packet, buf.Out, q)
+		f.rejectInside(packet, out, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping outbound packet",
-				"fwPacket", buf.FwPacket,
+				"fwPacket", fwPacket,
 				"reason", dropReason,
 			)
 		}
@@ -103,27 +102,27 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	}
 }

-func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *HostInfo, scratch []byte, buf *WireBuffer, q int) {
+func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *HostInfo, nb, out []byte, q int) {
 	if !f.firewall.OutSendReject {
 		return
 	}

-	rejectIP := iputil.CreateRejectPacket(packet, scratch)
-	if len(rejectIP) == 0 {
+	out = iputil.CreateRejectPacket(packet, out)
+	if len(out) == 0 {
 		return
 	}

-	if len(rejectIP) > iputil.MaxRejectPacketSize {
+	if len(out) > iputil.MaxRejectPacketSize {
 		if f.l.Enabled(context.Background(), slog.LevelInfo) {
 			f.l.Info("rejectOutside: packet too big, not sending",
 				"packet", packet,
-				"outPacket", rejectIP,
+				"outPacket", out,
 			)
 		}
 		return
 	}

-	f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, rejectIP, buf, q)
+	f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, out, nb, packet, q)
 }

 // Handshake will attempt to initiate a tunnel with the provided vpn address. This is a no-op if the tunnel is already established or being established
@@ -216,7 +215,7 @@ func (f *Interface) getOrHandshakeConsiderRouting(fwPacket *firewall.Packet, cac

 }

-func (f *Interface) sendMessageNow(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p []byte, buf *WireBuffer) {
+func (f *Interface) sendMessageNow(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p, nb, out []byte) {
 	fp := &firewall.Packet{}
 	err := newPacket(p, false, fp)
 	if err != nil {
@@ -236,12 +235,12 @@ func (f *Interface) sendMessageNow(t header.MessageType, st header.MessageSubTyp
 		return
 	}

-	f.sendNoMetrics(header.Message, st, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, p, buf, 0)
+	f.sendNoMetrics(header.Message, st, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, p, nb, out, 0)
 }

 // SendMessageToVpnAddr handles real addr:port lookup and sends to the current best known address for vpnAddr.
 // This function ignores myVpnNetworksTable, and will always attempt to treat the address as a vpnAddr
-func (f *Interface) SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnAddr netip.Addr, p []byte, buf *WireBuffer) {
+func (f *Interface) SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnAddr netip.Addr, p, nb, out []byte) {
 	hostInfo, ready := f.handshakeManager.GetOrHandshake(vpnAddr, func(hh *HandshakeHostInfo) {
 		hh.cachePacket(f.l, t, st, p, f.SendMessageToHostInfo, f.cachedPacketMetrics)
 	})
@@ -259,73 +258,113 @@ func (f *Interface) SendMessageToVpnAddr(t header.MessageType, st header.Message
 		return
 	}

-	f.SendMessageToHostInfo(t, st, hostInfo, p, buf)
+	f.SendMessageToHostInfo(t, st, hostInfo, p, nb, out)
 }

-func (f *Interface) SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hi *HostInfo, p []byte, buf *WireBuffer) {
-	f.send(t, st, hi.ConnectionState, hi, p, buf)
+func (f *Interface) SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hi *HostInfo, p, nb, out []byte) {
+	f.send(t, st, hi.ConnectionState, hi, p, nb, out)
 }

-func (f *Interface) send(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, p []byte, buf *WireBuffer) {
+func (f *Interface) send(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, p, nb, out []byte) {
 	f.messageMetrics.Tx(t, st, 1)
-	f.sendNoMetrics(t, st, ci, hostinfo, netip.AddrPort{}, p, buf, 0)
+	f.sendNoMetrics(t, st, ci, hostinfo, netip.AddrPort{}, p, nb, out, 0)
 }

-func (f *Interface) sendTo(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p []byte, buf *WireBuffer) {
+func (f *Interface) sendTo(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte) {
 	f.messageMetrics.Tx(t, st, 1)
-	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, buf, 0)
+	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, nb, out, 0)
 }

 // SendVia sends a payload through a Relay tunnel. No authentication or encryption is done
 // to the payload for the ultimate target host, making this a useful method for sending
 // handshake messages to peers through relay tunnels.
-//
-// via is the HostInfo through which the message is relayed. ad is staged into
-// the inner-payload slot of buf and then AAD-only sealed under via's key by
-// SealRelayInPlace. The sendNoMetrics relay-forward path skips this entry
-// point and calls sendViaInPlace directly because its inner ciphertext is
-// already in place from the encrypt step.
-func (f *Interface) SendVia(via *HostInfo, relay *Relay, ad []byte, buf *WireBuffer) {
-	if header.Len+len(ad)+via.ConnectionState.eKey.Overhead() > cap(buf.Out) {
+// via is the HostInfo through which the message is relayed.
+// ad is the plaintext data to authenticate, but not encrypt
+// nb is a buffer used to store the nonce value, re-used for performance reasons.
+// out is a buffer used to store the result of the Encrypt operation
+// q indicates which writer to use to send the packet.
+func (f *Interface) SendVia(via *HostInfo,
+	relay *Relay,
+	ad,
+	nb,
+	out []byte,
+	nocopy bool,
+) {
+	if noiseutil.EncryptLockNeeded {
+		// NOTE: for goboring AESGCMTLS we need to lock because of the nonce check
+		via.ConnectionState.writeLock.Lock()
+	}
+	c := via.ConnectionState.messageCounter.Add(1)
+
+	out = header.Encode(out, header.Version, header.Message, header.MessageRelay, relay.RemoteIndex, c)
+	f.connectionManager.Out(via)
+
+	// Authenticate the header and payload, but do not encrypt for this message type.
+	// The payload consists of the inner, unencrypted Nebula header, as well as the end-to-end encrypted payload.
+	if len(out)+len(ad)+via.ConnectionState.eKey.Overhead() > cap(out) {
+		if noiseutil.EncryptLockNeeded {
+			via.ConnectionState.writeLock.Unlock()
+		}
 		via.logger(f.l).Error("SendVia out buffer not large enough for relay",
-			"outCap", cap(buf.Out),
+			"outCap", cap(out),
 			"payloadLen", len(ad),
-			"headerLen", header.Len,
+			"headerLen", len(out),
 			"cipherOverhead", via.ConnectionState.eKey.Overhead(),
 		)
 		return
 	}
-	buf.StageRelayInner(ad)
-	f.sendViaInPlace(via, relay, len(ad), buf)
-}

-// sendViaInPlace stamps the outer relay header, AAD-seals over the [outer
-// header | inner-already-staged] region, and writes the result to via.remote.
-// Called from SendVia (after staging ad) and from sendNoMetrics' relay-forward
-// path (where the inner ciphertext is already in place from SealForRelay).
-func (f *Interface) sendViaInPlace(via *HostInfo, relay *Relay, innerLen int, buf *WireBuffer) {
-	f.connectionManager.Out(via)
-	out, err := buf.SealRelayInPlace(via.ConnectionState, relay.RemoteIndex, innerLen)
+	// The header bytes are written to the 'out' slice; Grow the slice to hold the header and associated data payload.
+	offset := len(out)
+	out = out[:offset+len(ad)]
+
+	// In one call path, the associated data _is_ already stored in out. In other call paths, the associated data must
+	// be copied into 'out'.
+	if !nocopy {
+		copy(out[offset:], ad)
+	}
+
+	var err error
+	out, err = via.ConnectionState.eKey.EncryptDanger(out, out, nil, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		via.ConnectionState.writeLock.Unlock()
+	}
 	if err != nil {
 		via.logger(f.l).Info("Failed to EncryptDanger in sendVia", "error", err)
 		return
 	}
-	if err := f.writers[0].WriteTo(out, via.remote); err != nil {
+	err = f.writers[0].WriteTo(out, via.remote)
+	if err != nil {
 		via.logger(f.l).Info("Failed to WriteTo in sendVia", "error", err)
 	}
 	f.connectionManager.RelayUsed(relay.LocalIndex)
 }

-// sendNoMetrics encrypts and writes one outbound nebula packet (data, control,
-// lighthouse, etc) using buf as the per-call wire scratch. When the hostinfo
-// has no direct remote we encrypt into the relay-reserved slot via
-// SealForRelay so sendViaInPlace can wrap it without an extra copy.
-func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p []byte, buf *WireBuffer, q int) {
+func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte, q int) {
 	if ci.eKey == nil {
 		return
 	}
 	useRelay := !remote.IsValid() && !hostinfo.remote.IsValid()
+	fullOut := out

+	if useRelay {
+		if len(out) < header.Len {
+			// out always has a capacity of mtu, but not always a length greater than the header.Len.
+			// Grow it to make sure the next operation works.
+			out = out[:header.Len]
+		}
+		// Save a header's worth of data at the front of the 'out' buffer.
+		out = out[header.Len:]
+	}
+
+	if noiseutil.EncryptLockNeeded {
+		// NOTE: for goboring AESGCMTLS we need to lock because of the nonce check
+		ci.writeLock.Lock()
+	}
+	c := ci.messageCounter.Add(1)
+
+	//l.WithField("trace", string(debug.Stack())).Error("out Header ", &Header{Version, t, st, 0, hostinfo.remoteIndexId, c}, p)
+	out = header.Encode(out, header.Version, t, st, hostinfo.remoteIndexId, c)
 	f.connectionManager.Out(hostinfo)

 	// Query our LH if we haven't since the last time we've been rebound, this will cause the remote to punch against
@@ -342,42 +381,49 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 		}
 	}

-	var out []byte
 	var err error
-	if useRelay {
-		out, err = buf.SealForRelay(ci, t, st, hostinfo.remoteIndexId, p)
-	} else {
-		out, err = buf.Seal(ci, t, st, hostinfo.remoteIndexId, p)
+	out, err = ci.eKey.EncryptDanger(out, out, p, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
 	}
 	if err != nil {
 		hostinfo.logger(f.l).Error("Failed to encrypt outgoing packet",
 			"error", err,
 			"udpAddr", remote,
+			"counter", c,
 		)
 		return
 	}

-	switch {
-	case remote.IsValid():
-		if err := f.writers[q].WriteTo(out, remote); err != nil {
-			hostinfo.logger(f.l).Error("Failed to write outgoing packet", "error", err, "udpAddr", remote)
+	if remote.IsValid() {
+		err = f.writers[q].WriteTo(out, remote)
+		if err != nil {
+			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
+				"error", err,
+				"udpAddr", remote,
+			)
 		}
-	case hostinfo.remote.IsValid():
-		if err := f.writers[q].WriteTo(out, hostinfo.remote); err != nil {
-			hostinfo.logger(f.l).Error("Failed to write outgoing packet", "error", err, "udpAddr", hostinfo.remote)
+	} else if hostinfo.remote.IsValid() {
+		err = f.writers[q].WriteTo(out, hostinfo.remote)
+		if err != nil {
+			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
+				"error", err,
+				"udpAddr", remote,
+			)
 		}
-	default:
-		// SealForRelay placed the inner ciphertext at buf.Out[header.Len:],
-		// so sendViaInPlace can wrap it with the outer relay header without
-		// an extra copy.
+	} else {
+		// Try to send via a relay
 		for _, relayIP := range hostinfo.relayState.CopyRelayIps() {
 			relayHostInfo, relay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relayIP)
 			if err != nil {
 				hostinfo.relayState.DeleteRelay(relayIP)
-				hostinfo.logger(f.l).Info("sendNoMetrics failed to find HostInfo", "relay", relayIP, "error", err)
+				hostinfo.logger(f.l).Info("sendNoMetrics failed to find HostInfo",
+					"relay", relayIP,
+					"error", err,
+				)
 				continue
 			}
-			f.sendViaInPlace(relayHostInfo, relay, len(out), buf)
+			f.SendVia(relayHostInfo, relay, out, nb, fullOut[:header.Len+len(out)], true)
 			break
 		}
 	}
@@ -7,6 +7,7 @@ import (
 	"io"
 	"log/slog"
 	"net/netip"
+	"slices"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -14,6 +15,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/rcrowley/go-metrics"

+	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
@@ -101,19 +103,19 @@ type Interface struct {
 	messageMetrics      *MessageMetrics
 	cachedPacketMetrics *cachedPacketMetrics

-	// bufAlloc hands out reusable WireBuffers sized for this interface's
-	// inside Device. All buf consumers (hot-path data-plane goroutines,
-	// long-lived workers, and cold callers) acquire from here so sizing
-	// is centralized and consistent. Long-lived owners just don't release.
-	bufAlloc WireBufferAllocator
-
 	l *slog.Logger
 }

 type EncWriter interface {
-	SendVia(via *HostInfo, relay *Relay, ad []byte, buf *WireBuffer)
-	SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnAddr netip.Addr, p []byte, buf *WireBuffer)
-	SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p []byte, buf *WireBuffer)
+	SendVia(via *HostInfo,
+		relay *Relay,
+		ad,
+		nb,
+		out []byte,
+		nocopy bool,
+	)
+	SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnAddr netip.Addr, p, nb, out []byte)
+	SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p, nb, out []byte)
 	Handshake(vpnAddr netip.Addr)
 	GetHostInfo(vpnAddr netip.Addr) *HostInfo
 	GetCertState() *CertState
@@ -204,8 +206,6 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 			dropped: metrics.GetOrRegisterCounter("hostinfo.cached_packets.dropped", nil),
 		},

-		bufAlloc: NewWireBufferPool(mtu, c.Inside.TunPrefixLen()),
-
 		l: c.l,
 	}

@@ -313,11 +313,13 @@ func (f *Interface) listenOut(i int) {

 	ctCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-	// Long-lived per-receive-goroutine buf; never released back to the pool.
-	buf := f.bufAlloc.Acquire()
+	plaintext := make([]byte, udp.MTU)
+	h := &header.H{}
+	fwPacket := &firewall.Packet{}
+	nb := make([]byte, 12, 12)

 	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, buf, payload, lhh, i, ctCache.Get())
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
 	})

 	if err != nil && !f.closed.Load() {
@@ -329,12 +331,15 @@ func (f *Interface) listenOut(i int) {
 }

 func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
-	// Long-lived per-tun-reader buf; never released back to the pool.
-	buf := f.bufAlloc.Acquire()
+	packet := make([]byte, mtu)
+	out := make([]byte, mtu)
+	fwPacket := &firewall.Packet{}
+	nb := make([]byte, 12, 12)
+
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)

 	for {
-		_, err := buf.ReadIPFromTUN(reader)
+		n, err := reader.Read(packet)
 		if err != nil {
 			if !f.closed.Load() {
 				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", i)
@@ -343,7 +348,7 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 			break
 		}

-		f.consumeInsidePacket(buf, i, conntrackCache.Get())
+		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get())
 	}

 	f.l.Debug("overlay reader is done", "reader", i)
@@ -372,13 +377,22 @@ func (f *Interface) reloadDisconnectInvalid(c *config.C) {
 }

 func (f *Interface) reloadFirewall(c *config.C) {
-	//TODO: need to trigger/detect if the certificate changed too
-	if c.HasChanged("firewall") == false {
+	cs := f.pki.getCertState()
+	curCert := cs.getCertificate(cert.Version2)
+	if curCert == nil {
+		curCert = cs.getCertificate(cert.Version1)
+	}
+
+	// The firewall builds its routableNetworks set from the certificate's UnsafeNetworks at construction.
+	// Check to see if that set has changed, and if so, rebuild the firewall.
+	certUnsafeChanged := curCert != nil && !slices.Equal(curCert.UnsafeNetworks(), f.firewall.unsafeNetworks)
+
+	if !c.HasChanged("firewall") && !certUnsafeChanged {
 		f.l.Debug("No firewall config change detected")
 		return
 	}

-	fw, err := NewFirewallFromConfig(f.l, f.pki.getCertState(), c)
+	fw, err := NewFirewallFromConfig(f.l, cs, c)
 	if err != nil {
 		f.l.Error("Error while creating firewall during reload", "error", err)
 		return
@@ -488,26 +502,34 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
 	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)

+	emit := func() {
+		f.firewall.EmitStats()
+		f.handshakeManager.EmitStats()
+		udpStats()
+
+		certState := f.pki.getCertState()
+		defaultCrt := certState.GetDefaultCertificate()
+		certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
+		certInitiatingVersion.Update(int64(defaultCrt.Version()))
+
+		// Report the max certificate version we are capable of using
+		if certState.v2Cert != nil {
+			certMaxVersion.Update(int64(certState.v2Cert.Version()))
+		} else {
+			certMaxVersion.Update(int64(certState.v1Cert.Version()))
+		}
+	}
+
+	// Prime gauges so a Prometheus scrape that lands before the first tick
+	// sees real values instead of the zero defaults (issue #907).
+	emit()
+
 	for {
 		select {
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			f.firewall.EmitStats()
-			f.handshakeManager.EmitStats()
-			udpStats()
-
-			certState := f.pki.getCertState()
-			defaultCrt := certState.GetDefaultCertificate()
-			certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
-			certInitiatingVersion.Update(int64(defaultCrt.Version()))
-
-			// Report the max certificate version we are capable of using
-			if certState.v2Cert != nil {
-				certMaxVersion.Update(int64(certState.v2Cert.Version()))
-			} else {
-				certMaxVersion.Update(int64(certState.v1Cert.Version()))
-			}
+			emit()
 		}
 	}
 }
@@ -0,0 +1,73 @@
+//go:build linux || darwin
+
+package nebula
+
+import (
+	"context"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/rcrowley/go-metrics"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/firewall"
+	"github.com/slackhq/nebula/overlay/overlaytest"
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// Test_emitStats_primesGauges covers issue #907: a Prometheus scrape that
+// landed before the first ticker fire used to read 0 for the cert gauges.
+// emitStats now primes the gauges before entering the ticker loop. We assert
+// the gauge is zero before the first call and non-zero after.
+func Test_emitStats_primesGauges(t *testing.T) {
+	defer metrics.DefaultRegistry.UnregisterAll()
+
+	l := test.NewLogger()
+	hostMap := newHostMap(l)
+	preferredRanges := []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}
+	hostMap.preferredRanges.Store(&preferredRanges)
+
+	notAfter := time.Now().Add(time.Hour)
+	cs := &CertState{
+		initiatingVersion: cert.Version1,
+		privateKey:        []byte{},
+		v1Cert:            &dummyCert{version: cert.Version1, notAfter: notAfter},
+		v1Credential:      nil,
+	}
+
+	lh := newTestLighthouse()
+	ifce := &Interface{
+		hostMap:          hostMap,
+		inside:           &overlaytest.NoopTun{},
+		outside:          &udp.NoopConn{},
+		firewall:         &Firewall{Conntrack: &FirewallConntrack{Conns: map[firewall.Packet]*conn{}}},
+		lightHouse:       lh,
+		pki:              &PKI{},
+		handshakeManager: NewHandshakeManager(l, hostMap, lh, &udp.NoopConn{}, defaultHandshakeConfig),
+		l:                l,
+		// On linux, udp.NewUDPStatsEmitter indexes writers[0] and asserts to
+		// *udp.StdConn. A zero value works: getMemInfo sees a nil rawConn,
+		// returns an error, and the emitter falls through to a no-op.
+		writers: []udp.Conn{&udp.StdConn{}},
+	}
+	ifce.pki.cs.Store(cs)
+
+	ttlGauge := metrics.GetOrRegisterGauge("certificate.ttl_seconds", nil)
+	require.Zero(t, ttlGauge.Value(), "gauge should be zero before emitStats runs")
+
+	// Pre-cancel the context so emitStats returns after priming the gauges
+	// without ever reading from ticker.C. The one hour interval is just a
+	// belt-and-suspenders, the test does not expect the ticker to fire.
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	ifce.emitStats(ctx, time.Hour)
+
+	ttl := ttlGauge.Value()
+	assert.Positive(t, ttl, "ttl gauge should be primed by emitStats before its first tick")
+	assert.LessOrEqual(t, ttl, int64(3600))
+	assert.Equal(t, int64(cert.Version1), metrics.GetOrRegisterGauge("certificate.initiating_version", nil).Value())
+	assert.Equal(t, int64(cert.Version1), metrics.GetOrRegisterGauge("certificate.max_version", nil).Value())
+}
@@ -0,0 +1,120 @@
+package nebula
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestReloadFirewall_CertUnsafeNetworksChanged verifies that reloadFirewall
+// rebuilds the firewall when only the certificate's UnsafeNetworks have changed,
+// even if the firewall section of the YAML has not.
+func TestReloadFirewall_CertUnsafeNetworksChanged(t *testing.T) {
+	l := test.NewLogger()
+
+	vpnNet := netip.MustParsePrefix("10.0.0.1/24")
+	initialUnsafe := []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")}
+
+	// dummyCert avoids dragging the real signing pipeline into a unit test.
+	c1 := &dummyCert{
+		version:        cert.Version2,
+		networks:       []netip.Prefix{vpnNet},
+		unsafeNetworks: initialUnsafe,
+	}
+	pki := &PKI{}
+	pki.cs.Store(&CertState{v2Cert: c1, initiatingVersion: cert.Version2})
+
+	rawYAML := `firewall:
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+`
+	cfg := config.NewC(l)
+	require.NoError(t, cfg.LoadString(rawYAML))
+
+	fw, err := NewFirewallFromConfig(l, pki.getCertState(), cfg)
+	require.NoError(t, err)
+	require.Equal(t, initialUnsafe, fw.unsafeNetworks)
+
+	f := &Interface{
+		pki:      pki,
+		firewall: fw,
+		l:        l,
+	}
+
+	// Swap the cert with a different UnsafeNetworks set.
+	newUnsafe := []netip.Prefix{
+		netip.MustParsePrefix("198.51.100.0/24"),
+		netip.MustParsePrefix("203.0.113.0/24"),
+	}
+	c2 := &dummyCert{
+		version:        cert.Version2,
+		networks:       []netip.Prefix{vpnNet},
+		unsafeNetworks: newUnsafe,
+	}
+	pki.cs.Store(&CertState{v2Cert: c2, initiatingVersion: cert.Version2})
+
+	// Reload with the same YAML so HasChanged("firewall") reports false.
+	require.NoError(t, cfg.ReloadConfigString(rawYAML))
+	require.False(t, cfg.HasChanged("firewall"))
+
+	f.reloadFirewall(cfg)
+
+	assert.NotSame(t, fw, f.firewall, "firewall pointer should have been replaced")
+	assert.Equal(t, newUnsafe, f.firewall.unsafeNetworks)
+	assert.True(t, f.firewall.routableNetworks.Contains(netip.MustParseAddr("203.0.113.5")))
+}
+
+// TestReloadFirewall_NoChange verifies that reloadFirewall is a no-op when
+// neither the firewall config nor the cert's UnsafeNetworks have changed.
+func TestReloadFirewall_NoChange(t *testing.T) {
+	l := test.NewLogger()
+
+	vpnNet := netip.MustParsePrefix("10.0.0.1/24")
+	unsafe := []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")}
+
+	c1 := &dummyCert{
+		version:        cert.Version2,
+		networks:       []netip.Prefix{vpnNet},
+		unsafeNetworks: unsafe,
+	}
+	pki := &PKI{}
+	pki.cs.Store(&CertState{v2Cert: c1, initiatingVersion: cert.Version2})
+
+	rawYAML := `firewall:
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+`
+	cfg := config.NewC(l)
+	require.NoError(t, cfg.LoadString(rawYAML))
+
+	fw, err := NewFirewallFromConfig(l, pki.getCertState(), cfg)
+	require.NoError(t, err)
+
+	f := &Interface{
+		pki:      pki,
+		firewall: fw,
+		l:        l,
+	}
+
+	require.NoError(t, cfg.ReloadConfigString(rawYAML))
+	f.reloadFirewall(cfg)
+
+	assert.Same(t, fw, f.firewall, "firewall should not have been replaced")
+}
@@ -4,26 +4,54 @@ import (
 	"encoding/binary"

 	"golang.org/x/net/ipv4"
+	"golang.org/x/net/ipv6"
 )

 const (
-	// Need 96 bytes for the largest reject packet:
+	// MaxIPv4RejectPacketSize is the largest IPv4 reject packet:
 	// - 20 byte ipv4 header
 	// - 8 byte icmpv4 header
 	// - 68 byte body (60 byte max orig ipv4 header + 8 byte orig icmpv4 header)
-	MaxRejectPacketSize = ipv4.HeaderLen + 8 + 60 + 8
+	maxIPv4RejectPacketSize = ipv4.HeaderLen + 8 + 60 + 8
+
+	// MaxRejectPacketSize is sized for the largest possible reject packet (IPv6):
+	// - 40 byte ipv6 header
+	// - 8 byte icmpv6 header
+	// - up to 1000 byte body (original packet, possibly truncated. We want to stay
+	//   under the MTU with Nebula overhead included)
+	maxIPv6RejectPacketSize = ipv6.HeaderLen + 8 + 1000
+
+	MaxRejectPacketSize = maxIPv6RejectPacketSize
 )

 func CreateRejectPacket(packet []byte, out []byte) []byte {
-	if len(packet) < ipv4.HeaderLen || int(packet[0]>>4) != ipv4.Version {
+	if len(packet) < 1 {
 		return nil
 	}

-	switch packet[9] {
-	case 6: // tcp
-		return ipv4CreateRejectTCPPacket(packet, out)
+	version := int(packet[0] >> 4)
+	switch version {
+	case ipv4.Version:
+		if len(packet) < ipv4.HeaderLen {
+			return nil
+		}
+		// Do not send reject packets for non-first fragments
+		if packet[6]&0x1f != 0 || packet[7] != 0 {
+			return nil
+		}
+		switch packet[9] {
+		case 6: // tcp
+			return ipv4CreateRejectTCPPacket(packet, out)
+		default:
+			return ipv4CreateRejectICMPPacket(packet, out)
+		}
+	case ipv6.Version:
+		if len(packet) < ipv6.HeaderLen {
+			return nil
+		}
+		return ipv6CreateRejectPacket(packet, out)
 	default:
-		return ipv4CreateRejectICMPPacket(packet, out)
+		return nil
 	}
 }

@@ -35,12 +63,17 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 		return nil
 	}

-	// ICMP reply includes original header and first 8 bytes of the packet
-	packetLen := len(packet)
-	if packetLen > ihl+8 {
-		packetLen = ihl + 8
+	// Do not generate ICMP errors in response to ICMP error packets
+	if packet[9] == 1 && len(packet) > ihl {
+		icmpType := packet[ihl]
+		if icmpType == 3 || icmpType == 4 || icmpType == 5 || icmpType == 11 || icmpType == 12 {
+			return nil
+		}
 	}

+	// ICMP reply includes original header and first 8 bytes of the packet
+	packetLen := min(len(packet), ihl+8)
+
 	outLen := ipv4.HeaderLen + 8 + packetLen
 	if outLen > cap(out) {
 		return nil
@@ -71,14 +104,14 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {

 	// ICMP Destination Unreachable
 	icmpOut := out[ipv4.HeaderLen:]
-	icmpOut[0] = 3 // type (Destination unreachable)
-	icmpOut[1] = 3 // code (Port unreachable error)
-	icmpOut[2] = 0 // checksum
-	icmpOut[3] = 0 //  .
-	icmpOut[4] = 0 // unused
-	icmpOut[5] = 0 //  .
-	icmpOut[6] = 0 //  .
-	icmpOut[7] = 0 //  .
+	icmpOut[0] = 3  // type (Destination unreachable)
+	icmpOut[1] = 13 // code (Communication administratively prohibited)
+	icmpOut[2] = 0  // checksum
+	icmpOut[3] = 0  //  .
+	icmpOut[4] = 0  // unused
+	icmpOut[5] = 0  //  .
+	icmpOut[6] = 0  //  .
+	icmpOut[7] = 0  //  .

 	// Copy original IP header and first 8 bytes as body
 	copy(icmpOut[8:], packet[:packetLen])
@@ -165,7 +198,193 @@ func ipv4CreateRejectTCPPacket(packet []byte, out []byte) []byte {
 	return out
 }

+func ipv6CreateRejectPacket(packet []byte, out []byte) []byte {
+	proto, offset, isFragment := ipv6FindUpperProtocol(packet)
+	if isFragment {
+		return nil
+	}
+	switch proto {
+	case 6: // tcp
+		return ipv6CreateRejectTCPPacket(packet, out, offset)
+	default:
+		return ipv6CreateRejectICMPPacket(packet, out, proto, offset)
+	}
+}
+
+func ipv6CreateRejectICMPPacket(packet []byte, out []byte, proto uint8, offset int) []byte {
+	// Do not generate ICMPv6 errors in response to ICMPv6 error packets
+	if proto == 58 && len(packet) > offset {
+		icmpType := packet[offset]
+		if icmpType >= 1 && icmpType <= 4 {
+			return nil
+		}
+	}
+
+	// Include as much of the original packet as possible, up to 1000 bytes,
+	// so the response fits comfortably within any tunnel MTU.
+	packetLen := min(len(packet), 1000)
+
+	outLen := ipv6.HeaderLen + 8 + packetLen
+	if outLen > cap(out) {
+		return nil
+	}
+
+	out = out[:outLen]
+
+	// IPv6 header
+	ipHdr := out[0:ipv6.HeaderLen]
+	ipHdr[0] = ipv6.Version << 4 // version, traffic class (high bits)
+	ipHdr[1] = 0                 // traffic class (low bits), flow label (high bits)
+	ipHdr[2] = 0                 // flow label
+	ipHdr[3] = 0                 // flow label
+
+	payloadLen := uint16(outLen - ipv6.HeaderLen)
+	binary.BigEndian.PutUint16(ipHdr[4:], payloadLen) // payload length
+	ipHdr[6] = 58                                     // next header (ICMPv6)
+	ipHdr[7] = 64                                     // hop limit
+
+	// Swap dest / src IPs (each 16 bytes, src at 8, dst at 24)
+	copy(ipHdr[8:24], packet[24:40])
+	copy(ipHdr[24:40], packet[8:24])
+
+	// ICMPv6 Destination Unreachable
+	icmpOut := out[ipv6.HeaderLen:]
+	icmpOut[0] = 1 // type (Destination Unreachable)
+	icmpOut[1] = 1 // code (Communication with destination administratively prohibited)
+	icmpOut[2] = 0 // checksum
+	icmpOut[3] = 0 //  .
+	icmpOut[4] = 0 // unused
+	icmpOut[5] = 0 //  .
+	icmpOut[6] = 0 //  .
+	icmpOut[7] = 0 //  .
+
+	copy(icmpOut[8:], packet[:packetLen])
+
+	// ICMPv6 checksum uses a pseudo-header
+	csum := ipv6PseudoheaderChecksum(ipHdr[8:24], ipHdr[24:40], 58, uint32(payloadLen))
+	binary.BigEndian.PutUint16(icmpOut[2:], tcpipChecksum(icmpOut, csum))
+
+	return out
+}
+
+func ipv6CreateRejectTCPPacket(packet []byte, out []byte, offset int) []byte {
+	const tcpLen = 20
+
+	if len(packet) < offset+tcpLen {
+		return nil
+	}
+
+	outLen := ipv6.HeaderLen + tcpLen
+	if outLen > cap(out) {
+		return nil
+	}
+
+	out = out[:outLen]
+
+	// IPv6 header
+	ipHdr := out[0:ipv6.HeaderLen]
+	ipHdr[0] = ipv6.Version << 4 // version, traffic class (high bits)
+	ipHdr[1] = 0                 // traffic class (low bits), flow label (high bits)
+	ipHdr[2] = 0                 // flow label
+	ipHdr[3] = 0                 // flow label
+
+	binary.BigEndian.PutUint16(ipHdr[4:], tcpLen) // payload length
+	ipHdr[6] = 6                                  // next header (TCP)
+	ipHdr[7] = 64                                 // hop limit
+
+	// Swap dest / src IPs
+	copy(ipHdr[8:24], packet[24:40])
+	copy(ipHdr[24:40], packet[8:24])
+
+	// TCP RST
+	tcpIn := packet[offset:]
+	var ackSeq, seq uint32
+	outFlags := byte(0b00000100) // RST
+
+	inAck := tcpIn[13]&0b00010000 != 0
+	if inAck {
+		seq = binary.BigEndian.Uint32(tcpIn[8:])
+	} else {
+		inSyn := uint32((tcpIn[13] & 0b00000010) >> 1)
+		inFin := uint32(tcpIn[13] & 0b00000001)
+		ackSeq = binary.BigEndian.Uint32(tcpIn[4:]) + inSyn + inFin + uint32(len(tcpIn)) - uint32(tcpIn[12]>>4)<<2
+		outFlags |= 0b00010000 // ACK
+	}
+
+	tcpOut := out[ipv6.HeaderLen:]
+	// Swap dest / src ports
+	copy(tcpOut[0:2], tcpIn[2:4])
+	copy(tcpOut[2:4], tcpIn[0:2])
+	binary.BigEndian.PutUint32(tcpOut[4:], seq)
+	binary.BigEndian.PutUint32(tcpOut[8:], ackSeq)
+	tcpOut[12] = (tcpLen >> 2) << 4 // data offset, reserved, NS
+	tcpOut[13] = outFlags           // CWR, ECE, URG, ACK, PSH, RST, SYN, FIN
+	tcpOut[14] = 0                  // window size
+	tcpOut[15] = 0                  //  .
+	tcpOut[16] = 0                  // checksum
+	tcpOut[17] = 0                  //  .
+	tcpOut[18] = 0                  // URG Pointer
+	tcpOut[19] = 0                  //  .
+
+	// Calculate checksum with IPv6 pseudo-header
+	csum := ipv6PseudoheaderChecksum(ipHdr[8:24], ipHdr[24:40], 6, tcpLen)
+	binary.BigEndian.PutUint16(tcpOut[16:], tcpipChecksum(tcpOut, csum))
+
+	return out
+}
+
+func ipv6FindUpperProtocol(packet []byte) (nextHeader uint8, offset int, isFragment bool) {
+	nextHeader = packet[6]
+	offset = ipv6.HeaderLen
+
+	for {
+		switch nextHeader {
+		case 0, 43, 60: // Hop-by-Hop, Routing, Destination
+			if len(packet) < offset+2 {
+				return nextHeader, offset, isFragment
+			}
+			nextHeader = packet[offset]
+			offset += int(packet[offset+1]+1) << 3
+
+		case 44: // Fragment
+			if len(packet) < offset+8 {
+				return nextHeader, offset, isFragment
+			}
+			if packet[offset+2] != 0 || packet[offset+3]&0xf8 != 0 {
+				isFragment = true
+			}
+			nextHeader = packet[offset]
+			offset += 8
+
+		case 51: // AH
+			if len(packet) < offset+2 {
+				return nextHeader, offset, isFragment
+			}
+			nextHeader = packet[offset]
+			offset += int(packet[offset+1]+2) << 2
+
+		default:
+			return nextHeader, offset, isFragment
+		}
+	}
+}
+
 func CreateICMPEchoResponse(packet, out []byte) []byte {
+	if len(packet) < 1 {
+		return nil
+	}
+
+	switch packet[0] >> 4 {
+	case 4:
+		return createICMPv4EchoResponse(packet, out)
+	case 6:
+		return createICMPv6EchoResponse(packet, out)
+	default:
+		return nil
+	}
+}
+
+func createICMPv4EchoResponse(packet, out []byte) []byte {
 	// Return early if this is not a simple ICMP Echo Request
 	//TODO: make constants out of these
 	if !(len(packet) >= 28 && len(packet) <= 9001 && packet[0] == 0x45 && packet[9] == 0x01 && packet[20] == 0x08) {
@@ -199,6 +418,43 @@ func CreateICMPEchoResponse(packet, out []byte) []byte {
 	return out
 }

+func createICMPv6EchoResponse(packet, out []byte) []byte {
+	// IPv6 header (40 bytes) + ICMPv6 header (8 bytes minimum)
+	if len(packet) < ipv6.HeaderLen+8 || len(packet) > 9001 {
+		return nil
+	}
+
+	// Next Header must be ICMPv6 (58)
+	if packet[6] != 58 {
+		return nil
+	}
+
+	// ICMPv6 type must be Echo Request (128)
+	if packet[ipv6.HeaderLen] != 128 {
+		return nil
+	}
+
+	out = out[:len(packet)]
+	copy(out, packet)
+
+	// Swap src/dst addresses (bytes 8-23 and 24-39)
+	copy(out[8:24], packet[24:40])
+	copy(out[24:40], packet[8:24])
+
+	// Change ICMPv6 type to Echo Reply (129)
+	icmp := out[ipv6.HeaderLen:]
+	icmp[0] = 129
+	icmp[2] = 0
+	icmp[3] = 0
+
+	// ICMPv6 checksum uses a pseudo-header with src, dst, length, and next header
+	payloadLen := uint32(len(icmp))
+	csum := ipv6PseudoheaderChecksum(out[8:24], out[24:40], 58, payloadLen)
+	binary.BigEndian.PutUint16(icmp[2:], tcpipChecksum(icmp, csum))
+
+	return out
+}
+
 // calculates the TCP/IP checksum defined in rfc1071. The passed-in
 // csum is any initial checksum data that's already been computed.
 //
@@ -236,3 +492,18 @@ func ipv4PseudoheaderChecksum(src, dst []byte, proto, length uint32) (csum uint3
 	csum += length >> 16
 	return csum
 }
+
+// based on:
+// - https://github.com/google/gopacket/blob/v1.1.19/layers/tcpip.go#L37-L48
+func ipv6PseudoheaderChecksum(src, dst []byte, proto, length uint32) (csum uint32) {
+	for i := 0; i < 16; i += 2 {
+		csum += uint32(src[i]) << 8
+		csum += uint32(src[i+1])
+		csum += uint32(dst[i]) << 8
+		csum += uint32(dst[i+1])
+	}
+	csum += proto
+	csum += length & 0xffff
+	csum += length >> 16
+	return csum
+}
@@ -1,11 +1,13 @@
 package iputil

 import (
+	"encoding/binary"
 	"net"
 	"testing"

 	"github.com/stretchr/testify/assert"
 	"golang.org/x/net/ipv4"
+	"golang.org/x/net/ipv6"
 )

 func Test_CreateRejectPacket(t *testing.T) {
@@ -43,7 +45,7 @@ func Test_CreateRejectPacket(t *testing.T) {
 	}
 	b = append(b, []byte{0, 3, 0, 4, 0, 0, 0, 0}...)

-	expectedLen = MaxRejectPacketSize
+	expectedLen = maxIPv4RejectPacketSize
 	out = make([]byte, MaxRejectPacketSize)
 	rejectPacket = CreateRejectPacket(b, out)
 	assert.NotNil(t, rejectPacket)
@@ -71,3 +73,404 @@ func Test_CreateRejectPacket(t *testing.T) {
 	assert.NotNil(t, rejectPacket)
 	assert.Len(t, rejectPacket, expectedLen)
 }
+
+func Test_CreateRejectPacket_NoFragment(t *testing.T) {
+	out := make([]byte, MaxRejectPacketSize)
+
+	// IPv4: non-zero fragment offset should not generate reject packet
+	h := ipv4.Header{
+		Len:      20,
+		Src:      net.IPv4(10, 0, 0, 1),
+		Dst:      net.IPv4(10, 0, 0, 2),
+		Protocol: 17, // UDP
+	}
+	b, err := h.Marshal()
+	if err != nil {
+		t.Fatalf("h.Marshal: %v", err)
+	}
+	b = append(b, make([]byte, 8)...)
+	// Set fragment offset to non-zero (byte 6-7, offset in 8-byte units)
+	b[6] = 0x00
+	b[7] = 0x01
+	assert.Nil(t, CreateRejectPacket(b, out))
+
+	// MF flag with zero offset (first fragment) should still generate reject
+	b[6] = 0x20 // MF flag set
+	b[7] = 0x00
+	assert.NotNil(t, CreateRejectPacket(b, out))
+
+	// Non-fragment should still generate reject packet
+	b[6] = 0x00
+	b[7] = 0x00
+	assert.NotNil(t, CreateRejectPacket(b, out))
+
+	// DF flag only (not a fragment) should still generate reject packet
+	b[6] = 0x40
+	b[7] = 0x00
+	assert.NotNil(t, CreateRejectPacket(b, out))
+}
+
+func Test_CreateRejectPacketIPv6_NoFragment(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+	out := make([]byte, MaxRejectPacketSize)
+
+	// IPv6 with Fragment header and non-zero offset should not generate reject
+	fragHeader := []byte{
+		17,   // next header: UDP
+		0,    // reserved
+		0, 9, // fragment offset=1 (shifted left 3), M=1
+		0, 0, 0, 1, // identification
+	}
+	udpPayload := make([]byte, 8)
+	payload := append(fragHeader, udpPayload...)
+	packet := makeIPv6Packet(src, dst, 44, payload) // next header 44 = Fragment
+	assert.Nil(t, CreateRejectPacket(packet, out))
+
+	// Fragment header with zero offset (first fragment) should still generate reject
+	fragHeader[2] = 0
+	fragHeader[3] = 1 // offset=0, M=1
+	payload = append(fragHeader, udpPayload...)
+	packet = makeIPv6Packet(src, dst, 44, payload)
+	assert.NotNil(t, CreateRejectPacket(packet, out))
+}
+
+func Test_CreateRejectPacket_NoICMPError(t *testing.T) {
+	out := make([]byte, MaxRejectPacketSize)
+
+	// ICMP error types should not generate reject packets
+	icmpErrorTypes := []byte{3, 4, 5, 11, 12}
+	for _, icmpType := range icmpErrorTypes {
+		h := ipv4.Header{
+			Len:      20,
+			Src:      net.IPv4(10, 0, 0, 1),
+			Dst:      net.IPv4(10, 0, 0, 2),
+			Protocol: 1, // ICMP
+		}
+
+		b, err := h.Marshal()
+		if err != nil {
+			t.Fatalf("h.Marshal: %v", err)
+		}
+		b = append(b, icmpType, 0, 0, 0, 0, 0, 0, 0)
+
+		rejectPacket := CreateRejectPacket(b, out)
+		assert.Nil(t, rejectPacket, "ICMP type %d should not generate a reject packet", icmpType)
+	}
+
+	// ICMP non-error types should still generate reject packets
+	icmpNonErrorTypes := []byte{0, 8, 13, 14}
+	for _, icmpType := range icmpNonErrorTypes {
+		h := ipv4.Header{
+			Len:      20,
+			Src:      net.IPv4(10, 0, 0, 1),
+			Dst:      net.IPv4(10, 0, 0, 2),
+			Protocol: 1, // ICMP
+		}
+
+		b, err := h.Marshal()
+		if err != nil {
+			t.Fatalf("h.Marshal: %v", err)
+		}
+		b = append(b, icmpType, 0, 0, 0, 0, 0, 0, 0)
+
+		rejectPacket := CreateRejectPacket(b, out)
+		assert.NotNil(t, rejectPacket, "ICMP type %d should generate a reject packet", icmpType)
+	}
+}
+
+func makeIPv6Packet(src, dst net.IP, nextHeader uint8, payload []byte) []byte {
+	b := make([]byte, ipv6.HeaderLen+len(payload))
+	b[0] = ipv6.Version << 4
+	binary.BigEndian.PutUint16(b[4:], uint16(len(payload)))
+	b[6] = nextHeader
+	b[7] = 64
+	copy(b[8:24], src.To16())
+	copy(b[24:40], dst.To16())
+	copy(b[ipv6.HeaderLen:], payload)
+	return b
+}
+
+func Test_CreateRejectPacketIPv6_ICMP(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+
+	// Small UDP packet: entire original included in body
+	udpPayload := make([]byte, 20)
+	udpPayload[0] = 0x00 // src port high
+	udpPayload[1] = 0x50 // src port low (80)
+	udpPayload[2] = 0x01 // dst port high
+	udpPayload[3] = 0xBB // dst port low (443)
+	packet := makeIPv6Packet(src, dst, 17, udpPayload)
+
+	out := make([]byte, MaxRejectPacketSize)
+	rejectPacket := CreateRejectPacket(packet, out)
+	assert.NotNil(t, rejectPacket)
+
+	// Small packet fits entirely: 40 (ipv6 hdr) + 8 (icmpv6 hdr) + 60 (original)
+	expectedLen := ipv6.HeaderLen + 8 + len(packet)
+	assert.Len(t, rejectPacket, expectedLen)
+
+	// Verify version
+	assert.Equal(t, byte(ipv6.Version<<4), rejectPacket[0]&0xf0)
+	// Verify next header is ICMPv6 (58)
+	assert.Equal(t, byte(58), rejectPacket[6])
+	// Verify src/dst are swapped
+	assert.Equal(t, dst.To16(), net.IP(rejectPacket[8:24]))
+	assert.Equal(t, src.To16(), net.IP(rejectPacket[24:40]))
+	// Verify ICMPv6 type=1 (Dest Unreachable), code=1 (Administratively prohibited)
+	assert.Equal(t, byte(1), rejectPacket[ipv6.HeaderLen])
+	assert.Equal(t, byte(1), rejectPacket[ipv6.HeaderLen+1])
+	// Verify entire original packet is included in body
+	assert.Equal(t, packet, rejectPacket[ipv6.HeaderLen+8:])
+
+	// Large packet: body is truncated to 1000 bytes
+	largePkt := makeIPv6Packet(src, dst, 17, make([]byte, 1200))
+	rejectPacket = CreateRejectPacket(largePkt, out)
+	assert.NotNil(t, rejectPacket)
+	assert.Len(t, rejectPacket, ipv6.HeaderLen+8+1000)
+	assert.Equal(t, largePkt[:1000], rejectPacket[ipv6.HeaderLen+8:])
+}
+
+func Test_CreateRejectPacketIPv6_TCP(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+
+	// TCP SYN packet (next header 6)
+	tcpPayload := make([]byte, 20)
+	tcpPayload[0] = 0x00                             // src port high
+	tcpPayload[1] = 0x50                             // src port low (80)
+	tcpPayload[2] = 0x01                             // dst port high
+	tcpPayload[3] = 0xBB                             // dst port low (443)
+	binary.BigEndian.PutUint32(tcpPayload[4:], 1000) // seq
+	binary.BigEndian.PutUint32(tcpPayload[8:], 0)    // ack seq
+	tcpPayload[12] = (20 >> 2) << 4                  // data offset
+	tcpPayload[13] = 0b00000010                      // SYN flag
+
+	packet := makeIPv6Packet(src, dst, 6, tcpPayload)
+
+	out := make([]byte, MaxRejectPacketSize)
+	rejectPacket := CreateRejectPacket(packet, out)
+	assert.NotNil(t, rejectPacket)
+
+	// Expected: 40 (ipv6 hdr) + 20 (tcp RST)
+	expectedLen := ipv6.HeaderLen + 20
+	assert.Len(t, rejectPacket, expectedLen)
+
+	// Verify version
+	assert.Equal(t, byte(ipv6.Version<<4), rejectPacket[0]&0xf0)
+	// Verify next header is TCP (6)
+	assert.Equal(t, byte(6), rejectPacket[6])
+	// Verify src/dst are swapped
+	assert.Equal(t, dst.To16(), net.IP(rejectPacket[8:24]))
+	assert.Equal(t, src.To16(), net.IP(rejectPacket[24:40]))
+	// Verify ports are swapped
+	tcpOut := rejectPacket[ipv6.HeaderLen:]
+	assert.Equal(t, uint16(443), binary.BigEndian.Uint16(tcpOut[0:2]))
+	assert.Equal(t, uint16(80), binary.BigEndian.Uint16(tcpOut[2:4]))
+	// RST+ACK flags (since input was SYN without ACK)
+	assert.Equal(t, byte(0b00010100), tcpOut[13])
+	// ack_seq = original seq (1000) + SYN (1) + FIN (0) + segment data (0)
+	assert.Equal(t, uint32(1001), binary.BigEndian.Uint32(tcpOut[8:]))
+}
+
+func Test_CreateRejectPacketIPv6_TCPWithACK(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+
+	// TCP packet with ACK set
+	tcpPayload := make([]byte, 20)
+	tcpPayload[0] = 0x00
+	tcpPayload[1] = 0x50
+	tcpPayload[2] = 0x01
+	tcpPayload[3] = 0xBB
+	binary.BigEndian.PutUint32(tcpPayload[4:], 1000) // seq
+	binary.BigEndian.PutUint32(tcpPayload[8:], 2000) // ack seq
+	tcpPayload[12] = (20 >> 2) << 4                  // data offset
+	tcpPayload[13] = 0b00010000                      // ACK flag
+
+	packet := makeIPv6Packet(src, dst, 6, tcpPayload)
+
+	out := make([]byte, MaxRejectPacketSize)
+	rejectPacket := CreateRejectPacket(packet, out)
+	assert.NotNil(t, rejectPacket)
+
+	tcpOut := rejectPacket[ipv6.HeaderLen:]
+	// RST only (no ACK) since input had ACK
+	assert.Equal(t, byte(0b00000100), tcpOut[13])
+	// seq = original ack_seq
+	assert.Equal(t, uint32(2000), binary.BigEndian.Uint32(tcpOut[4:]))
+}
+
+func Test_CreateRejectPacketIPv6_NoICMPError(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+	out := make([]byte, MaxRejectPacketSize)
+
+	// ICMPv6 error types (1-4) should not generate reject packets
+	for icmpType := byte(1); icmpType <= 4; icmpType++ {
+		payload := make([]byte, 8)
+		payload[0] = icmpType
+		packet := makeIPv6Packet(src, dst, 58, payload)
+
+		rejectPacket := CreateRejectPacket(packet, out)
+		assert.Nil(t, rejectPacket, "ICMPv6 type %d should not generate a reject packet", icmpType)
+	}
+
+	// ICMPv6 non-error types should still generate reject packets
+	nonErrorTypes := []byte{128, 129, 133, 134}
+	for _, icmpType := range nonErrorTypes {
+		payload := make([]byte, 8)
+		payload[0] = icmpType
+		packet := makeIPv6Packet(src, dst, 58, payload)
+
+		rejectPacket := CreateRejectPacket(packet, out)
+		assert.NotNil(t, rejectPacket, "ICMPv6 type %d should generate a reject packet", icmpType)
+	}
+}
+
+func Test_CreateRejectPacketIPv6_TooShort(t *testing.T) {
+	// Packet too short to be valid IPv6
+	out := make([]byte, MaxRejectPacketSize)
+	assert.Nil(t, CreateRejectPacket([]byte{0x60}, out))
+	assert.Nil(t, CreateRejectPacket(make([]byte, 39), out))
+}
+
+func Test_CreateRejectPacketIPv6_ExtensionHeaders(t *testing.T) {
+	src := net.ParseIP("fd00::1")
+	dst := net.ParseIP("fd00::2")
+
+	// IPv6 + Hop-by-Hop extension header + TCP
+	hopByHop := []byte{
+		6,    // next header: TCP
+		0,    // length (8 bytes total)
+		0, 0, // padding
+		0, 0, 0, 0,
+	}
+	tcpPayload := make([]byte, 20)
+	tcpPayload[0] = 0x00
+	tcpPayload[1] = 0x50
+	tcpPayload[2] = 0x01
+	tcpPayload[3] = 0xBB
+	binary.BigEndian.PutUint32(tcpPayload[4:], 1000)
+	binary.BigEndian.PutUint32(tcpPayload[8:], 2000)
+	tcpPayload[12] = (20 >> 2) << 4
+	tcpPayload[13] = 0b00010000 // ACK
+
+	payload := append(hopByHop, tcpPayload...)
+	packet := makeIPv6Packet(src, dst, 0, payload) // next header 0 = Hop-by-Hop
+
+	out := make([]byte, MaxRejectPacketSize)
+	rejectPacket := CreateRejectPacket(packet, out)
+	assert.NotNil(t, rejectPacket)
+
+	// Should produce TCP RST
+	expectedLen := ipv6.HeaderLen + 20
+	assert.Len(t, rejectPacket, expectedLen)
+	assert.Equal(t, byte(6), rejectPacket[6]) // next header is TCP
+	tcpOut := rejectPacket[ipv6.HeaderLen:]
+	assert.Equal(t, byte(0b00000100), tcpOut[13]) // RST only
+}
+
+func TestCreateICMPEchoResponse_IPv4(t *testing.T) {
+	// Build a simple IPv4 ICMP Echo Request
+	packet := make([]byte, 28)
+	packet[0] = 0x45                                   // version 4, IHL 5
+	binary.BigEndian.PutUint16(packet[2:], uint16(28)) // total length
+	packet[8] = 64                                     // TTL
+	packet[9] = 1                                      // protocol ICMP
+	copy(packet[12:16], net.IPv4(10, 0, 0, 1).To4())   // src
+	copy(packet[16:20], net.IPv4(10, 0, 0, 2).To4())   // dst
+	packet[20] = 8                                     // ICMP Echo Request
+
+	out := make([]byte, len(packet))
+	result := CreateICMPEchoResponse(packet, out)
+	assert.NotNil(t, result)
+	assert.Equal(t, byte(0x45), result[0])
+	// src/dst swapped
+	assert.Equal(t, net.IPv4(10, 0, 0, 2).To4(), net.IP(result[12:16]))
+	assert.Equal(t, net.IPv4(10, 0, 0, 1).To4(), net.IP(result[16:20]))
+	// ICMP Echo Reply
+	assert.Equal(t, byte(0), result[20])
+}
+
+func TestCreateICMPEchoResponse_IPv6(t *testing.T) {
+	src := net.ParseIP("fd00::1").To16()
+	dst := net.ParseIP("fd00::2").To16()
+
+	// Build an IPv6 ICMPv6 Echo Request packet
+	// IPv6 header (40 bytes) + ICMPv6 (8 bytes)
+	packet := make([]byte, 48)
+	packet[0] = 0x60        // version 6
+	payloadLen := uint16(8) // ICMPv6 header only
+	binary.BigEndian.PutUint16(packet[4:], payloadLen)
+	packet[6] = 58           // Next Header: ICMPv6
+	packet[7] = 64           // Hop Limit
+	copy(packet[8:24], src)  // src address
+	copy(packet[24:40], dst) // dst address
+
+	// ICMPv6 Echo Request
+	icmp := packet[40:]
+	icmp[0] = 128                           // type: Echo Request
+	icmp[1] = 0                             // code
+	binary.BigEndian.PutUint16(icmp[4:], 1) // identifier
+	binary.BigEndian.PutUint16(icmp[6:], 1) // sequence number
+
+	// Compute correct checksum for the request
+	csum := ipv6PseudoheaderChecksum(src, dst, 58, uint32(payloadLen))
+	binary.BigEndian.PutUint16(icmp[2:], tcpipChecksum(icmp, csum))
+
+	out := make([]byte, len(packet))
+	result := CreateICMPEchoResponse(packet, out)
+	assert.NotNil(t, result)
+
+	// Version should still be 6
+	assert.Equal(t, byte(6), result[0]>>4)
+	// src/dst swapped
+	assert.Equal(t, dst, net.IP(result[8:24]))
+	assert.Equal(t, src, net.IP(result[24:40]))
+	// ICMPv6 Echo Reply type
+	assert.Equal(t, byte(129), result[40])
+
+	// Verify checksum is valid (tcpipChecksum returns 0 when data+checksum is correct)
+	respIcmp := result[40:]
+	verifyCsum := ipv6PseudoheaderChecksum(result[8:24], result[24:40], 58, uint32(payloadLen))
+	assert.Equal(t, uint16(0), tcpipChecksum(respIcmp, verifyCsum))
+}
+
+func TestCreateICMPEchoResponse_IPv6_NotEchoRequest(t *testing.T) {
+	src := net.ParseIP("fd00::1").To16()
+	dst := net.ParseIP("fd00::2").To16()
+
+	packet := make([]byte, 48)
+	packet[0] = 0x60
+	binary.BigEndian.PutUint16(packet[4:], 8)
+	packet[6] = 58
+	packet[7] = 64
+	copy(packet[8:24], src)
+	copy(packet[24:40], dst)
+
+	// ICMPv6 type 1 (Destination Unreachable) - not Echo Request
+	packet[40] = 1
+
+	out := make([]byte, len(packet))
+	result := CreateICMPEchoResponse(packet, out)
+	assert.Nil(t, result)
+}
+
+func TestCreateICMPEchoResponse_IPv6_NotICMPv6(t *testing.T) {
+	src := net.ParseIP("fd00::1").To16()
+	dst := net.ParseIP("fd00::2").To16()
+
+	packet := make([]byte, 48)
+	packet[0] = 0x60
+	binary.BigEndian.PutUint16(packet[4:], 8)
+	packet[6] = 6 // TCP, not ICMPv6
+	packet[7] = 64
+	copy(packet[8:24], src)
+	copy(packet[24:40], dst)
+
+	out := make([]byte, len(packet))
+	result := CreateICMPEchoResponse(packet, out)
+	assert.Nil(t, result)
+}
@@ -15,7 +15,6 @@ import (
 	"time"

 	"github.com/gaissmai/bart"
-	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/header"
@@ -35,7 +34,6 @@ type LightHouse struct {

 	myVpnNetworks      []netip.Prefix
 	myVpnNetworksTable *bart.Lite
-	punchConn          udp.Conn
 	punchy             *Punchy

 	// Local cache of answers from light houses
@@ -63,11 +61,7 @@ type LightHouse struct {
 	interval     atomic.Int64
 	updateCancel context.CancelFunc
 	ifce         EncWriter
-	// bufAlloc lets the lighthouse query/update workers, request handlers
-	// and punchback goroutines acquire correctly sized WireBuffers from
-	// the same pool as the data plane. Set by main.go alongside ifce.
-	bufAlloc   WireBufferAllocator
-	nebulaPort uint32 // 32 bits because protobuf does not have a uint16
+	nebulaPort   uint32 // 32 bits because protobuf does not have a uint16

 	advertiseAddrs atomic.Pointer[[]netip.AddrPort]

@@ -79,9 +73,8 @@ type LightHouse struct {

 	calculatedRemotes atomic.Pointer[bart.Table[[]*calculatedRemote]] // Maps VpnAddr to []*calculatedRemote

-	metrics           *MessageMetrics
-	metricHolepunchTx metrics.Counter
-	l                 *slog.Logger
+	metrics *MessageMetrics
+	l       *slog.Logger
 }

 // NewLightHouseFromConfig will build a Lighthouse struct from the values provided in the config object
@@ -109,15 +102,10 @@ func NewLightHouseFromConfig(ctx context.Context, l *slog.Logger, c *config.C, c
 		myVpnNetworksTable: cs.myVpnNetworksTable,
 		addrMap:            make(map[netip.Addr]*RemoteList),
 		nebulaPort:         nebulaPort,
-		punchConn:          pc,
 		punchy:             p,
 		updateTrigger:      make(chan struct{}, 1),
 		queryChan:          make(chan netip.Addr, c.GetUint32("handshakes.query_buffer", 64)),
-		// Default to a no-prefix pool so the query/update workers and
-		// request handlers have a working WireBufferAllocator before
-		// main.go wires up the real one from the Interface.
-		bufAlloc: NewWireBufferPool(mtu, 0),
-		l:        l,
+		l:                  l,
 	}
 	lighthouses := make([]netip.Addr, 0)
 	h.lighthouses.Store(&lighthouses)
@@ -126,9 +114,6 @@ func NewLightHouseFromConfig(ctx context.Context, l *slog.Logger, c *config.C, c

 	if c.GetBool("stats.lighthouse_metrics", false) {
 		h.metrics = newLighthouseMetrics()
-		h.metricHolepunchTx = metrics.GetOrRegisterCounter("messages.tx.holepunch", nil)
-	} else {
-		h.metricHolepunchTx = metrics.NilCounter{}
 	}

 	err := h.reload(c, true)
@@ -287,16 +272,18 @@ func (lh *LightHouse) reload(c *config.C, initial bool) error {
 	//NOTE: many things will get much simpler when we combine static_host_map and lighthouse.hosts in config
 	if initial || c.HasChanged("static_host_map") || c.HasChanged("static_map.cadence") || c.HasChanged("static_map.network") || c.HasChanged("static_map.lookup_timeout") {
 		// Clean up. Entries still in the static_host_map will be re-built.
-		// Entries no longer present must have their (possible) background DNS goroutines stopped.
-		if existingStaticList := lh.staticList.Load(); existingStaticList != nil {
+		ourselves := lh.myVpnNetworks[0].Addr()
+		oldStaticList := lh.staticList.Load()
+		if oldStaticList != nil {
 			lh.RLock()
-			for staticVpnAddr := range *existingStaticList {
+			for staticVpnAddr := range *oldStaticList {
 				if am, ok := lh.addrMap[staticVpnAddr]; ok && am != nil {
-					am.hr.Cancel()
+					am.ResetForOwner(ourselves)
 				}
 			}
 			lh.RUnlock()
 		}
+
 		// Build a new list based on current config.
 		staticList := make(map[netip.Addr]struct{})
 		err := lh.loadStaticMap(c, staticList)
@@ -304,6 +291,21 @@ func (lh *LightHouse) reload(c *config.C, initial bool) error {
 			return err
 		}

+		// For entries removed from static_host_map, stop the DNS goroutine and drop the cached addrs.
+		// All addrs must come from the lighthouses now that it's no longer a static host.
+		if oldStaticList != nil {
+			lh.RLock()
+			for staticVpnAddr := range *oldStaticList {
+				if _, stillStatic := staticList[staticVpnAddr]; stillStatic {
+					continue
+				}
+				if am, ok := lh.addrMap[staticVpnAddr]; ok && am != nil {
+					am.ClearHostnameResults()
+				}
+			}
+			lh.RUnlock()
+		}
+
 		lh.staticList.Store(&staticList)
 		if !initial {
 			if c.HasChanged("static_host_map") {
@@ -766,22 +768,21 @@ func (lh *LightHouse) startQueryWorker() {
 	}

 	go func() {
-		// Long-lived per-worker WireBuffer; reused for every lighthouse query
-		// this worker issues for the life of the goroutine.
-		buf := lh.bufAlloc.Acquire()
+		nb := make([]byte, 12, 12)
+		out := make([]byte, mtu)

 		for {
 			select {
 			case <-lh.ctx.Done():
 				return
 			case addr := <-lh.queryChan:
-				lh.innerQueryServer(addr, buf)
+				lh.innerQueryServer(addr, nb, out)
 			}
 		}
 	}()
 }

-func (lh *LightHouse) innerQueryServer(addr netip.Addr, buf *WireBuffer) {
+func (lh *LightHouse) innerQueryServer(addr netip.Addr, nb, out []byte) {
 	if lh.IsLighthouseAddr(addr) {
 		return
 	}
@@ -830,7 +831,7 @@ func (lh *LightHouse) innerQueryServer(addr netip.Addr, buf *WireBuffer) {
 				}
 			}

-			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v1Query, buf)
+			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v1Query, nb, out)
 			queried++

 		} else if v == cert.Version2 {
@@ -849,7 +850,7 @@ func (lh *LightHouse) innerQueryServer(addr netip.Addr, buf *WireBuffer) {
 				}
 			}

-			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v2Query, buf)
+			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v2Query, nb, out)
 			queried++

 		} else {
@@ -878,12 +879,8 @@ func (lh *LightHouse) StartUpdateWorker() {
 	go func() {
 		defer clockSource.Stop()

-		// Long-lived per-worker WireBuffer; reused across every periodic
-		// update for the life of this goroutine.
-		buf := lh.bufAlloc.Acquire()
-
 		for {
-			lh.sendUpdate(buf)
+			lh.SendUpdate()

 			select {
 			case <-updateCtx.Done():
@@ -897,15 +894,6 @@ func (lh *LightHouse) StartUpdateWorker() {
 	}()
 }

-// SendUpdate is the public entry point that triggers a one-shot lighthouse
-// update outside the worker loop (e.g. tests or reload paths). It allocates
-// its own WireBuffer since callers don't already own one.
-func (lh *LightHouse) SendUpdate() {
-	buf := lh.bufAlloc.Acquire()
-	defer lh.bufAlloc.Release(buf)
-	lh.sendUpdate(buf)
-}
-
 // TriggerUpdate requests an immediate lighthouse update. This is a non-blocking
 // operation intended to be called after a handshake completes with a lighthouse,
 // so the lighthouse has our current addresses without waiting for the next
@@ -917,7 +905,7 @@ func (lh *LightHouse) TriggerUpdate() {
 	}
 }

-func (lh *LightHouse) sendUpdate(buf *WireBuffer) {
+func (lh *LightHouse) SendUpdate() {
 	var v4 []*V4AddrPort
 	var v6 []*V6AddrPort

@@ -943,6 +931,9 @@ func (lh *LightHouse) sendUpdate(buf *WireBuffer) {
 		}
 	}

+	nb := make([]byte, 12, 12)
+	out := make([]byte, mtu)
+
 	var v1Update, v2Update []byte
 	var err error
 	updated := 0
@@ -993,7 +984,7 @@ func (lh *LightHouse) sendUpdate(buf *WireBuffer) {
 				}
 			}

-			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v1Update, buf)
+			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v1Update, nb, out)
 			updated++

 		} else if v == cert.Version2 {
@@ -1022,7 +1013,7 @@ func (lh *LightHouse) sendUpdate(buf *WireBuffer) {
 				}
 			}

-			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v2Update, buf)
+			lh.ifce.SendMessageToVpnAddr(header.LightHouse, 0, lhVpnAddr, v2Update, nb, out)
 			updated++

 		} else {
@@ -1038,11 +1029,9 @@ func (lh *LightHouse) sendUpdate(buf *WireBuffer) {
 }

 type LightHouseHandler struct {
-	lh *LightHouse
-	// buf is the long-lived per-handler wire scratch. NewRequestHandler is
-	// called once per data-plane receive goroutine, so buf is owned by that
-	// goroutine and reused for every lighthouse send the handler issues.
-	buf  *WireBuffer
+	lh   *LightHouse
+	nb   []byte
+	out  []byte
 	pb   []byte
 	meta *NebulaMeta
 	l    *slog.Logger
@@ -1051,7 +1040,8 @@ type LightHouseHandler struct {
 func (lh *LightHouse) NewRequestHandler() *LightHouseHandler {
 	lhh := &LightHouseHandler{
 		lh:  lh,
-		buf: lh.bufAlloc.Acquire(),
+		nb:  make([]byte, 12, 12),
+		out: make([]byte, mtu),
 		l:   lh.l,
 		pb:  make([]byte, mtu),

@@ -1188,7 +1178,7 @@ func (lhh *LightHouseHandler) handleHostQuery(n *NebulaMeta, fromVpnAddrs []neti
 	}

 	lhh.lh.metricTx(NebulaMeta_HostQueryReply, 1)
-	w.SendMessageToVpnAddr(header.LightHouse, 0, fromVpnAddrs[0], lhh.pb[:ln], lhh.buf)
+	w.SendMessageToVpnAddr(header.LightHouse, 0, fromVpnAddrs[0], lhh.pb[:ln], lhh.nb, lhh.out[:0])

 	lhh.sendHostPunchNotification(n, fromVpnAddrs, queryVpnAddr, w)
 }
@@ -1248,7 +1238,7 @@ func (lhh *LightHouseHandler) sendHostPunchNotification(n *NebulaMeta, fromVpnAd
 	}

 	lhh.lh.metricTx(NebulaMeta_HostPunchNotification, 1)
-	w.SendMessageToVpnAddr(header.LightHouse, 0, punchNotifDest, lhh.pb[:ln], lhh.buf)
+	w.SendMessageToVpnAddr(header.LightHouse, 0, punchNotifDest, lhh.pb[:ln], lhh.nb, lhh.out[:0])
 }

 func (lhh *LightHouseHandler) coalesceAnswers(v cert.Version, c *cache, n *NebulaMeta) {
@@ -1405,7 +1395,7 @@ func (lhh *LightHouseHandler) handleHostUpdateNotification(n *NebulaMeta, fromVp
 	}

 	lhh.lh.metricTx(NebulaMeta_HostUpdateNotificationAck, 1)
-	w.SendMessageToVpnAddr(header.LightHouse, 0, fromVpnAddrs[0], lhh.pb[:ln], lhh.buf)
+	w.SendMessageToVpnAddr(header.LightHouse, 0, fromVpnAddrs[0], lhh.pb[:ln], lhh.nb, lhh.out[:0])
 }

 func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpnAddrs []netip.Addr, w EncWriter) {
@@ -1426,61 +1416,25 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 		return
 	}

-	empty := []byte{0}
-	punch := func(vpnPeer netip.AddrPort, logVpnAddr netip.Addr) {
-		if !vpnPeer.IsValid() {
-			return
-		}
-
-		go func() {
-			time.Sleep(lhh.lh.punchy.GetDelay())
-			lhh.lh.metricHolepunchTx.Inc(1)
-			lhh.lh.punchConn.WriteTo(empty, vpnPeer)
-		}()
-
-		if lhh.l.Enabled(context.Background(), slog.LevelDebug) {
-			lhh.l.Debug("Punching",
-				"vpnPeer", vpnPeer,
-				"logVpnAddr", logVpnAddr,
-			)
-		}
-	}
-
 	remoteAllowList := lhh.lh.GetRemoteAllowList()
 	for _, a := range n.Details.V4AddrPorts {
 		b := protoV4AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
-			punch(b, detailsVpnAddr)
+			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
 		}
 	}

 	for _, a := range n.Details.V6AddrPorts {
 		b := protoV6AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
-			punch(b, detailsVpnAddr)
+			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
 		}
 	}

 	// This sends a nebula test packet to the host trying to contact us. In the case
 	// of a double nat or other difficult scenario, this may help establish
-	// a tunnel.
-	if lhh.lh.punchy.GetRespond() {
-		go func() {
-			time.Sleep(lhh.lh.punchy.GetRespondDelay())
-			if lhh.l.Enabled(context.Background(), slog.LevelDebug) {
-				lhh.l.Debug("Sending a nebula test packet",
-					"vpnAddr", detailsVpnAddr,
-				)
-			}
-			// We acquire and release a fresh buf within this goroutine so it
-			// returns to the pool once the punchback send completes. We
-			// should move this into a timerwheel or a single goroutine
-			// managed by a channel.
-			pbuf := lhh.lh.bufAlloc.Acquire()
-			defer lhh.lh.bufAlloc.Release(pbuf)
-			w.SendMessageToVpnAddr(header.Test, header.TestRequest, detailsVpnAddr, []byte(""), pbuf)
-		}()
-	}
+	// a tunnel. ScheduleRespond is a no-op when punchy.respond is disabled.
+	lhh.lh.punchy.ScheduleRespond(detailsVpnAddr)
 }

 func protoAddrToNetAddr(addr *Addr) netip.Addr {
@@ -303,6 +303,132 @@ func TestLighthouse_reload(t *testing.T) {
 	require.NoError(t, err)
 }

+// TestLighthouse_reloadStaticHostMap verifies that reloading static_host_map applies the new
+// config rather than appending to it. See issue #718.
+func TestLighthouse_reloadStaticHostMap(t *testing.T) {
+	l := test.NewLogger()
+	c := config.NewC(l)
+	c.Settings["lighthouse"] = map[string]any{"am_lighthouse": true}
+	c.Settings["listen"] = map[string]any{"port": 4242}
+	c.Settings["static_host_map"] = map[string]any{
+		"10.128.0.2": []any{"1.1.1.1:4242"},
+	}
+
+	myVpnNet := netip.MustParsePrefix("10.128.0.1/24")
+	nt := new(bart.Lite)
+	nt.Insert(myVpnNet)
+	cs := &CertState{
+		myVpnNetworks:      []netip.Prefix{myVpnNet},
+		myVpnNetworksTable: nt,
+	}
+
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
+	require.NoError(t, err)
+
+	staticHost := netip.MustParseAddr("10.128.0.2")
+	otherHost := netip.MustParseAddr("10.128.0.3")
+
+	// Capture the RemoteList pointer up front; an in-flight handshake would hold the same one
+	// on hostinfo.remotes, so it must reflect every reload below.
+	pinned := lh.Query(staticHost)
+	require.NotNil(t, pinned)
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, pinned.CopyAddrs([]netip.Prefix{}))
+
+	// Replace the remote address. The new address should be the only entry.
+	nc := map[string]any{
+		"static_host_map": map[string]any{
+			"10.128.0.2": []any{"2.2.2.2:4242"},
+		},
+	}
+	rc, err := yaml.Marshal(nc)
+	require.NoError(t, err)
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	rl := lh.Query(staticHost)
+	require.NotNil(t, rl)
+	assert.Same(t, pinned, rl, "RemoteList pointer must stay stable so in-flight handshakes pick up the change")
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("2.2.2.2:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+
+	// Reload back to the original IP. Mirrors the round-trip in issue #718 step 6-8 where
+	// the buggy reload produced [1.1.1.1, 2.2.2.2, 1.1.1.1] instead of [1.1.1.1].
+	nc = map[string]any{
+		"static_host_map": map[string]any{
+			"10.128.0.2": []any{"1.1.1.1:4242"},
+		},
+	}
+	rc, err = yaml.Marshal(nc)
+	require.NoError(t, err)
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	rl = lh.Query(staticHost)
+	require.NotNil(t, rl)
+	assert.Same(t, pinned, rl)
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+
+	// Reload with the same config. An unchanged entry must not duplicate.
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	rl = lh.Query(staticHost)
+	require.NotNil(t, rl)
+	assert.Same(t, pinned, rl)
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+
+	// Switch back to 2.2.2.2 so the rest of the test continues against a known address.
+	nc = map[string]any{
+		"static_host_map": map[string]any{
+			"10.128.0.2": []any{"2.2.2.2:4242"},
+		},
+	}
+	rc, err = yaml.Marshal(nc)
+	require.NoError(t, err)
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	// Add a second host alongside the first. Both should be present, neither duplicated.
+	nc = map[string]any{
+		"static_host_map": map[string]any{
+			"10.128.0.2": []any{"2.2.2.2:4242"},
+			"10.128.0.3": []any{"3.3.3.3:4242"},
+		},
+	}
+	rc, err = yaml.Marshal(nc)
+	require.NoError(t, err)
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	rl = lh.Query(staticHost)
+	require.NotNil(t, rl)
+	assert.Same(t, pinned, rl, "adding a sibling entry must not displace the existing RemoteList")
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("2.2.2.2:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+
+	rl = lh.Query(otherHost)
+	require.NotNil(t, rl)
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("3.3.3.3:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+
+	// Drop the first host entirely. The vpnAddr is no longer marked static, our owner
+	// contribution is cleared, but the addrMap entry stays in place so non-static cache
+	// data (from lighthouse queries) on the same RemoteList isn't lost. In-flight handshakes
+	// that already had the pointer see an empty address list rather than retrying stale ones.
+	nc = map[string]any{
+		"static_host_map": map[string]any{
+			"10.128.0.3": []any{"3.3.3.3:4242"},
+		},
+	}
+	rc, err = yaml.Marshal(nc)
+	require.NoError(t, err)
+	require.NoError(t, c.ReloadConfigString(string(rc)))
+
+	_, isStatic := lh.GetStaticHostList()[staticHost]
+	assert.False(t, isStatic)
+
+	rl = lh.Query(staticHost)
+	require.NotNil(t, rl)
+	assert.Same(t, pinned, rl)
+	assert.Empty(t, rl.CopyAddrs([]netip.Prefix{}))
+
+	rl = lh.Query(otherHost)
+	require.NotNil(t, rl)
+	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("3.3.3.3:4242")}, rl.CopyAddrs([]netip.Prefix{}))
+}
+
 func newLHHostRequest(fromAddr netip.AddrPort, myVpnIp, queryVpnIp netip.Addr, lhh *LightHouseHandler) testLhReply {
 	req := &NebulaMeta{
 		Type:    NebulaMeta_HostQuery,
@@ -372,12 +498,12 @@ type testEncWriter struct {
 	protocolVersion cert.Version
 }

-func (tw *testEncWriter) SendVia(via *HostInfo, relay *Relay, ad []byte, buf *WireBuffer) {
+func (tw *testEncWriter) SendVia(via *HostInfo, relay *Relay, ad, nb, out []byte, nocopy bool) {
 }
 func (tw *testEncWriter) Handshake(vpnIp netip.Addr) {
 }

-func (tw *testEncWriter) SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p []byte, _ *WireBuffer) {
+func (tw *testEncWriter) SendMessageToHostInfo(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p, _, _ []byte) {
 	msg := &NebulaMeta{}
 	err := msg.Unmarshal(p)
 	if tw.metaFilter == nil || msg.Type == *tw.metaFilter {
@@ -394,7 +520,7 @@ func (tw *testEncWriter) SendMessageToHostInfo(t header.MessageType, st header.M
 	}
 }

-func (tw *testEncWriter) SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnIp netip.Addr, p []byte, _ *WireBuffer) {
+func (tw *testEncWriter) SendMessageToVpnAddr(t header.MessageType, st header.MessageSubType, vpnIp netip.Addr, p, _, _ []byte) {
 	msg := &NebulaMeta{}
 	err := msg.Unmarshal(p)
 	if tw.metaFilter == nil || msg.Type == *tw.metaFilter {
@@ -55,7 +55,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	}
 	l.Info("Firewall started", "firewallHashes", fw.GetRuleHashes())

-	ssh, err := sshd.NewSSHServer(l.With("subsystem", "sshd"))
+	ssh, err := sshd.NewSSHServer(ctx, l.With("subsystem", "sshd"))
 	if err != nil {
 		return nil, util.ContextualizeIfNeeded("Error while creating SSH server", err)
 	}
@@ -170,7 +170,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	}

 	hostMap := NewHostMapFromConfig(l, c)
-	punchy := NewPunchyFromConfig(l, c)
+	punchy := NewPunchyFromConfig(l, c, udpConns[0])
 	connManager := newConnectionManagerFromConfig(l, c, hostMap, punchy)
 	lightHouse, err := NewLightHouseFromConfig(ctx, l, c, pki.getCertState(), udpConns[0], punchy)
 	if err != nil {
@@ -194,7 +194,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	handshakeManager := NewHandshakeManager(l, hostMap, lightHouse, udpConns[0], handshakeConfig)
 	lightHouse.handshakeTrigger = handshakeManager.trigger

-	ds, err := newDnsServerFromConfig(ctx, l, pki.getCertState(), hostMap, c)
+	ds, err := newDnsServerFromConfig(ctx, l, pki, hostMap, c)
 	if err != nil {
 		l.Warn("Failed to start DNS responder", "error", err)
 	}
@@ -232,7 +232,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev

 		ifce.writers = udpConns
 		lightHouse.ifce = ifce
-		lightHouse.bufAlloc = ifce.bufAlloc

 		ifce.RegisterConfigChangeCallbacks(c)
 		ifce.reloadDisconnectInvalid(c)
@@ -241,6 +240,8 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev

 		handshakeManager.f = ifce
 		go handshakeManager.Run(ctx)
+
+		punchy.Start(ctx, ifce, hostMap, lightHouse)
 	}

 	stats, err := newStatsServerFromConfig(ctx, l, c, buildVersion, configTest)
@@ -13,6 +13,8 @@ type MessageMetrics struct {

 	rxUnknown metrics.Counter
 	txUnknown metrics.Counter
+
+	rxInvalid metrics.Counter
 }

 func (m *MessageMetrics) Rx(t header.MessageType, s header.MessageSubType, i int64) {
@@ -33,6 +35,11 @@ func (m *MessageMetrics) Tx(t header.MessageType, s header.MessageSubType, i int
 		}
 	}
 }
+func (m *MessageMetrics) RxInvalid(i int64) {
+	if m != nil && m.rxInvalid != nil {
+		m.rxInvalid.Inc(i)
+	}
+}

 func newMessageMetrics() *MessageMetrics {
 	gen := func(t string) [][]metrics.Counter {
@@ -56,6 +63,7 @@ func newMessageMetrics() *MessageMetrics {

 		rxUnknown: metrics.GetOrRegisterCounter("messages.rx.other", nil),
 		txUnknown: metrics.GetOrRegisterCounter("messages.tx.other", nil),
+		rxInvalid: metrics.GetOrRegisterCounter("messages.rx.invalid", nil),
 	}
 }

@@ -1,86 +0,0 @@
-package nebula
-
-import (
-	"crypto/cipher"
-	"encoding/binary"
-	"errors"
-
-	"github.com/flynn/noise"
-)
-
-type endianness interface {
-	PutUint64(b []byte, v uint64)
-}
-
-var noiseEndianness endianness = binary.BigEndian
-
-// NonceSize is the AEAD nonce length used by all ciphers nebula supports
-// today (AES-GCM and ChaCha20-Poly1305 both use 96-bit nonces). Encrypt-
-// and DecryptDanger lay out the nonce as 4 zero bytes followed by an 8-byte
-// big-endian counter; if a future cipher with a different nonce size is
-// added, this constant and those layouts must change together.
-const NonceSize = 12
-
-// AEADOverhead is the AEAD authentication tag length the ciphers nebula
-// supports append to ciphertext. Both AES-GCM and ChaCha20-Poly1305 use
-// 128-bit tags. NebulaCipherState.Overhead() returns this dynamically from
-// the cipher; the constant is for sizing buffers at construction time.
-const AEADOverhead = 16
-
-type NebulaCipherState struct {
-	c cipher.AEAD
-}
-
-func NewNebulaCipherState(s *noise.CipherState) *NebulaCipherState {
-	x := s.Cipher()
-	return &NebulaCipherState{c: x.(cipher.AEAD)}
-}
-
-// EncryptDanger encrypts and authenticates a given payload.
-//
-// out is a destination slice to hold the output of the EncryptDanger operation.
-//   - ad is additional data, which will be authenticated and appended to out, but not encrypted.
-//   - plaintext is encrypted, authenticated and appended to out.
-//   - n is a nonce value which must never be re-used with this key.
-//   - nb is a buffer used for temporary storage in the implementation of this call, which should
-//     be re-used by callers to minimize garbage collection.
-func (s *NebulaCipherState) EncryptDanger(out, ad, plaintext []byte, n uint64, nb []byte) ([]byte, error) {
-	if s != nil {
-		// TODO: Is this okay now that we have made messageCounter atomic?
-		// Alternative may be to split the counter space into ranges
-		//if n <= s.n {
-		//	return nil, errors.New("CRITICAL: a duplicate counter value was used")
-		//}
-		//s.n = n
-		nb[0] = 0
-		nb[1] = 0
-		nb[2] = 0
-		nb[3] = 0
-		noiseEndianness.PutUint64(nb[4:], n)
-		out = s.c.Seal(out, nb, plaintext, ad)
-		//l.Debugf("Encryption: outlen: %d, nonce: %d, ad: %s, plainlen %d", len(out), n, ad, len(plaintext))
-		return out, nil
-	} else {
-		return nil, errors.New("no cipher state available to encrypt")
-	}
-}
-
-func (s *NebulaCipherState) DecryptDanger(out, ad, ciphertext []byte, n uint64, nb []byte) ([]byte, error) {
-	if s != nil {
-		nb[0] = 0
-		nb[1] = 0
-		nb[2] = 0
-		nb[3] = 0
-		noiseEndianness.PutUint64(nb[4:], n)
-		return s.c.Open(out, nb, ciphertext, ad)
-	} else {
-		return []byte{}, nil
-	}
-}
-
-func (s *NebulaCipherState) Overhead() int {
-	if s != nil {
-		return s.c.Overhead()
-	}
-	return 0
-}
@@ -0,0 +1,53 @@
+package noiseutil
+
+import (
+	"crypto/cipher"
+	"encoding/binary"
+	"errors"
+
+	"github.com/flynn/noise"
+)
+
+// CipherStateAESGCM is the data-plane wrapper for the AES-GCM AEAD cipher.
+// AES-GCM uses big-endian nonce encoding per the Noise spec.
+type CipherStateAESGCM struct {
+	c cipher.AEAD
+}
+
+// NewCipherStateAESGCM extracts the underlying AEAD from the post-handshake noise.CipherState.
+// The caller is responsible for ensuring the noise cipher is actually AES-GCM,
+// otherwise the type assertion still succeeds but the nonce endianness will be wrong on the wire.
+func NewCipherStateAESGCM(s *noise.CipherState) *CipherStateAESGCM {
+	return &CipherStateAESGCM{c: s.Cipher().(cipher.AEAD)}
+}
+
+func (s *CipherStateAESGCM) EncryptDanger(out, ad, plaintext []byte, n uint64, nb []byte) ([]byte, error) {
+	if s == nil {
+		return nil, errors.New("no cipher state available to encrypt")
+	}
+	nb[0] = 0
+	nb[1] = 0
+	nb[2] = 0
+	nb[3] = 0
+	binary.BigEndian.PutUint64(nb[4:], n)
+	return s.c.Seal(out, nb, plaintext, ad), nil
+}
+
+func (s *CipherStateAESGCM) DecryptDanger(out, ad, ciphertext []byte, n uint64, nb []byte) ([]byte, error) {
+	if s == nil {
+		return []byte{}, nil
+	}
+	nb[0] = 0
+	nb[1] = 0
+	nb[2] = 0
+	nb[3] = 0
+	binary.BigEndian.PutUint64(nb[4:], n)
+	return s.c.Open(out, nb, ciphertext, ad)
+}
+
+func (s *CipherStateAESGCM) Overhead() int {
+	if s == nil {
+		return 0
+	}
+	return s.c.Overhead()
+}
@@ -0,0 +1,52 @@
+package noiseutil
+
+import (
+	"crypto/cipher"
+	"encoding/binary"
+	"errors"
+
+	"github.com/flynn/noise"
+)
+
+// CipherStateChaChaPoly is the data-plane wrapper for the ChaCha20-Poly1305 AEAD cipher.
+// ChaCha20-Poly1305 uses little-endian nonce encoding per the Noise spec.
+type CipherStateChaChaPoly struct {
+	c cipher.AEAD
+}
+
+// NewCipherStateChaChaPoly extracts the underlying AEAD from the post-handshake noise.CipherState.
+// The caller is responsible for ensuring the noise cipher is actually ChaCha20-Poly1305.
+func NewCipherStateChaChaPoly(s *noise.CipherState) *CipherStateChaChaPoly {
+	return &CipherStateChaChaPoly{c: s.Cipher().(cipher.AEAD)}
+}
+
+func (s *CipherStateChaChaPoly) EncryptDanger(out, ad, plaintext []byte, n uint64, nb []byte) ([]byte, error) {
+	if s == nil {
+		return nil, errors.New("no cipher state available to encrypt")
+	}
+	nb[0] = 0
+	nb[1] = 0
+	nb[2] = 0
+	nb[3] = 0
+	binary.LittleEndian.PutUint64(nb[4:], n)
+	return s.c.Seal(out, nb, plaintext, ad), nil
+}
+
+func (s *CipherStateChaChaPoly) DecryptDanger(out, ad, ciphertext []byte, n uint64, nb []byte) ([]byte, error) {
+	if s == nil {
+		return []byte{}, nil
+	}
+	nb[0] = 0
+	nb[1] = 0
+	nb[2] = 0
+	nb[3] = 0
+	binary.LittleEndian.PutUint64(nb[4:], n)
+	return s.c.Open(out, nb, ciphertext, ad)
+}
+
+func (s *CipherStateChaChaPoly) Overhead() int {
+	if s == nil {
+		return 0
+	}
+	return s.c.Overhead()
+}
@@ -0,0 +1,40 @@
+package noiseutil
+
+import (
+	"fmt"
+
+	"github.com/flynn/noise"
+)
+
+// CipherState is the post-handshake AEAD cipher used for the data plane.
+// Each supported cipher has its own concrete implementation in this package with the nonce endianness hardcoded,
+// so the encrypt/decrypt fast path avoids interface dispatch on the byte order.
+type CipherState interface {
+	// EncryptDanger encrypts and authenticates a given payload.
+	//
+	// out is a destination slice to hold the output of the EncryptDanger operation.
+	//   - ad is additional data, which will be authenticated and appended to out, but not encrypted.
+	//   - plaintext is encrypted, authenticated and appended to out.
+	//   - n is a nonce value which must never be re-used with this key.
+	//   - nb is a scratch buffer used to assemble the nonce.
+	EncryptDanger(out, ad, plaintext []byte, n uint64, nb []byte) ([]byte, error)
+
+	// DecryptDanger authenticates and decrypts a given payload, with the same argument shape as EncryptDanger.
+	DecryptDanger(out, ad, ciphertext []byte, n uint64, nb []byte) ([]byte, error)
+
+	// Overhead returns the AEAD tag size, or 0 if the receiver is nil.
+	Overhead() int
+}
+
+// NewCipherState wraps the post-handshake noise.CipherState in the per-cipher type that matches cipherFunc.
+// cipherFunc must be the same cipher used to build the noise CipherSuite that produced s.
+func NewCipherState(s *noise.CipherState, cipherFunc noise.CipherFunc) CipherState {
+	switch cipherFunc.CipherName() {
+	case CipherAESGCM.CipherName():
+		return NewCipherStateAESGCM(s)
+	case noise.CipherChaChaPoly.CipherName():
+		return NewCipherStateChaChaPoly(s)
+	default:
+		panic(fmt.Sprintf("noiseutil: unsupported cipher %q", cipherFunc.CipherName()))
+	}
+}
@@ -0,0 +1,166 @@
+package noiseutil
+
+import (
+	"testing"
+
+	"github.com/flynn/noise"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCipherStateAESGCMRoundtrip(t *testing.T) {
+	enc, dec := buildCipherStates(t, CipherAESGCM)
+	roundtrip(t, NewCipherStateAESGCM(enc), NewCipherStateAESGCM(dec))
+}
+
+func TestCipherStateChaChaPolyRoundtrip(t *testing.T) {
+	enc, dec := buildCipherStates(t, noise.CipherChaChaPoly)
+	roundtrip(t, NewCipherStateChaChaPoly(enc), NewCipherStateChaChaPoly(dec))
+}
+
+func TestNewCipherStateDispatch(t *testing.T) {
+	encA, _ := buildCipherStates(t, CipherAESGCM)
+	encC, _ := buildCipherStates(t, noise.CipherChaChaPoly)
+
+	assert.IsType(t, &CipherStateAESGCM{}, NewCipherState(encA, CipherAESGCM))
+	assert.IsType(t, &CipherStateChaChaPoly{}, NewCipherState(encC, noise.CipherChaChaPoly))
+}
+
+func TestNewCipherStateUnsupportedPanics(t *testing.T) {
+	enc, _ := buildCipherStates(t, CipherAESGCM)
+	assert.Panics(t, func() {
+		NewCipherState(enc, fakeCipher{})
+	})
+}
+
+type fakeCipher struct{}
+
+func (fakeCipher) Cipher(k [32]byte) noise.Cipher { return nil }
+func (fakeCipher) CipherName() string             { return "Fake" }
+
+// buildCipherStates runs an in-memory NN handshake with the requested cipher
+// to produce a pair of post-handshake CipherStates that share keys.
+func buildCipherStates(t *testing.T, c noise.CipherFunc) (*noise.CipherState, *noise.CipherState) {
+	t.Helper()
+	suite := noise.NewCipherSuite(noise.DH25519, c, noise.HashSHA256)
+	cfg := noise.Config{CipherSuite: suite, Pattern: noise.HandshakeNN}
+	cfg.Initiator = true
+	hsI, err := noise.NewHandshakeState(cfg)
+	require.NoError(t, err)
+	cfg.Initiator = false
+	hsR, err := noise.NewHandshakeState(cfg)
+	require.NoError(t, err)
+
+	msg, _, _, err := hsI.WriteMessage(nil, nil)
+	require.NoError(t, err)
+	_, _, _, err = hsR.ReadMessage(nil, msg)
+	require.NoError(t, err)
+
+	msg, dR, _, err := hsR.WriteMessage(nil, nil)
+	require.NoError(t, err)
+	_, eI, _, err := hsI.ReadMessage(nil, msg)
+	require.NoError(t, err)
+	require.NotNil(t, eI)
+	require.NotNil(t, dR)
+
+	// noise returns (cs1, cs2) where cs1 is the initiator->responder cipher.
+	return eI, dR
+}
+
+func roundtrip(t *testing.T, enc, dec CipherState) {
+	t.Helper()
+	plaintext := []byte("nebula cipher state roundtrip")
+	ad := []byte("aad")
+	nb := make([]byte, 12)
+
+	ct, err := enc.EncryptDanger(nil, ad, plaintext, 1, nb)
+	require.NoError(t, err)
+	assert.NotEqual(t, plaintext, ct)
+
+	pt, err := dec.DecryptDanger(nil, ad, ct, 1, nb)
+	require.NoError(t, err)
+	assert.Equal(t, plaintext, pt)
+
+	// Wrong nonce must fail authentication.
+	_, err = dec.DecryptDanger(nil, ad, ct, 2, nb)
+	require.Error(t, err)
+
+	assert.Equal(t, enc.Overhead(), dec.Overhead())
+	assert.Equal(t, 16, enc.Overhead())
+}
+
+func BenchmarkCipherStateEncryptAESGCM(b *testing.B) {
+	enc, _ := buildCipherStatesB(b, CipherAESGCM)
+	benchEncryptCipherState(b, NewCipherState(enc, CipherAESGCM))
+}
+
+func BenchmarkCipherStateEncryptChaChaPoly(b *testing.B) {
+	enc, _ := buildCipherStatesB(b, noise.CipherChaChaPoly)
+	benchEncryptCipherState(b, NewCipherState(enc, noise.CipherChaChaPoly))
+}
+
+func benchEncryptCipherState(b *testing.B, cs CipherState) {
+	plaintext := make([]byte, 1280)
+	ad := make([]byte, 16)
+	nb := make([]byte, 12)
+	out := make([]byte, 0, len(plaintext)+cs.Overhead())
+	b.ResetTimer()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		var err error
+		out, err = cs.EncryptDanger(out[:0], ad, plaintext, uint64(i+1), nb)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func buildCipherStatesB(b *testing.B, c noise.CipherFunc) (*noise.CipherState, *noise.CipherState) {
+	b.Helper()
+	suite := noise.NewCipherSuite(noise.DH25519, c, noise.HashSHA256)
+	cfg := noise.Config{CipherSuite: suite, Pattern: noise.HandshakeNN}
+	cfg.Initiator = true
+	hsI, err := noise.NewHandshakeState(cfg)
+	if err != nil {
+		b.Fatal(err)
+	}
+	cfg.Initiator = false
+	hsR, err := noise.NewHandshakeState(cfg)
+	if err != nil {
+		b.Fatal(err)
+	}
+	msg, _, _, err := hsI.WriteMessage(nil, nil)
+	if err != nil {
+		b.Fatal(err)
+	}
+	if _, _, _, err := hsR.ReadMessage(nil, msg); err != nil {
+		b.Fatal(err)
+	}
+	msg, dR, _, err := hsR.WriteMessage(nil, nil)
+	if err != nil {
+		b.Fatal(err)
+	}
+	_, eI, _, err := hsI.ReadMessage(nil, msg)
+	if err != nil {
+		b.Fatal(err)
+	}
+	return eI, dR
+}
+
+func TestCipherStateNilSafety(t *testing.T) {
+	var aes *CipherStateAESGCM
+	_, err := aes.EncryptDanger(nil, nil, nil, 0, make([]byte, 12))
+	require.Error(t, err)
+	out, err := aes.DecryptDanger(nil, nil, nil, 0, make([]byte, 12))
+	require.NoError(t, err)
+	assert.Empty(t, out)
+	assert.Equal(t, 0, aes.Overhead())
+
+	var cc *CipherStateChaChaPoly
+	_, err = cc.EncryptDanger(nil, nil, nil, 0, make([]byte, 12))
+	require.Error(t, err)
+	out, err = cc.DecryptDanger(nil, nil, nil, 0, make([]byte, 12))
+	require.NoError(t, err)
+	assert.Empty(t, out)
+	assert.Equal(t, 0, cc.Overhead())
+}
@@ -20,24 +20,46 @@ const (
 	minFwPacketLen = 4
 )

-func (f *Interface) readOutsidePackets(via ViaSender, buf *WireBuffer, packet []byte, lhf *LightHouseHandler, q int, localCache firewall.ConntrackCache) {
-	h := buf.H
+var ErrOutOfWindow = errors.New("out of window packet")
+
+func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
 	err := h.Parse(packet)
 	if err != nil {
 		// Hole punch packets are 0 or 1 byte big, so lets ignore printing those errors
+		// TODO: record metrics for rx holepunch/punchy packets?
 		if len(packet) > 1 {
-			f.l.Info("Error while parsing inbound packet",
-				"from", via,
-				"error", err,
-				"packet", packet,
-			)
+			f.messageMetrics.RxInvalid(1)
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				f.l.Debug("Error while parsing inbound packet",
+					"from", via,
+					"error", err,
+					"packet", packet,
+				)
+			}
+		}
+		return
+	}
+
+	if h.Version != header.Version {
+		f.messageMetrics.RxInvalid(1)
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("Unexpected header version received", "from", via)
+		}
+		return
+	}
+
+	// Check before processing to see if this is a expected type/subtype
+	if !h.IsValidSubType() {
+		f.messageMetrics.RxInvalid(1)
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("Unexpected packet received", "from", via)
 		}
 		return
 	}

-	//l.Error("in packet ", header, packet[HeaderLen:])
 	if !via.IsRelayed {
 		if f.myVpnNetworksTable.Contains(via.UdpAddr.Addr()) {
+			f.messageMetrics.RxInvalid(1)
 			if f.l.Enabled(context.Background(), slog.LevelDebug) {
 				f.l.Debug("Refusing to process double encrypted packet", "from", via)
 			}
@@ -45,217 +67,198 @@ func (f *Interface) readOutsidePackets(via ViaSender, buf *WireBuffer, packet []
 		}
 	}

+	// don't keep Rx metrics for message type, since you can see those in the tun metrics
+	if h.Type != header.Message {
+		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
+	}
+
+	// Unencrypted packets
+	switch h.Type {
+	case header.Handshake:
+		f.handshakeManager.HandleIncoming(via, packet, h)
+		return
+
+	case header.RecvError:
+		f.handleRecvError(via.UdpAddr, h)
+		return
+	}
+
+	// Relay packets are special
+	isMessageRelay := (h.Type == header.Message && h.Subtype == header.MessageRelay)
+
 	var hostinfo *HostInfo
-	// verify if we've seen this index before, otherwise respond to the handshake initiation
-	if h.Type == header.Message && h.Subtype == header.MessageRelay {
+	if isMessageRelay {
 		hostinfo = f.hostMap.QueryRelayIndex(h.RemoteIndex)
 	} else {
 		hostinfo = f.hostMap.QueryIndex(h.RemoteIndex)
 	}

-	var ci *ConnectionState
-	if hostinfo != nil {
-		ci = hostinfo.ConnectionState
+	// At this point we should have a valid existing tunnel, verify and send
+	// recvError if necessary
+	if hostinfo == nil || hostinfo.ConnectionState == nil {
+		if !via.IsRelayed {
+			f.maybeSendRecvError(via.UdpAddr, h.RemoteIndex)
+		}
+		return
 	}

+	// All remaining packets are encrypted
+	ci := hostinfo.ConnectionState
+	if !ci.window.Check(f.l, h.MessageCounter) {
+		return
+	}
+
+	// Relay packets are special
+	if isMessageRelay {
+		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
+
+		return
+	}
+
+	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
+	if err != nil {
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			hostinfo.logger(f.l).Debug("Failed to decrypt packet",
+				"error", err,
+				"from", via,
+				"header", h,
+			)
+		}
+		return
+	}
+
+	// Roam before we respond
+	f.handleHostRoaming(hostinfo, via)
+	f.connectionManager.In(hostinfo)
+
 	switch h.Type {
 	case header.Message:
-		if !f.handleEncrypted(ci, via, h) {
-			return
-		}
-
 		switch h.Subtype {
 		case header.MessageNone:
-			if !f.decryptToTun(hostinfo, h.MessageCounter, buf, packet, q, localCache) {
-				return
-			}
-		case header.MessageRelay:
-			// The entire body is sent as AD, not encrypted.
-			// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
-			// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
-			// otherwise malformed (meaning, there is no trailing 16 byte AEAD value), then this will result in at worst a 0-length slice
-			// which will gracefully fail in the DecryptDanger call.
-			signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-			signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-			// AAD-only validation: passing dst=nil since there's no plaintext
-			// to recover (ciphertext is just the trailing AEAD tag).
-			if _, err = hostinfo.ConnectionState.dKey.DecryptDanger(nil, signedPayload, signatureValue, h.MessageCounter, buf.NB); err != nil {
-				return
-			}
-			// Successfully validated the thing. Get rid of the Relay header.
-			signedPayload = signedPayload[header.Len:]
-			// Pull the Roaming parts up here, and return in all call paths.
-			f.handleHostRoaming(hostinfo, via)
-			// Track usage of both the HostInfo and the Relay for the received & authenticated packet
-			f.connectionManager.In(hostinfo)
-			f.connectionManager.RelayUsed(h.RemoteIndex)
-
-			relay, ok := hostinfo.relayState.QueryRelayForByIdx(h.RemoteIndex)
-			if !ok {
-				// The only way this happens is if hostmap has an index to the correct HostInfo, but the HostInfo is missing
-				// its internal mapping. This should never happen.
-				hostinfo.logger(f.l).Error("HostInfo missing remote relay index",
-					"vpnAddrs", hostinfo.vpnAddrs,
-					"remoteIndex", h.RemoteIndex,
-				)
-				return
-			}
-
-			switch relay.Type {
-			case TerminalType:
-				// If I am the target of this relay, process the unwrapped packet
-				// From this recursive point, all these variables are 'burned'. We shouldn't rely on them again.
-				via = ViaSender{
-					UdpAddr:   via.UdpAddr,
-					relayHI:   hostinfo,
-					remoteIdx: relay.RemoteIndex,
-					relay:     relay,
-					IsRelayed: true,
-				}
-				buf.Reset()
-				f.readOutsidePackets(via, buf, signedPayload, lhf, q, localCache)
-				return
-			case ForwardingType:
-				// Find the target HostInfo relay object
-				targetHI, targetRelay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relay.PeerAddr)
-				if err != nil {
-					hostinfo.logger(f.l).Info("Failed to find target host info by ip",
-						"relayTo", relay.PeerAddr,
-						"error", err,
-						"hostinfo.vpnAddrs", hostinfo.vpnAddrs,
-					)
-					return
-				}
-
-				// If that relay is Established, forward the payload through it
-				if targetRelay.State == Established {
-					switch targetRelay.Type {
-					case ForwardingType:
-						// Forward this packet through the relay tunnel
-						// Find the target HostInfo
-						f.SendVia(targetHI, targetRelay, signedPayload, buf)
-						return
-					case TerminalType:
-						hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
-					}
-				} else {
-					hostinfo.logger(f.l).Info("Unexpected target relay state",
-						"relayTo", relay.PeerAddr,
-						"relayFrom", hostinfo.vpnAddrs[0],
-						"targetRelayState", targetRelay.State,
-					)
-					return
-				}
-			}
+			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache)
+		default:
+			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected message subtype seen", "from", via, "header", h)
+			return
 		}

 	case header.LightHouse:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
-		if !f.handleEncrypted(ci, via, h) {
-			return
-		}
-
-		d, err := f.decrypt(hostinfo, h.MessageCounter, buf, packet, h)
-		if err != nil {
-			hostinfo.logger(f.l).Error("Failed to decrypt lighthouse packet",
-				"error", err,
-				"from", via,
-				"packet", packet,
-			)
-			return
-		}
-
 		//TODO: assert via is not relayed
-		lhf.HandleRequest(via.UdpAddr, hostinfo.vpnAddrs, d, f)
-
-		// Fallthrough to the bottom to record incoming traffic
+		lhf.HandleRequest(via.UdpAddr, hostinfo.vpnAddrs, out, f)

 	case header.Test:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
-		if !f.handleEncrypted(ci, via, h) {
+		switch h.Subtype {
+		case header.TestReply:
+			// No-op, useful for the Roaming and connectionManager side-effects above
+		case header.TestRequest:
+			f.send(header.Test, header.TestReply, ci, hostinfo, out, nb, out)
+		default:
+			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected test subtype seen", "from", via, "header", h)
 			return
 		}

-		d, err := f.decrypt(hostinfo, h.MessageCounter, buf, packet, h)
-		if err != nil {
-			hostinfo.logger(f.l).Error("Failed to decrypt test packet",
-				"error", err,
-				"from", via,
-				"packet", packet,
-			)
-			return
-		}
-
-		if h.Subtype == header.TestRequest {
-			// This testRequest might be from TryPromoteBest, so we should roam
-			// to the new IP address before responding.
-			f.handleHostRoaming(hostinfo, via)
-			f.send(header.Test, header.TestReply, ci, hostinfo, d, buf)
-		}
-
-		// Fallthrough to the bottom to record incoming traffic
-
-		// Non encrypted messages below here, they should not fall through to avoid tracking incoming traffic since they
-		// are unauthenticated
-
-	case header.Handshake:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
-		f.handshakeManager.HandleIncoming(via, packet, h)
-		return
-
-	case header.RecvError:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
-		f.handleRecvError(via.UdpAddr, h)
-		return
-
 	case header.CloseTunnel:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
-		if !f.handleEncrypted(ci, via, h) {
-			return
-		}
-		_, err = f.decrypt(hostinfo, h.MessageCounter, buf, packet, h)
-		if err != nil {
-			hostinfo.logger(f.l).Error("Failed to decrypt CloseTunnel packet",
-				"error", err,
-				"from", via,
-				"packet", packet,
-			)
-			return
-		}
-
 		hostinfo.logger(f.l).Info("Close tunnel received, tearing down.", "from", via)
-
 		f.closeTunnel(hostinfo)
-		return

 	case header.Control:
-		if !f.handleEncrypted(ci, via, h) {
-			return
-		}
-
-		d, err := f.decrypt(hostinfo, h.MessageCounter, buf, packet, h)
-		if err != nil {
-			hostinfo.logger(f.l).Error("Failed to decrypt Control packet",
-				"error", err,
-				"from", via,
-				"packet", packet,
-			)
-			return
-		}
-
-		f.relayManager.HandleControlMsg(hostinfo, d, f)
+		f.relayManager.HandleControlMsg(hostinfo, out, f)

 	default:
-		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
+		hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected message type seen", "from", via, "header", h)
+	}
+}
+
+func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
+	// The entire body is sent as AD, not encrypted.
+	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
+	// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
+	// otherwise malformed (meaning, there is no trailing 16 byte AEAD value), then this will result in at worst a 0-length slice
+	// which will gracefully fail in the DecryptDanger call.
+	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
+	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
+	var err error
+	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, signedPayload, signatureValue, h.MessageCounter, nb)
+	if err != nil {
+		return
+	}
+	// Advance the replay window now that the frame is authenticated
+	if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Unexpected packet received", "from", via)
+			hostinfo.logger(f.l).Debug("dropping out of window relay packet", "header", h)
 		}
 		return
 	}
-
+	// Successfully validated the thing. Get rid of the Relay header.
+	signedPayload = signedPayload[header.Len:]
+	// Pull the Roaming parts up here, and return in all call paths.
 	f.handleHostRoaming(hostinfo, via)
-
+	// Track usage of both the HostInfo and the Relay for the received & authenticated packet
 	f.connectionManager.In(hostinfo)
+	f.connectionManager.RelayUsed(h.RemoteIndex)
+
+	relay, ok := hostinfo.relayState.QueryRelayForByIdx(h.RemoteIndex)
+	if !ok {
+		// The only way this happens is if hostmap has an index to the correct HostInfo, but the HostInfo is missing
+		// its internal mapping. This should never happen.
+		hostinfo.logger(f.l).Error("HostInfo missing remote relay index",
+			"relayRemoteIndex", h.RemoteIndex,
+		)
+		return
+	}
+
+	switch relay.Type {
+	case TerminalType:
+		// If I am the target of this relay, process the unwrapped packet
+		// From this recursive point, all these variables are 'burned'. We shouldn't rely on them again.
+		via = ViaSender{
+			UdpAddr:   via.UdpAddr,
+			relayHI:   hostinfo,
+			remoteIdx: relay.RemoteIndex,
+			relay:     relay,
+			IsRelayed: true,
+		}
+		f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
+	case ForwardingType:
+		// Find the target HostInfo relay object
+		targetHI, targetRelay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relay.PeerAddr)
+		if err != nil {
+			hostinfo.logger(f.l).Info("Failed to find target host info by ip",
+				"relayTo", relay.PeerAddr,
+				"relayFrom", hostinfo.vpnAddrs[0],
+				"error", err,
+			)
+			return
+		}
+
+		// If that relay is Established, forward the payload through it
+		if targetRelay.State == Established {
+			switch targetRelay.Type {
+			case ForwardingType:
+				// Forward this packet through the relay tunnel
+				// Find the target HostInfo
+				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
+			case TerminalType:
+				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
+				return
+			default:
+				if f.l.Enabled(context.Background(), slog.LevelDebug) {
+					hostinfo.logger(f.l).Debug("Unexpected targetRelay Type", "from", via, "relayType", targetRelay.Type)
+				}
+				return
+			}
+		} else {
+			hostinfo.logger(f.l).Info("Unexpected target relay state",
+				"relayTo", relay.PeerAddr,
+				"relayFrom", hostinfo.vpnAddrs[0],
+				"targetRelayState", targetRelay.State,
+			)
+			return
+		}
+	default:
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			hostinfo.logger(f.l).Debug("Unexpected relay type", "from", via, "relayType", relay.Type)
+		}
+	}
 }

 // closeTunnel closes a tunnel locally, it does not send a closeTunnel packet to the remote
@@ -269,9 +272,7 @@ func (f *Interface) closeTunnel(hostInfo *HostInfo) {

 // sendCloseTunnel is a helper function to send a proper close tunnel packet to a remote
 func (f *Interface) sendCloseTunnel(h *HostInfo) {
-	buf := f.bufAlloc.Acquire()
-	defer f.bufAlloc.Release(buf)
-	f.send(header.CloseTunnel, 0, h.ConnectionState, h, []byte{}, buf)
+	f.send(header.CloseTunnel, 0, h.ConnectionState, h, []byte{}, make([]byte, 12, 12), make([]byte, mtu))
 }

 func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
@@ -305,23 +306,6 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {

 }

-// handleEncrypted returns true if a packet should be processed, false otherwise
-func (f *Interface) handleEncrypted(ci *ConnectionState, via ViaSender, h *header.H) bool {
-	// If connectionstate does not exist, send a recv error, if possible, to encourage a fast reconnect
-	if ci == nil {
-		if !via.IsRelayed {
-			f.maybeSendRecvError(via.UdpAddr, h.RemoteIndex)
-		}
-		return false
-	}
-	// If the window check fails, refuse to process the packet, but don't send a recv error
-	if !ci.window.Check(f.l, h.MessageCounter) {
-		return false
-	}
-
-	return true
-}
-
 var (
 	ErrPacketTooShort          = errors.New("packet is too short")
 	ErrUnknownIPVersion        = errors.New("packet is an unknown ip version")
@@ -520,65 +504,48 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 	return nil
 }

-func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, buf *WireBuffer, packet []byte, h *header.H) ([]byte, error) {
-	plaintext, err := buf.DecryptForHandler(hostinfo.ConnectionState, packet, mc)
+func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []byte, h *header.H, nb []byte) ([]byte, error) {
+	var err error
+	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], mc, nb)
 	if err != nil {
 		return nil, err
 	}

 	if !hostinfo.ConnectionState.window.Update(f.l, mc) {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("dropping out of window packet", "header", h)
-		}
-		return nil, errors.New("out of window packet")
+		return nil, ErrOutOfWindow
 	}

-	return plaintext, nil
+	return out, nil
 }

-func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, buf *WireBuffer, packet []byte, q int, localCache firewall.ConntrackCache) bool {
-	if err := buf.DecryptDatagram(hostinfo.ConnectionState, packet, messageCounter); err != nil {
-		hostinfo.logger(f.l).Error("Failed to decrypt packet", "error", err)
-		return false
-	}
-
-	ipPacket := buf.IPPacket()
-	if err := newPacket(ipPacket, true, buf.FwPacket); err != nil {
+func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) {
+	err := newPacket(out, true, fwPacket)
+	if err != nil {
 		hostinfo.logger(f.l).Warn("Error while validating inbound packet",
 			"error", err,
-			"packet", ipPacket,
+			"packet", out,
 		)
-		return false
+		return
 	}

-	if !hostinfo.ConnectionState.window.Update(f.l, messageCounter) {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("dropping out of window packet", "fwPacket", buf.FwPacket)
-		}
-		return false
-	}
-
-	dropReason := f.firewall.Drop(*buf.FwPacket, true, hostinfo, f.pki.GetCAPool(), localCache)
+	dropReason := f.firewall.Drop(*fwPacket, true, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason != nil {
-		// NOTE: We hand `packet` (the original UDP ciphertext we already
-		// decrypted from) as the reject-IP scratch since we no longer
-		// need its ciphertext, and it's disjoint from buf.Out where
-		// sendNoMetrics will encrypt the wire packet.
-		f.rejectOutside(ipPacket, hostinfo.ConnectionState, hostinfo, packet, buf, q)
+		// NOTE: We give `packet` as the `out` here since we already decrypted from it and we don't need it anymore
+		// This gives us a buffer to build the reject packet in
+		f.rejectOutside(out, hostinfo.ConnectionState, hostinfo, nb, packet, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping inbound packet",
-				"fwPacket", buf.FwPacket,
+				"fwPacket", fwPacket,
 				"reason", dropReason,
 			)
 		}
-		return false
+		return
 	}

-	f.connectionManager.In(hostinfo)
-	if _, err := buf.WriteIPToTUN(f.readers[q]); err != nil {
+	_, err = f.readers[q].Write(out)
+	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}
-	return true
 }

 func (f *Interface) maybeSendRecvError(endpoint netip.AddrPort, index uint32) {
@@ -15,7 +15,4 @@ type Device interface {
 	RoutesFor(netip.Addr) routing.Gateways
 	SupportsMultiqueue() bool
 	NewMultiQueueReader() (io.ReadWriteCloser, error)
-	// TunPrefixLen reports the number of bytes the device prepends to every IP packet on the wire.
-	// Currently only non zero for the BSD tun devices.
-	TunPrefixLen() int
 }
@@ -0,0 +1,358 @@
+//go:build !e2e_testing
+// +build !e2e_testing
+
+package overlay
+
+import (
+	"errors"
+	"fmt"
+	"log/slog"
+	"runtime"
+	"strings"
+	"syscall"
+	"time"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// networkCategory mirrors NLM_NETWORK_CATEGORY from netlistmgr.h.
+type networkCategory int32
+
+const (
+	networkCategoryPublic              networkCategory = 0
+	networkCategoryPrivate             networkCategory = 1
+	networkCategoryDomainAuthenticated networkCategory = 2
+)
+
+func (c networkCategory) String() string {
+	switch c {
+	case networkCategoryPublic:
+		return "public"
+	case networkCategoryPrivate:
+		return "private"
+	case networkCategoryDomainAuthenticated:
+		return "domain"
+	}
+	return fmt.Sprintf("unknown(%d)", c)
+}
+
+// parseNetworkCategory accepts the user-supplied tun.network_category. A
+// second return of false means "leave the category alone".
+func parseNetworkCategory(s string) (networkCategory, bool, error) {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "", "unset":
+		return 0, false, nil
+	case "public":
+		return networkCategoryPublic, true, nil
+	case "private":
+		return networkCategoryPrivate, true, nil
+	case "domain", "domainauthenticated":
+		return networkCategoryDomainAuthenticated, true, nil
+	}
+	return 0, false, fmt.Errorf("unknown tun.network_category %q (expected public, private, domain, or unset)", s)
+}
+
+// CLSID_NetworkListManager {DCB00C01-570F-4A9B-8D69-199FDBA5723B}
+var clsidNetworkListManager = windows.GUID{
+	Data1: 0xDCB00C01, Data2: 0x570F, Data3: 0x4A9B,
+	Data4: [8]byte{0x8D, 0x69, 0x19, 0x9F, 0xDB, 0xA5, 0x72, 0x3B},
+}
+
+// IID_INetworkListManager {DCB00000-570F-4A9B-8D69-199FDBA5723B}
+var iidINetworkListManager = windows.GUID{
+	Data1: 0xDCB00000, Data2: 0x570F, Data3: 0x4A9B,
+	Data4: [8]byte{0x8D, 0x69, 0x19, 0x9F, 0xDB, 0xA5, 0x72, 0x3B},
+}
+
+// x/sys/windows doesn't expose CoCreateInstance, so we bind it ourselves.
+var procCoCreateInstance = windows.NewLazySystemDLL("ole32.dll").NewProc("CoCreateInstance")
+
+const clsCtxAll = windows.CLSCTX_INPROC_SERVER | windows.CLSCTX_INPROC_HANDLER |
+	windows.CLSCTX_LOCAL_SERVER | windows.CLSCTX_REMOTE_SERVER
+
+const (
+	hrSFALSE          = 0x00000001
+	hrRPCEChangedMode = 0x80010106
+)
+
+type hresult uint32
+
+func (h hresult) failed() bool { return int32(h) < 0 }
+func (h hresult) String() string {
+	return fmt.Sprintf("HRESULT 0x%08x", uint32(h))
+}
+
+var errAdapterNotFound = errors.New("adapter not present in network connections enumeration")
+
+// Vtable layouts. Slot order must match the declaration order in netlistmgr.h.
+// All NLM interfaces here derive from IDispatch, which derives from IUnknown.
+
+type iUnknownVtbl struct {
+	QueryInterface uintptr
+	AddRef         uintptr
+	Release        uintptr
+}
+
+type iDispatchVtbl struct {
+	iUnknownVtbl
+	GetTypeInfoCount uintptr
+	GetTypeInfo      uintptr
+	GetIDsOfNames    uintptr
+	Invoke           uintptr
+}
+
+type iNetworkListManagerVtbl struct {
+	iDispatchVtbl
+	GetNetworks           uintptr
+	GetNetwork            uintptr
+	GetNetworkConnections uintptr
+	GetNetworkConnection  uintptr
+	IsConnectedToInternet uintptr
+	IsConnected           uintptr
+	GetConnectivity       uintptr
+}
+
+type iNetworkListManager struct{ Vtbl *iNetworkListManagerVtbl }
+
+func (n *iNetworkListManager) Release() {
+	syscall.SyscallN(n.Vtbl.Release, uintptr(unsafe.Pointer(n)))
+}
+
+func (n *iNetworkListManager) GetNetworkConnections() (*iEnumNetworkConnections, error) {
+	var enum *iEnumNetworkConnections
+	r1, _, _ := syscall.SyscallN(n.Vtbl.GetNetworkConnections,
+		uintptr(unsafe.Pointer(n)), uintptr(unsafe.Pointer(&enum)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return nil, fmt.Errorf("INetworkListManager.GetNetworkConnections: %s", hr)
+	}
+	return enum, nil
+}
+
+type iEnumNetworkConnectionsVtbl struct {
+	iDispatchVtbl
+	NewEnum uintptr
+	Next    uintptr
+	Skip    uintptr
+	Reset   uintptr
+	Clone   uintptr
+}
+
+type iEnumNetworkConnections struct{ Vtbl *iEnumNetworkConnectionsVtbl }
+
+func (e *iEnumNetworkConnections) Release() {
+	syscall.SyscallN(e.Vtbl.Release, uintptr(unsafe.Pointer(e)))
+}
+
+// Next returns the next connection, or (nil, nil) at the end of the enumeration.
+func (e *iEnumNetworkConnections) Next() (*iNetworkConnection, error) {
+	var conn *iNetworkConnection
+	var fetched uint32
+	r1, _, _ := syscall.SyscallN(e.Vtbl.Next,
+		uintptr(unsafe.Pointer(e)), 1,
+		uintptr(unsafe.Pointer(&conn)), uintptr(unsafe.Pointer(&fetched)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return nil, fmt.Errorf("IEnumNetworkConnections.Next: %s", hr)
+	}
+	if fetched == 0 {
+		return nil, nil
+	}
+	return conn, nil
+}
+
+type iNetworkConnectionVtbl struct {
+	iDispatchVtbl
+	GetNetwork            uintptr
+	IsConnectedToInternet uintptr
+	IsConnected           uintptr
+	GetConnectivity       uintptr
+	GetConnectionId       uintptr
+	GetAdapterId          uintptr
+	GetDomainType         uintptr
+}
+
+type iNetworkConnection struct{ Vtbl *iNetworkConnectionVtbl }
+
+func (c *iNetworkConnection) Release() {
+	syscall.SyscallN(c.Vtbl.Release, uintptr(unsafe.Pointer(c)))
+}
+
+func (c *iNetworkConnection) GetAdapterId() (windows.GUID, error) {
+	var g windows.GUID
+	r1, _, _ := syscall.SyscallN(c.Vtbl.GetAdapterId,
+		uintptr(unsafe.Pointer(c)), uintptr(unsafe.Pointer(&g)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return windows.GUID{}, fmt.Errorf("INetworkConnection.GetAdapterId: %s", hr)
+	}
+	return g, nil
+}
+
+func (c *iNetworkConnection) GetNetwork() (*iNetwork, error) {
+	var net *iNetwork
+	r1, _, _ := syscall.SyscallN(c.Vtbl.GetNetwork,
+		uintptr(unsafe.Pointer(c)), uintptr(unsafe.Pointer(&net)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return nil, fmt.Errorf("INetworkConnection.GetNetwork: %s", hr)
+	}
+	return net, nil
+}
+
+type iNetworkVtbl struct {
+	iDispatchVtbl
+	GetName                    uintptr
+	SetName                    uintptr
+	GetDescription             uintptr
+	SetDescription             uintptr
+	GetNetworkId               uintptr
+	GetDomainType              uintptr
+	GetNetworkConnections      uintptr
+	GetTimeCreatedAndConnected uintptr
+	IsConnectedToInternet      uintptr
+	IsConnected                uintptr
+	GetConnectivity            uintptr
+	GetCategory                uintptr
+	SetCategory                uintptr
+}
+
+type iNetwork struct{ Vtbl *iNetworkVtbl }
+
+func (n *iNetwork) Release() {
+	syscall.SyscallN(n.Vtbl.Release, uintptr(unsafe.Pointer(n)))
+}
+
+func (n *iNetwork) GetCategory() (networkCategory, error) {
+	var c networkCategory
+	r1, _, _ := syscall.SyscallN(n.Vtbl.GetCategory,
+		uintptr(unsafe.Pointer(n)), uintptr(unsafe.Pointer(&c)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return 0, fmt.Errorf("INetwork.GetCategory: %s", hr)
+	}
+	return c, nil
+}
+
+func (n *iNetwork) SetCategory(c networkCategory) error {
+	r1, _, _ := syscall.SyscallN(n.Vtbl.SetCategory,
+		uintptr(unsafe.Pointer(n)), uintptr(int32(c)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return fmt.Errorf("INetwork.SetCategory: %s", hr)
+	}
+	return nil
+}
+
+// coInit initializes COM for the current OS thread. The returned function must
+// be deferred to balance a successful init. RPC_E_CHANGED_MODE means COM is
+// already initialized in a different mode on this thread, which is still fine
+// for our calls but we must not Uninitialize in that case.
+func coInit() (func(), error) {
+	err := windows.CoInitializeEx(0, windows.COINIT_MULTITHREADED)
+	if err == nil {
+		return windows.CoUninitialize, nil
+	}
+	if e, ok := err.(syscall.Errno); ok {
+		switch uint32(e) {
+		case hrSFALSE:
+			return windows.CoUninitialize, nil
+		case hrRPCEChangedMode:
+			return func() {}, nil
+		}
+	}
+	return nil, fmt.Errorf("CoInitializeEx: %w", err)
+}
+
+func createNetworkListManager() (*iNetworkListManager, error) {
+	var nlm *iNetworkListManager
+	r1, _, _ := procCoCreateInstance.Call(
+		uintptr(unsafe.Pointer(&clsidNetworkListManager)),
+		0,
+		uintptr(clsCtxAll),
+		uintptr(unsafe.Pointer(&iidINetworkListManager)),
+		uintptr(unsafe.Pointer(&nlm)),
+	)
+	if hr := hresult(r1); hr.failed() {
+		return nil, fmt.Errorf("CoCreateInstance(NetworkListManager): %s", hr)
+	}
+	return nlm, nil
+}
+
+// setNetworkCategory locates the network connection bound to adapterGUID and
+// sets the category of its parent network. Returns errAdapterNotFound if the
+// adapter is not yet visible in the NLM enumeration.
+func setNetworkCategory(adapterGUID windows.GUID, cat networkCategory) error {
+	deinit, err := coInit()
+	if err != nil {
+		return err
+	}
+	defer deinit()
+
+	nlm, err := createNetworkListManager()
+	if err != nil {
+		return err
+	}
+	defer nlm.Release()
+
+	enum, err := nlm.GetNetworkConnections()
+	if err != nil {
+		return err
+	}
+	defer enum.Release()
+
+	for {
+		conn, err := enum.Next()
+		if err != nil {
+			return err
+		}
+		if conn == nil {
+			return errAdapterNotFound
+		}
+
+		guid, err := conn.GetAdapterId()
+		if err != nil || guid != adapterGUID {
+			conn.Release()
+			continue
+		}
+
+		net, err := conn.GetNetwork()
+		conn.Release()
+		if err != nil {
+			return err
+		}
+		err = net.SetCategory(cat)
+		net.Release()
+		return err
+	}
+}
+
+// applyNetworkCategory polls until the wintun adapter shows up in the NLM
+// enumeration, then sets the category. Intended to run in its own goroutine.
+func applyNetworkCategory(l *slog.Logger, adapterGUID windows.GUID, cat networkCategory) {
+	// COM Init/Uninit must be paired on the same OS thread.
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	const (
+		attempts = 30
+		interval = 500 * time.Millisecond
+	)
+	for i := 0; i < attempts; i++ {
+		err := setNetworkCategory(adapterGUID, cat)
+		if err == nil {
+			l.Info("Set Windows network category", "category", cat.String())
+			return
+		}
+		if !errors.Is(err, errAdapterNotFound) {
+			l.Warn("Failed to set Windows network category", "error", err, "category", cat.String())
+			return
+		}
+		time.Sleep(interval)
+	}
+	l.Warn("Gave up waiting for adapter to appear in NLM enumeration; network category not set",
+		"category", cat.String(),
+		"waited", time.Duration(attempts)*interval,
+	)
+}
@@ -0,0 +1,109 @@
+//go:build !e2e_testing
+// +build !e2e_testing
+
+package overlay
+
+import (
+	"testing"
+)
+
+func Test_parseNetworkCategory(t *testing.T) {
+	cases := []struct {
+		in        string
+		wantCat   networkCategory
+		wantApply bool
+		wantErr   bool
+	}{
+		{"", 0, false, false},
+		{"unset", 0, false, false},
+		{"  UNSET  ", 0, false, false},
+		{"private", networkCategoryPrivate, true, false},
+		{"Private", networkCategoryPrivate, true, false},
+		{"  PRIVATE  ", networkCategoryPrivate, true, false},
+		{"public", networkCategoryPublic, true, false},
+		{"PUBLIC", networkCategoryPublic, true, false},
+		{"domain", networkCategoryDomainAuthenticated, true, false},
+		{"DomainAuthenticated", networkCategoryDomainAuthenticated, true, false},
+		{"garbage", 0, false, true},
+		{"privates", 0, false, true},
+	}
+	for _, tc := range cases {
+		cat, apply, err := parseNetworkCategory(tc.in)
+		if (err != nil) != tc.wantErr {
+			t.Errorf("parseNetworkCategory(%q) err=%v, wantErr=%v", tc.in, err, tc.wantErr)
+			continue
+		}
+		if cat != tc.wantCat || apply != tc.wantApply {
+			t.Errorf("parseNetworkCategory(%q) = (%v, %v), want (%v, %v)", tc.in, cat, apply, tc.wantCat, tc.wantApply)
+		}
+	}
+}
+
+// Test_NLM_round_trip exercises every COM call path used by setNetworkCategory
+// without mutating the host's network state. It validates the CLSID/IID
+// constants and every vtable index by enumerating connections, fetching the
+// adapter id and parent network, reading the current category, and writing it
+// back unchanged.
+//
+// Requires Windows but does not require admin or the wintun driver. Skips if
+// no network connections are available (unlikely outside of an isolated
+// container).
+func Test_NLM_round_trip(t *testing.T) {
+	deinit, err := coInit()
+	if err != nil {
+		t.Fatalf("coInit: %v", err)
+	}
+	defer deinit()
+
+	nlm, err := createNetworkListManager()
+	if err != nil {
+		t.Fatalf("createNetworkListManager: %v", err)
+	}
+	defer nlm.Release()
+
+	enum, err := nlm.GetNetworkConnections()
+	if err != nil {
+		t.Fatalf("GetNetworkConnections: %v", err)
+	}
+	defer enum.Release()
+
+	saw := 0
+	for {
+		conn, err := enum.Next()
+		if err != nil {
+			t.Fatalf("EnumNetworkConnections.Next: %v", err)
+		}
+		if conn == nil {
+			break
+		}
+		saw++
+
+		if _, err := conn.GetAdapterId(); err != nil {
+			conn.Release()
+			t.Fatalf("INetworkConnection.GetAdapterId: %v", err)
+		}
+
+		net, err := conn.GetNetwork()
+		conn.Release()
+		if err != nil {
+			t.Fatalf("INetworkConnection.GetNetwork: %v", err)
+		}
+
+		cat, err := net.GetCategory()
+		if err != nil {
+			net.Release()
+			t.Fatalf("INetwork.GetCategory: %v", err)
+		}
+		// Set to the current value so the host's NLM state is unchanged but
+		// SetCategory's vtable slot is still validated end-to-end.
+		if err := net.SetCategory(cat); err != nil {
+			net.Release()
+			t.Fatalf("INetwork.SetCategory(%v): %v", cat, err)
+		}
+		net.Release()
+	}
+
+	if saw == 0 {
+		t.Skip("no NLM network connections available; skipping round-trip")
+	}
+}
@@ -50,5 +50,3 @@ func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 func (NoopTun) Close() error {
 	return nil
 }
-
-func (NoopTun) TunPrefixLen() int { return 0 }
@@ -102,5 +102,3 @@ func (t *tun) SupportsMultiqueue() bool {
 func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
 }
-
-func (t *tun) TunPrefixLen() int { return 0 }
@@ -1,29 +0,0 @@
-//go:build (darwin || ios || freebsd || openbsd || netbsd) && !e2e_testing
-
-package overlay
-
-import (
-	"fmt"
-	"syscall"
-)
-
-// StampTunPrefix writes the 4-byte AF_INET / AF_INET6 protocol-family marker into buf[0:4] in place,
-// picking the family from the first byte of the IP packet at buf[4].
-func StampTunPrefix(buf []byte) error {
-	if len(buf) < 5 {
-		return fmt.Errorf("tun write buffer too small for prefix")
-	}
-	ipVer := buf[4] >> 4
-	buf[0] = 0
-	buf[1] = 0
-	buf[2] = 0
-	switch ipVer {
-	case 4:
-		buf[3] = syscall.AF_INET
-	case 6:
-		buf[3] = syscall.AF_INET6
-	default:
-		return fmt.Errorf("unable to determine IP version from packet")
-	}
-	return nil
-}
@@ -0,0 +1,23 @@
+//go:build (amd64 || arm64) && !e2e_testing
+// +build amd64 arm64
+// +build !e2e_testing
+
+package overlay
+
+import (
+	"log/slog"
+
+	"github.com/slackhq/nebula/wfp"
+)
+
+// installInterfaceBypass installs a WFP PERMIT filter scoped to the wintun interface LUID so inbound traffic on the
+// nebula adapter bypasses Windows Defender Firewall.
+func installInterfaceBypass(l *slog.Logger, luid uint64) closer {
+	s, err := wfp.PermitInterface(luid)
+	if err != nil {
+		l.Warn("Failed to install WFP bypass filters on nebula interface", "error", err)
+		return nil
+	}
+	l.Info("Installed WFP filters bypassing Windows Defender Firewall on nebula interface")
+	return s
+}
@@ -0,0 +1,11 @@
+//go:build !e2e_testing
+// +build !e2e_testing
+
+package overlay
+
+import "log/slog"
+
+// installInterfaceBypass is a no-op on windows-386 because we don't currently build for it.
+func installInterfaceBypass(_ *slog.Logger, _ uint64) closer {
+	return nil
+}
@@ -11,6 +11,7 @@ import (
 	"net/netip"
 	"os"
 	"sync/atomic"
+	"syscall"
 	"unsafe"

 	"github.com/gaissmai/bart"
@@ -30,6 +31,9 @@ type tun struct {
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	linkAddr    *netroute.LinkAddr
 	l           *slog.Logger
+
+	// cache out buffer since we need to prepend 4 bytes for tun metadata
+	out []byte
 }

 type ifReq struct {
@@ -498,6 +502,44 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }

+func (t *tun) Read(to []byte) (int, error) {
+	buf := make([]byte, len(to)+4)
+
+	n, err := t.ReadWriteCloser.Read(buf)
+
+	copy(to, buf[4:])
+	return n - 4, err
+}
+
+// Write is only valid for single threaded use
+func (t *tun) Write(from []byte) (int, error) {
+	buf := t.out
+	if cap(buf) < len(from)+4 {
+		buf = make([]byte, len(from)+4)
+		t.out = buf
+	}
+	buf = buf[:len(from)+4]
+
+	if len(from) == 0 {
+		return 0, syscall.EIO
+	}
+
+	// Determine the IP Family for the NULL L2 Header
+	ipVer := from[0] >> 4
+	if ipVer == 4 {
+		buf[3] = syscall.AF_INET
+	} else if ipVer == 6 {
+		buf[3] = syscall.AF_INET6
+	} else {
+		return 0, fmt.Errorf("unable to determine IP version from packet")
+	}
+
+	copy(buf[4:], from)
+
+	n, err := t.ReadWriteCloser.Write(buf)
+	return n - 4, err
+}
+
 func (t *tun) Networks() []netip.Prefix {
 	return t.vpnNetworks
 }
@@ -513,7 +555,3 @@ func (t *tun) SupportsMultiqueue() bool {
 func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
 }
-
-// TunPrefixLen reports the 4-byte BSD AF_INET / AF_INET6 protocol-family
-// marker the kernel prepends on read and expects on write.
-func (t *tun) TunPrefixLen() int { return 4 }
@@ -136,5 +136,3 @@ func (p prettyPacket) String() string {

 	return s.String()
 }
-
-func (t *disabledTun) TunPrefixLen() int { return 0 }
@@ -158,43 +158,74 @@ func (t *tun) blockOnWrite() error {
 }

 func (t *tun) Read(to []byte) (int, error) {
+	// first 4 bytes is protocol family, in network byte order
+	var head [4]byte
+	iovecs := [2]syscall.Iovec{
+		{&head[0], 4},
+		{&to[0], uint64(len(to))},
+	}
 	for {
-		n, err := unix.Read(t.fd, to)
-		if err == nil {
-			return n, nil
+		n, _, errno := syscall.Syscall(syscall.SYS_READV, uintptr(t.fd), uintptr(unsafe.Pointer(&iovecs[0])), 2)
+		if errno == 0 {
+			bytesRead := int(n)
+			if bytesRead < 4 {
+				return 0, nil
+			}
+			return bytesRead - 4, nil
 		}
-		switch err {
+		switch errno {
 		case unix.EAGAIN:
-			if berr := t.blockOnRead(); berr != nil {
-				return 0, berr
+			if err := t.blockOnRead(); err != nil {
+				return 0, err
 			}
 		case unix.EINTR:
 			// retry
 		case unix.EBADF:
 			return 0, os.ErrClosed
 		default:
-			return 0, err
+			return 0, errno
 		}
 	}
 }

+// Write is only valid for single threaded use
 func (t *tun) Write(from []byte) (int, error) {
+	if len(from) <= 1 {
+		return 0, syscall.EIO
+	}
+
+	ipVer := from[0] >> 4
+	var head [4]byte
+	// first 4 bytes is protocol family, in network byte order
+	switch ipVer {
+	case 4:
+		head[3] = syscall.AF_INET
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
+		return 0, fmt.Errorf("unable to determine IP version from packet")
+	}
+
+	iovecs := [2]syscall.Iovec{
+		{&head[0], 4},
+		{&from[0], uint64(len(from))},
+	}
 	for {
-		n, err := unix.Write(t.fd, from)
-		if err == nil {
-			return n, nil
+		n, _, errno := syscall.Syscall(syscall.SYS_WRITEV, uintptr(t.fd), uintptr(unsafe.Pointer(&iovecs[0])), 2)
+		if errno == 0 {
+			return int(n) - 4, nil
 		}
-		switch err {
+		switch errno {
 		case unix.EAGAIN:
-			if berr := t.blockOnWrite(); berr != nil {
-				return 0, berr
+			if err := t.blockOnWrite(); err != nil {
+				return 0, err
 			}
 		case unix.EINTR:
 			// retry
 		case unix.EBADF:
 			return 0, os.ErrClosed
 		default:
-			return 0, err
+			return 0, errno
 		}
 	}
 }
@@ -701,7 +732,3 @@ func getLinkAddr(name string) (*netroute.LinkAddr, error) {

 	return nil, nil
 }
-
-// TunPrefixLen reports the 4-byte BSD AF_INET / AF_INET6 protocol-family
-// marker the kernel prepends on read and expects on write.
-func (t *tun) TunPrefixLen() int { return 4 }
@@ -4,12 +4,15 @@
 package overlay

 import (
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"net/netip"
 	"os"
+	"sync"
 	"sync/atomic"
+	"syscall"

 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
@@ -33,7 +36,7 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
 		vpnNetworks:     vpnNetworks,
-		ReadWriteCloser: file,
+		ReadWriteCloser: &tunReadCloser{f: file},
 		l:               l,
 	}

@@ -82,6 +85,64 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }

+// The following is hoisted up from water, we do this so we can inject our own fd on iOS
+type tunReadCloser struct {
+	f io.ReadWriteCloser
+
+	rMu  sync.Mutex
+	rBuf []byte
+
+	wMu  sync.Mutex
+	wBuf []byte
+}
+
+func (tr *tunReadCloser) Read(to []byte) (int, error) {
+	tr.rMu.Lock()
+	defer tr.rMu.Unlock()
+
+	if cap(tr.rBuf) < len(to)+4 {
+		tr.rBuf = make([]byte, len(to)+4)
+	}
+	tr.rBuf = tr.rBuf[:len(to)+4]
+
+	n, err := tr.f.Read(tr.rBuf)
+	copy(to, tr.rBuf[4:])
+	return n - 4, err
+}
+
+func (tr *tunReadCloser) Write(from []byte) (int, error) {
+	if len(from) == 0 {
+		return 0, syscall.EIO
+	}
+
+	tr.wMu.Lock()
+	defer tr.wMu.Unlock()
+
+	if cap(tr.wBuf) < len(from)+4 {
+		tr.wBuf = make([]byte, len(from)+4)
+	}
+	tr.wBuf = tr.wBuf[:len(from)+4]
+
+	// Determine the IP Family for the NULL L2 Header
+	ipVer := from[0] >> 4
+	if ipVer == 4 {
+		tr.wBuf[3] = syscall.AF_INET
+	} else if ipVer == 6 {
+		tr.wBuf[3] = syscall.AF_INET6
+	} else {
+		return 0, errors.New("unable to determine IP version from packet")
+	}
+
+	copy(tr.wBuf[4:], from)
+
+	n, err := tr.f.Write(tr.wBuf)
+	return n - 4, err
+}
+
+func (tr *tunReadCloser) Close() error {
+	return tr.f.Close()
+}
+
 func (t *tun) Networks() []netip.Prefix {
 	return t.vpnNetworks
 }
@@ -97,7 +158,3 @@ func (t *tun) SupportsMultiqueue() bool {
 func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
 }
-
-// TunPrefixLen reports the 4-byte BSD AF_INET / AF_INET6 protocol-family
-// marker the kernel prepends on read and expects on write.
-func (t *tun) TunPrefixLen() int { return 4 }
@@ -907,5 +907,3 @@ func (t *tun) Close() error {
 	}
 	return err
 }
-
-func (t *tun) TunPrefixLen() int { return 0 }
@@ -58,13 +58,13 @@ type addrLifetime struct {
 }

 type tun struct {
-	io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	MTU         int
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
+	f           *os.File
 	fd          int
 }

@@ -96,12 +96,12 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}

 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
-		fd:              fd,
-		Device:          deviceName,
-		vpnNetworks:     vpnNetworks,
-		MTU:             c.GetInt("tun.mtu", DefaultMTU),
-		l:               l,
+		f:           os.NewFile(uintptr(fd), ""),
+		fd:          fd,
+		Device:      deviceName,
+		vpnNetworks: vpnNetworks,
+		MTU:         c.GetInt("tun.mtu", DefaultMTU),
+		l:           l,
 	}

 	err = t.reload(c, true)
@@ -120,12 +120,12 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 }

 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
-		if err := t.ReadWriteCloser.Close(); err != nil {
+	if t.f != nil {
+		if err := t.f.Close(); err != nil {
 			return fmt.Errorf("error closing tun file: %w", err)
 		}

-		// Close on the os.File should have handled the fd for us but let's be extra sure
+		// t.f.Close should have handled it for us but let's be extra sure
 		_ = unix.Close(t.fd)

 		s, err := syscall.Socket(syscall.AF_INET, syscall.SOCK_DGRAM, syscall.IPPROTO_IP)
@@ -141,6 +141,99 @@ func (t *tun) Close() error {
 	return nil
 }

+func (t *tun) Read(to []byte) (int, error) {
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, fmt.Errorf("failed to get syscall conn for tun: %w", err)
+	}
+
+	var errno syscall.Errno
+	var n uintptr
+	err = rc.Read(func(fd uintptr) bool {
+		// first 4 bytes is protocol family, in network byte order
+		head := [4]byte{}
+		iovecs := []syscall.Iovec{
+			{&head[0], 4},
+			{&to[0], uint64(len(to))},
+		}
+
+		n, _, errno = syscall.Syscall(syscall.SYS_READV, fd, uintptr(unsafe.Pointer(&iovecs[0])), uintptr(2))
+		if errno.Temporary() {
+			// We got an EAGAIN, EINTR, or EWOULDBLOCK, go again
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		if err == syscall.EBADF || err.Error() == "use of closed file" {
+			// Go doesn't export poll.ErrFileClosing but happily reports it to us so here we are
+			// https://github.com/golang/go/blob/master/src/internal/poll/fd_poll_runtime.go#L121
+			return 0, os.ErrClosed
+		}
+		return 0, fmt.Errorf("failed to make read call for tun: %w", err)
+	}
+
+	if errno != 0 {
+		return 0, fmt.Errorf("failed to make inner read call for tun: %w", errno)
+	}
+
+	// fix bytes read number to exclude header
+	bytesRead := int(n)
+	if bytesRead < 0 {
+		return bytesRead, nil
+	} else if bytesRead < 4 {
+		return 0, nil
+	} else {
+		return bytesRead - 4, nil
+	}
+}
+
+// Write is only valid for single threaded use
+func (t *tun) Write(from []byte) (int, error) {
+	if len(from) <= 1 {
+		return 0, syscall.EIO
+	}
+
+	ipVer := from[0] >> 4
+	var head [4]byte
+	// first 4 bytes is protocol family, in network byte order
+	if ipVer == 4 {
+		head[3] = syscall.AF_INET
+	} else if ipVer == 6 {
+		head[3] = syscall.AF_INET6
+	} else {
+		return 0, fmt.Errorf("unable to determine IP version from packet")
+	}
+
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
+
+	var errno syscall.Errno
+	var n uintptr
+	err = rc.Write(func(fd uintptr) bool {
+		iovecs := []syscall.Iovec{
+			{&head[0], 4},
+			{&from[0], uint64(len(from))},
+		}
+
+		n, _, errno = syscall.Syscall(syscall.SYS_WRITEV, fd, uintptr(unsafe.Pointer(&iovecs[0])), uintptr(2))
+		// According to NetBSD documentation for TUN, writes will only return errors in which
+		// this packet will never be delivered so just go on living life.
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+
+	if errno != 0 {
+		return 0, errno
+	}
+
+	return int(n) - 4, err
+}
+
 func (t *tun) addIp(cidr netip.Prefix) error {
 	if cidr.Addr().Is4() {
 		var req ifreqAlias4
@@ -458,7 +551,3 @@ func delRoute(prefix netip.Prefix, gateways []netip.Prefix) error {

 	return nil
 }
-
-// TunPrefixLen reports the 4-byte BSD AF_INET / AF_INET6 protocol-family
-// marker the kernel prepends on read and expects on write.
-func (t *tun) TunPrefixLen() int { return 4 }
@@ -1,10 +0,0 @@
-//go:build (!darwin && !ios && !freebsd && !openbsd && !netbsd) || e2e_testing
-
-package overlay
-
-// StampTunPrefix is a no-op on platforms whose tun devices have no
-// protocol-family marker. WireBuffer only invokes it when its prefixLen
-// is non-zero, so this should never be reached on these platforms.
-func StampTunPrefix(buf []byte) error {
-	return nil
-}
@@ -49,14 +49,16 @@ type ifreq struct {
 }

 type tun struct {
-	io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	MTU         int
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
+	f           *os.File
 	fd          int
+	// cache out buffer since we need to prepend 4 bytes for tun metadata
+	out []byte
 }

 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -87,12 +89,12 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}

 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
-		fd:              fd,
-		Device:          deviceName,
-		vpnNetworks:     vpnNetworks,
-		MTU:             c.GetInt("tun.mtu", DefaultMTU),
-		l:               l,
+		f:           os.NewFile(uintptr(fd), ""),
+		fd:          fd,
+		Device:      deviceName,
+		vpnNetworks: vpnNetworks,
+		MTU:         c.GetInt("tun.mtu", DefaultMTU),
+		l:           l,
 	}

 	err = t.reload(c, true)
@@ -111,17 +113,55 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 }

 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
-		if err := t.ReadWriteCloser.Close(); err != nil {
+	if t.f != nil {
+		if err := t.f.Close(); err != nil {
 			return fmt.Errorf("error closing tun file: %w", err)
 		}

-		// Close on the os.File should have handled the fd for us but let's be extra sure
+		// t.f.Close should have handled it for us but let's be extra sure
 		_ = unix.Close(t.fd)
 	}
 	return nil
 }

+func (t *tun) Read(to []byte) (int, error) {
+	buf := make([]byte, len(to)+4)
+
+	n, err := t.f.Read(buf)
+
+	copy(to, buf[4:])
+	return n - 4, err
+}
+
+// Write is only valid for single threaded use
+func (t *tun) Write(from []byte) (int, error) {
+	buf := t.out
+	if cap(buf) < len(from)+4 {
+		buf = make([]byte, len(from)+4)
+		t.out = buf
+	}
+	buf = buf[:len(from)+4]
+
+	if len(from) == 0 {
+		return 0, syscall.EIO
+	}
+
+	// Determine the IP Family for the NULL L2 Header
+	ipVer := from[0] >> 4
+	if ipVer == 4 {
+		buf[3] = syscall.AF_INET
+	} else if ipVer == 6 {
+		buf[3] = syscall.AF_INET6
+	} else {
+		return 0, fmt.Errorf("unable to determine IP version from packet")
+	}
+
+	copy(buf[4:], from)
+
+	n, err := t.f.Write(buf)
+	return n - 4, err
+}
+
 func (t *tun) addIp(cidr netip.Prefix) error {
 	if cidr.Addr().Is4() {
 		var req ifreqAlias4
@@ -431,7 +471,3 @@ func delRoute(prefix netip.Prefix, gateways []netip.Prefix) error {

 	return nil
 }
-
-// TunPrefixLen reports the 4-byte BSD AF_INET / AF_INET6 protocol-family
-// marker the kernel prepends on read and expects on write.
-func (t *tun) TunPrefixLen() int { return 4 }
@@ -184,5 +184,3 @@ func (t *TestTun) SupportsMultiqueue() bool {
 func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented")
 }
-
-func (t *TestTun) TunPrefixLen() int { return 0 }
@@ -25,15 +25,24 @@ import (
 	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
 )

+type closer interface {
+	Close()
+}
+
 const tunGUIDLabel = "Fixed Nebula Windows GUID v1"

 type winTun struct {
-	Device      string
-	vpnNetworks []netip.Prefix
-	MTU         int
-	Routes      atomic.Pointer[[]Route]
-	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
-	l           *slog.Logger
+	Device          string
+	vpnNetworks     []netip.Prefix
+	MTU             int
+	Routes          atomic.Pointer[[]Route]
+	routeTree       atomic.Pointer[bart.Table[routing.Gateways]]
+	guid            windows.GUID
+	networkCategory networkCategory
+	setCategory     bool
+	bypassWDF       bool
+	wdfBypass       closer
+	l               *slog.Logger

 	tun *wintun.NativeTun
 }
@@ -54,11 +63,20 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*w
 		return nil, fmt.Errorf("generate GUID failed: %w", err)
 	}

+	cat, setCat, err := parseNetworkCategory(c.GetString("tun.network_category", "private"))
+	if err != nil {
+		return nil, err
+	}
+
 	t := &winTun{
-		Device:      deviceName,
-		vpnNetworks: vpnNetworks,
-		MTU:         c.GetInt("tun.mtu", DefaultMTU),
-		l:           l,
+		Device:          deviceName,
+		vpnNetworks:     vpnNetworks,
+		MTU:             c.GetInt("tun.mtu", DefaultMTU),
+		guid:            *guid,
+		networkCategory: cat,
+		setCategory:     setCat,
+		bypassWDF:       c.GetBool("tun.windows_bypass_wdf", true),
+		l:               l,
 	}

 	err = t.reload(c, true)
@@ -142,6 +160,17 @@ func (t *winTun) Activate() error {
 		return err
 	}

+	if t.setCategory {
+		// The wintun adapter takes a moment to register with the Network List
+		// Manager, so we apply the category in the background and retry until
+		// it shows up.
+		go applyNetworkCategory(t.l, t.guid, t.networkCategory)
+	}
+
+	if t.bypassWDF {
+		t.wdfBypass = installInterfaceBypass(t.l, uint64(t.tun.LUID()))
+	}
+
 	return nil
 }

@@ -156,11 +185,8 @@ func (t *winTun) addRoutes(logErrors bool) error {
 			continue
 		}

-		// Add our unsafe route
-		// Windows does not support multipath routes natively, so we install only a single route.
-		// This is not a problem as traffic will always be sent to Nebula which handles the multipath routing internally.
-		// In effect this provides multipath routing support to windows supporting loadbalancing and redundancy.
-		err := luid.AddRoute(r.Cidr, r.Via[0].Addr(), uint32(r.Metric))
+		// Add our unsafe route as an on-link route to the nebula tun device.
+		err := luid.AddRoute(r.Cidr, unspecifiedNextHop(r.Cidr), uint32(r.Metric))
 		if err != nil {
 			retErr := util.NewContextualError("Failed to add route", map[string]any{"route": r}, err)
 			if logErrors {
@@ -206,7 +232,7 @@ func (t *winTun) removeRoutes(routes []Route) error {
 		}

 		// See comment on luid.AddRoute
-		err := luid.DeleteRoute(r.Cidr, r.Via[0].Addr())
+		err := luid.DeleteRoute(r.Cidr, unspecifiedNextHop(r.Cidr))
 		if err != nil {
 			t.l.Error("Failed to remove route", "error", err, "route", r)
 		} else {
@@ -258,9 +284,21 @@ func (t *winTun) Close() error {
 	_ = luid.FlushDNS(windows.AF_INET)
 	_ = luid.FlushDNS(windows.AF_INET6)

+	if t.wdfBypass != nil {
+		t.wdfBypass.Close()
+		t.wdfBypass = nil
+	}
+
 	return t.tun.Close()
 }

+func unspecifiedNextHop(p netip.Prefix) netip.Addr {
+	if p.Addr().Is4() {
+		return netip.IPv4Unspecified()
+	}
+	return netip.IPv6Unspecified()
+}
+
 func generateGUIDByDeviceName(name string) (*windows.GUID, error) {
 	// GUID is 128 bit
 	hash := crypto.MD5.New()
@@ -296,5 +334,3 @@ func checkWinTunExists() error {
 	_, err = syscall.LoadDLL(filepath.Join(filepath.Dir(myPath), "dist", "windows", "wintun", "bin", arch, "wintun.dll"))
 	return err
 }
-
-func (t *winTun) TunPrefixLen() int { return 0 }
@@ -69,5 +69,3 @@ func (d *UserDevice) Close() error {
 	d.outboundWriter.Close()
 	return nil
 }
-
-func (d *UserDevice) TunPrefixLen() int { return 0 }
@@ -99,12 +99,10 @@ func (p *PKI) reloadCerts(c *config.C, initial bool) *util.ContextualError {
 	var currentState *CertState
 	if initial {
 		cipher = c.GetString("cipher", "aes")
-		//TODO: this sucks and we should make it not a global
 		switch cipher {
-		case "aes":
-			noiseEndianness = binary.BigEndian
-		case "chachapoly":
-			noiseEndianness = binary.LittleEndian
+		case "aes", "chachapoly":
+			// Each post-handshake CipherState in noiseutil hardcodes its own
+			// nonce endianness now, so there's nothing to set up here.
 		default:
 			return util.NewContextualError(
 				"unknown cipher",
@@ -1,24 +1,70 @@
 package nebula

 import (
+	"context"
 	"log/slog"
+	"net/netip"
 	"sync/atomic"
 	"time"

+	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
 )

+// holepunchQueueSize buffers the channel that pending holepunchJobs land on after their delay timer fires.
+const holepunchQueueSize = 64
+
+// holepunchJob is one scheduled item delivered to the worker goroutine.
+//   - target valid -> send a UDP punch to target. vpnAddr, if set, is the peer's vpn addr carried for log context.
+//   - target invalid, vpnAddr valid -> send an encrypted test packet to vpnAddr (a "punchback").
+type holepunchJob struct {
+	target  netip.AddrPort
+	vpnAddr netip.Addr
+}
+
+// lighthouseChecker is the slice of LightHouse that Punchy actually needs.
+// Defined here so Punchy doesn't take a *LightHouse dependency (LightHouse
+// already holds a *Punchy, and the bidirectional pointer reference is awkward
+// even within the same package). Tests can also substitute a fake.
+type lighthouseChecker interface {
+	IsAnyLighthouseAddr(vpnAddrs []netip.Addr) bool
+}
+
 type Punchy struct {
 	punch           atomic.Bool
 	respond         atomic.Bool
 	delay           atomic.Int64
 	respondDelay    atomic.Int64
 	punchEverything atomic.Bool
-	l               *slog.Logger
+
+	sched             *Scheduler[holepunchJob]
+	punchConn         udp.Conn
+	metricHolepunchTx metrics.Counter
+	metricPunchyTx    metrics.Counter
+
+	ctx  context.Context
+	ifce EncWriter
+	hm   *HostMap
+	lh   lighthouseChecker
+
+	l *slog.Logger
 }

-func NewPunchyFromConfig(l *slog.Logger, c *config.C) *Punchy {
-	p := &Punchy{l: l}
+func NewPunchyFromConfig(l *slog.Logger, c *config.C, punchConn udp.Conn) *Punchy {
+	p := &Punchy{
+		l:              l,
+		punchConn:      punchConn,
+		sched:          NewScheduler[holepunchJob](holepunchQueueSize),
+		metricPunchyTx: metrics.GetOrRegisterCounter("messages.tx.punchy", nil),
+	}
+
+	if c.GetBool("stats.lighthouse_metrics", false) {
+		p.metricHolepunchTx = metrics.GetOrRegisterCounter("messages.tx.holepunch", nil)
+	} else {
+		p.metricHolepunchTx = metrics.NilCounter{}
+	}

 	p.reload(c, true)
 	c.RegisterReloadCallback(func(c *config.C) {
@@ -29,7 +75,7 @@ func NewPunchyFromConfig(l *slog.Logger, c *config.C) *Punchy {
 }

 func (p *Punchy) reload(c *config.C, initial bool) {
-	if initial {
+	if initial || c.HasChanged("punchy.punch") || c.HasChanged("punchy") {
 		var yes bool
 		if c.IsSet("punchy.punch") {
 			yes = c.GetBool("punchy.punch", false)
@@ -38,16 +84,15 @@ func (p *Punchy) reload(c *config.C, initial bool) {
 			yes = c.GetBool("punchy", false)
 		}

-		p.punch.Store(yes)
-		if yes {
+		old := p.punch.Swap(yes)
+		switch {
+		case initial && yes:
 			p.l.Info("punchy enabled")
-		} else {
+		case initial:
 			p.l.Info("punchy disabled")
+		case old != yes:
+			p.l.Info("punchy.punch changed", "punch", yes)
 		}
-
-	} else if c.HasChanged("punchy.punch") || c.HasChanged("punchy") {
-		//TODO: it should be relatively easy to support this, just need to be able to cancel the goroutine and boot it up from here
-		p.l.Warn("Changing punchy.punch with reload is not supported, ignoring.")
 	}

 	if initial || c.HasChanged("punchy.respond") || c.HasChanged("punch_back") {
@@ -59,52 +104,132 @@ func (p *Punchy) reload(c *config.C, initial bool) {
 			yes = c.GetBool("punch_back", false)
 		}

-		p.respond.Store(yes)
-
-		if !initial {
-			p.l.Info("punchy.respond changed", "respond", p.GetRespond())
+		old := p.respond.Swap(yes)
+		if !initial && old != yes {
+			p.l.Info("punchy.respond changed", "respond", yes)
 		}
 	}

 	//NOTE: this will not apply to any in progress operations, only the next one
 	if initial || c.HasChanged("punchy.delay") {
-		p.delay.Store((int64)(c.GetDuration("punchy.delay", time.Second)))
-		if !initial {
-			p.l.Info("punchy.delay changed", "delay", p.GetDelay())
+		newDelay := int64(c.GetDuration("punchy.delay", time.Second))
+		old := p.delay.Swap(newDelay)
+		if !initial && old != newDelay {
+			p.l.Info("punchy.delay changed", "delay", time.Duration(newDelay))
 		}
 	}

 	if initial || c.HasChanged("punchy.target_all_remotes") {
-		p.punchEverything.Store(c.GetBool("punchy.target_all_remotes", false))
-		if !initial {
-			p.l.Info("punchy.target_all_remotes changed", "target_all_remotes", p.GetTargetEverything())
+		yes := c.GetBool("punchy.target_all_remotes", false)
+		old := p.punchEverything.Swap(yes)
+		if !initial && old != yes {
+			p.l.Info("punchy.target_all_remotes changed", "target_all_remotes", yes)
 		}
 	}

 	if initial || c.HasChanged("punchy.respond_delay") {
-		p.respondDelay.Store((int64)(c.GetDuration("punchy.respond_delay", 5*time.Second)))
-		if !initial {
-			p.l.Info("punchy.respond_delay changed", "respond_delay", p.GetRespondDelay())
+		newDelay := int64(c.GetDuration("punchy.respond_delay", 5*time.Second))
+		old := p.respondDelay.Swap(newDelay)
+		if !initial && old != newDelay {
+			p.l.Info("punchy.respond_delay changed", "respond_delay", time.Duration(newDelay))
 		}
 	}
 }

-func (p *Punchy) GetPunch() bool {
-	return p.punch.Load()
+// Schedule queues a punch packet to target, to be sent after the configured delay.
+// vpnAddr is the peer's vpn addr, used for log context when the packet actually fires.
+// No-op if target is not a valid AddrPort or if Start has not yet been called. Safe to call from any goroutine.
+func (p *Punchy) Schedule(target netip.AddrPort, vpnAddr netip.Addr) {
+	if !target.IsValid() || p.ctx == nil {
+		return
+	}
+	p.scheduleJob(holepunchJob{target: target, vpnAddr: vpnAddr}, time.Duration(p.delay.Load()))
 }

-func (p *Punchy) GetRespond() bool {
-	return p.respond.Load()
+// ScheduleRespond queues a punchback test packet to vpnAddr after the configured respond delay,
+// gated on punchy.respond. No-op when respond is disabled or before Start has been called.
+func (p *Punchy) ScheduleRespond(vpnAddr netip.Addr) {
+	if !p.respond.Load() || p.ctx == nil {
+		return
+	}
+	p.scheduleJob(holepunchJob{vpnAddr: vpnAddr}, time.Duration(p.respondDelay.Load()))
 }

-func (p *Punchy) GetDelay() time.Duration {
-	return (time.Duration)(p.delay.Load())
+// scheduleJob delegates to the pooled Scheduler.
+// The callback observes p.ctx so a job that becomes due after Stop is dropped instead of queued.
+func (p *Punchy) scheduleJob(job holepunchJob, delay time.Duration) {
+	p.sched.Schedule(p.ctx, job, delay)
 }

-func (p *Punchy) GetRespondDelay() time.Duration {
-	return (time.Duration)(p.respondDelay.Load())
+// SendPunch sends an immediate keepalive punch for an idle hostinfo.
+// The configured punchy.target_all_remotes mode picks the targets. Gated on punchy.punch and the lighthouse-skip rule
+// (lighthouses don't get keepalive punches because the regular update interval keeps their NAT state warm).
+func (p *Punchy) SendPunch(hostinfo *HostInfo) {
+	if !p.punch.Load() {
+		return
+	}
+	if p.lh.IsAnyLighthouseAddr(hostinfo.vpnAddrs) {
+		return
+	}
+
+	if p.punchEverything.Load() {
+		p.sendPunchToAllRemotes(hostinfo)
+	} else if hostinfo.remote.IsValid() {
+		p.metricPunchyTx.Inc(1)
+		p.punchConn.WriteTo([]byte{1}, hostinfo.remote)
+	}
 }

-func (p *Punchy) GetTargetEverything() bool {
-	return p.punchEverything.Load()
+// SendPunchToAll punches every known remote for hostinfo, but only when punchy.target_all_remotes is enabled.
+// The connection manager calls this during outbound-only traffic: the outbound traffic itself keeps the primary's
+// NAT state warm, but non-primary remotes need separate refresh, so we fan out to all of them (the redundant
+// primary punch is harmless). Gated on punchy.punch and the lighthouse-skip rule.
+func (p *Punchy) SendPunchToAll(hostinfo *HostInfo) {
+	if !p.punchEverything.Load() {
+		return
+	}
+	if !p.punch.Load() {
+		return
+	}
+	if p.lh.IsAnyLighthouseAddr(hostinfo.vpnAddrs) {
+		return
+	}
+	p.sendPunchToAllRemotes(hostinfo)
+}
+
+func (p *Punchy) sendPunchToAllRemotes(hostinfo *HostInfo) {
+	hostinfo.remotes.ForEach(p.hm.GetPreferredRanges(), func(addr netip.AddrPort, preferred bool) {
+		p.metricPunchyTx.Inc(1)
+		p.punchConn.WriteTo([]byte{1}, addr)
+	})
+}
+
+// Start wires the runtime dependencies and spawns the scheduler worker.
+func (p *Punchy) Start(ctx context.Context, ifce EncWriter, hm *HostMap, lh lighthouseChecker) {
+	p.ctx = ctx
+	p.ifce = ifce
+	p.hm = hm
+	p.lh = lh
+
+	nb := make([]byte, 12, 12)
+	out := make([]byte, mtu)
+	empty := []byte{0}
+
+	go p.sched.Run(ctx, func(job holepunchJob) {
+		switch {
+		case job.target.IsValid():
+			if p.l.Enabled(context.Background(), slog.LevelDebug) {
+				p.l.Debug("Punching", "target", job.target, "vpnAddr", job.vpnAddr)
+			}
+			p.metricHolepunchTx.Inc(1)
+			p.punchConn.WriteTo(empty, job.target)
+		case job.vpnAddr.IsValid():
+			// A nebula test packet to the host trying to contact us.
+			// In the case of a double nat or other difficult scenario, this may help establish a tunnel.
+			if p.l.Enabled(context.Background(), slog.LevelDebug) {
+				p.l.Debug("Sending a nebula test packet", "vpnAddr", job.vpnAddr)
+			}
+			p.ifce.SendMessageToVpnAddr(header.Test, header.TestRequest, job.vpnAddr, []byte(""), nb, out)
+		}
+	})
 }
@@ -17,42 +17,42 @@ func TestNewPunchyFromConfig(t *testing.T) {
 	c := config.NewC(l)

 	// Test defaults
-	p := NewPunchyFromConfig(test.NewLogger(), c)
-	assert.False(t, p.GetPunch())
-	assert.False(t, p.GetRespond())
-	assert.Equal(t, time.Second, p.GetDelay())
-	assert.Equal(t, 5*time.Second, p.GetRespondDelay())
+	p := NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.False(t, p.punch.Load())
+	assert.False(t, p.respond.Load())
+	assert.Equal(t, time.Second, time.Duration(p.delay.Load()))
+	assert.Equal(t, 5*time.Second, time.Duration(p.respondDelay.Load()))

 	// punchy deprecation
 	c.Settings["punchy"] = true
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.True(t, p.GetPunch())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.True(t, p.punch.Load())

 	// punchy.punch
 	c.Settings["punchy"] = map[string]any{"punch": true}
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.True(t, p.GetPunch())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.True(t, p.punch.Load())

 	// punch_back deprecation
 	c.Settings["punch_back"] = true
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.True(t, p.GetRespond())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.True(t, p.respond.Load())

 	// punchy.respond
 	c.Settings["punchy"] = map[string]any{"respond": true}
 	c.Settings["punch_back"] = false
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.True(t, p.GetRespond())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.True(t, p.respond.Load())

 	// punchy.delay
 	c.Settings["punchy"] = map[string]any{"delay": "1m"}
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.Equal(t, time.Minute, p.GetDelay())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.Equal(t, time.Minute, time.Duration(p.delay.Load()))

 	// punchy.respond_delay
 	c.Settings["punchy"] = map[string]any{"respond_delay": "1m"}
-	p = NewPunchyFromConfig(test.NewLogger(), c)
-	assert.Equal(t, time.Minute, p.GetRespondDelay())
+	p = NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.Equal(t, time.Minute, time.Duration(p.respondDelay.Load()))
 }

 func TestPunchy_reload(t *testing.T) {
@@ -61,35 +61,34 @@ func TestPunchy_reload(t *testing.T) {
 	delay, _ := time.ParseDuration("1m")
 	require.NoError(t, c.LoadString(`
 punchy:
+  punch: false
  delay: 1m
  respond: false
 `))
-	p := NewPunchyFromConfig(test.NewLogger(), c)
-	assert.Equal(t, delay, p.GetDelay())
-	assert.False(t, p.GetRespond())
+	p := NewPunchyFromConfig(test.NewLogger(), c, nil)
+	assert.False(t, p.punch.Load())
+	assert.Equal(t, delay, time.Duration(p.delay.Load()))
+	assert.False(t, p.respond.Load())

 	newDelay, _ := time.ParseDuration("10m")
 	require.NoError(t, c.ReloadConfigString(`
 punchy:
+  punch: true
  delay: 10m
  respond: true
 `))
 	p.reload(c, false)
-	assert.Equal(t, newDelay, p.GetDelay())
-	assert.True(t, p.GetRespond())
+	assert.True(t, p.punch.Load())
+	assert.Equal(t, newDelay, time.Duration(p.delay.Load()))
+	assert.True(t, p.respond.Load())
 }

 // The tests below pin the shape of each log line Punchy produces so changes
 // cannot silently break whatever operators are grepping for. The assertions
 // are on the structured message + attrs (e.g. "punchy.respond changed" with
-// a respond=true field) rather than a formatted string.
-//
-// Punchy.reload also emits a spurious "Changing punchy.punch with reload is
-// not supported" warning whenever any key under punchy changes, because of
-// the c.HasChanged("punchy") fallback kept for the deprecated top-level
-// punchy form. The tests filter by message rather than asserting total
-// entry counts so that warning is tolerated without being locked into
-// the format.
+// a respond=true field) rather than a formatted string. Tests filter by
+// message rather than asserting total entry counts so unrelated info lines
+// are tolerated without being locked into the format.

 type capturedEntry struct {
 	Level slog.Level
@@ -145,7 +144,7 @@ func TestPunchy_LogFormat_InitialEnabled(t *testing.T) {
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {punch: true}`))

-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)

 	entry := findEntry(t, hook.entries, "punchy enabled")
 	assert.Equal(t, slog.LevelInfo, entry.Level)
@@ -157,32 +156,32 @@ func TestPunchy_LogFormat_InitialDisabled(t *testing.T) {
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {punch: false}`))

-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)

 	entry := findEntry(t, hook.entries, "punchy disabled")
 	assert.Equal(t, slog.LevelInfo, entry.Level)
 	assert.Empty(t, entry.Attrs)
 }

-func TestPunchy_LogFormat_ReloadPunchUnsupported(t *testing.T) {
+func TestPunchy_LogFormat_ReloadPunch(t *testing.T) {
 	l, hook := newCapturingPunchyLogger(t)
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {punch: false}`))
-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)
 	hook.entries = nil

 	require.NoError(t, c.ReloadConfigString(`punchy: {punch: true}`))

-	entry := findEntry(t, hook.entries, "Changing punchy.punch with reload is not supported, ignoring.")
-	assert.Equal(t, slog.LevelWarn, entry.Level)
-	assert.Empty(t, entry.Attrs)
+	entry := findEntry(t, hook.entries, "punchy.punch changed")
+	assert.Equal(t, slog.LevelInfo, entry.Level)
+	assert.Equal(t, map[string]any{"punch": true}, entry.Attrs)
 }

 func TestPunchy_LogFormat_ReloadRespond(t *testing.T) {
 	l, hook := newCapturingPunchyLogger(t)
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {respond: false}`))
-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)
 	hook.entries = nil

 	require.NoError(t, c.ReloadConfigString(`punchy: {respond: true}`))
@@ -196,7 +195,7 @@ func TestPunchy_LogFormat_ReloadDelay(t *testing.T) {
 	l, hook := newCapturingPunchyLogger(t)
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {delay: 1s}`))
-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)
 	hook.entries = nil

 	require.NoError(t, c.ReloadConfigString(`punchy: {delay: 10s}`))
@@ -210,7 +209,7 @@ func TestPunchy_LogFormat_ReloadTargetAllRemotes(t *testing.T) {
 	l, hook := newCapturingPunchyLogger(t)
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {target_all_remotes: false}`))
-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)
 	hook.entries = nil

 	require.NoError(t, c.ReloadConfigString(`punchy: {target_all_remotes: true}`))
@@ -224,7 +223,7 @@ func TestPunchy_LogFormat_ReloadRespondDelay(t *testing.T) {
 	l, hook := newCapturingPunchyLogger(t)
 	c := config.NewC(test.NewLogger())
 	require.NoError(t, c.LoadString(`punchy: {respond_delay: 5s}`))
-	NewPunchyFromConfig(l, c)
+	NewPunchyFromConfig(l, c, nil)
 	hook.entries = nil

 	require.NoError(t, c.ReloadConfigString(`punchy: {respond_delay: 15s}`))
--- a/Show More
+++ b/Show More