diff --git a/docs/android-redroid-runbook.md b/docs/android-redroid-runbook.md
new file mode 100644
index 0000000..0583484
--- /dev/null
+++ b/docs/android-redroid-runbook.md
@@ -0,0 +1,172 @@
+# Running Android (redroid) inside a matchlock sandbox VM
+
+Proven end-to-end 2026-08-04: Android 15 arm64 boots inside an rsv2 sandbox VM,
+the RSHomebuyerApp debug APK builds *inside the same VM*, and the app signs into
+that instance's Rails — no emulator binary, no KVM, no host-side Android
+tooling. This runbook records the full recipe and every trap we hit.
+
+## Why redroid and not the emulator
+
+Google ships **no linux-aarch64 Android emulator** in the SDK channel
+(repository2-3.xml has linux_x64 / darwin_x64 / darwin_aarch64 / windows_x64
+only). A CI build (`emulator-linux_aarch64_gfxstream`) exists on ci.android.com
+but is unfetchable without authenticated `fetch_artifact`. Combined with no
+`/dev/kvm` in the guest, emulator-based Android is a dead end. redroid runs
+Android as a container on the guest kernel at native ARM speed.
+
+## 1. Kernel (this PR)
+
+redroid needs guest-kernel features the matchlock kernel didn't have. All are
+`=y` (CONFIG_MODULES=n):
+
+- `CONFIG_ANDROID_BINDER_IPC` / `CONFIG_ANDROID_BINDERFS` /
+  `CONFIG_ANDROID_BINDER_DEVICES="binder,hwbinder,vndbinder"` — Android IPC.
+  binderfs allows multiple Android instances per kernel. ashmem is gone since
+  5.18; modern redroid needs binder only.
+- `CONFIG_BINFMT_MISC` — lets qemu-user run the x86_64-only Android SDK
+  tooling (aapt2, NDK clang) in the arm64 guest.
+- `CONFIG_SECURITY` + `SECURITY_NETWORK` + `AUDIT` + `SECURITY_SELINUX` +
+  `SELINUX_DEVELOP` — Android init expects /sys/fs/selinux; DEVELOP boots
+  permissive. SECURITY_SELINUX's Kconfig deps drag the rest in.
+- `CONFIG_MEMCG_V1` / `CPUSETS_V1` — Android's libprocessgroup sets task
+  profiles against cgroup v1 (/dev/cpuctl, /dev/cpuset). Note 6.19 has **no**
+  `CONFIG_CGROUP_V1`; that symbol was sitting in the config doing nothing.
+
+`olddefconfig` drops unknown/unmet options **silently** — we found two phantom
+symbols (`CONFIG_CGROUP_V1`, `CONFIG_SELINUX`) that read as authoritative but
+did nothing. `guest/kernel/verify-config.sh` now fails the Docker kernel build
+on any *new* drift; `known-drift.txt` holds the untriaged pre-existing drops.
+
+Build: `KERNEL_VERSION=6.19.8-dmcrypt-vfs-binder2 KERNEL_SOURCE_VERSION=6.19.8
+ARCH=arm64 ./scripts/build-kernel.sh` → lands in
+`~/.cache/matchlock/kernels/<version>/kernel-arm64`.
+
+## 2. Boot an instance on the new kernel (no fleet impact)
+
+rsv2 pins the kernel in `qa/rsv2-cli/cmd/setup.go` (`sandboxKernelVersion`);
+per-launch overrides avoid touching the pin:
+
+```sh
+RSV2_SANDBOX_KERNEL=$HOME/.cache/matchlock/kernels/6.19.8-dmcrypt-vfs-binder2/kernel-arm64 \
+RSV2_SANDBOX_MEMORY_MB=9216 \
+rsv2 sandbox 1 --detach
+```
+
+Restarting a VM destroys `/opt` and `/root` (ephemeral overlay). Salvage
+first, over virtiofs, from a detached container so the exec channel stays free:
+
+```sh
+matchlock exec <vm> -- sh -c 'docker run -d --name salvage \
+  -v /workspace/app/tmp:/out -v /opt:/s/opt:ro -v /root/.gradle:/s/root/.gradle:ro \
+  alpine tar -C /s -cf /out/toolchain.tar opt/android-sdk opt/node20 opt/work root/.gradle'
+```
+
+**Per-VM CA trap:** every VM boot mints a fresh matchlock MITM CA. A salvaged
+Java cacerts carries the old CA and Gradle fails with PKIX "signature check
+failed" (while curl works — it uses the system store guest-init refreshes).
+After any VM swap, re-import:
+
+```sh
+keytool -delete -cacerts -storepass changeit -alias matchlock
+keytool -importcert -noprompt -cacerts -storepass changeit -alias matchlock \
+  -file /etc/ssl/certs/matchlock-ca.crt
+```
+
+## 3. redroid
+
+```sh
+docker run -d --privileged --name redroid -v /data/redroid:/data -p 5555:5555 \
+  redroid/redroid:15.0.0_64only-latest androidboot.redroid_gpu_mode=guest
+```
+
+- Boots to `sys.boot_completed=1` in ~2 min with ≥4 CPU / ≥6 GB (it failed
+  mysteriously at 1 CPU / 512 MB — check resources before debugging Android).
+- **Black screencap ≠ GPU bug.** redroid starts no Home app, so nothing draws.
+  `am start -a android.intent.action.MAIN -c android.intent.category.HOME`
+  starts the launcher. Read the GPU mode back as `ro.boot.redroid_gpu_mode`
+  (NOT `redroid.gpu.mode`). Byte-size check: black 720x1280 PNG ≈ 6 KB, real
+  content ≥ 50 KB.
+- **redroid's netd programs no kernel routes** — policy-routing tables 1003/99
+  are empty, so every app socket dies ENETUNREACH. Fix inside the container:
+
+  ```sh
+  ip rule add from all lookup main pref 20000
+  ip route add default via 172.17.0.1 dev eth0 onlink   # only for egress
+  ```
+
+- No adb needed: `docker cp` + `docker exec redroid pm install`, `am start`,
+  `input tap/text`, `screencap -p | base64`.
+- No Google Play services (AOSP image): the app's map view is dead. Everything
+  else works. A GApps/microG redroid variant exists if maps ever matter.
+
+## 4. In-VM networking (the shield changes the rules)
+
+- Containers **cannot** reach VM host-netns ports (shield INPUT chain) and
+  **cannot** hairpin to published ports via the gateway. Same-bridge
+  container-to-container traffic is the only reliable path.
+- Put everything the app needs on redroid's bridge:
+  `docker network connect bridge rsv2-qa-1-web-1`, run Metro as a bridge
+  container. Talk container-IP to container-IP.
+- Rails host authorization only admits `.rslocal.test` — point the name at the
+  web container's bridge IP inside redroid (`/system/etc/hosts` works; bionic
+  reads it) and give the app `http://www.rslocal.test:3000`.
+
+## 5. Building the APK inside the VM
+
+- **binfmt_misc is per-user-namespace, and every matchlock exec AND every
+  docker container here gets its own instance.** A qemu registration made
+  anywhere else is invisible to the build. Register inside the same exec that
+  runs Gradle (F flag pins the interpreter at registration):
+
+  ```sh
+  mount -t binfmt_misc binfmt_misc /proc/sys/fs/binfmt_misc
+  printf ':qemu-x86_64:M::\x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x3e\x00:\xff\xff\xff\xff\xff\xfe\xfe\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff:/usr/bin/qemu-x86_64-static:F\n' \
+    > /proc/sys/fs/binfmt_misc/register
+  ```
+
+- Dynamic x86_64 binaries need amd64 multiarch libs:
+  `dpkg --add-architecture amd64`, add an archive.ubuntu.com amd64 source
+  (arm64 stays on ports.ubuntu.com), then
+  `apt-get install libc6:amd64 libstdc++6:amd64 zlib1g:amd64`. dpkg's
+  configure step fails without VM-wide binfmt — harmless; the unpacked files
+  are what matters.
+- AGP demands the aapt2 override be a file named exactly `aapt2`
+  (`android.aapt2FromMavenOverride=/opt/aapt2-shim/aapt2`, a one-line
+  qemu-exec wrapper).
+- `reactNativeArchitectures=arm64-v8a` in gradle.properties — the default
+  builds 4 ABIs; under emulation that's 4x the C++ for nothing.
+- Shield allowlist needed: `dl.google.com maven.google.com
+  repo.maven.apache.org plugins.gradle.org plugins-artifacts.gradle.org
+  services.gradle.org archive.ubuntu.com ports.ubuntu.com`
+  (`matchlock allow-list add <vm> ...`).
+- Run the build as ONE long exec logging to virtiofs
+  (`> /workspace/app/tmp/build.log`) and read the log from the host — execs
+  serialize, and polling from a second exec collides with the build.
+- Expo debug builds bundle no JS: run Metro
+  (`npx expo start --offline --port 8081`) as a bridge-network container and
+  aim the dev client with
+  `am start -a android.intent.action.VIEW -d
+  "rshomebuyerapp://expo-development-client/?url=http%3A%2F%2F<metro-ip>%3A8081"`.
+
+## 6. Demo-data gotchas (rsv2 seed wizard)
+
+- Seeded listings live in one MLS (RECSFAR here, all "san francisco") — a user
+  whose saved searches target another MLS sees zero matches. jordan.reyes's QA
+  searches target RECSFAR; harry.homebuyer's don't.
+- Favorites are `Verdict` rows (`prose: 'saved'`); create them via Rails
+  console to pre-populate the interested tab.
+- The seed generator writes **relative** `/stub_photos/...` photo URLs — fine
+  in a browser, opaque to React Native's `Image`. Absolutize in ES
+  (`_update_by_query` prepending `http://www.rslocal.test:3000`) and photos
+  render, served by the in-VM web container.
+
+## Next direction: Rosetta instead of qemu
+
+The one real tax left is qemu TCG emulation of the x86_64 NDK toolchain.
+Virtualization.framework can inject Apple's Rosetta into a Linux guest as a
+read-only directory share (`VZLinuxRosettaDirectoryShare`); registering
+`/media/rosetta/rosetta` as the x86_64 binfmt handler gives near-native speed —
+the mechanism Docker Desktop uses for fast amd64 containers. Security-neutral
+relative to qemu: translation runs entirely in-guest, no host channel, one more
+read-only virtiofs device. Requires host-side plumbing in the VZ config, which
+we own in this fork.
diff --git a/docs/android/build-apk.sh b/docs/android/build-apk.sh
new file mode 100644
index 0000000..69aa68d
--- /dev/null
+++ b/docs/android/build-apk.sh
@@ -0,0 +1,20 @@
+#!/bin/sh
+# Runs inside one matchlock exec. binfmt_misc instances are per-userns and
+# every exec gets a fresh userns, so the qemu registration must happen in
+# the same exec that runs the build — it cannot be done once VM-wide.
+LOG=/workspace/app/tmp/apk-build.log
+{
+    set -ex
+    mount -t binfmt_misc binfmt_misc /proc/sys/fs/binfmt_misc 2>/dev/null || true
+    printf ':qemu-x86_64:M::\x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x3e\x00:\xff\xff\xff\xff\xff\xfe\xfe\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff:/usr/bin/qemu-x86_64-static:F\n' > /proc/sys/fs/binfmt_misc/register
+    /opt/android-sdk/build-tools/36.0.0/aapt2 version
+    export JAVA_HOME=/usr/lib/jvm/java-17-openjdk-arm64
+    export ANDROID_HOME=/opt/android-sdk
+    export ANDROID_SDK_ROOT=/opt/android-sdk
+    export PATH="/opt/node20/bin:$JAVA_HOME/bin:$PATH"
+    node --version
+    cd /opt/work/RSHomebuyerApp/android
+    ./gradlew assembleDebug
+    cp app/build/outputs/apk/debug/app-debug.apk /workspace/app/tmp/RSHomebuyerApp-debug.apk
+    echo BUILD_SCRIPT_DONE
+} > "$LOG" 2>&1
diff --git a/docs/android/register-binfmt.sh b/docs/android/register-binfmt.sh
new file mode 100644
index 0000000..36e0703
--- /dev/null
+++ b/docs/android/register-binfmt.sh
@@ -0,0 +1,16 @@
+#!/bin/sh
+# Register qemu-x86_64 as the binfmt_misc handler for x86_64 ELF binaries.
+# Must run in the VM's init user namespace (e.g. a privileged docker
+# container) — matchlock execs get their own userns, whose binfmt instance
+# evaporates with the exec. The F flag pins the interpreter at registration
+# so callers never need /usr/bin/qemu-x86_64-static in their own mount ns.
+set -e
+mount -t binfmt_misc binfmt_misc /proc/sys/fs/binfmt_misc 2>/dev/null || true
+if [ -e /proc/sys/fs/binfmt_misc/qemu-x86_64 ]; then
+    echo "already registered:"
+    head -3 /proc/sys/fs/binfmt_misc/qemu-x86_64
+    exit 0
+fi
+printf ':qemu-x86_64:M::\x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x3e\x00:\xff\xff\xff\xff\xff\xfe\xfe\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff:/usr/bin/qemu-x86_64-static:F\n' > /proc/sys/fs/binfmt_misc/register
+echo "registered:"
+head -3 /proc/sys/fs/binfmt_misc/qemu-x86_64
diff --git a/guest/kernel/Dockerfile b/guest/kernel/Dockerfile
index 1580095..17ddf5a 100644
--- a/guest/kernel/Dockerfile
+++ b/guest/kernel/Dockerfile
@@ -44,10 +44,14 @@ RUN set -eux; \
 # ---------------------------------------------------------------------------
 FROM base AS build-x86_64
 
-COPY x86_64.config /tmp/linux-src/.config
+COPY x86_64.config /tmp/requested.config
+COPY verify-config.sh /usr/local/bin/verify-config.sh
+COPY known-drift.txt /tmp/known-drift.txt
 
 RUN cd /tmp/linux-src \
+    && cp /tmp/requested.config .config \
     && make olddefconfig \
+    && sh /usr/local/bin/verify-config.sh /tmp/requested.config .config /tmp/known-drift.txt \
     && make -j"$(nproc)" vmlinux
 
 # ---------------------------------------------------------------------------
@@ -55,10 +59,14 @@ RUN cd /tmp/linux-src \
 # ---------------------------------------------------------------------------
 FROM base AS build-arm64
 
-COPY arm64.config /tmp/linux-src/.config
+COPY arm64.config /tmp/requested.config
+COPY verify-config.sh /usr/local/bin/verify-config.sh
+COPY known-drift.txt /tmp/known-drift.txt
 
 RUN cd /tmp/linux-src \
+    && cp /tmp/requested.config .config \
     && make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- olddefconfig \
+    && sh /usr/local/bin/verify-config.sh /tmp/requested.config .config /tmp/known-drift.txt \
     && make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- -j"$(nproc)" Image
 
 # ---------------------------------------------------------------------------
diff --git a/guest/kernel/arm64.config b/guest/kernel/arm64.config
index 92afae3..1a4d1aa 100644
--- a/guest/kernel/arm64.config
+++ b/guest/kernel/arm64.config
@@ -142,6 +142,26 @@ CONFIG_BLK_DEV_INITRD=y
 CONFIG_RD_GZIP=y
 CONFIG_BINFMT_ELF=y
 CONFIG_BINFMT_SCRIPT=y
+# Lets qemu-user register foreign-arch binaries, so x86_64-only vendor tooling
+# (Android SDK aapt2/adb ship no linux-aarch64 build) runs in an arm64 guest.
+CONFIG_BINFMT_MISC=y
+
+# Android binder — runs containerised Android (redroid) in the guest, which
+# needs no KVM and no vendor emulator binary. Both are bool, so =y is required
+# under CONFIG_MODULES=n.
+CONFIG_ANDROID_BINDER_IPC=y
+CONFIG_ANDROID_BINDERFS=y
+CONFIG_ANDROID_BINDER_DEVICES="binder,hwbinder,vndbinder"
+
+# SELinux — Android's init and system_server expect /sys/fs/selinux to exist.
+# SECURITY_SELINUX depends on SECURITY_NETWORK && AUDIT && NET && INET, so all
+# of these have to come along; SELINUX_DEVELOP boots permissive, which is what
+# we want (Android supplies its own policy, we are not enforcing a host one).
+CONFIG_SECURITY=y
+CONFIG_SECURITY_NETWORK=y
+CONFIG_AUDIT=y
+CONFIG_SECURITY_SELINUX=y
+CONFIG_SECURITY_SELINUX_DEVELOP=y
 
 # Kernel options
 CONFIG_PREEMPT_NONE=y
@@ -178,7 +198,11 @@ CONFIG_BLK_CGROUP=y
 CONFIG_CGROUP_WRITEBACK=y
 CONFIG_CGROUP_BPF=y
 CONFIG_CGROUP_CPUACCT=y
-CONFIG_CGROUP_V1=y
+# 6.19 has no CONFIG_CGROUP_V1 — the legacy hierarchies split per controller.
+# Android's libprocessgroup sets task profiles against cgroup v1 (/dev/cpuctl,
+# /dev/cpuset), so redroid needs these.
+CONFIG_MEMCG_V1=y
+CONFIG_CPUSETS_V1=y
 
 # Namespaces (all types required for rootless BuildKit / runc)
 CONFIG_NAMESPACES=y
@@ -235,9 +259,6 @@ CONFIG_DEBUG_INFO=n
 CONFIG_DEBUG_KERNEL=n
 CONFIG_WATCHDOG=n
 CONFIG_INPUT=n
-CONFIG_SELINUX=n
-CONFIG_SECURITY=n
-CONFIG_AUDIT=n
 CONFIG_PROFILING=n
 CONFIG_WEXT_CORE=n
 CONFIG_WEXT_PROC=n
diff --git a/guest/kernel/known-drift.txt b/guest/kernel/known-drift.txt
new file mode 100644
index 0000000..c1b1661
--- /dev/null
+++ b/guest/kernel/known-drift.txt
@@ -0,0 +1,34 @@
+# Options that olddefconfig already drops from the configs on 6.19.8, from
+# before verify-config.sh existed — so the kernels we ship today are already
+# built without them. Listed here to keep the build moving while still failing
+# on any NEW drift.
+#
+# These have NOT been triaged. Two different things are mixed together:
+#   - Options that no longer exist on this arch/version (EXT4_FS_XATTR was
+#     folded into ext4; EARLY_PRINTK is largely x86) — harmless.
+#   - Options with genuinely unmet dependencies. The netfilter entries are
+#     this kind, and Docker NAT depends on them.
+# Shrinking this list is the actual fix; every line is a known gap.
+#
+# One entry has already been triaged off this list: CONFIG_CGROUP_V1 was a
+# symbol that does not exist in 6.19 (the legacy hierarchies split into
+# per-controller MEMCG_V1/CPUSETS_V1), so it read as authoritative while
+# silently doing nothing. Expect others here to be the same kind of phantom —
+# check the symbol exists for this version before assuming it is a real gap.
+
+CONFIG_IP_NF_FILTER
+CONFIG_IP_NF_MANGLE
+CONFIG_IP_NF_RAW
+CONFIG_IP_NF_NAT
+CONFIG_IP_NF_TARGET_MASQUERADE
+CONFIG_IP_NF_TARGET_REDIRECT
+CONFIG_IP6_NF_FILTER
+CONFIG_IP6_NF_MANGLE
+CONFIG_IP6_NF_RAW
+CONFIG_IP6_NF_NAT
+CONFIG_IP6_NF_TARGET_MASQUERADE
+CONFIG_EXT4_FS_XATTR
+CONFIG_EARLY_PRINTK
+CONFIG_ARCH_VIRT
+CONFIG_HYPERVISOR_GUEST
+CONFIG_BLK_MQ_VIRTIO
diff --git a/guest/kernel/verify-config.sh b/guest/kernel/verify-config.sh
new file mode 100755
index 0000000..b0f5cc1
--- /dev/null
+++ b/guest/kernel/verify-config.sh
@@ -0,0 +1,53 @@
+#!/bin/sh
+# Fail the kernel build when olddefconfig silently drops a requested option.
+#
+# olddefconfig resolves unmet dependencies by dropping the option rather than
+# erroring, so a config can build "successfully" without features the guest
+# needs — this bit us with netfilter (Docker NAT) on 6.19.8. Comparing the
+# requested config against the resolved .config is the only way to catch it.
+#
+# Usage: verify-config.sh <requested.config> <resolved .config> [known-drift.txt]
+#
+# known-drift.txt lists options already being dropped before this check
+# existed. They are reported but do not fail the build, so pre-existing gaps
+# stay visible without blocking every kernel build on a backlog.
+
+set -eu
+
+requested="$1"
+resolved="$2"
+drift="${3:-}"
+
+known() {
+    [ -n "$drift" ] && [ -f "$drift" ] || return 1
+    grep -v '^#' "$drift" | grep -qxF "$1"
+}
+
+missing=""
+tolerated=""
+while IFS= read -r line; do
+    grep -qxF "$line" "$resolved" && continue
+    if known "${line%=y}"; then
+        tolerated="${tolerated}  ${line}
+"
+    else
+        missing="${missing}${line}
+"
+    fi
+done <<EOF
+$(grep -E '^CONFIG_[A-Za-z0-9_]+=y$' "$requested")
+EOF
+
+if [ -n "$tolerated" ]; then
+    echo "verify-config: known pre-existing drift (see $drift):"
+    printf '%s' "$tolerated"
+fi
+
+if [ -n "$missing" ]; then
+    echo "ERROR: olddefconfig dropped options requested in $requested:" >&2
+    printf '%s' "$missing" >&2
+    echo "Check each option's Kconfig dependencies against this kernel version." >&2
+    exit 1
+fi
+
+echo "verify-config: all requested =y options survived olddefconfig."
diff --git a/guest/kernel/x86_64.config b/guest/kernel/x86_64.config
index 11ea9d8..abddc02 100644
--- a/guest/kernel/x86_64.config
+++ b/guest/kernel/x86_64.config
@@ -148,6 +148,26 @@ CONFIG_BLK_DEV_INITRD=y
 CONFIG_RD_GZIP=y
 CONFIG_BINFMT_ELF=y
 CONFIG_BINFMT_SCRIPT=y
+# Lets qemu-user register foreign-arch binaries, so aarch64-only tooling runs
+# in an x86_64 guest.
+CONFIG_BINFMT_MISC=y
+
+# Android binder — runs containerised Android (redroid) in the guest, which
+# needs no KVM and no vendor emulator binary. Both are bool, so =y is required
+# under CONFIG_MODULES=n.
+CONFIG_ANDROID_BINDER_IPC=y
+CONFIG_ANDROID_BINDERFS=y
+CONFIG_ANDROID_BINDER_DEVICES="binder,hwbinder,vndbinder"
+
+# SELinux — Android's init and system_server expect /sys/fs/selinux to exist.
+# SECURITY_SELINUX depends on SECURITY_NETWORK && AUDIT && NET && INET, so all
+# of these have to come along; SELINUX_DEVELOP boots permissive, which is what
+# we want (Android supplies its own policy, we are not enforcing a host one).
+CONFIG_SECURITY=y
+CONFIG_SECURITY_NETWORK=y
+CONFIG_AUDIT=y
+CONFIG_SECURITY_SELINUX=y
+CONFIG_SECURITY_SELINUX_DEVELOP=y
 
 # Kernel options
 CONFIG_PREEMPT_NONE=y
@@ -184,7 +204,11 @@ CONFIG_BLK_CGROUP=y
 CONFIG_CGROUP_WRITEBACK=y
 CONFIG_CGROUP_BPF=y
 CONFIG_CGROUP_CPUACCT=y
-CONFIG_CGROUP_V1=y
+# 6.19 has no CONFIG_CGROUP_V1 — the legacy hierarchies split per controller.
+# Android's libprocessgroup sets task profiles against cgroup v1 (/dev/cpuctl,
+# /dev/cpuset), so redroid needs these.
+CONFIG_MEMCG_V1=y
+CONFIG_CPUSETS_V1=y
 
 # Namespaces (all types required for rootless BuildKit / runc)
 CONFIG_NAMESPACES=y
@@ -232,9 +256,6 @@ CONFIG_DEBUG_INFO=n
 CONFIG_DEBUG_KERNEL=n
 CONFIG_WATCHDOG=n
 CONFIG_INPUT=n
-CONFIG_SELINUX=n
-CONFIG_SECURITY=n
-CONFIG_AUDIT=n
 CONFIG_PROFILING=n
 CONFIG_WEXT_CORE=n
 CONFIG_WEXT_PROC=n
diff --git a/repro-kernel-config-drift.sh b/repro-kernel-config-drift.sh
new file mode 100755
index 0000000..8556ec9
--- /dev/null
+++ b/repro-kernel-config-drift.sh
@@ -0,0 +1,104 @@
+#!/usr/bin/env bash
+# Reproduction: kernel 6.19.8 olddefconfig silently drops netfilter options
+# required for Docker bridge/NAT networking.
+#
+# This runs inside Docker — no host toolchain needed.
+# Requires: Docker with BuildKit.
+#
+# Expected result:
+#   6.1.137: all 6 iptables options preserved after olddefconfig
+#   6.19.8:  all 6 iptables options silently dropped after olddefconfig
+#
+# See: https://github.com/jingkaihe/matchlock/issues/XXX
+
+set -euo pipefail
+
+CONFIG_DIR="$(cd "$(dirname "$0")/guest/kernel" && pwd)"
+ARM64_CONFIG="$CONFIG_DIR/arm64.config"
+
+if [ ! -f "$ARM64_CONFIG" ]; then
+  echo "ERROR: arm64.config not found at $ARM64_CONFIG" >&2
+  exit 1
+fi
+
+# Explicitly set in arm64.config, and all `default n` with nothing selecting
+# them — so olddefconfig will drop them silently rather than erroring.
+# Netfilter: Docker iptables-legacy NAT/filter/mangle support.
+# Binder: containerised Android (redroid) in the guest.
+# BINFMT_MISC: qemu-user registration for foreign-arch binaries.
+# CONFIG_ANDROID_BINDER_DEVICES is omitted deliberately — it is a string
+# option, not =y, and its upstream default is already the value we want.
+CRITICAL_OPTIONS=(
+  CONFIG_IP_NF_FILTER
+  CONFIG_IP_NF_MANGLE
+  CONFIG_IP_NF_NAT
+  CONFIG_IP6_NF_FILTER
+  CONFIG_IP6_NF_MANGLE
+  CONFIG_IP6_NF_NAT
+  CONFIG_BINFMT_MISC
+  CONFIG_ANDROID_BINDER_IPC
+  CONFIG_ANDROID_BINDERFS
+)
+
+# Derived so the grep below cannot drift from the list above.
+CRITICAL_REGEX=$(IFS='|'; echo "${CRITICAL_OPTIONS[*]}")
+
+check_kernel() {
+  local version="$1"
+  local source_version="$version"
+
+  # kernel.org strips trailing .0 from URLs
+  case "$source_version" in
+    *.*.0) source_version="${source_version%.0}" ;;
+  esac
+
+  echo ""
+  echo "=== Kernel $version ==="
+
+  local result
+  result=$(docker run --rm \
+    -v "$ARM64_CONFIG:/config/arm64.config:ro" \
+    ubuntu:22.04 bash -c "
+      apt-get update -qq >/dev/null 2>&1
+      apt-get install -y -qq --no-install-recommends \
+        build-essential bc flex bison libelf-dev libssl-dev wget ca-certificates \
+        gcc-aarch64-linux-gnu >/dev/null 2>&1
+
+      wget -qO /tmp/linux.tar.xz \
+        https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-${source_version}.tar.xz
+      mkdir /tmp/linux-src
+      tar xf /tmp/linux.tar.xz -C /tmp/linux-src --strip-components=1
+
+      cp /config/arm64.config /tmp/linux-src/.config
+      cd /tmp/linux-src
+      make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- olddefconfig >/dev/null 2>&1
+
+      grep -E '^(${CRITICAL_REGEX})=' .config || true
+    " 2>&1)
+
+  echo "$result"
+
+  local dropped=0
+  for opt in "${CRITICAL_OPTIONS[@]}"; do
+    if ! echo "$result" | grep -q "${opt}=y"; then
+      echo "  DROPPED: ${opt}"
+      dropped=$((dropped + 1))
+    fi
+  done
+
+  if [ $dropped -eq 0 ]; then
+    echo "  All ${#CRITICAL_OPTIONS[@]} critical options preserved."
+  else
+    echo "  $dropped/${#CRITICAL_OPTIONS[@]} critical options silently dropped."
+  fi
+}
+
+echo "Reproduction: arm64.config default-n options dropped by olddefconfig on 6.19.8"
+echo "Config: $ARM64_CONFIG"
+
+check_kernel "6.1.137"
+check_kernel "6.19.8"
+
+echo ""
+echo "If 6.19.8 shows dropped options, the bug is confirmed."
+echo "Docker fails with: iptables ... can't initialize iptables table 'nat'"