ci: split into fast analyze + slow release build with path gating

analyze runs on every push/PR (fast check).
release-apk only fires on push to master or workflow_dispatch (slow build).
Doc-only pushes no longer trigger any CI run.
workflow_dispatch triggers both jobs unconditionally.

.github/workflows/ci.yaml

@@ -3,20 +3,39 @@ name: CI
 on:
   push:
     branches: [ master ]
+    paths:
+      - 'lib/**.dart'
+      - 'test/**.dart'
+      - 'android/**'
+      - 'tool/**'
+      - 'pubspec.yaml'
+      - 'pubspec.lock'
   pull_request:
     branches: [ master ]
+  workflow_dispatch:
+
+# Both jobs target a self-hosted runner, by capability rather than by hostname.
+# See tool/setup_ci_runner.sh for how that runner is provisioned and why it is
+# not the Gitea host: Flutter cannot build an Android APK on a Linux arm64 host
+# (flutter/flutter#189724), and the Gitea box is aarch64. The runner is
+# registered at android-stuff org scope so any repo in the org can use it.
+#
+# The runner is host-mode, so Flutter, the JDK and the Android SDK/NDK are the
+# machine's own — the same toolchain that produced the verified release, and no
+# multi-hundred-MB download per run. Nothing here installs a toolchain; the
+# Toolchain step fails loudly if one is missing.
 
 jobs:
   analyze:
     name: Analyze & Test
-    runs-on: ubuntu-latest
+    runs-on: flutter-android
     steps:
       - uses: actions/checkout@v4
 
-      - uses: subosito/flutter-action@v2
+      - name: Toolchain
-        with:
+        run: |
-          flutter-version: '3.47.1'
+          flutter --version
-          channel: 'stable'
+          echo "ANDROID_HOME=${ANDROID_HOME:?not set on this runner}"
 
       - name: Get dependencies
         run: flutter pub get
@@ -24,5 +43,36 @@ jobs:
       - name: Analyze
         run: flutter analyze
 
+      # Includes test/architecture_test.dart, which fails the build if a layer
+      # boundary is crossed.
       - name: Run tests
         run: flutter test
+
+  release-apk:
+    name: Release APK
+    runs-on: flutter-android
+    needs: [analyze]
+    if: github.event_name == 'push' || github.event_name == 'workflow_dispatch'
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Get dependencies
+        run: flutter pub get
+
+      # The debug build does not minify, so it cannot catch an R8 problem. This
+      # job exists because a release APK that aborted on every test shipped
+      # once while the debug build passed every check.
+      - name: Build release APK
+        run: flutter build apk --release
+
+      # Building is not enough on its own: a renamed JNI callback still
+      # produces a successful build and a broken APK. This reads R8's mapping
+      # and the DEX string pool to confirm the names JNI resolves survived.
+      - name: Check JNI names survived minification
+        run: ./tool/check_jni_keep_rules.sh
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: app-release-apk
+          path: build/app/outputs/flutter-apk/app-release.apk
+          if-no-files-found: error

LICENSE

@@ -0,0 +1,26 @@
+MIT License
+
+Copyright (c) 2024–2026 Mikkeli Matlock
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+---
+
+Note: This project bundles iperf3 v3.19, which is licensed under the
+BSD-3-Clause license. See the bundled iperf3 source for its full license text.

README.md

@@ -40,9 +40,6 @@ domain, so shipping an iperf3 ELF in assets and running it via `Process.start()`
 fails with "Permission denied". iperf3 is therefore compiled as a shared library
 (`libiperf3jni.so`) and called in-process over JNI.
 
-`assets/iperf3/` still holds the pre-compiled binaries from that earlier
-approach. They are not declared in `pubspec.yaml`, so they are not packaged into
-the APK, and nothing in the app reads them.
 
 ## Building
 

android/app/build.gradle.kts

@@ -37,6 +37,13 @@ android {
             // TODO: Add your own signing config for the release build.
             // Signing with the debug keys for now, so `flutter run --release` works.
             signingConfig = signingConfigs.getByName("debug")
+
+            // proguard-rules.pro keeps the names JNI resolves at runtime.
+            // Without it a release build aborts on the first callback.
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro",
+            )
         }
     }
 

android/app/proguard-rules.pro

@@ -0,0 +1,21 @@
+# JNI reaches these by name. R8 renames by default, so without these rules a
+# release build resolves nothing: GetMethodID("onOutput", ...) returns NULL,
+# raises NoSuchMethodError, and the next JNI call aborts the process with
+#   NoSuchMethodError: no non-static method "Lp0;.onOutput(Ljava/lang/String;)V"
+# Debug builds do not minify, so this only ever appears in release.
+
+# The plugin object holds the native method declarations and is looked up from
+# C by its fully qualified name (Java_com_novoyuuparosk_iperf3gui_IperfRunner_*).
+-keep class com.novoyuuparosk.iperf3gui.IperfRunner { *; }
+
+# The callback interface and everything implementing it: iperf_jni.c resolves
+# onOutput/onError/onComplete by name on the concrete object's class, so the
+# implementors' method names must survive too, not just the interface's.
+-keep interface com.novoyuuparosk.iperf3gui.IperfRunner$IperfCallback { *; }
+-keep class * implements com.novoyuuparosk.iperf3gui.IperfRunner$IperfCallback { *; }
+
+# Belt and braces for any future native method: keep the name of any class
+# declaring one, and the native methods themselves.
+-keepclasseswithmembernames,includedescriptorclasses class * {
+    native <methods>;
+}

android/app/src/main/cpp/iperf_jni.c

@@ -86,6 +86,21 @@ struct iperf_run_state {
  * iperf_run_client() and cleared after it returns. */
 static struct iperf_run_state *g_current_state = NULL;
 
+/* Resolve a method ID, clearing the exception a failed lookup leaves behind.
+ *
+ * GetMethodID raises NoSuchMethodError on failure, and any further JNI call
+ * made while an exception is pending aborts the process. Returning NULL and
+ * clearing lets the caller fail on its own terms. */
+static jmethodID resolve_method(JNIEnv *env, jclass cls,
+                                const char *name, const char *sig) {
+    jmethodID mid = (*env)->GetMethodID(env, cls, name, sig);
+    if ((*env)->ExceptionCheck(env)) {
+        (*env)->ExceptionClear(env);
+        return NULL;
+    }
+    return mid;
+}
+
 /* Helper: invoke the output callback via CallVoidMethod */
 static void signal_output(JNIEnv *env, struct iperf_run_state *state,
                           const char *msg) {
@@ -272,13 +287,25 @@ Java_com_novoyuuparosk_iperf3gui_IperfRunner_runIperfLive(
 
     /* Resolve callback method IDs up front so failures can be reported. */
     jclass callback_cls = (*env)->GetObjectClass(env, callback);
-    jmethodID mid_output = (*env)->GetMethodID(
+    jmethodID mid_output = resolve_method(
             env, callback_cls, "onOutput", "(Ljava/lang/String;)V");
-    jmethodID mid_error = (*env)->GetMethodID(
+    jmethodID mid_error = resolve_method(
             env, callback_cls, "onError", "(Ljava/lang/String;)V");
-    jmethodID mid_complete = (*env)->GetMethodID(
+    jmethodID mid_complete = resolve_method(
             env, callback_cls, "onComplete", "()V");
 
+    /* If the names did not resolve there is no way to report anything: the
+     * reporting path is the thing that is missing. Bail out loudly instead.
+     *
+     * In practice this means R8 renamed the callback in a release build and
+     * android/app/proguard-rules.pro has stopped covering it. */
+    if (!mid_output || !mid_error || !mid_complete) {
+        LOGE("Callback methods did not resolve (onOutput=%p onError=%p "
+             "onComplete=%p) — check the -keep rules for IperfCallback",
+             mid_output, mid_error, mid_complete);
+        return;
+    }
+
     /* Report a startup failure directly on the callback object and finish,
      * so Dart's completer always resolves instead of waiting for a timeout. */
     #define FAIL_EARLY(msg) do {                                              \

docs/README.md

@@ -183,6 +183,62 @@ Dart uses `Completer<Iperf3Result>` to await native completion. The MethodChanne
 | `isTestRunning()` | `isTestRunning` | Check if test is in progress |
 | `setTempDir(String)` | `setTempDir` | Point `mkstemp()` at the app cache dir |
 
+## CI
+
+Both jobs run on a **self-hosted runner registered at `android-stuff` org scope**,
+targeted by the capability label `flutter-android` rather than by hostname.
+`tool/setup_ci_runner.sh` provisions it.
+
+It does not run on the Gitea host. Gitea lives on an aarch64 box, and Flutter
+cannot build an Android APK on a Linux arm64 host at all — the SDK ships no
+arm64-host `gen_snapshot` ([flutter#189724](https://github.com/flutter/flutter/issues/189724),
+[#75864](https://github.com/flutter/flutter/issues/75864), open as of July 2026).
+An x86_64 runner is a requirement, not a preference.
+
+The runner is **host mode**: jobs use the machine's own Flutter, JDK and Android
+SDK/NDK. CI therefore builds with the same toolchain that produced the verified
+release, and no NDK is downloaded per run. The cost is that there is no
+container boundary, which is what decides the next paragraph.
+
+Jobs run as a dedicated unprivileged account, `gitea-runner`, in the
+`flutterdev` group. In host mode the service account *is* the sandbox — there
+is no other — and a workflow file is code that runs as whoever the service runs
+as. A human account would hand every workflow that user's home directory and
+SSH keys; root would be worse and would also miss the warm caches in that home,
+re-downloading dependencies every run. The dedicated account keeps its own
+Gradle and pub caches under `/var/lib/gitea-runner`, so the first Android build
+is slow by design and later ones are not.
+
+`flutterdev` is what grants write access to `/opt/flutter` and
+`/opt/android-sdk`; both carry ACLs with `group:flutterdev:rwx` and matching
+`default:` entries, so new files inherit it. The setup script proves that access
+as the service account rather than assuming group membership is sufficient.
+
+It installs a **system** service, not a user service: root-managed lifecycle,
+starts at boot, no lingering session. That the service is root-managed is
+unrelated to the jobs being unprivileged — `User=` in the unit settles the
+second, and conflating the two is easy to do.
+
+`sudo ./tool/setup_ci_runner.sh --check` validates this machine's toolchain
+without creating anything.
+
+Two jobs:
+
+| Job | What it protects |
+|---|---|
+| `analyze` | `flutter analyze`, then `flutter test` — which includes `architecture_test.dart`, failing the build if a layer boundary is crossed |
+| `release-apk` | builds the **minified** APK and runs `tool/check_jni_keep_rules.sh` |
+
+The second exists because of a specific escape. A release APK that aborted on
+every test shipped once, while every on-device check passed — because those
+checks all used the debug build, which does not minify. Building the release APK
+is necessary but not sufficient: a renamed JNI callback still produces a
+*successful build* and a broken app. So the guard reads R8's `mapping.txt` for
+the callback classes and methods, and the DEX string pool for the native entry
+points. It is verified to fail: delete the keep rules from
+`android/app/proguard-rules.pro`, rebuild, and it reports `MainThreadCallback`,
+`onOutput`, `onError` and `onComplete` renamed.
+
 ## Building
 
 Prerequisites:

lib/domain/models/test_configuration.dart

@@ -7,7 +7,7 @@ class TestConfiguration {
   final int parallelStreams;
 
   const TestConfiguration({
-    this.server = 'iperf.es.net',
+    this.server = 'speedtest.uztelecom.uz',
     this.port = 5201,
     this.duration = const Duration(seconds: 10),
     this.interval = 1,

lib/main_headless.dart

@@ -25,7 +25,7 @@ import 'package:iperf3gui/domain/services/iperf3_runner.dart';
 
 /// Test parameters, supplied at build time with --dart-define.
 const String _server = String.fromEnvironment('IPERF_SERVER',
-    defaultValue: 'iperf.es.net');
+    defaultValue: 'speedtest.uztelecom.uz');
 const int _port = int.fromEnvironment('IPERF_PORT', defaultValue: 5201);
 const int _seconds = int.fromEnvironment('IPERF_SECONDS', defaultValue: 10);
 const int _streams = int.fromEnvironment('IPERF_STREAMS', defaultValue: 1);

lib/ui/features/test_run/test_run_screen.dart

@@ -110,7 +110,7 @@ class _TestRunContentState extends State<_TestRunContent> {
               autocorrect: false,
               decoration: const InputDecoration(
                 labelText: 'Server',
-                hintText: 'e.g. iperf.es.net or 192.168.1.1',
+                hintText: 'e.g. speedtest.uztelecom.uz or 192.168.1.1',
               ),
             ),
             const SizedBox(height: 12),

pubspec.yaml

@@ -2,7 +2,7 @@ name: iperf3gui
 description: "A Flutter-based Android GUI wrapper for iperf3."
 publish_to: 'none'
 
-version: 1.0.0+1
+version: 0.1.0+1
 
 environment:
   sdk: ^3.13.1

tool/check_jni_keep_rules.sh

@@ -0,0 +1,112 @@
+#!/usr/bin/env bash
+# Assert that R8 left the names JNI resolves at runtime alone.
+#
+# iperf_jni.c looks up the callback methods by name with GetMethodID. R8 renames
+# by default, and a renamed callback does not fail the build — it produces an
+# APK that aborts on the first callback of every test:
+#
+#   NoSuchMethodError: no non-static method "Lp0;.onOutput(Ljava/lang/String;)V"
+#
+# That shipped once. Building the release APK is not enough to catch it, because
+# the build succeeds; this reads R8's own mapping file and checks the names
+# survived. Run after `flutter build apk --release`.
+
+set -euo pipefail
+
+MAPPING="${1:-build/app/outputs/mapping/release/mapping.txt}"
+APK="${2:-build/app/outputs/flutter-apk/app-release.apk}"
+
+if [[ ! -f "$MAPPING" ]]; then
+    echo "FAIL: no R8 mapping at $MAPPING"
+    echo "      Run 'flutter build apk --release' first. If the release build"
+    echo "      no longer minifies, this check is meaningless — say so"
+    echo "      explicitly rather than deleting it."
+    exit 1
+fi
+
+failed=0
+
+# A class kept by name maps to itself: "a.b.C -> a.b.C:".
+require_class_kept() {
+    local fqcn="$1"
+    if grep -qF -- "$fqcn -> $fqcn:" "$MAPPING"; then
+        echo "  ok    class  $fqcn"
+    else
+        echo "  FAIL  class  $fqcn was renamed or removed"
+        grep -F -- "$fqcn ->" "$MAPPING" | head -3 | sed 's/^/          got: /' || true
+        failed=1
+    fi
+}
+
+# A method kept by name ends with "-> <name>".
+require_method_kept() {
+    local signature="$1" name="$2"
+    if grep -qE -- "$signature.*-> ${name}\$" "$MAPPING"; then
+        echo "  ok    method $name"
+    else
+        echo "  FAIL  method $name was renamed"
+        grep -E -- "$signature" "$MAPPING" | head -3 | sed 's/^/          got: /' || true
+        failed=1
+    fi
+}
+
+echo "Checking JNI-visible names in $MAPPING"
+
+require_class_kept 'com.novoyuuparosk.iperf3gui.IperfRunner'
+require_class_kept 'com.novoyuuparosk.iperf3gui.IperfRunner$IperfCallback'
+require_class_kept 'com.novoyuuparosk.iperf3gui.IperfRunner$MainThreadCallback'
+
+# Resolved by iperf_jni.c on the concrete callback object's class.
+require_method_kept 'void onOutput\(java\.lang\.String\)' 'onOutput'
+require_method_kept 'void onError\(java\.lang\.String\)' 'onError'
+require_method_kept 'void onComplete\(\)' 'onComplete'
+
+# Native methods are NOT checked against mapping.txt: R8 only emits entries for
+# members it has line-number info or a rename for, and a native method has no
+# body, so it is absent whether or not it was renamed. Absence would prove
+# nothing. Check the DEX string pool instead — a surviving name is interned
+# there, a renamed one is not.
+check_dex_names() {
+    local apk="$1"
+    shift
+    if [[ ! -f "$apk" ]]; then
+        echo "  skip  DEX name check: no APK at $apk"
+        return
+    fi
+    local tmp
+    tmp="$(mktemp -d)"
+    trap 'rm -rf "$tmp"' RETURN
+    if ! unzip -o -q "$apk" 'classes*.dex' -d "$tmp" 2>/dev/null; then
+        echo "  FAIL  could not extract classes*.dex from $apk"
+        failed=1
+        return
+    fi
+    # Dump the string pool once to a file, then grep the file. Never
+    # `... | strings | grep -q`: grep -q exits at the first match, SIGPIPEs
+    # strings, and `set -o pipefail` reports the whole pipeline as failed — so
+    # a name found early "fails" and one found late passes. That flakiness cost
+    # an afternoon; leave the intermediate file in place.
+    cat "$tmp"/classes*.dex | strings > "$tmp/dex-strings.txt"
+
+    local name
+    for name in "$@"; do
+        if grep -qxF -- "$name" "$tmp/dex-strings.txt"; then
+            echo "  ok    native $name"
+        else
+            echo "  FAIL  native $name is absent from the DEX string pool"
+            failed=1
+        fi
+    done
+}
+
+check_dex_names "${2:-build/app/outputs/flutter-apk/app-release.apk}" \
+    runIperfLive forceStopIperfTest setTempDir isTestRunning
+
+if [[ $failed -ne 0 ]]; then
+    echo
+    echo "R8 renamed something JNI resolves by name. The release APK will abort"
+    echo "on the first callback. Fix android/app/proguard-rules.pro."
+    exit 1
+fi
+
+echo "All JNI-visible names survived minification."

tool/setup_ci_runner.sh

@@ -0,0 +1,288 @@
+#!/usr/bin/env bash
+# Provision this machine as a Gitea Actions runner for the android-stuff org.
+#
+#   sudo ./tool/setup_ci_runner.sh --check      # pre-flight only, changes nothing
+#   sudo RUNNER_TOKEN=xxxx ./tool/setup_ci_runner.sh
+#
+# Get an ORG-scoped token from
+#   <gitea>/org/android-stuff/settings/actions/runners → "Create new runner"
+# Org scope, not repo scope, so every android-stuff repo shares one runner.
+#
+# ── WHY THIS MACHINE, NOT THE GITEA HOST ────────────────────────────────────
+# Gitea runs on an aarch64 box. Flutter cannot build an Android APK on a Linux
+# arm64 host at all: the SDK ships no arm64-host gen_snapshot
+# (flutter/flutter#189724, #75864, open as of July 2026). An x86_64 runner is a
+# requirement here, not a preference.
+#
+# ── WHY HOST MODE, NOT A CONTAINER ──────────────────────────────────────────
+# This machine has a working Flutter + Android SDK + NDK. In host mode jobs use
+# it directly, so CI builds with the same toolchain that produced the verified
+# release and skips a per-run NDK download. The cost is that there is no
+# container boundary — which is exactly why the service account below matters.
+#
+# ── WHY A DEDICATED USER, AND A SYSTEM SERVICE ──────────────────────────────
+# Host mode means the service account IS the sandbox; there is no other. A
+# workflow file is code, and it runs as whoever this service runs as. Running
+# as a human user would give every workflow read access to that user's home,
+# SSH keys included. Running as root would be worse, and would also miss the
+# warm Gradle/pub caches in that home (HOME=/root), re-downloading dependencies
+# every run for no benefit.
+#
+# So: a dedicated unprivileged account, in the flutterdev group for SDK access,
+# with its own HOME so its caches are its own. A *system* service rather than a
+# user service, so it starts at boot under plain systemctl with no lingering
+# session — the service being root-managed is unrelated to the jobs being
+# unprivileged, and conflating those two was the earlier mistake here.
+
+set -euo pipefail
+
+VERSION="${RUNNER_VERSION:-3.3.0}"
+INSTANCE="${GITEA_INSTANCE:-http://pi5-16.local:3005}"
+RUNNER_USER="${RUNNER_USER:-gitea-runner}"
+SDK_GROUP="${SDK_GROUP:-flutterdev}"
+RUNNER_HOME="${RUNNER_HOME:-/var/lib/gitea-runner}"
+BIN="${RUNNER_BIN:-/usr/local/bin/gitea-runner}"
+UNIT="/etc/systemd/system/gitea-runner.service"
+RUNNER_NAME="${RUNNER_NAME:-$(hostname)-android}"
+
+# Capability label, deliberately not a hostname: it says what the runner can
+# do, so a workflow does not encode which machine happens to provide it.
+LABEL="${RUNNER_LABEL:-flutter-android}"
+
+CHECK_ONLY=0
+[[ "${1:-}" == "--check" ]] && CHECK_ONLY=1
+
+[[ $EUID -eq 0 ]] || { echo "Run with sudo: sudo RUNNER_TOKEN=... $0" >&2; exit 1; }
+# --check is for validating this machine before committing to anything, so it
+# deliberately does not require a token.
+if [[ $CHECK_ONLY -eq 0 ]]; then
+    : "${RUNNER_TOKEN:?set RUNNER_TOKEN to an org registration token (see header)}"
+fi
+
+say() { printf '\n==> %s\n' "$1"; }
+
+# ── Pre-flight, before anything is created ─────────────────────────────────
+say "Pre-flight"
+missing=0
+for cmd in flutter dart java cmake ninja node git unzip curl; do
+    # Resolved from systemd's own PATH, not an interactive shell's, since that
+    # is what the service will actually have.
+    if PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin \
+        command -v "$cmd" >/dev/null; then
+        printf '  ok       %s\n' "$cmd"
+    else
+        printf '  MISSING  %s (not on systemd default PATH)\n' "$cmd"
+        missing=1
+    fi
+done
+
+SDK="${ANDROID_HOME:-${ANDROID_SDK_ROOT:-/opt/android-sdk}}"
+for d in "$SDK/ndk" "$SDK/build-tools" "$SDK/platforms"; do
+    [[ -d "$d" ]] && printf '  ok       %s\n' "$d" \
+        || { printf '  MISSING  %s\n' "$d"; missing=1; }
+done
+getent group "$SDK_GROUP" >/dev/null \
+    && printf '  ok       group %s\n' "$SDK_GROUP" \
+    || { printf '  MISSING  group %s\n' "$SDK_GROUP"; missing=1; }
+
+if [[ $missing -ne 0 ]]; then
+    echo
+    echo "Refusing to register: a runner advertising '$LABEL' must actually" >&2
+    echo "provide these, or every job fails confusingly instead of here." >&2
+    exit 1
+fi
+
+if [[ $CHECK_ONLY -eq 1 ]]; then
+    say "Pre-flight passed"
+    echo "  Nothing was created. Re-run with RUNNER_TOKEN set to provision."
+    echo "  Note: SDK write access is verified as $RUNNER_USER, which can only"
+    echo "  be checked once that account exists — so it is not covered here."
+    exit 0
+fi
+
+# ── Service account ────────────────────────────────────────────────────────
+say "Service account: $RUNNER_USER"
+if id "$RUNNER_USER" &>/dev/null; then
+    echo "  exists; ensuring membership of $SDK_GROUP"
+    usermod -aG "$SDK_GROUP" "$RUNNER_USER"
+else
+    # --shell /bin/bash, not nologin: host-mode steps are executed as shell
+    # scripts, and a login shell costs nothing here — the account has no
+    # password, so it is only reachable via root anyway.
+    useradd --system \
+            --create-home --home-dir "$RUNNER_HOME" \
+            --shell /bin/bash \
+            --groups "$SDK_GROUP" \
+            --comment "Gitea Actions runner ($LABEL)" \
+            "$RUNNER_USER"
+    echo "  created"
+fi
+install -d -o "$RUNNER_USER" -g "$SDK_GROUP" -m 0750 \
+    "$RUNNER_HOME" "$RUNNER_HOME/workspace"
+id "$RUNNER_USER"
+
+# Group membership is worthless if the ACLs disagree, so prove it rather than
+# assume it — as the service user, against the paths a build actually writes.
+#
+# By actually writing a file, not by asking `test -w`. This machine's
+# /usr/bin/test is uutils coreutils 0.8.0, whose -w does not credit
+# supplementary group membership: it reports /opt/android-sdk unwritable for
+# any account whose write access comes from the flutterdev group rather than
+# from the owner bits — which is every account that should be running builds
+# here. (-r and -x pass only because `other` has r-x.) bash's builtin `[ -w ]`
+# gets it right, but sudo execs the external binary, so the builtin is not
+# available here. A write probe is immune to the whole question.
+say "Verifying $RUNNER_USER can write the SDKs"
+write_ok=1
+for d in "/opt/flutter/bin/cache" "$SDK"; do
+    if sudo -u "$RUNNER_USER" env probe="$d/.gitea-runner-probe.$$" \
+        sh -c 'touch "$probe" && rm -f "$probe"' 2>/dev/null; then
+        printf '  ok       writable: %s\n' "$d"
+    else
+        printf '  FAIL     not writable: %s\n' "$d"
+        sudo -u "$RUNNER_USER" rm -f "$d/.gitea-runner-probe.$$" 2>/dev/null || true
+        write_ok=0
+    fi
+done
+if [[ $write_ok -ne 1 ]]; then
+    echo
+    echo "$RUNNER_USER is in $SDK_GROUP but still cannot write. Check:" >&2
+    echo "  getfacl /opt/flutter /opt/android-sdk" >&2
+    echo "  id $RUNNER_USER" >&2
+    exit 1
+fi
+
+# ── Binary ─────────────────────────────────────────────────────────────────
+say "Installing gitea-runner $VERSION"
+if [[ -x "$BIN" ]] && "$BIN" --version 2>/dev/null | grep -qF "$VERSION"; then
+    echo "  already at $VERSION"
+else
+    url="https://dl.gitea.com/gitea-runner/${VERSION}/gitea-runner-${VERSION}-linux-amd64"
+    echo "  $url"
+    curl -fSL --retry 3 -o "$BIN.tmp" "$url"
+    chmod 0755 "$BIN.tmp"
+    mv "$BIN.tmp" "$BIN"
+fi
+"$BIN" --version
+
+# ── Config ─────────────────────────────────────────────────────────────────
+say "Config"
+CONFIG="$RUNNER_HOME/config.yaml"
+if [[ -f "$CONFIG" && -f "$RUNNER_HOME/.runner" ]]; then
+    # Registered: this config is in use, do not touch it.
+    echo "  $CONFIG exists and the runner is registered — leaving it alone"
+else
+    if [[ -f "$CONFIG" ]]; then
+        echo "  $CONFIG exists but nothing is registered yet — rewriting it"
+    fi
+    cat > "$CONFIG" <<YAML
+log:
+  level: info
+
+runner:
+  # Absolute on purpose: a relative path resolves against whatever directory
+  # the command was invoked from, which for `register` is wherever the admin
+  # happened to be standing — typically a repo this account cannot write.
+  file: ${RUNNER_HOME}/.runner
+  # One at a time: a 2-ABI NDK compile plus Gradle will saturate this box, and
+  # concurrent jobs would only contend for it.
+  capacity: 1
+  timeout: 45m
+  # ':host' runs steps directly on this machine as ${RUNNER_USER}, using the
+  # toolchain verified in pre-flight, rather than in a container that would
+  # have to fetch its own.
+  labels:
+    - "${LABEL}:host"
+
+container:
+  # Unused in host mode. Kept so switching to container mode later is an edit
+  # rather than a rediscovery — it would need the Docker network that the
+  # Gitea host's own runner also needs.
+  network: ""
+  privileged: false
+
+host:
+  workdir_parent: ${RUNNER_HOME}/workspace
+YAML
+    chown "$RUNNER_USER:$SDK_GROUP" "$CONFIG"
+    chmod 0640 "$CONFIG"
+    echo "  wrote $CONFIG"
+fi
+
+# ── Registration ───────────────────────────────────────────────────────────
+say "Registering with $INSTANCE"
+if [[ -f "$RUNNER_HOME/.runner" ]]; then
+    echo "  already registered ($RUNNER_HOME/.runner) — skipping."
+    echo "  To re-register: delete that file AND remove the runner in Gitea."
+else
+    # Run from the runner's own directory: register writes .runner and
+    # .runner.lock relative to cwd unless told otherwise, and this account
+    # cannot write wherever the admin invoked the script from.
+    # --labels is deliberately absent: gitea-runner ignores it when the config
+    # file defines labels, and warns that it is doing so.
+    sudo -u "$RUNNER_USER" env HOME="$RUNNER_HOME" sh -c '
+        cd "$1" || exit 1
+        exec "$2" register --no-interactive \
+            --instance "$3" --token "$4" --name "$5" --config "$6"
+    ' sh "$RUNNER_HOME" "$BIN" "$INSTANCE" "$RUNNER_TOKEN" "$RUNNER_NAME" "$CONFIG"
+    # register writes .runner into the cwd it was given; make sure it landed.
+    [[ -f "$RUNNER_HOME/.runner" ]] || {
+        echo "  registration produced no .runner in $RUNNER_HOME" >&2
+        exit 1
+    }
+fi
+
+# ── Service ────────────────────────────────────────────────────────────────
+say "systemd unit"
+cat > "$UNIT" <<UNITFILE
+[Unit]
+Description=Gitea Actions runner (${LABEL}, host mode)
+Documentation=https://gitea.com/gitea/act_runner
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=${RUNNER_USER}
+Group=${SDK_GROUP}
+WorkingDirectory=${RUNNER_HOME}
+ExecStart=${BIN} daemon --config ${CONFIG}
+Restart=always
+RestartSec=5
+
+# HOME must be explicit: it decides where Gradle and pub put their caches, and
+# a service does not inherit a login environment.
+Environment=HOME=${RUNNER_HOME}
+Environment=GRADLE_USER_HOME=${RUNNER_HOME}/.gradle
+Environment=PUB_CACHE=${RUNNER_HOME}/.pub-cache
+Environment=ANDROID_HOME=${SDK}
+Environment=ANDROID_SDK_ROOT=${SDK}
+
+[Install]
+WantedBy=multi-user.target
+UNITFILE
+echo "  wrote $UNIT"
+
+systemctl daemon-reload
+systemctl enable --now gitea-runner
+sleep 2
+systemctl --no-pager --lines=15 status gitea-runner || true
+
+cat <<EOF
+
+==> Done.
+
+    Logs:    journalctl -u gitea-runner -f
+    Stop:    systemctl stop gitea-runner
+    Remove:  systemctl disable --now gitea-runner
+             rm $UNIT $BIN && rm -rf $RUNNER_HOME
+             userdel $RUNNER_USER
+             ...and delete the runner in the Gitea org settings.
+
+    The runner should now appear under
+      ${INSTANCE}/org/android-stuff/settings/actions/runners
+    with label '$LABEL'. The queued run for this repo should start on its own.
+
+    First Android build will be slow: this account's Gradle and pub caches
+    start empty by design. Subsequent runs reuse them.
+EOF