feat(benchmark): per-k latency sweep infrastructure + GPU/CPU report

app/android/app/src/main/AndroidManifest.xml

@@ -4,6 +4,12 @@
     <uses-permission android:name="android.permission.FOREGROUND_SERVICE"/>
     <!-- Required on Android 14+ for network-data foreground services -->
     <uses-permission android:name="android.permission.FOREGROUND_SERVICE_DATA_SYNC"/>
+    <!-- BenchmarkForegroundService acquires a PARTIAL_WAKE_LOCK so the CPU
+         keeps running when the screen is off or locked. Without this,
+         multi-hour benchmarks stall silently when the device idles.
+         (BenchmarkActivity is now a thin launcher that just starts the
+         service; the wake lock lives in the service.) -->
+    <uses-permission android:name="android.permission.WAKE_LOCK"/>
 
     <application
         android:label="MAM-AI"
@@ -45,6 +51,26 @@
             android:foregroundServiceType="dataSync"
             android:exported="false" />
 
+        <!-- Foreground service for the on-device latency benchmark. Holds a
+             PARTIAL_WAKE_LOCK + sticky notification so the work survives
+             screen-off and device-lock through hours-long k-sweeps. Runs
+             in its own :benchmark process to keep the main app isolated.
+             DEV-ONLY: foregroundServiceType="dataSync" is technically a
+             misuse here (no actual data sync) — Google Play would reject
+             this declaration. Acceptable because BenchmarkForegroundService
+             is launched only via `adb shell am start` for in-house
+             benchmarking; it never appears in any user-facing flow and
+             this manifest entry should be stripped from any Play Store
+             build. If we ever need to ship benchmark capabilities, switch
+             to foregroundServiceType="specialUse" and add the corresponding
+             android.permission.FOREGROUND_SERVICE_SPECIAL_USE permission
+             plus the PROPERTY_SPECIAL_USE_FGS_SUBTYPE property. -->
+        <service
+            android:name=".BenchmarkForegroundService"
+            android:foregroundServiceType="dataSync"
+            android:exported="false"
+            android:process=":benchmark" />
+
         <!-- FileProvider for sharing PDF files from getExternalFilesDir with viewer apps -->
         <provider
             android:name="androidx.core.content.FileProvider"

app/android/app/src/main/kotlin/com/example/app/RagPipeline.kt

@@ -219,14 +219,20 @@ class RagPipeline(application: Application) {
         }
     }
 
-    /** Generates the response from the LLM with conversation history support. */
+    /** Generates the response from the LLM with conversation history support.
+     *
+     *  [retrieveKOverride] — when non-null, replaces `retrievalConfig.top_k`
+     *  for this call only. Used by [BenchmarkActivity] for the per-k latency
+     *  sweep; production callers leave it null and inherit the runtime config.
+     */
     suspend fun generateResponse(
         prompt: String,
         history: List<Map<String, String>>,
         useRetrieval: Boolean = true,
         language: String = "en",
         retrievalListener: (docs: List<RetrievedDoc>) -> Unit,
         generationListener: (partial: String, done: Boolean) -> Unit,
+        retrieveKOverride: Int? = null,
     ): String =
         coroutineScope {
             awaitLlmReady()
@@ -235,10 +241,11 @@ class RagPipeline(application: Application) {
             val qStart = System.currentTimeMillis()
 
             val docs = if (useRetrieval) {
+                val effectiveTopK = retrieveKOverride ?: retrievalConfig.getInt("top_k")
                 val retrievalRequest = RetrievalRequest.create(
                     prompt,
                     RetrievalConfig.create(
-                        retrievalConfig.getInt("top_k"),
+                        effectiveTopK,
                         retrievalConfig.getDouble("similarity_threshold").toFloat(),
                         TaskType.RETRIEVAL_QUERY,
                     ),