Profiler: cache disk benchmark for a week instead of rewriting 32 MB every launch

app/src/main/java/io/github/abrarul/modularai/model/DeviceProfiler.kt

@@ -51,7 +51,7 @@ object DeviceProfiler {
             availRamMb = mi.availMem / 1_048_576,
             freeStorageMb = stat.availableBytes / 1_048_576,
             vulkan = ctx.packageManager.hasSystemFeature(PackageManager.FEATURE_VULKAN_HARDWARE_LEVEL),
-            storageMBps = benchStorage(ctx.cacheDir),
+            storageMBps = cachedBench(ctx),
             tier = when {
                 total < 5_000 -> Tier.ULTRA_LIGHT
                 total < 10_000 -> Tier.BALANCED
@@ -60,6 +60,15 @@ object DeviceProfiler {
         )
     }
 
+    private fun cachedBench(ctx: Context): Double {
+        val p = ctx.getSharedPreferences("profiler", Context.MODE_PRIVATE)
+        val age = System.currentTimeMillis() - p.getLong("bench_at", 0)
+        if (age < 7L * 24 * 3600 * 1000 && p.contains("bench")) return p.getFloat("bench", 0f).toDouble()
+        val v = runCatching { benchStorage(ctx.cacheDir) }.getOrDefault(0.0)
+        p.edit().putFloat("bench", v.toFloat()).putLong("bench_at", System.currentTimeMillis()).apply()
+        return v
+    }
+
     /** 32 MB sequential write benchmark. */
     private fun benchStorage(dir: File): Double {
         val f = File(dir, "bench.tmp")
@@ -73,8 +82,7 @@ object DeviceProfiler {
 
     /**
      * Heuristic: a Q4_K_M GGUF needs roughly 0.6 GB per billion params plus KV cache.
-     * Uses ~60% of currently available RAM as the budget. Tokens/s is a rough estimate only;
-     * real numbers come from the in-app benchmark.
+     * Uses ~60% of currently available RAM as the budget. Tokens/s is a rough estimate only.
      */
     fun recommend(p: Profile): Recommendation {
         val budgetGb = p.availRamMb * 0.6 / 1024.0