fix(runtime): start a queued game after leaving the current one

Emulated images keep a running total of their decoded and ETC2/EAC sizes over every mip and layer. --profile writes both as counters, and summarize_profile_trace.py prints their peaks.

CHANGELOG.md

@@ -19,9 +19,19 @@ Release notes on GitHub are taken from the matching section here.
   maintains it.
 - The log now names each Roblox Java call that Nightcap has no handler for, once per call, as
   `[JNI] no handler for <class>.<method><signature>`. This shows what a new Roblox version needs.
+- `--profile` traces now show how much GPU memory decoded ETC2 textures take, as the `emulated texture (MB)`
+  and `compressed texture (MB)` counters. `scripts/summarize_profile_trace.py` prints the peak of each.
+
+### Fixed
+
+- Joining a game from a website link while already in a game no longer freezes Nightcap. It leaves the
+  current game and joins the new one, and leaving that game returns to the home screen.
 
 ### Changed
 
+- The home screen and avatar editor have fewer long frames. PC profiles compile the app's menu scripts to
+  native code and turn this off while you are in a game, so games run as before. Set
+  `FFlagDebugLuauCodegenAll` in `fflags.json` to choose it yourself.
 - Nightcap answers Roblox's Java calls with less work per call, and the audio thread no longer allocates
   memory for each block of sound it plays.
 

CMakeLists.txt

@@ -279,6 +279,7 @@ add_library(mocktail_runtime STATIC
   src/runtime/game_session_coordinator.cc
   src/runtime/graphics_launch_policy.cc
   src/runtime/http_client_policy.cc
+  src/runtime/luau_codegen_switch.cc
   src/runtime/memory_limit.cc
   src/runtime/owned_pthread.cc
   src/runtime/performance_policy.cc

README.md

@@ -136,6 +136,11 @@ caused by `Color3` errors in the newer SDUI page. An explicit
 PC profiles also set `FIntFlyoutTestVariantRollout12` to `2`, which shows the
 desktop side panel menu. An explicit override takes precedence.
 
+PC profiles compile the app's menus to native code with
+`FFlagDebugLuauCodegenAll`, and turn it off while you are in a game. Setting
+it yourself turns the switching off and your value applies everywhere. See
+[docs/PERFORMANCE.md](docs/PERFORMANCE.md#native-code-in-the-menus).
+
 ```json
 {
   "FFlagExample": "True",

docs/BENCHMARKING.md

@@ -68,6 +68,12 @@ the CPU time of every thread together, so dividing it by the frame interval
 gives the number of cores busy. A long frame with little present-thread CPU
 was spent waiting.
 
+`emulated texture (MB)` is the VRAM held by the decoded images that stand in
+for live ETC2/EAC textures, over every mip level and layer, upscaled where
+small-texture upscale applies. `compressed texture (MB)` is the size the same
+textures would take as ETC2/EAC. Both change only when such an image is
+created or destroyed. `summarize_profile_trace.py` prints each one's peak.
+
 ## The benchmark runs
 
 Run every scenario before and after the change, on the same build type

docs/PERFORMANCE.md

@@ -129,14 +129,23 @@ runs cooler for a few frames per second less; measure before keeping it.
 
 ## Small texture upscaling
 
-`MOCKTAIL_SMALL_TEXTURE_UPSCALE` (default `4`) redraws ETC2 textures of 64px or
+`MOCKTAIL_SMALL_TEXTURE_UPSCALE` (default `4`) stores ETC2 textures of 64px or
 smaller at N times their width and height. At the default that is 16 times the
 pixels per affected texture.
 
+Upscaling alone adds no detail. Each original pixel becomes an N×N block of the
+same colour, and the GPU's filtering already blends a small texture that much
+when it draws it. The extra size only shows when a replacement PNG in
+`~/.config/mocktail/textures` has more detail than the original.
+
 The cost lands in two places. Decoding and resampling is CPU work on up to 8
 threads while textures load, which shows up as a burst rather than a steady
 load. The enlarged textures then stay in GPU memory for the session and cost
-extra sampling bandwidth on every frame that uses them.
+extra sampling bandwidth on every frame that uses them. In Blade Ball the
+memory share is small: decoded textures came to 5.1 times their ETC2 size both
+with the default and with upscaling off. The `emulated texture (MB)` and
+`compressed texture (MB)` counters in a `--profile` trace give the figures for
+any game.
 
 That decode does not run on the thread that submits the frame. On a device
 with a timeline semaphore, uploads decode on a background dispatcher and the
@@ -151,6 +160,37 @@ Set it to `1` to turn upscaling off:
 MOCKTAIL_SMALL_TEXTURE_UPSCALE=1 mocktail
 ```
 
+## Native code in the menus
+
+The home screen and avatar editor run Roblox's Lua app in the Luau interpreter,
+and its long frames come from that thread. PC profiles set
+`FFlagDebugLuauCodegenAll`, which compiles every script to native code. That
+makes the menus' Lua about 2.5 times cheaper for about 56 MB of compiled code.
+
+In a game the same flag compiles every game script too: Blade Ball grew from 7
+MB to about 300 MB and hit Roblox's native code limit, which shows warnings in
+F9. So Nightcap turns the flag off when a game starts joining and back on when
+you return to the menus. Roblox reads the flag each time it compiles a script,
+and a game's compiled code is freed when you leave it. Games run with Roblox's
+own per-script choice, the same as without the flag.
+
+The log shows each switch:
+
+```text
+[luau-codegen] menus-only switch at +0x<offset in libroblox.so>
+[luau-codegen] native code for all scripts off
+[luau-codegen] native code for all scripts on
+```
+
+Setting `FFlagDebugLuauCodegenAll` in `fflags.json` turns the switching off
+and your value applies everywhere. To keep the interpreter in the menus too:
+
+```json
+{
+  "FFlagDebugLuauCodegenAll": false
+}
+```
+
 ## Frame rate and present mode
 
 `graphics.frame_rate_limit` (default `-1`) leaves Roblox's own in-game

include/mocktail/graphics/vulkan_etc2_emulation.h

@@ -30,6 +30,18 @@ bool Etc2UploadByteCounts(EtcFormat format, const VkExtent3D& extent,
                           std::uint32_t layer_count, VkDeviceSize* compressed,
                           VkDeviceSize* decoded);
 
+// Sums every mip level and layer; `scale` multiplies the host image's
+// level-0 width and height only.
+bool Etc2ImageByteCounts(EtcFormat format, const VkExtent3D& extent,
+                         std::uint32_t mip_levels, std::uint32_t layers,
+                         std::uint32_t scale, VkDeviceSize* compressed,
+                         VkDeviceSize* decoded);
+
+struct EmulatedTextureBytes {
+  VkDeviceSize compressed = 0;
+  VkDeviceSize decoded = 0;
+};
+
 // Whether emulated ETC2 is reported to the application as a supported
 // feature, from the MOCKTAIL_ADVERTISE_ETC2 value: "0" hides it, anything
 // else (or unset) reports it. Hidden ETC2 still decodes ETC2 images the
@@ -152,6 +164,10 @@ class VulkanEtc2Emulation final {
   // pending, after waiting for any decode still writing into them.
   void ReleaseCommandBuffer(VkCommandBuffer command_buffer);
 
+  // Traced as the "emulated texture (MB)" and "compressed texture (MB)"
+  // counters.
+  EmulatedTextureBytes TextureBytes() const;
+
   struct State;
 
  private:

include/runtime/luau_codegen_switch.h

@@ -0,0 +1,68 @@
+#ifndef MOCKTAIL_RUNTIME_LUAU_CODEGEN_SWITCH_H_
+#define MOCKTAIL_RUNTIME_LUAU_CODEGEN_SWITCH_H_
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <mutex>
+#include <string_view>
+#include <vector>
+
+#include "runtime/roblox_experience_presence.h"
+
+namespace mocktail {
+namespace runtime {
+
+// Nightcap defaults this on and keeps it on only outside games; a user value
+// set in fflags.json or the overrides environment disables the switching.
+inline constexpr std::string_view kLuauCodegenAllClientSetting =
+    "FFlagDebugLuauCodegenAll";
+inline constexpr std::string_view kLuauCodegenAllFastFlagName =
+    "DebugLuauCodegenAll";
+
+struct LoadedImageSegment {
+  std::uintptr_t address = 0;
+  std::size_t size = 0;
+  bool writable = false;
+};
+
+std::vector<LoadedImageSegment> LoadedImageSegments(std::uintptr_t image_base);
+
+// Record layout: {u8 value; pad; const char* name; size_t length}, in a
+// writable segment filled by libroblox's static constructors. Returns null
+// unless exactly one record names a NUL-terminated copy of `name`.
+std::uint8_t* FindFastFlagValue(
+    const std::vector<LoadedImageSegment>& segments, std::string_view name);
+
+bool ClientSettingsSetLuauCodegenAll(std::string_view overrides_json);
+
+bool LuauCodegenAllForPhase(RobloxExperiencePresencePhase phase);
+
+// Locates the flag on the first notification; a byte that is not 0 or 1 is
+// left alone.
+class LuauCodegenMenuSwitch {
+ public:
+  using LocateFlagFn = std::function<std::uint8_t*()>;
+
+  LuauCodegenMenuSwitch(LocateFlagFn locate_flag,
+                        RobloxExperiencePresenceObserver next);
+
+  RobloxExperiencePresenceObserver observer();
+  void OnPhase(RobloxExperiencePresencePhase phase,
+               const RobloxExperienceLaunchRequest* request);
+
+ private:
+  static void Notify(void* context, RobloxExperiencePresencePhase phase,
+                     const RobloxExperienceLaunchRequest* request);
+
+  std::mutex mutex_;
+  LocateFlagFn locate_flag_;
+  bool located_ = false;
+  std::uint8_t* flag_ = nullptr;
+  const RobloxExperiencePresenceObserver next_;
+};
+
+}  // namespace runtime
+}  // namespace mocktail
+
+#endif  // MOCKTAIL_RUNTIME_LUAU_CODEGEN_SWITCH_H_

include/runtime/roblox_text_font_resolver.h

@@ -3,6 +3,7 @@
 
 #include <cstdint>
 #include <string>
+#include <vector>
 
 namespace mocktail {
 namespace runtime {
@@ -10,8 +11,10 @@ namespace runtime {
 // Android's RbxKeyboard resolves the Roblox Font enum through
 // assets/android/fonts/font-mappings.json. The selected file lives below
 // assets/content/fonts while legacy SourceSansPro faces live in assets/fonts.
+// Emoji come from the color fonts Roblox ships next to the mapped faces.
 struct RobloxTextFontSelection {
   std::string primary_file;
+  std::vector<std::string> emoji_files;
   float size_scale = 0.795F;
   bool exact_mapping = false;
 };

scripts/summarize_profile_trace.py

@@ -6,8 +6,9 @@
 Usage: summarize_profile_trace.py TRACE [TRACE ...]
 
 Each trace gets its own column, so `before.json after.json` compares them.
-Frame times come from the "frame interval (ms)" counter; every other row is
+Frame times come from the "frame interval (ms)" counter, peaks from the
-a slice name recorded by the Vulkan adapter.
+texture counters; every other row is a slice name recorded by the Vulkan
+adapter.
 """
 
 from __future__ import annotations
@@ -20,6 +21,7 @@ import sys
 from typing import Iterable
 
 FRAME_COUNTER = "frame interval (ms)"
+PEAK_COUNTERS = ("emulated texture (MB)", "compressed texture (MB)")
 
 
 @dataclass(frozen=True)
@@ -37,6 +39,7 @@ class Distribution:
 class Summary:
     frames: Distribution | None
     slices: dict[str, Distribution]
+    peaks: dict[str, float]
 
 
 def _nearest_rank(ordered: list[float], percent: float) -> float:
@@ -57,16 +60,21 @@ def _distribution(values: list[float]) -> Distribution:
 def summarize(events: Iterable[dict]) -> Summary:
     frames: list[float] = []
     slices: dict[str, list[float]] = {}
+    peaks: dict[str, float] = {}
     for event in events:
         phase = event.get("ph")
         if phase == "C" and event.get("name") == FRAME_COUNTER:
             frames.append(float(event["args"]["value"]))
+        elif phase == "C" and event.get("name") in PEAK_COUNTERS:
+            value = float(event["args"]["value"])
+            peaks[event["name"]] = max(peaks.get(event["name"], value), value)
         elif phase == "X":
             slices.setdefault(event["name"], []).append(event["dur"] / 1000)
     return Summary(
         frames=_distribution(frames) if frames else None,
         slices={name: _distribution(values)
-                for name, values in slices.items()})
+                for name, values in slices.items()},
+        peaks=peaks)
 
 
 def _row(label: str, cells: list[str]) -> str:
@@ -96,6 +104,11 @@ def to_markdown(traces: list[tuple[str, Summary]]) -> str:
     if any(frames):
         lines += [_metric_row(f"frame interval {field} (ms)", frames, field)
                   for field in ("mean", "p50", "p95", "p99", "max")]
+    for name in PEAK_COUNTERS:
+        if any(name in summary.peaks for summary in summaries):
+            lines.append(_row(f"{name} peak", [
+                f"{summary.peaks[name]:.2f}" if name in summary.peaks else "-"
+                for summary in summaries]))
 
     # Largest total time first, taken from whichever trace spent the most.
     totals: dict[str, float] = {}

src/graphics/vulkan_etc2_emulation.cc

@@ -92,6 +92,34 @@ bool Etc2UploadByteCounts(EtcFormat format, const VkExtent3D& extent,
   return true;
 }
 
+bool Etc2ImageByteCounts(EtcFormat format, const VkExtent3D& extent,
+                         std::uint32_t mip_levels, std::uint32_t layers,
+                         std::uint32_t scale, VkDeviceSize* compressed,
+                         VkDeviceSize* decoded) {
+  if (extent.width == 0 || extent.height == 0 || extent.depth == 0 ||
+      mip_levels == 0 || layers == 0 || scale == 0 || compressed == nullptr ||
+      decoded == nullptr) {
+    return false;
+  }
+  const auto level = [](std::uint32_t size, std::uint32_t mip) {
+    return std::max<VkDeviceSize>(1, static_cast<VkDeviceSize>(size) >> mip);
+  };
+  VkDeviceSize compressed_total = 0;
+  VkDeviceSize decoded_total = 0;
+  for (std::uint32_t mip = 0; mip < mip_levels && mip < 32; ++mip) {
+    const VkDeviceSize depth = level(extent.depth, mip);
+    compressed_total += ((level(extent.width, mip) + 3) / 4) *
+                        ((level(extent.height, mip) + 3) / 4) * depth *
+                        EtcBlockBytes(format);
+    decoded_total += level(extent.width * scale, mip) *
+                     level(extent.height * scale, mip) * depth *
+                     EtcDecodedTexelBytes(format);
+  }
+  *compressed = compressed_total * layers;
+  *decoded = decoded_total * layers;
+  return true;
+}
+
 namespace {
 
 bool EmulationDisabledByEnvironment() {
@@ -207,8 +235,21 @@ struct ImageRecord {
   // Decoded level-0 texels of a scaled image; every host mip is resampled
   // from them so the texture keeps its detail at every draw size.
   std::shared_ptr<const RgbaImage> base;
+  EmulatedTextureBytes bytes;
 };
 
+// `now` is read under the lock that produced `totals`, so concurrent changes
+// reach the trace in the order they happened.
+void TraceTextureBytes(const EmulatedTextureBytes& totals, std::uint64_t now) {
+  ChromeTraceWriter* const trace = ActiveProfileTrace();
+  if (trace == nullptr) {
+    return;
+  }
+  constexpr double kMegabyte = 1024.0 * 1024.0;
+  trace->Counter("emulated texture (MB)", now, totals.decoded / kMegabyte);
+  trace->Counter("compressed texture (MB)", now, totals.compressed / kMegabyte);
+}
+
 bool IsColorFormat(EtcFormat format) {
   return format == EtcFormat::kEtc2Rgb8 || format == EtcFormat::kEtc2Rgb8A1 ||
          format == EtcFormat::kEtc2Rgba8;
@@ -669,6 +710,8 @@ struct VulkanEtc2Emulation::State {
   std::unordered_map<VkPhysicalDevice, bool> physical_devices;
   std::vector<std::unique_ptr<HostDevice>> devices;
   std::unordered_map<VkImage, ImageRecord> images;
+  // Sum over `images`; guarded by `mutex`.
+  EmulatedTextureBytes texture_bytes;
   std::unordered_map<VkBuffer, BufferBinding> buffers;
   std::unordered_map<VkDeviceMemory, Mapping> mappings;
   std::unordered_map<VkCommandBuffer, CommandRecord> commands;
@@ -1245,6 +1288,8 @@ void VulkanEtc2Emulation::DestroyDevice(VkDevice device) {
   // Nothing may free staging or the semaphore while a decode still writes it.
   state_->DrainDecodes();
   std::vector<std::unique_ptr<StagingBlock>> doomed;
+  EmulatedTextureBytes totals;
+  std::uint64_t now = 0;
   {
     std::lock_guard<std::mutex> lock(state_->mutex);
     for (auto it = state_->commands.begin(); it != state_->commands.end();) {
@@ -1265,11 +1310,20 @@ void VulkanEtc2Emulation::DestroyDevice(VkDevice device) {
       }
     }
     for (auto it = state_->images.begin(); it != state_->images.end();) {
-      it = it->second.device == device ? state_->images.erase(it) : std::next(it);
+      if (it->second.device == device) {
+        state_->texture_bytes.compressed -= it->second.bytes.compressed;
+        state_->texture_bytes.decoded -= it->second.bytes.decoded;
+        it = state_->images.erase(it);
+      } else {
+        ++it;
+      }
     }
+    totals = state_->texture_bytes;
+    now = TraceClockNanos();
     state_->has_commands.store(!state_->commands.empty(),
                                std::memory_order_release);
   }
+  TraceTextureBytes(totals, now);
   if (const HostDevice* dev = state_->Find(device); dev != nullptr) {
     for (const auto& block : doomed) {
       DestroyStagingBlock(*dev, *block);
@@ -1316,6 +1370,14 @@ VkResult VulkanEtc2Emulation::CreateImage(VkDevice device,
   }
   const VkResult result = dev->create_image(device, &host_info, allocator, image);
   if (result == VK_SUCCESS && image != nullptr) {
+    EmulatedTextureBytes bytes;
+    if (!Etc2ImageByteCounts(format.etc_format, create_info->extent,
+                             create_info->mipLevels, create_info->arrayLayers,
+                             scale, &bytes.compressed, &bytes.decoded)) {
+      bytes = {};
+    }
+    EmulatedTextureBytes totals;
+    std::uint64_t now = 0;
     {
       std::lock_guard<std::mutex> lock(state_->mutex);
       state_->images[*image] = {device,
@@ -1324,8 +1386,14 @@ VkResult VulkanEtc2Emulation::CreateImage(VkDevice device,
                                 create_info->extent.width,
                                 create_info->extent.height,
                                 nullptr,
-                                nullptr};
+                                nullptr,
+                                bytes};
+      state_->texture_bytes.compressed += bytes.compressed;
+      state_->texture_bytes.decoded += bytes.decoded;
+      totals = state_->texture_bytes;
+      now = TraceClockNanos();
     }
+    TraceTextureBytes(totals, now);
     if (ShouldLog(&g_image_logs, 4)) {
       std::fprintf(stderr,
                    "  [vulkan] ETC2 image %ux%u mips=%u format=%d -> host "
@@ -1346,12 +1414,33 @@ void VulkanEtc2Emulation::DestroyImage(VkDevice device, VkImage image,
     return;
   }
   if (dev->emulated) {
-    std::lock_guard<std::mutex> lock(state_->mutex);
+    EmulatedTextureBytes totals;
-    state_->images.erase(image);
+    std::uint64_t now = 0;
+    bool changed = false;
+    {
+      std::lock_guard<std::mutex> lock(state_->mutex);
+      const auto it = state_->images.find(image);
+      if (it != state_->images.end()) {
+        state_->texture_bytes.compressed -= it->second.bytes.compressed;
+        state_->texture_bytes.decoded -= it->second.bytes.decoded;
+        state_->images.erase(it);
+        totals = state_->texture_bytes;
+        now = TraceClockNanos();
+        changed = true;
+      }
+    }
+    if (changed) {
+      TraceTextureBytes(totals, now);
+    }
   }
   dev->destroy_image(device, image, allocator);
 }
 
+EmulatedTextureBytes VulkanEtc2Emulation::TextureBytes() const {
+  std::lock_guard<std::mutex> lock(state_->mutex);
+  return state_->texture_bytes;
+}
+
 VkResult VulkanEtc2Emulation::CreateImageView(
     VkDevice device, const VkImageViewCreateInfo* create_info,
     const VkAllocationCallbacks* allocator, VkImageView* view) {

src/legacy/legacy_runtime.cc

@@ -66,6 +66,7 @@
 #include "runtime/display_size.h"
 #include "runtime/environment.h"
 #include "runtime/discord_rpc.h"
+#include "runtime/luau_codegen_switch.h"
 #include "runtime/jnivm_platform_web_callbacks.h"
 #include "runtime/owned_pthread.h"
 #include "runtime/platform_cache_migration.h"
@@ -4656,6 +4657,27 @@ int mocktail::legacy::Run(const runtime::CommandLineOptions& options,
                 << "; continuing without Rich Presence\n";
     }
   }
+  mocktail::runtime::LuauCodegenMenuSwitch luau_codegen_switch(
+      [&process_environment]() -> std::uint8_t* {
+        if (process_environment.GetOr("MOCKTAIL_LUAU_CODEGEN_MENU_SWITCH",
+                                      "0") != "1") {
+          return nullptr;
+        }
+        const std::uintptr_t base = g_libroblox_base;
+        std::uint8_t* flag = mocktail::runtime::FindFastFlagValue(
+            mocktail::runtime::LoadedImageSegments(base),
+            mocktail::runtime::kLuauCodegenAllFastFlagName);
+        if (flag == nullptr) {
+          std::cerr << "  [luau-codegen] DebugLuauCodegenAll not found; "
+                       "leaving it on in games\n";
+        } else {
+          std::cout << "  [luau-codegen] menus-only switch at +0x" << std::hex
+                    << (reinterpret_cast<std::uintptr_t>(flag) - base)
+                    << std::dec << '\n';
+        }
+        return flag;
+      },
+      discord_rpc.observer());
   const mocktail::runtime::InputCapabilityConfig& input_capabilities =
       runtime_config.input_capabilities();
   if (!runtime_config.frame_rate().valid()) {
@@ -6834,7 +6856,7 @@ int mocktail::legacy::Run(const runtime::CommandLineOptions& options,
               jni_factory, present_boundary, std::move(surface_config),
               &dependencies.roblox_credential(),
               mocktail::runtime::RobloxExperienceSurfaceProvider{},
-              discord_rpc.observer(),
+              luau_codegen_switch.observer(),
               dependencies.clear_persisted_web_view_cookie(),
               runtime_config.microphone_enabled(),
               platform_web_symbols.account);

src/main.cc

@@ -33,6 +33,7 @@
 #include "runtime/fleasion.h"
 #include "runtime/game_mode.h"
 #include "runtime/graphics_launch_policy.h"
+#include "runtime/luau_codegen_switch.h"
 #include "runtime/memory_limit.h"
 #include "runtime/payload_update_preflight.h"
 #include "runtime/performance_policy.h"
@@ -1012,6 +1013,7 @@ int main(int argc, char* argv[]) {
       std::cout
           << "  [auth] explicit guest identity selected for production VM\n";
     }
+    unsetenv("MOCKTAIL_LUAU_CODEGEN_MENU_SWITCH");
     if (runtime_config.config.desktop_playability()) {
       std::filesystem::path assets = paths.DefaultAssetPath();
       if (environment.HasNonEmpty("MOCKTAIL_ASSET_PATH")) {
@@ -1034,10 +1036,13 @@ int main(int argc, char* argv[]) {
         return EXIT_FAILURE;
       }
       std::string desktop_client_settings;
+      const std::string user_client_settings =
+          environment.GetOr("MOCKTAIL_CLIENT_SETTINGS_OVERRIDES_JSON", "{}");
+      const bool user_set_luau_codegen =
+          mocktail::runtime::ClientSettingsSetLuauCodegenAll(
+              user_client_settings);
       if (!mocktail::runtime::MergeDesktopAppPolicyClientSettingsOverride(
-              desktop_policy.policy_json,
+              desktop_policy.policy_json, user_client_settings,
-              environment.GetOr("MOCKTAIL_CLIENT_SETTINGS_OVERRIDES_JSON",
-                                "{}"),
               &desktop_client_settings, &command_line_error)) {
         std::cerr << "[FATAL] Cannot activate desktop Roblox app-policy: "
                   << command_line_error << '\n';
@@ -1048,6 +1053,8 @@ int main(int argc, char* argv[]) {
         std::cerr << "[FATAL] Cannot export desktop Roblox app-policy\n";
         return EXIT_FAILURE;
       }
+      setenv("MOCKTAIL_LUAU_CODEGEN_MENU_SWITCH",
+             user_set_luau_codegen ? "0" : "1", 1);
       std::cout << "  [runtime] desktop app-policy ready: normalized="
                 << desktop_policy.normalized_policy_count
                 << " updated=" << (desktop_policy.updated ? 1 : 0)

src/runtime/luau_codegen_switch.cc

@@ -0,0 +1,150 @@
+#include "runtime/luau_codegen_switch.h"
+
+#include <elf.h>
+
+#include <algorithm>
+#include <cstring>
+#include <iostream>
+#include <string>
+#include <utility>
+
+#include <nlohmann/json.hpp>
+
+namespace mocktail {
+namespace runtime {
+namespace {
+
+constexpr std::size_t kNameSlotOffset = 8;
+constexpr std::size_t kRecordTailSize = 16;
+
+std::string_view SegmentBytes(const LoadedImageSegment& segment) {
+  return {reinterpret_cast<const char*>(segment.address), segment.size};
+}
+
+std::vector<std::uintptr_t> NameAddresses(
+    const std::vector<LoadedImageSegment>& segments, std::string_view name) {
+  std::string needle(name);
+  needle.push_back('\0');
+  std::vector<std::uintptr_t> addresses;
+  for (const LoadedImageSegment& segment : segments) {
+    const std::string_view bytes = SegmentBytes(segment);
+    for (std::size_t at = bytes.find(needle); at != std::string_view::npos;
+         at = bytes.find(needle, at + 1)) {
+      addresses.push_back(segment.address + at);
+    }
+  }
+  return addresses;
+}
+
+}  // namespace
+
+std::vector<LoadedImageSegment> LoadedImageSegments(std::uintptr_t image_base) {
+  std::vector<LoadedImageSegment> segments;
+  if (image_base == 0) {
+    return segments;
+  }
+  const auto* header = reinterpret_cast<const Elf64_Ehdr*>(image_base);
+  if (std::memcmp(header->e_ident, ELFMAG, SELFMAG) != 0 ||
+      header->e_ident[EI_CLASS] != ELFCLASS64 || header->e_phoff == 0) {
+    return segments;
+  }
+  const auto* program = reinterpret_cast<const Elf64_Phdr*>(
+      image_base + static_cast<std::uintptr_t>(header->e_phoff));
+  for (std::size_t index = 0; index < header->e_phnum; ++index) {
+    if (program[index].p_type != PT_LOAD || program[index].p_memsz == 0) {
+      continue;
+    }
+    segments.push_back({image_base + program[index].p_vaddr,
+                        static_cast<std::size_t>(program[index].p_memsz),
+                        (program[index].p_flags & PF_W) != 0});
+  }
+  return segments;
+}
+
+std::uint8_t* FindFastFlagValue(const std::vector<LoadedImageSegment>& segments,
+                                std::string_view name) {
+  const std::vector<std::uintptr_t> names = NameAddresses(segments, name);
+  if (names.empty()) {
+    return nullptr;
+  }
+  std::uint8_t* found = nullptr;
+  int matches = 0;
+  for (const LoadedImageSegment& segment : segments) {
+    if (!segment.writable || segment.size < kNameSlotOffset + kRecordTailSize) {
+      continue;
+    }
+    const std::uintptr_t first =
+        (segment.address + kNameSlotOffset + 7) & ~std::uintptr_t{7};
+    const std::uintptr_t end = segment.address + segment.size;
+    for (std::uintptr_t slot = first; slot + kRecordTailSize <= end;
+         slot += 8) {
+      std::uintptr_t pointer = 0;
+      std::memcpy(&pointer, reinterpret_cast<const void*>(slot),
+                  sizeof(pointer));
+      if (std::find(names.begin(), names.end(), pointer) == names.end()) {
+        continue;
+      }
+      std::uint64_t length = 0;
+      std::memcpy(&length, reinterpret_cast<const void*>(slot + 8),
+                  sizeof(length));
+      if (length != name.size()) {
+        continue;
+      }
+      found = reinterpret_cast<std::uint8_t*>(slot - kNameSlotOffset);
+      ++matches;
+    }
+  }
+  return matches == 1 ? found : nullptr;
+}
+
+bool ClientSettingsSetLuauCodegenAll(std::string_view overrides_json) {
+  const nlohmann::json overrides =
+      nlohmann::json::parse(overrides_json, nullptr, false, true);
+  return overrides.is_object() &&
+         overrides.contains(std::string(kLuauCodegenAllClientSetting));
+}
+
+bool LuauCodegenAllForPhase(RobloxExperiencePresencePhase phase) {
+  return phase == RobloxExperiencePresencePhase::kBrowsing;
+}
+
+LuauCodegenMenuSwitch::LuauCodegenMenuSwitch(
+    LocateFlagFn locate_flag, RobloxExperiencePresenceObserver next)
+    : locate_flag_(std::move(locate_flag)), next_(next) {}
+
+RobloxExperiencePresenceObserver LuauCodegenMenuSwitch::observer() {
+  return {this, &LuauCodegenMenuSwitch::Notify};
+}
+
+void LuauCodegenMenuSwitch::OnPhase(
+    RobloxExperiencePresencePhase phase,
+    const RobloxExperienceLaunchRequest* request) {
+  {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!located_) {
+      located_ = true;
+      flag_ = locate_flag_ ? locate_flag_() : nullptr;
+    }
+    if (flag_ != nullptr) {
+      const std::uint8_t current = __atomic_load_n(flag_, __ATOMIC_RELAXED);
+      const std::uint8_t wanted = LuauCodegenAllForPhase(phase) ? 1 : 0;
+      if (current <= 1 && current != wanted) {
+        __atomic_store_n(flag_, wanted, __ATOMIC_RELAXED);
+        std::cout << "  [luau-codegen] native code for all scripts "
+                  << (wanted != 0 ? "on" : "off") << '\n';
+      }
+    }
+  }
+  if (next_.valid()) {
+    next_.notify(next_.context, phase, request);
+  }
+}
+
+void LuauCodegenMenuSwitch::Notify(
+    void* context, RobloxExperiencePresencePhase phase,
+    const RobloxExperienceLaunchRequest* request) {
+  static_cast<LuauCodegenMenuSwitch*>(context)->OnPhase(phase, request);
+}
+
+}  // namespace runtime
+}  // namespace mocktail

src/runtime/roblox_desktop_app_policy.cc

@@ -14,6 +14,7 @@
 #include <string_view>
 #include <utility>
 
+#include "runtime/luau_codegen_switch.h"
 #include "runtime/runtime_paths.h"
 
 namespace mocktail {
@@ -357,6 +358,9 @@ bool MergeDesktopAppPolicyClientSettingsOverride(std::string_view policy_json,
     return false;
   }
   overrides[std::string(kDesktopAppPolicyOverride)] = policy_json;
+  // Native Luau code makes the app's menus faster. LuauCodegenMenuSwitch turns
+  // it off in games, which keep Roblox's own per-script choice.
+  overrides.emplace(std::string(kLuauCodegenAllClientSetting), "True");
   // The SDUI Charts page in 2.736/2.738 can pass a table to TextColor3 and
   // leave the whole LuaApp blank. Default to the working legacy page while
   // preserving explicit user overrides for testing newer UI patches.

src/runtime/roblox_experience_composition.cc

@@ -1446,8 +1446,9 @@ Status RobloxExperienceComposition::DrainLaunchRequests() {
       return Status::Ok();
     if (!launch_in_progress_ && active_launch_ == nullptr && game_active_ &&
         !pending_launch_requests_.empty()) {
-      // Start the queued launch only after consuming this leave callback, so
+      // Blocks a reentrant drain while the leave pumps main-thread messages.
-      // it cannot reach the next controller.
+      // The leave ends in the Native stage with no Lua app return, so the
+      // queued launch starts on the next drain.
       controlled_switch_waiting_for_return_ = true;
       switch_active_game = true;
     }
@@ -1456,7 +1457,7 @@ Status RobloxExperienceComposition::DrainLaunchRequests() {
     std::fprintf(stderr,
                  "[experience] leaving active game for queued launch\n");
     const Status leave_status = LeaveGame();
-    if (!leave_status.ok()) {
+    {
       std::lock_guard<std::mutex> lock(mutex_);
       controlled_switch_waiting_for_return_ = false;
     }

src/runtime/roblox_text_font_resolver.cc

@@ -10,6 +10,7 @@
 #include <string>
 #include <unordered_map>
 #include <utility>
+#include <vector>
 
 namespace mocktail {
 namespace runtime {
@@ -101,6 +102,24 @@ std::string LegacyFontName(int32_t font) {
   return "SourceSansPro-Regular.ttf";
 }
 
+// TwemojiMozilla covers Unicode emoji; RobloxEmoji holds the private-use
+// Roblox and Robux symbols.
+constexpr const char* kEmojiFontNames[] = {"TwemojiMozilla.ttf",
+                                           "RobloxEmoji.ttf"};
+
+std::vector<std::string> EmojiFontFiles(
+    const std::filesystem::path& asset_content_root) {
+  std::vector<std::string> files;
+  for (const char* name : kEmojiFontNames) {
+    const std::filesystem::path candidate =
+        asset_content_root / "fonts" / name;
+    if (IsRegularFile(candidate)) {
+      files.push_back(candidate.lexically_normal().string());
+    }
+  }
+  return files;
+}
+
 }  // namespace
 
 RobloxTextFontSelection ResolveRobloxTextFont(
@@ -111,6 +130,7 @@ RobloxTextFontSelection ResolveRobloxTextFont(
   }
 
   const std::filesystem::path content(asset_content_root);
+  selection.emoji_files = EmojiFontFiles(content);
   const std::optional<FontMapping> mapped = MappingFor(font, content);
   if (mapped.has_value()) {
     const std::filesystem::path candidate = content / "fonts" / mapped->file;

src/runtime/roblox_text_surface_overlay.cc

@@ -26,7 +26,7 @@ namespace {
 
 constexpr std::size_t kMaximumRasterTextBytes = 4096;
 constexpr std::uint64_t kMaximumOverlayPixels = 16ULL * 1024ULL * 1024ULL;
-constexpr std::size_t kMaximumFallbackFonts = 6;
+constexpr std::size_t kMaximumFallbackFonts = 8;
 
 std::mutex g_active_overlay_mutex;
 RobloxTextSurfaceOverlay* g_active_overlay = nullptr;
@@ -121,10 +121,35 @@ std::vector<std::string> ResolveFontFiles(
   } else if (!selection.primary_file.empty()) {
     files.push_back(selection.primary_file);
   }
+  files.insert(files.end(), selection.emoji_files.begin(),
+               selection.emoji_files.end());
   if (!FcInit()) {
     return files;
   }
 
+  // The sorted sans list rarely reaches an emoji font before the cap, so ask
+  // for the system emoji font by name ahead of it.
+  FcPattern* emoji = FcNameParse(reinterpret_cast<const FcChar8*>("emoji"));
+  if (emoji != nullptr) {
+    FcConfigSubstitute(nullptr, emoji, FcMatchPattern);
+    FcDefaultSubstitute(emoji);
+    FcResult emoji_result = FcResultNoMatch;
+    FcPattern* matched = FcFontMatch(nullptr, emoji, &emoji_result);
+    FcPatternDestroy(emoji);
+    FcChar8* file = nullptr;
+    if (matched != nullptr &&
+        FcPatternGetString(matched, FC_FILE, 0, &file) == FcResultMatch &&
+        file != nullptr && file[0] != '\0') {
+      const std::string candidate(reinterpret_cast<const char*>(file));
+      if (std::find(files.begin(), files.end(), candidate) == files.end()) {
+        files.push_back(candidate);
+      }
+    }
+    if (matched != nullptr) {
+      FcPatternDestroy(matched);
+    }
+  }
+
   FcPattern* pattern = FcNameParse(reinterpret_cast<const FcChar8*>("sans"));
   if (pattern == nullptr) {
     return files;

tests/CMakeLists.txt

@@ -279,6 +279,13 @@ target_link_libraries(roblox_desktop_app_policy_test PRIVATE
 )
 mocktail_apply_compile_options(roblox_desktop_app_policy_test)
 
+add_executable(luau_codegen_switch_test luau_codegen_switch_test.cc)
+target_link_libraries(luau_codegen_switch_test PRIVATE
+  Mocktail::Runtime
+  GTest::gtest_main
+)
+mocktail_apply_compile_options(luau_codegen_switch_test)
+
 add_executable(browser_tracker_service_test browser_tracker_service_test.cc)
 target_link_libraries(browser_tracker_service_test PRIVATE
   Mocktail::Services
@@ -1143,6 +1150,7 @@ if(MOCKTAIL_ENABLE_UPSTREAM_JNIVM)
 endif()
 gtest_discover_tests(libc_shim_test)
 gtest_discover_tests(discord_rpc_test)
+gtest_discover_tests(luau_codegen_switch_test)
 gtest_discover_tests(asset_manager_test)
 gtest_discover_tests(android_looper_test)
 gtest_discover_tests(android_native_window_test)

tests/luau_codegen_switch_test.cc

@@ -0,0 +1,162 @@
+#include "runtime/luau_codegen_switch.h"
+
+#include <elf.h>
+#include <gtest/gtest.h>
+
+#include <cstdint>
+#include <cstring>
+#include <string>
+#include <vector>
+
+namespace mocktail::runtime {
+namespace {
+
+constexpr char kStrings[] = "\0FFlagOther\0DebugLuauCodegenAll\0tail";
+
+struct alignas(8) FlagRecord {
+  std::uint8_t value = 1;
+  std::uint8_t padding[7] = {};
+  const char* name = nullptr;
+  std::uint64_t name_length = 0;
+  std::uint64_t next = 0;
+};
+
+const char* NameInStrings() {
+  return kStrings + 12;
+}
+
+std::vector<LoadedImageSegment> Segments(std::vector<FlagRecord>& records) {
+  return {
+      {reinterpret_cast<std::uintptr_t>(kStrings), sizeof(kStrings), false},
+      {reinterpret_cast<std::uintptr_t>(records.data()),
+       records.size() * sizeof(FlagRecord), true},
+  };
+}
+
+TEST(LuauCodegenSwitchTest, FindsTheFlagRecordByName) {
+  ASSERT_EQ(std::string(NameInStrings()), "DebugLuauCodegenAll");
+  std::vector<FlagRecord> records(3);
+  records[1].name = NameInStrings();
+  records[1].name_length = 19;
+  const auto segments = Segments(records);
+  EXPECT_EQ(FindFastFlagValue(segments, kLuauCodegenAllFastFlagName),
+            &records[1].value);
+}
+
+TEST(LuauCodegenSwitchTest, RejectsWrongLengthMissingOrDuplicateRecords) {
+  std::vector<FlagRecord> records(3);
+  records[1].name = NameInStrings();
+  records[1].name_length = 18;
+  auto segments = Segments(records);
+  EXPECT_EQ(FindFastFlagValue(segments, kLuauCodegenAllFastFlagName), nullptr);
+
+  records[1].name_length = 19;
+  EXPECT_EQ(FindFastFlagValue(segments, "DebugLuauCodegen"), nullptr);
+
+  records[2].name = NameInStrings();
+  records[2].name_length = 19;
+  EXPECT_EQ(FindFastFlagValue(segments, kLuauCodegenAllFastFlagName), nullptr);
+}
+
+TEST(LuauCodegenSwitchTest, IgnoresRecordsInReadOnlySegments) {
+  std::vector<FlagRecord> records(1);
+  records[0].name = NameInStrings();
+  records[0].name_length = 19;
+  auto segments = Segments(records);
+  segments[1].writable = false;
+  EXPECT_EQ(FindFastFlagValue(segments, kLuauCodegenAllFastFlagName), nullptr);
+}
+
+TEST(LuauCodegenSwitchTest, ReadsLoadSegmentsFromTheElfHeader) {
+  alignas(8) std::uint8_t image[0x400] = {};
+  auto* header = reinterpret_cast<Elf64_Ehdr*>(image);
+  std::memcpy(header->e_ident, ELFMAG, SELFMAG);
+  header->e_ident[EI_CLASS] = ELFCLASS64;
+  header->e_phoff = sizeof(Elf64_Ehdr);
+  header->e_phnum = 3;
+  auto* program = reinterpret_cast<Elf64_Phdr*>(image + header->e_phoff);
+  program[0] = {PT_LOAD, PF_R | PF_X, 0, 0, 0, 0x200, 0x200, 0x1000};
+  program[1] = {PT_DYNAMIC, PF_R | PF_W, 0, 0x200, 0, 0x10, 0x10, 8};
+  program[2] = {PT_LOAD, PF_R | PF_W, 0, 0x200, 0, 0x80, 0x180, 0x1000};
+  const auto base = reinterpret_cast<std::uintptr_t>(image);
+  const auto segments = LoadedImageSegments(base);
+  ASSERT_EQ(segments.size(), 2u);
+  EXPECT_EQ(segments[0].address, base);
+  EXPECT_EQ(segments[0].size, 0x200u);
+  EXPECT_FALSE(segments[0].writable);
+  EXPECT_EQ(segments[1].address, base + 0x200);
+  EXPECT_EQ(segments[1].size, 0x180u);
+  EXPECT_TRUE(segments[1].writable);
+
+  std::memset(image, 0, SELFMAG);
+  EXPECT_TRUE(LoadedImageSegments(base).empty());
+}
+
+TEST(LuauCodegenSwitchTest, DetectsAUserSetting) {
+  EXPECT_FALSE(ClientSettingsSetLuauCodegenAll("{}"));
+  EXPECT_FALSE(ClientSettingsSetLuauCodegenAll(R"({"FIntOther":1})"));
+  EXPECT_FALSE(ClientSettingsSetLuauCodegenAll("not-json"));
+  EXPECT_TRUE(
+      ClientSettingsSetLuauCodegenAll(R"({"FFlagDebugLuauCodegenAll":true})"));
+  EXPECT_TRUE(
+      ClientSettingsSetLuauCodegenAll(R"({"FFlagDebugLuauCodegenAll":"False"})"));
+}
+
+TEST(LuauCodegenSwitchTest, EnablesCodegenOnlyWhileBrowsing) {
+  EXPECT_TRUE(LuauCodegenAllForPhase(RobloxExperiencePresencePhase::kBrowsing));
+  EXPECT_FALSE(LuauCodegenAllForPhase(RobloxExperiencePresencePhase::kJoining));
+  EXPECT_FALSE(LuauCodegenAllForPhase(RobloxExperiencePresencePhase::kPlaying));
+}
+
+struct Forwarded {
+  int count = 0;
+  RobloxExperiencePresencePhase phase = RobloxExperiencePresencePhase::kBrowsing;
+};
+
+void RecordForward(void* context, RobloxExperiencePresencePhase phase,
+                   const RobloxExperienceLaunchRequest*) {
+  auto* forwarded = static_cast<Forwarded*>(context);
+  ++forwarded->count;
+  forwarded->phase = phase;
+}
+
+TEST(LuauCodegenSwitchTest, WritesTheFlagPerPhaseAndForwards) {
+  std::uint8_t flag = 1;
+  int locate_calls = 0;
+  Forwarded forwarded;
+  LuauCodegenMenuSwitch menu_switch(
+      [&] {
+        ++locate_calls;
+        return &flag;
+      },
+      {&forwarded, &RecordForward});
+  const RobloxExperiencePresenceObserver observer = menu_switch.observer();
+  ASSERT_TRUE(observer.valid());
+
+  observer.notify(observer.context, RobloxExperiencePresencePhase::kJoining,
+                  nullptr);
+  EXPECT_EQ(flag, 0);
+  observer.notify(observer.context, RobloxExperiencePresencePhase::kPlaying,
+                  nullptr);
+  EXPECT_EQ(flag, 0);
+  observer.notify(observer.context, RobloxExperiencePresencePhase::kBrowsing,
+                  nullptr);
+  EXPECT_EQ(flag, 1);
+  EXPECT_EQ(locate_calls, 1);
+  EXPECT_EQ(forwarded.count, 3);
+  EXPECT_EQ(forwarded.phase, RobloxExperiencePresencePhase::kBrowsing);
+}
+
+TEST(LuauCodegenSwitchTest, LeavesUnexpectedValuesAndMissingFlagsAlone) {
+  std::uint8_t flag = 7;
+  LuauCodegenMenuSwitch menu_switch([&] { return &flag; }, {});
+  menu_switch.OnPhase(RobloxExperiencePresencePhase::kJoining, nullptr);
+  EXPECT_EQ(flag, 7);
+
+  LuauCodegenMenuSwitch missing([] { return nullptr; }, {});
+  missing.OnPhase(RobloxExperiencePresencePhase::kJoining, nullptr);
+  EXPECT_TRUE(missing.observer().valid());
+}
+
+}  // namespace
+}  // namespace mocktail::runtime

tests/roblox_desktop_app_policy_test.cc

@@ -236,6 +236,22 @@ TEST(RobloxDesktopAppPolicyTest, DefaultsDesktopFlyoutSidePanelUnlessOverridden)
             "0");
 }
 
+TEST(RobloxDesktopAppPolicyTest, DefaultsLuauCodegenAllUnlessOverridden) {
+  const std::string policy =
+      nlohmann::json({{"PlatformGroup", "Desktop"}}).dump();
+  std::string merged;
+  std::string error;
+  ASSERT_TRUE(MergeDesktopAppPolicyClientSettingsOverride(policy, "{}", &merged,
+                                                          &error))
+      << error;
+  EXPECT_EQ(nlohmann::json::parse(merged)["FFlagDebugLuauCodegenAll"], "True");
+
+  ASSERT_TRUE(MergeDesktopAppPolicyClientSettingsOverride(
+      policy, R"({"FFlagDebugLuauCodegenAll":false})", &merged, &error))
+      << error;
+  EXPECT_EQ(nlohmann::json::parse(merged)["FFlagDebugLuauCodegenAll"], false);
+}
+
 TEST(RobloxDesktopAppPolicyTest, RejectsMalformedOrSymlinkedStorage) {
   TemporaryDirectory temporary;
   const std::filesystem::path default_path = temporary.path() / "default.json";

tests/roblox_experience_composition_test.cc

@@ -816,13 +816,9 @@ TEST(RobloxExperienceCompositionTest,
   ASSERT_TRUE(composition.DrainLaunchRequests().ok());
   EXPECT_EQ(probe.leaves, 1);
   EXPECT_EQ(probe.starts.load(), 1);
-  ASSERT_TRUE(composition.DrainLaunchRequests().ok());
-  EXPECT_EQ(probe.leaves, 1);
-  EXPECT_EQ(probe.starts.load(), 1);
 
-  // The exact gameDidLeave callback is the safe boundary between the old
+  // nativeAppBridgeV2LeaveGame ends in the Native stage and Roblox sends no
-  // controller and the replacement server launch.
+  // Lua app return, so the queued launch starts once the leave has returned.
-  composition.NotifyLuaAppDidReturn();
   for (int attempt = 0; attempt < 1000 && probe.starts.load() < 2; ++attempt) {
     ASSERT_TRUE(composition.DrainLaunchRequests().ok());
     std::this_thread::sleep_for(std::chrono::milliseconds(1));

tests/roblox_text_font_resolver_test.cc

@@ -97,6 +97,23 @@ TEST(RobloxTextFontResolverTest, RejectsMappingPathTraversal) {
       (assets.root() / "assets/fonts/SourceSansPro-Regular.ttf").string());
 }
 
+TEST(RobloxTextFontResolverTest, AddsBundledEmojiFontsAsFallbacks) {
+  TemporaryFontAssets assets;
+  ASSERT_FALSE(assets.root().empty());
+  ASSERT_TRUE(assets.CreateFile("assets/fonts/SourceSansPro-Regular.ttf"));
+  ASSERT_TRUE(assets.CreateFile("assets/content/fonts/RobloxEmoji.ttf"));
+  ASSERT_TRUE(assets.CreateFile("assets/content/fonts/TwemojiMozilla.ttf"));
+
+  const RobloxTextFontSelection selection =
+      ResolveRobloxTextFont(3, assets.content().string());
+
+  ASSERT_EQ(selection.emoji_files.size(), 2U);
+  EXPECT_EQ(selection.emoji_files[0],
+            (assets.content() / "fonts/TwemojiMozilla.ttf").string());
+  EXPECT_EQ(selection.emoji_files[1],
+            (assets.content() / "fonts/RobloxEmoji.ttf").string());
+}
+
 }  // namespace
 }  // namespace runtime
 }  // namespace mocktail

tests/summarize_profile_trace_test.py

@@ -89,6 +89,20 @@ class SummarizeTest(unittest.TestCase):
         self.assertIn("| frame interval p50 (ms) | 20.00 | 10.00 |", lines)
         self.assertIn("| a mean (ms) | 1.00 | 0.50 |", lines)
 
+    def test_texture_counters_report_their_peak(self) -> None:
+        events = [
+            {"name": "emulated texture (MB)", "ph": "C", "ts": ts, "pid": 1,
+             "tid": 1, "args": {"value": value}}
+            for ts, value in ((1, 10.0), (2, 300.0), (3, 120.0))
+        ]
+        with_textures = summary.summarize(events)
+        without = summary.summarize([])
+
+        self.assertEqual(with_textures.peaks["emulated texture (MB)"], 300.0)
+        table = summary.to_markdown([("one", with_textures), ("two", without)])
+        self.assertIn("| emulated texture (MB) peak | 300.00 | - |",
+                      table.splitlines())
+
     def test_markdown_marks_missing_values(self) -> None:
         with_slice = summary.summarize([slice_event("a", 1_000)])
         without = summary.summarize([])

tests/vulkan_etc2_emulation_test.cc

@@ -102,6 +102,37 @@ TEST(VulkanEtc2EmulationTest, CountsUploadBytesPerRegion) {
   EXPECT_EQ(decoded, 2U);
 }
 
+TEST(VulkanEtc2EmulationTest, CountsWholeImageBytesAcrossMipsAndLayers) {
+  VkDeviceSize compressed = 0;
+  VkDeviceSize decoded = 0;
+  // 64x64 with a full chain: 64, 32, 16, 8, 4, 2, 1.
+  ASSERT_TRUE(Etc2ImageByteCounts(EtcFormat::kEtc2Rgb8, {64, 64, 1}, 7, 1, 1,
+                                  &compressed, &decoded));
+  EXPECT_EQ(compressed, 343U * 8U);
+  EXPECT_EQ(decoded, 5461U * 4U);
+
+  // Two layers of EAC R11, 2 bytes per decoded texel.
+  ASSERT_TRUE(Etc2ImageByteCounts(EtcFormat::kEacR11, {8, 4, 1}, 1, 2, 1,
+                                  &compressed, &decoded));
+  EXPECT_EQ(compressed, 2U * 2U * 8U);
+  EXPECT_EQ(decoded, 2U * 32U * 2U);
+}
+
+TEST(VulkanEtc2EmulationTest, CountsDecodedBytesAtTheUpscaledExtent) {
+  VkDeviceSize compressed = 0;
+  VkDeviceSize decoded = 0;
+  // The host image is 256x256 with the same 7 levels: 256 down to 4.
+  ASSERT_TRUE(Etc2ImageByteCounts(EtcFormat::kEtc2Rgba8, {64, 64, 1}, 7, 1, 4,
+                                  &compressed, &decoded));
+  EXPECT_EQ(compressed, 343U * 16U);
+  EXPECT_EQ(decoded, 87376U * 4U);
+
+  EXPECT_FALSE(Etc2ImageByteCounts(EtcFormat::kEtc2Rgb8, {0, 4, 1}, 1, 1, 1,
+                                   &compressed, &decoded));
+  EXPECT_FALSE(Etc2ImageByteCounts(EtcFormat::kEtc2Rgb8, {4, 4, 1}, 0, 1, 1,
+                                   &compressed, &decoded));
+}
+
 TEST(VulkanEtc2EmulationTest, RejectsUnsupportedRegions) {
   VkDeviceSize compressed = 0;
   VkDeviceSize decoded = 0;
@@ -147,6 +178,9 @@ VKAPI_ATTR VkResult VKAPI_CALL FakeCreateImage(VkDevice,
   return VK_SUCCESS;
 }
 
+VKAPI_ATTR void VKAPI_CALL FakeDestroyImage(VkDevice, VkImage,
+                                            const VkAllocationCallbacks*) {}
+
 VKAPI_ATTR void VKAPI_CALL FakeCopyImage(VkCommandBuffer, VkImage,
                                          VkImageLayout, VkImage, VkImageLayout,
                                          std::uint32_t, const VkImageCopy*) {
@@ -295,6 +329,7 @@ VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL FakeGetDeviceProcAddr(
   if (is("vkDestroySemaphore"))
     return reinterpret_cast<PFN_vkVoidFunction>(FakeDestroySemaphore);
   if (is("vkCreateImage")) return reinterpret_cast<PFN_vkVoidFunction>(FakeCreateImage);
+  if (is("vkDestroyImage")) return reinterpret_cast<PFN_vkVoidFunction>(FakeDestroyImage);
   if (is("vkCreateBuffer")) return reinterpret_cast<PFN_vkVoidFunction>(FakeCreateBuffer);
   if (is("vkGetBufferMemoryRequirements"))
     return reinterpret_cast<PFN_vkVoidFunction>(FakeBufferRequirements);
@@ -313,6 +348,47 @@ VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL FakeGetDeviceProcAddr(
   return nullptr;
 }
 
+TEST(VulkanEtc2EmulationTest, TotalsTheBytesOfLiveEmulatedImages) {
+  const VkDevice device = FakeHandle<VkDevice>(0x12);
+  VkPhysicalDeviceMemoryProperties memory{};
+  VulkanEtc2Emulation emulation;
+  emulation.RegisterDevice(device, FakeHandle<VkPhysicalDevice>(0x20), true,
+                           memory, FakeGetDeviceProcAddr);
+
+  VkImageCreateInfo image_info{};
+  image_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
+  image_info.imageType = VK_IMAGE_TYPE_3D;
+  image_info.format = VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK;
+  image_info.extent = {8, 8, 1};
+  image_info.mipLevels = 1;
+  image_info.arrayLayers = 1;
+  VkImage first = VK_NULL_HANDLE;
+  VkImage second = VK_NULL_HANDLE;
+  ASSERT_EQ(emulation.CreateImage(device, &image_info, nullptr, &first),
+            VK_SUCCESS);
+  ASSERT_EQ(emulation.CreateImage(device, &image_info, nullptr, &second),
+            VK_SUCCESS);
+  image_info.format = VK_FORMAT_R8G8B8A8_UNORM;
+  VkImage plain = VK_NULL_HANDLE;
+  ASSERT_EQ(emulation.CreateImage(device, &image_info, nullptr, &plain),
+            VK_SUCCESS);
+
+  EmulatedTextureBytes totals = emulation.TextureBytes();
+  EXPECT_EQ(totals.compressed, 2U * 4U * 8U);
+  EXPECT_EQ(totals.decoded, 2U * 64U * 4U);
+
+  emulation.DestroyImage(device, first, nullptr);
+  emulation.DestroyImage(device, plain, nullptr);
+  totals = emulation.TextureBytes();
+  EXPECT_EQ(totals.compressed, 4U * 8U);
+  EXPECT_EQ(totals.decoded, 64U * 4U);
+
+  emulation.DestroyDevice(device);
+  totals = emulation.TextureBytes();
+  EXPECT_EQ(totals.compressed, 0U);
+  EXPECT_EQ(totals.decoded, 0U);
+}
+
 TEST(VulkanEtc2EmulationTest, SkipsUploadOutsideTheMappedWindow) {
   const VkDevice device = FakeHandle<VkDevice>(0x10);
   const VkBuffer source = FakeHandle<VkBuffer>(0x200);