[CI] Use uname for Arch runner hostname

[CI] Add Arch WSL GPU runner smoke workflow and configuration
[Fix] (MG_Remote): bound socket buffers in log flood tests
2026-11-07 10:03:23 +09:00 · 2026-10-05 08:57:17 -04:00 · 2026-10-05 08:56:27 -04:00 · 2026-10-05 07:13:39 -04:00 · 2026-10-05 06:21:42 -04:00 · 2026-10-05 05:12:42 -04:00
582 changed files with 72748 additions and 3931 deletions
@@ -0,0 +1,3 @@
+# Everything here is executed by bash/sh on Linux, Android or WSL: keep LF even on a Windows checkout
+# (core.autocrlf=true would otherwise hand sh a script full of $'\r').
+* text=auto eol=lf
@@ -0,0 +1,183 @@
+---
+name: anland-mobilegl-plasma
+description: Bring up, rebuild, deploy, switch backends on, test and debug the "MobileGL + anland + KDE Plasma" stack - a rooted Android tablet whose anland APK embeds the MobileGL server, plus a Droidspaces Arch Linux ARM container running KWin/plasmashell/Chrome through the MobileGL client library and GBM backend - driven from this Windows machine (Git Bash, WSL archlinux, adb). Use it after a phone reboot, when changing server (NDK) or client (container glibc) MobileGL code or anland's KWin backend/patch, when switching Espryt/Magma, when validating zero-copy dma-buf presentation, and when chasing crashes, black windows, noise or software-rendering fallbacks in the Plasma session.
+---
+
+# MobileGL on anland KDE Plasma
+
+**Setting this up on a new machine or phone from source: `reproduce.md` next to this file.**
+Human-readable runbook (Chinese, same content): `docs/Disaggregated/notes/anland/runbook-plasma.md`.
+Background: `docs/Disaggregated/notes/anland/handoff-legacy-mobilegl-unified.md` (state, commits, known
+issues) and `plan-ahb-dmabuf.md` (zero-copy design).  All scripts are in `scripts/` next to this file;
+parameters are documented at the top of each script and in `scripts/env.sh`.
+
+**The phone may be shared with another session.** Before anything that restarts the app, the daemon,
+the container session or replaces a library, run the read-only `scripts/device/status.sh` and make sure
+nobody else is mid-test.  Use your own `ct.sh` channel names (others use c, c2, c3, xd, xr, gbm).
+
+## Architecture (10 lines)
+
+1. Phone: Lenovo TB321FU (Y700, Adreno 750, Android 15, kernel 6.1), rooted with **KernelSU** (`/data/adb/ksu`, `ksud`; `su -c` works). USB serial `HA27Q3LQ`, or wireless adb `ip:port`.
+2. anland APK `com.anland.consumer.mobilegl` (debuggable build of anland branch `legacy-mobilegl-unified`): UI process + private service process `:mobilegl` (`MobileGLWorker`), which loads the **server** `libMobileGL.so` (NDK, from MobileGL `feat/disaggregated`) and listens on abstract socket `@anland-mobilegl` from the first attached window on. One process serves every client (KWin, plasmashell, ksplash, Chrome...), one backend per process. `MobileGLWorker` is a **foreground service** (notification "Linux desktop is running" / action "Stop desktop"), so the server and the session outlive a hidden, locked or swiped-away window.
+3. Backend = the app setting "Renderer backend" (Settings > Connection > MobileGL desktop; file `/data/data/<pkg>/files/mobilegl-backend`, default DirectGLES; also `am start ... --es mobilegl_backend X`). Espryt = DirectGLES, Magma = DirectVulkan. The server reads it when its process starts; the app publishes it as `/data/local/tmp/anland-mobilegl/backend` and the session copies that into the container's `/etc/mobilegl/backend` (what every client reports; mismatch = "Hello.backendType disagrees with pinned backend"). `debug.mobilegl.backend`, if someone sets it, still overrides the setting until the next reboot.
+4. Experiment display daemon `/data/adb/modules/anland-daemon/display_daemon /data/local/tmp/anland-mobilegl/display.sock` (display/input broker), started by the app (below). The module only starts the original one on `/data/local/tmp/display_daemon.sock` at boot; leave that alone.
+5. Container `arch-kde-mgl` (Droidspaces 6.4.5, Arch Linux ARM aarch64 glibc, host net, `rootfs.img` mounted at `/mnt/Droidspaces/arch-kde-mgl` on the Android side). Bind mounts: `/data/local/tmp/anland-mobilegl` -> `/run/anland-mobilegl`, `/data/local/tmp/display_daemon.sock` -> `/run/display.sock`.
+6. `desktop-session.service` (User=swung0x48, **enabled**: the container boots straight into it, as the original image does) has drop-in `/etc/systemd/system/desktop-session.service.d/mobilegl.conf` (anland `producers/kde/Arch_v5/desktop-session-mobilegl.conf`, installed + enabled by `sync-build-kwin.sh`): `ExecStartPre=mobilegl-startup.sh wait` (until `@anland-mobilegl` listens in `/proc/net/unix` and the daemon socket exists), `ExecStartPre=+... sync-backend`, `Restart=always`, then `ExecStart=/opt/mobilegl/bin/mobilegl-startup.sh plasma`, which deletes the sticky software-renderer key, sets `MOBILEGL_IPC_PROBE_TIMEOUT_MS=0` session-wide (MobileGL never declines, so glvnd never falls through to another GL stack), writes a user drop-in for `plasma-kwin_wayland.service` (kwin_wayland_wrapper -> `/opt/mobilegl/kwin/bin/kwin_wayland` + `/opt/mobilegl/kwin/lib/libkwin.so.6.7.4`, `ANLAND_MOBILEGL=1`, `MOBILEGL_IPC_SURFACE=server`) and runs `startplasma-wayland`.
+7. KWin 6.7.4 with anland's `anland` backend renders straight into the anland Android Surface through the server. While the window is not on screen (consumer disconnected) the backend powers its output down (DPMS Off + RenderLoop inhibited: no compositing, no frame callbacks) and back up when the window returns. Every other GL process is `MOBILEGL_IPC_SURFACE=offscreen` and presents through KWin.
+8. Client = `/opt/mobilegl/lib/libMobileGL.so` (glibc aarch64), the container's **system-wide** GL vendor (section below): glvnd EGL vendor `/usr/share/glvnd/egl_vendor.d/10_mobilegl.json` (sorts before the distribution's `50_*.json`, which serves whenever MobileGL declines), GLX vendor `/usr/lib/libGLX_mobilegl.so.0`, settings from `/etc/mobilegl/client.conf` (`MOBILEGL_TRANSPORT=spawn MOBILEGL_IPC_DATA=shm MOBILEGL_IPC_CONTROL=unix:@anland-mobilegl`) + `/etc/mobilegl/backend`. No per-process environment is needed.
+9. Zero-copy windows: client `wl_egl_window` frames are server-allocated AHardwareBuffer "shared images" exported as dma-buf fds over `zwp_linux_dmabuf_v1`; KWin imports them with EGL and the server recognises its own buffer by inode. KWin gets a DRM device via `GBM_BACKEND=mobilegl` -> `/usr/lib/gbm/mobilegl_gbm.so` (node is only an identity, never ioctl'd). Without linux-dmabuf the client falls back to wl_shm readback.
+10. X11 (`docs/Disaggregated/notes/anland/plan-x11-gpu.md`): Xwayland draws with **glamor on MobileGL** (an offscreen session; `GBM_BACKEND=mobilegl` on KWin's render node; window pixmaps are shared images handed to KWin as linux-dmabuf buffers; its `glFlush` publishes them through the server, `/tmp/mgl-xwayland.client.log`); `ANLAND_XWAYLAND_GLAMOR=0` in KWin's drop-in brings back the software `-shm` Xwayland. GLX windows present their frames as shared images through **DRI3 + Present** (log `presented through DRI3+Present shared images`), falling back to readback + MIT-SHM / `xcb_put_image` when the X server has no DRI3 (`MOBILEGL_GLX_PRESENT=readback` forces it per app). Chrome runs ANGLE-GLES over MobileGL EGL with the GPU in its own process (`mobilegl-startup.sh chrome`; `MOBILEGL_CHROME_GPU=in-process` opts out), presenting GBM dma-bufs over linux-dmabuf; after a device loss Chrome restarts that process and repaints.
+
+## First-time prerequisites (Windows host)
+
+- Git Bash, `adb` (SDK platform-tools), CMake >= 3.22 and Ninja on PATH (seen: CMake 3.27.6, Ninja 1.13.2), Python 3 + Pillow (only `rec.sh`).
+- Android NDK **27.2.12479018** at `%LOCALAPPDATA%/Android/Sdk/ndk/27.2.12479018` (override `ANDROID_NDK`).
+- WSL distro **archlinux** with clang, lld, llvm, cmake, ninja. **lld's major version must equal clang/llvm-libs'** (lld 23 on llvm 22 broke `ld.lld`; working set: all 22.1.6). Then export the sysroot once: `bash scripts/xsysroot/refresh-sysroot.sh` (-> `~/sysroots/arch-kde-mgl` + `~/sysroots/aarch64-arch.cmake`; re-run after container package upgrades, then `xbuild.sh --reconfigure`).
+- MobileGL worktree (never build in the shared main checkout): `GIT_LFS_SKIP_SMUDGE=1 git worktree add ...`, `git submodule update --init --recursive`, and copy `3rdparty/glslang/External/spirv-tools` from the main checkout (the submodule leaves it empty).
+- anland worktree on `legacy-mobilegl-unified` (KWin patch/backend, startup script, `mobilegl-tools/`).
+- Phone side that already exists and is NOT recreated by this skill: the APK (build: anland `consumers/anland_v5/android_consumer/MOBILEGL.md`, gradle `:app:assemblePlainDebug` with `MOBILEGL_DIST`, `-PmobileglApplicationId=com.anland.consumer.mobilegl -PanlandDefaultSocket=/data/local/tmp/anland-mobilegl/display.sock`, from Git Bash with `MSYS_NO_PATHCONV=1`; signed with this machine's `~/.android/debug.keystore`, so `adb install -r` works; reinstalling with another key changes the UID and needs the KernelSU grant again - `mobilegl-tools/ksu_grant.c`), the anland KernelSU modules (`anland-daemon`, `anland-awl`), the container with KWin's configured build tree `/root/mobilegl-build/kwin-6.7.4/build` and pristine `/root/gbm-dev/kwin-a/src`.
+- Install the helpers once (container must be running): `bash scripts/push-tools.sh` (device/*.sh -> `/data/local/tmp/anl`, `mgrun` -> container `/usr/local/bin`, `kwprep.sh` -> container `/root`).
+
+Run host scripts from Git Bash as `bash scripts/<x>.sh`.  Raw adb with device paths needs `MSYS_NO_PATHCONV=1` and the whole root command in one quoted string: `adb shell 'su -c "..."'`.  A device script can also be run without pushing: `MSYS_NO_PATHCONV=1 adb shell 'su -c sh' < scripts/device/status.sh`.
+
+## Starting, idling, stopping - no manual steps
+
+Nothing starts at phone boot (container `run_at_boot=0`). **Opening the app** (launcher icon = `am start -n com.anland.consumer.mobilegl/com.anland.consumer.MainActivity`, no extras; the APK's default socket is the experiment's, `-PanlandDefaultSocket`) calls `MobileGLDesktop.ensureStarted()` - the single trigger - which starts the foreground service and runs the APK's `assets/mobilegl-desktop.sh up` as root (`su`): display daemon if absent (the script leaves the app's cgroup first, so it survives the app) -> backend published -> container started if stopped -> `systemctl start --no-block desktop-session`. The window waits for the daemon socket; the server listens once the Surface is attached; the session's `wait` step then lets Plasma start. A reopen with everything up only reattaches the window. Log: `/data/local/tmp/anland-mobilegl/desktop.log`, logcat `AnlandMobileGL` ("MobileGL desktop up done in N ms").
+
+Hidden window, screen off, lock screen or swiped-away task: server and session stay (foreground service); KWin turns the workspace's DPMS off (logs "viewer gone: output powered down"; windows get xdg_toplevel `suspended`, no frame callbacks) and back on ("viewer back"). Client swaps are paced by frame callbacks (`MG_Impl/EGLImpl/FrameThrottle.h`): interval >= 1 waits like any Wayland EGL; interval 0 (Chrome) drops to ~1 fps once a callback goes unanswered for 200 ms.
+
+Measured 2026-10-03 (Espryt, Chrome on webglsamples aquarium; % of one core, phone = all of Android): open-to-desktop cold 7.5-12.5 s (daemon 0.5, server 1.0, container 1.5, kwin ~5, plasmashell ~6-7), after a phone reboot 8.5 s, reopen of a running desktop 0.2 s; visible phone 249% / kwin 13 / chrome 99 / server 57 / GPU 22%, hidden 54 / 0 / 6 / 2 / 0%, locked 38 / 0 / 8 / 3 / 0%. Stop desktop 1.7 s. Wireless adb does not come back by itself after a reboot (re-enable it on the phone). **Stop desktop** (notification action, or `am startservice -n <pkg>/com.anland.consumer.MobileGLWorker -a com.anland.consumer.mobilegl.STOP`) runs `mobilegl-desktop.sh down` (session + container stopped, daemon stays), closes the windows, ends `:mobilegl`.
+
+```sh
+adb connect <ip>:<port>                        # only for wireless debugging
+export MSYS_NO_PATHCONV=1
+adb shell 'su -c "sh /data/local/tmp/anl/bringup.sh"'     # after a reboot: wake, dismiss keyguard, open the app, print the timeline
+bash scripts/desktop-verify.sh push cold reopen idle      # timed cold start, swipe-away + reopen, CPU/GPU visible/hidden/locked
+adb shell 'su -c "sh /data/local/tmp/anl/status.sh"'
+```
+
+`droidspaces` has no `status` command; `droidspaces show` lists running containers. `device/run-plasma.sh` is a developer restart (session + logind user, then the app), needed after a rebuild because KWin does not reconnect to a new server process. `device/desktop-reset.sh` takes everything down to the post-reboot state. The old manual sequence (setprop, start container, `restart-daemon.sh`, `run-plasma.sh`) is gone; `restart-daemon.sh` remains for daemon debugging only.
+
+## Build and deploy loops
+
+**Server (Android, NDK) -> APK lib**
+```sh
+bash scripts/build-android.sh          # -> <worktree>/build-android/libMobileGL.so (RelWithDebInfo, unstripped)
+bash scripts/cyc.sh s1                 # clear sticky SW-renderer key, push, swap into APK, restart session, screenshot $ANL_DIR/s1.png, print fatals
+```
+`device/swap.sh` overwrites `/data/app/.../lib/arm64/libMobileGL.so` keeping owner/mode/SELinux context (original saved as `/data/local/tmp/anland-mobilegl/libMobileGL.so.orig`); no re-signing. To swap without the session restart: push to `/data/local/tmp/anland-mobilegl/libMobileGL.new.so`, `su -c "sh /data/local/tmp/anl/swap.sh"`, then `run-plasma.sh` anyway (KWin does not reconnect to a new server process).
+
+**Client + GBM backend (container glibc) via WSL cross-build**
+```sh
+bash scripts/xbuild.sh [--reconfigure]  # WSL ~/mgl-xbuild-a64; ~150 s full, seconds incremental
+bash scripts/xdeploy.sh                 # rename-into-place install of libMobileGL.so + mobilegl_gbm.so, aliases, vendor JSONs, client.conf, GLX/GBM env drop-ins
+```
+Running clients keep the old inode; restart the client, or `run-plasma.sh` for KWin/plasmashell.  Build server and client from the **same commit** (both stamp `git rev-parse HEAD`), else the logs WARN "compatible wire with different build".  Fallback in-container build (slow): `scripts/sync-src.sh` + anland `mobilegl-tools/anland-build-client.sh`.
+
+**KWin (anland backend / kwin.patch) - built inside the container**
+```sh
+bash scripts/kwin-stage.sh build        # stage CR-stripped patch+backend+startup script, kwprep.sh, background sync-build-kwin.sh
+printf 'tail -5 /root/kwin-build.log\n' | bash scripts/ct.sh kw     # wait for "== ALL DONE =="
+adb shell 'su -c "sh /data/local/tmp/anl/run-plasma.sh"'
+```
+`sync-build-kwin.sh` (anland `producers/kde/Arch_v5/`) copies only the backend sources, then `ninja -j4`, installs `kwin_wayland`, `libkwin.so.6.7.4`, the patched screencast plugin and `/opt/mobilegl/bin/mobilegl-startup.sh` by atomic rename.  `kwprep.sh` is what applies changed `kwin.patch` hunks to core KWin files.  Repo-root `libdisplay_producer/` is not synced by either.  `kwin-stage.sh` is assembled from those two scripts and was not run as one unit yet.
+
+**Unit tests (no device)** - in WSL (the MSVC build of this branch is broken):
+```sh
+MSYS2_ARG_CONV_EXCL='*' wsl -d archlinux -- bash /mnt/c/<worktree>/.claude/skills/anland-mobilegl-plasma/scripts/wsl-test.sh MultiSessionTest TwinOwnershipTest
+MSYS2_ARG_CONV_EXCL='*' wsl -d archlinux -- bash -lc 'cd ~/mgl-anl-bld && ctest -R "MultiSession|TwinOwnership" -E DirectVulkan --output-on-failure'
+```
+`wsl-all.sh` builds and runs every unit test.  Never run DirectVulkan integration tests; never two ctest runs at once.
+
+## Switching backends
+
+In the app: Settings > Connection > MobileGL desktop > Renderer backend, then Stop desktop and open the app. Or `adb shell 'su -c "sh /data/local/tmp/anl/switch.sh DirectGLES"'`, which does the same (saves the setting through the launch extra, STOPs the service, reopens; clears a stale `debug.mobilegl.backend`). No setprop, no file to edit in the container: the session copies the published backend into `/etc/mobilegl/backend` on its next start. `cyc.sh` keeps whatever is set.
+
+## System-wide vendor and fallback
+
+Installed by `xdeploy.sh` (and anland's `anland-build-client.sh`), so any process in the container gets MobileGL through the stock libglvnd/libgbm:
+
+| File | Does |
+|---|---|
+| `/usr/share/glvnd/egl_vendor.d/10_mobilegl.json` | EGL vendor, asked before the system's `50_*.json` |
+| `/opt/mobilegl/share/glvnd/egl_vendor.d/50_mobilegl.json` | same JSON, for per-process forcing with `__EGL_VENDOR_LIBRARY_FILENAMES` |
+| `/etc/mobilegl/client.conf` | `KEY=value` defaults for every client (transport, data, endpoint); rewritten on deploy |
+| `/etc/mobilegl/backend` | the backend (the session's `sync-backend` step copies the app's choice in); wins over a `MOBILEGL_BACKEND_TYPE` line in client.conf |
+| `/etc/environment.d/10-mobilegl.conf` | `__GLX_VENDOR_LIBRARY_NAME=mobilegl`, `GBM_BACKEND=mobilegl` for the systemd user manager (all Plasma services and what they start, terminals included) |
+| `/etc/profile.d/mobilegl.sh` | the same two for login shells (`bash -l`, `su -l`, ssh, `droidspaces run bash -l`) |
+
+Precedence is environment > `/etc/mobilegl/backend` (backend only) > client.conf > built-in; `MOBILEGL_CONFIG_FILE` / `MOBILEGL_BACKEND_FILE` name other files (empty = off); a server process never reads them.
+
+**Fallback.** Before bringing anything up the client probes the configured endpoint (one non-blocking connect, `MOBILEGL_IPC_PROBE_TIMEOUT_MS`, default 250, 0 = off; a refused abstract socket answers in ~0.2 ms). With no server: the EGL vendor returns `EGL_NO_DISPLAY` and no devices (glvnd moves on to the other vendor), the GLX vendor fails `__glx_Main` (libGLX falls back to the vendor Xwayland names), and `mobilegl_gbm` fails `create_device` (libgbm falls back to the device's own backend). Client queries (`eglQueryString(EGL_NO_DISPLAY)`, `eglGetProcAddress`) never bring a session up. MobileGL also steps aside for a GBM device another backend created and for `LIBGL_ALWAYS_SOFTWARE=1` - that is how a software Xwayland (`ANLAND_XWAYLAND_GLAMOR=0`, KWin's patch sets it then) keeps the software stack. MobileGL never opens an X connection to a display its own process serves (`MG_Util/X11/DisplayGuard.h`; `MOBILEGL_X11_DIAL=0` forbids all), which is what made a glamor Xwayland deadlock on itself before. Show it without stopping the real server:
+```sh
+sed 's|^MOBILEGL_IPC_CONTROL=.*|MOBILEGL_IPC_CONTROL=unix:@nobody|' /etc/mobilegl/client.conf > /tmp/down.conf
+sudo -u swung0x48 env -i HOME=/home/swung0x48 PATH=/usr/bin XDG_RUNTIME_DIR=/run/user/1000 WAYLAND_DISPLAY=wayland-0 \
+  MOBILEGL_CONFIG_FILE=/tmp/down.conf eglinfo -B -p wayland     # vendor = the other one, ~0.1 s
+```
+
+## Test clients and how to verify
+
+X11 on the GPU has its own driver, `scripts/x11-verify.sh` (preflight, measure before/after with glxgears at `vblank_mode=0` on both GLX paths (presentation throughput only: glxgears is fixed-function GL, which MobileGL does not implement, so its window is black by design), deploy-client/-kwin/-server, check, apps, stop); `scripts/device/x11-measure.sh` refuses to load the phone below 40% battery.
+
+Inside the container as root (`printf '...' | bash scripts/ct.sh <chan>`), `mgrun offscreen <cmd>` runs a client as swung0x48 in the live session through a login shell, with nothing but the surface mode and the log path set (log `/tmp/mgl-run.client.log`, override base with `MGLOG=/tmp/x.log`); a plain `env -i ... bash -l -c '<cmd>'` as swung0x48 gets MobileGL the same way:
+```sh
+mgrun offscreen eglinfo -B ; mgrun offscreen glxinfo -B
+timeout 25 mgrun offscreen glmark2-es2-wayland -s 1280x720 --run-forever > /tmp/gm.out 2>&1 &
+sleep 8; grep -a "presented through" /tmp/mgl-run.client.log
+```
+Chrome: `bash scripts/chrome-launch.sh <tag> [flags]` (log `/tmp/mgl-chrome-<tag>.client.log`, screenshot `ch-<tag>.png`; `CHROME_GPU=in-process` puts the GPU back in the browser, which cannot recover from a device loss).
+Video: Chrome decodes in hardware. The GPU process's render node has a VA-API driver in the container (`/usr/lib/dri/msm_drm_drv_video.so`, DroidSpaces' V4L2 driver): Chrome allocates its frames with `vaCreateSurfaces` and exports them (`vaExportSurfaceHandle`, one fd, R8 + GR88 layers) - the driver's own buffers, never GBM ones - and imports each as an NV12 (P010 for 10-bit) EGLImage bound to `GL_TEXTURE_EXTERNAL_OES`. MobileGL takes such a FOREIGN YUV dma-buf through its labelled CPU-copy fallback: the server maps it and copies it into a YUV AHardwareBuffer of its own (`AHardwareBuffer_lockPlanes`) at the reader's first use in each frame (any present, publishing flush or native fence), then converts that into the texture's RGBA level on the GPU (Espryt: `GL_EXT_YUV_target` with the import's colour hints, `MOBILEGL_YUV_DRIVER_CONVERSION=1` for the driver's external sampler; Magma: `VkSamplerYcbcrConversion` on the AHB's external format). A server-allocated NV12/P010 buffer (GBM `gbm_bo_create(NV12)`) needs no copy. `MOBILEGL_CHROME_VIDEO_DECODE=software` in `mobilegl-startup.sh chrome` brings back `--disable-accelerated-video-decode`. Check a video with `scripts/probes/cdpmedia.py <port> <s>` in the container (CDP `Media.enable`: `kVideoDecoderName = VaapiVideoDecoder`, `kIsPlatformVideoDecoder = true`; `CDP_NAVIGATE=<url>` / `CDP_EXPR=<js>`), which needs Chrome started with `--remote-debugging-port` and a non-default `--user-data-dir` (a fresh profile also needs `--password-store=basic`, else page loads wait forever on KWallet), and grep Chrome's log for `exit_code` (a GPU-process restart). `scripts/probes/yuvprobe.c` pixel-checks foreign NV12/P010 imports, their per-frame refresh and a GBM NV12 buffer (`mgrun offscreen mgl-yuvprobe`, exit 0 = pass). KWin's own NV12 path (it splits a YUV buffer into R8 + GR88 imports) is not served: MobileGL imports no single-channel dma-bufs, so KWin advertises no YUV format over linux-dmabuf. A vendor-agnostic replacement driver (MobileVA: VA-API on Android MediaCodec, decoding into server YUV images with no CPU copy, the current driver kept as an A/B fallback) is planned, not built: `docs/MobileVA/plan.md`.
+Device loss: `setprop debug.mobilegl.inject_device_lost_pid <pid>` (as root) loses the server session of the client whose pid - as it sees itself in the container - is `<pid>`, once per value, on any backend and whether or not it presents; `debug.mobilegl.inject_device_lost_at N` instead loses whichever session makes the N-th device check (Espryt presents and read-backs only). `bash scripts/chrome-recover.sh <in-process|process> [rounds]` checks that Chrome repaints after such a loss without a restart (battery-guarded, stops below 40%). `bash scripts/kwin-recover.sh [rounds]` loses KWin's own session (default twice) and checks that KWin restarts compositing on a fresh session - same KWin, plasmashell and client processes, the screen still animating (glmark2 on Wayland, glxgears on X11, `CHROME=1` adds Chrome) - which needs the anland backend + kwin.patch reset handling deployed (`kwin-stage.sh build`).
+Probes: `scripts/probes/texfmt.c` (texture formats keep alpha?), `dmahash.c` (hash a process's dma-bufs), `yuvprobe.c` (YUV dma-buf sampling, `gcc -O1 /root/yuvprobe.c -o /usr/local/bin/mgl-yuvprobe -lEGL -lGLESv2 -lgbm -lm`), `cdpmedia.py` (which decoder a Chrome page uses); build in the container, e.g. `gcc -O1 /root/texfmt.c -o /usr/local/bin/mgl-texfmt -lEGL`, run with `mgrun offscreen`.
+GPU hangs: a shader that never ends stops the WHOLE GPU (every process, MobileGL or not) until the kernel resets it - ~2 s, or ~10 s ("Preemption Fault" in dmesg) when a higher-priority context waits; nothing in user space shortens one stall. What the server does is end the guilty session so it cannot hang the GPU again (Magma: event-bracketed submissions + `mgl-gpu-watch`, budget `MOBILEGL_GPU_HANG_BUDGET_MS` / `debug.mobilegl.gpu_hang_budget_ms`, default 2000; Espryt: the driver's GUILTY reset status, which it reports after the ~2 s reset only). Repro: `scripts/probes/hangclient.c` (Wayland + GL 3.3 core; `-H 60 -N 1 -E 60` hangs every 60 frames, `-H -1` is a frame-gap observer; build with wayland-scanner's xdg-shell files, `-lwayland-client -lwayland-egl -lEGL -lOpenGL -lm`), and `gpuprobe.c` (NDK, plain adb shell; `gpuprobe 30 high` is a high-priority GLES context outside MobileGL - it both measures the device-wide stall and forces the 10 s preemption path).
+
+A result counts only if ALL hold:
+- client log says `presented through linux-dmabuf shared images` (not `wl_shm`; `could not be presented through a linux-dmabuf shared image` = fell back);
+- `bash scripts/shot.sh g 2 1` gives two different hashes while the client animates (a "fast" score can mean frames are not shown);
+- `adb logcat -b crash -d` and `adb logcat -d | grep -E "Fatal signal|Abort message|E MobileGL"` are clean (a frozen window is usually a dead server);
+- for plasmashell verdicts, the sticky software renderer is off (below) and `/tmp/mgl-plasmashell.client.log` is freshly written.
+Reference (glmark2-es2-wayland 1280x720, zero-copy vs wl_shm): Espryt 187 vs 49, Magma 158 vs 34.  Rerun any surprising pass.
+
+## Logs
+
+| What | Where |
+|---|---|
+| KWin client side | container `/tmp/mobilegl-compositor.client.log` |
+| plasmashell / ksplash | `/tmp/mgl-plasmashell.client.log`, `/tmp/mgl-ksplash.client.log` (user drop-ins `~/.config/systemd/user/plasma-{plasmashell,ksplash}.service.d/mobilegl-log.conf`) |
+| mgrun / Chrome | `/tmp/mgl-run.client.log`, `/tmp/mgl-chrome-<tag>.client.log` (+ `/tmp/chrome-<tag>-out.log`) |
+| Server | `adb logcat -s MobileGL`; file `/data/data/com.anland.consumer.mobilegl/files/mobilegl-server.server.log` |
+| Native crashes | `adb logcat -b crash -d`; container `coredumpctl list` |
+| Session / KWin stderr | container `journalctl -u desktop-session --no-pager`, `journalctl --no-pager --since -2min \| grep kwin_wayland` |
+| Display daemon | `/data/local/tmp/anland-mobilegl/display-daemon.log` |
+| Desktop start/stop (app's su script) | `/data/local/tmp/anland-mobilegl/desktop.log`, `container-start.log`; logcat `-s AnlandMobileGL` |
+| Session wait / backend sync | container `journalctl -u desktop-session --no-pager` ("waiting for the MobileGL server ...") |
+
+`MOBILEGL_LOG_FILE_PATH=/tmp/x.log` produces `/tmp/x.client.log` (and `.server.log` for forwarded lines); files are truncated on open, so one base per process.  Client builds log at INFO.  Symbolize server frames with `$ANDROID_NDK/toolchains/llvm/prebuilt/windows-x86_64/bin/llvm-addr2line.exe -Cfe build-android/libMobileGL.so <offset>`.
+
+## Pitfalls
+
+- **pkill -f self-match**: a `ct.sh`/`droidspaces run bash -lc "<script>"` process carries the script in its argv; `pkill -f chrome` kills the runner. Use `pkill -x`.
+- **Root-owned /tmp files**: a log created by a root-run client cannot be reopened by swung0x48 later (no log appears). `rm -f` it as root first.
+- **Sticky Plasma software renderer**: after any server crash Plasma writes `[QtQuickRendererSettings] SceneGraphBackend=software` into `~swung0x48/.config/kdeglobals`; plasmashell then renders on the CPU and screenshots still look fine (tray icon "正在使用的软件渲染器"). `mobilegl-startup.sh plasma` deletes it at every session start; within a running session that hit a crash, restart the session (`run-plasma.sh`) before any verdict.
+- **Lock screen**: autolock is off (`kscreenlockerrc [Daemon] Autolock=false`); a real lock can only be left by typing the password (logind lock/unlock is not supported for this session type - `device/unlock.sh` is unverified). Test the greeter with `/usr/lib/kscreenlocker_greet --testing` and kill it.
+- **CRLF**: `core.autocrlf=true` here; anland's backend sources and `sync-build-kwin.sh` are CRLF in the Windows checkout. Strip CR before anything runs on Linux/Android (`tr -d '\r'`; `ct.sh`, `push-tools.sh`, `kwin-stage.sh` do). This skill has a `.gitattributes` forcing LF.
+- **Display daemon**: the app starts it; the session's `wait` step keeps KWin from starting without it (a KWin that cannot reach it logs "failed to connect to display daemon at /run/anland-mobilegl/display.sock" and exits in ~3 s). Anything started from the app's `su` must leave the app's cgroup (`echo $$ > /sys/fs/cgroup/cgroup.procs`) or a force-stop kills it.
+- **Device power/connection**: low battery or the screen turning off (30 s default) can kill the SurfaceView - `svc power stayon true`; wireless adb `ip:port` changes per pairing - never pin `-s HA27Q3LQ`, use `ANDROID_SERIAL` if several devices are attached.
+- **Container vs phone paths**: container `/tmp` is a different place from the phone's `/data/local/tmp`; move files via `/mnt/Droidspaces/arch-kde-mgl/root` (container `/root`) or `/data/local/tmp/anland-mobilegl` (= container `/run/anland-mobilegl`).
+- **Never overwrite a mapped .so in place** (container or APK while running): copy to a new inode and `mv -f`; the APK swap force-stops the app first.
+- **Input**: plain `adb shell input` is blocked - use `su -c input`; anland grabs the touchscreen, so if `input tap` does nothing use `input mouse tap x y`. Never `debuggerd -b` a container process (signal 35 kills glibc processes).
+- **linux-dmabuf**: the client must use `create` + roundtrip, never `create_immed` (KWin answers a failed immediate import with a protocol error that kills the client).
+- **Cross-session bugs**: noise / another window's pixels after some client starts or exits = process-global state in the multi-session server (handles are per client, FBO/VAO per context). Repro with `kscreenlocker_greet --testing`.
+- `/etc/environment` leaks the baseline session's driver variables through pam_env (and, via `/usr/lib/environment.d/99-environment.conf`, into the user manager); `mobilegl-startup.sh` unsets them - keep that when editing it.
+- **Shared log path**: Xwayland inherits KWin's `MOBILEGL_LOG_FILE_PATH`; any line a process logs opens (truncates) that file. MobileGL stays silent in processes it does not serve, but give any extra client its own `MOBILEGL_LOG_FILE_PATH`.
+- **Stale per-process overrides**: an `__EGL_VENDOR_LIBRARY_FILENAMES` / `MOBILEGL_*` left in a unit drop-in or the user manager's environment (`systemctl --user show-environment`) still wins over the system files; `run-plasma.sh` (terminate-user) clears the manager's.
+- `handoff` mentions a `kwin_wayland` logging shim and `/tmp/mgl-kwin-<pid>.log`: stale - `/opt/mobilegl/kwin/bin/kwin_wayland` is now the real binary and logs to `/tmp/mobilegl-compositor.client.log`.
+
+## House rules
+
+- Work in a git worktree, never in the shared main checkout or its build dirs; worktrees need `GIT_LFS_SKIP_SMUDGE=1`.
+- Commit with `git commit --no-verify`, one short sentence, **no** `Co-Authored-By` trailer.
+- Never run the DirectVulkan integration tests (`ctest -E DirectVulkan`).
+- No external project names ("Mesa", "Zink", ...) in source code or comments; describe the technique.
+- Vendor-agnostic only on the zero-copy path: AHB, EGL_ANDROID_*, VK_ANDROID AHB import, sync_file, dma-buf, libgbm ABI - no kgsl/gralloc layouts/QCOM modifiers.
+- Identity caches key on monotonic lifetime ids, never GL names or pointers.
@@ -0,0 +1,183 @@
+# Reproducing the MobileGL + anland + KDE Plasma setup from source
+
+This is the from-scratch path for a NEW development machine and/or a NEW phone.  `SKILL.md` covers the
+day-to-day loop once everything below exists.  Paths and names that are specific to the original
+machine are listed in step 9: set them in `scripts/env.sh` (or the environment) before running anything.
+
+## 0. What you need
+
+**Phone.**
+- arm64 Android 11+ with **root**. KernelSU is what this was built on; the scripts only need `su -c`.
+- A Vulkan 1.1 driver with `VK_ANDROID_external_memory_android_hardware_buffer` for Magma, and/or GLES 3.2 + EGL_ANDROID_native_fence_sync for Espryt. Nothing vendor-specific is used: no kgsl and no vendor ioctls, only AHardwareBuffer, sync_file and the vendor's public GLES/Vulkan.
+- About 20 GB free for the container image.
+
+**Development host.**
+- The scripts are written for **Windows + Git Bash** with WSL (`archlinux` distro) for the unit tests and the cross build.
+- A Linux host works too: drop the `cygpath`/`MSYS_NO_PATHCONV` handling in `env.sh`, and run the WSL parts natively.
+- Tools:
+  - `adb`, CMake >= 3.22, Ninja, Python 3 (+ Pillow for screenshots/recording).
+  - Android NDK **27.2.12479018**.
+  - JDK 17 and Android SDK platform 36 (for the APK).
+  - In the Linux environment: clang/lld/llvm of ONE major version (a mismatched lld breaks `ld.lld`), cmake, ninja.
+
+## 1. Sources
+
+```sh
+# MobileGL: branch feat/disaggregated (everything here is on it)
+git clone https://github.com/MobileGL-Dev/MobileGL.git && cd MobileGL
+git checkout feat/disaggregated
+GIT_LFS_SKIP_SMUDGE=1 git submodule update --init --recursive
+#   3rdparty/glslang is the MobileGL-Dev fork (main); its External/spirv-tools must be populated
+#   (glslang's own update_glslang_sources.py, or copy it from another checkout).
+# Keep builds out of a shared checkout: work in a worktree (GIT_LFS_SKIP_SMUDGE=1 git worktree add ...).
+
+# anland: branch legacy-mobilegl-unified (KWin backend + patch, startup script, APK, daemon module)
+git clone https://github.com/MobileGL-Dev/anland.git && cd anland   # the fork that carries this branch
+git checkout legacy-mobilegl-unified
+```
+
+Build the server (APK) and the client (container) from the **same MobileGL commit**: both stamp
+`git rev-parse HEAD`, and a mismatch logs "compatible wire with different build".
+
+## 2. Phone base: root, Droidspaces, display daemon
+
+1. **Root** with KernelSU (or another `su` provider).
+2. **Droidspaces:** install the module and its app. Containers live under `/data/local/Droidspaces/Containers/<name>/`, and the CLI is `/data/local/Droidspaces/bin/droidspaces`.
+3. **anland display daemon** (KernelSU/Magisk module `anland-daemon`):
+   - Build it from the anland repo with `build_daemon_android.sh`, then package it with `build_magisk_module.sh`, which uses `magisk_module/`. Flash it and reboot.
+   - The module's `service.sh` starts the original daemon on `/data/local/tmp/display_daemon.sock`.
+   - The MobileGL session uses a SECOND instance of the same binary on `/data/local/tmp/anland-mobilegl/display.sock`: see `scripts/device/restart-daemon.sh`, which also chmods the directory and socket.
+
+## 3. The container
+
+Use an **Arch Linux ARM (aarch64, glibc) KDE Plasma Wayland rootfs** that has a systemd `desktop-session.service` running the desktop as a normal user with `PAMName=login`. A ready-made one is the Arch KDE "Wayland" image from Droidspaces RootFS Desktop Builder (see anland `doc/UserManual/anland_guide.md`, "Use prebuilt packages or a RootFS"). Or install `plasma-desktop`, `plasma-workspace`, `libglvnd`, `xorg-xwayland`, `base-devel`, `cmake` and `ninja` on plain Arch Linux ARM and write the service yourself.
+
+Import it into Droidspaces as `arch-kde-mgl`. The default name is in `env.sh`. Reference `container.config` keys:
+```
+enable_gpu_mode=1
+net_mode=host
+bind_mounts=/data/local/tmp/display_daemon.sock:/run/display.sock,/data/local/tmp/anland-mobilegl:/run/anland-mobilegl
+run_at_boot=0
+```
+- **Host networking:** it is required. The MobileGL server listens on the ABSTRACT socket `@anland-mobilegl`, which crosses into the container only when the network namespace is shared.
+- **The anland-mobilegl directory:** create `/data/local/tmp/anland-mobilegl` on the phone before the first start, mode 777.
+- **Desktop user:** the default is `swung0x48` (`MGL_DESKTOP_USER`). Create your own and point `desktop-session.service` at it.
+- **Starting it:** `droidspaces -C .../container.config start`. Android root sees the rootfs at `/mnt/Droidspaces/<name>`, which `CT_ROOT` in `env.sh` relies on.
+- **Container tools:** install them with `bash scripts/push-tools.sh` once the container runs.
+
+## 4. MobileGL server and the anland APK
+
+```sh
+bash scripts/build-android.sh          # -> <worktree>/build-android/libMobileGL.so
+```
+The script runs, with the NDK toolchain:
+```
+-DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-30 -DCMAKE_BUILD_TYPE=RelWithDebInfo
+-DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_BUILD_DISAGGREGATED=ON -DMOBILEGL_BUILD_DISAGGREGATED_INPROC=ON
+-DMOBILEGL_BUILD_TEST=OFF -DMOBILEGL_BUILD_BENCHMARK=OFF
+```
+
+APK, from anland `consumers/anland_v5/android_consumer` (details in its `MOBILEGL.md`):
+```sh
+mkdir -p dist/arm64-v8a && cp <worktree>/build-android/libMobileGL.so dist/arm64-v8a/
+MOBILEGL_DIST=$PWD/dist ./gradlew :app:assemblePlainDebug \
+  -PanlandNdkVersion=27.2.12479018 -PmobileglApplicationId=com.anland.consumer.mobilegl \
+  -PanlandDefaultSocket=/data/local/tmp/anland-mobilegl/display.sock
+adb install -r app/build/outputs/apk/plain/debug/app-plain-debug.apk
+```
+- **Grant root:** the app needs root for its helper; approve it in KernelSU.
+- **Reinstall with another signing key:** that changes the UID, so grant root again (anland `mobilegl-tools/ksu_grant.c` does it from a shell). Don't uninstall to fix a signature clash: that wipes the app data.
+- **Later server-only changes:** swap just the library into the installed APK (`scripts/cyc.sh` / `device/swap.sh`) instead of rebuilding the APK.
+
+## 5. MobileGL client in the container
+
+Pick one of these two:
+- **Cross build** (fast, from WSL or Linux):
+  1. `bash scripts/xsysroot/refresh-sysroot.sh` exports the container's headers and libraries to `~/sysroots/arch-kde-mgl` and installs `~/sysroots/aarch64-arch.cmake`. Re-run it after container package upgrades.
+  2. `bash scripts/xbuild.sh`, then `bash scripts/xdeploy.sh`.
+- **Native build in the container** (slow, no WSL): `scripts/sync-src.sh` copies the worktree to `/root/mobilegl-unified-src`, then run anland `producers/kde/Arch_v5/mobilegl-tools/anland-build-client.sh` inside (g++, `-DMOBILEGL_BUILD_DISAGGREGATED=ON`, Release).
+
+Either route installs the same files, renamed into place so running clients keep their old mapping:
+
+| File | What it is |
+|---|---|
+| `/opt/mobilegl/lib/libMobileGL.so` (+ `libEGL`/`libGL` aliases) | the client library |
+| `/usr/share/glvnd/egl_vendor.d/10_mobilegl.json` | system-wide EGL vendor; sorts before the distribution's `50_*.json`, which serves whenever MobileGL declines (no server reachable) |
+| `/opt/mobilegl/share/glvnd/egl_vendor.d/50_mobilegl.json` | per-process forcing |
+| `/usr/lib/libGLX_mobilegl.so.0` | GLX vendor |
+| `/usr/lib/gbm/mobilegl_gbm.so` | GBM backend (zero-copy window buffers) |
+| `/etc/mobilegl/client.conf` | `MOBILEGL_TRANSPORT=spawn`, `MOBILEGL_IPC_DATA=shm`, `MOBILEGL_IPC_CONTROL=unix:@anland-mobilegl` |
+| `/etc/mobilegl/backend` | `DirectVulkan` (Magma) or `DirectGLES` (Espryt) |
+| `/etc/environment.d/10-mobilegl.conf`, `/etc/profile.d/mobilegl.sh` | `__GLX_VENDOR_LIBRARY_NAME=mobilegl`, `GBM_BACKEND=mobilegl` |
+
+## 6. KWin 6.7.4 with the anland backend, and Xwayland
+
+The MobileGL session runs a rebuilt KWin from `/opt/mobilegl/kwin`, so the packaged one stays untouched.
+
+1. **Source tree:** in the container, unpack KWin 6.7.4 to `/root/mobilegl-build/kwin-6.7.4`, apply anland `producers/kde/Arch_v5/kwin.patch`, and overlay `producers/kde/anland_backend_Arch_v5/src/backends/anland` into `src/backends/anland`. Also take the current `common/` and `libdisplay_producer/`, as `mobilegl.md` there says.
+2. **Configure:**
+   ```sh
+   cmake -S kwin-6.7.4 -B kwin-6.7.4/build -G Ninja -DCMAKE_BUILD_TYPE=Release \
+     -DCMAKE_INSTALL_PREFIX=/usr -DCMAKE_INSTALL_LIBDIR=lib -DBUILD_TESTING=OFF
+   ```
+3. **Build and install:** run `producers/kde/Arch_v5/sync-build-kwin.sh` (`ninja -j4`). It installs `kwin_wayland` + `libkwin.so.6.7.4` into `/opt/mobilegl/kwin/{bin,lib}`, the patched screencast plugin in place (the original is kept beside it), and `/opt/mobilegl/bin/mobilegl-startup.sh`.
+   - From the dev host, `bash scripts/kwin-stage.sh build` does steps 1–3 incrementally: `kwprep.sh` applies changed patch hunks.
+4. **Xwayland:** the anland-patched Xwayland 24.1.13 comes from `producers/kde/Arch_v5/build.sh` (makepkg, as a normal user; it builds the anland KWin and Xwayland packages). A prebuilt anland rootfs already has it.
+5. **Session hook:** add `/etc/systemd/system/desktop-session.service.d/mobilegl.conf`:
+   ```ini
+   [Service]
+   ExecStart=
+   ExecStart=/opt/mobilegl/bin/mobilegl-startup.sh plasma
+   Environment=ANLAND_SOCKET=/run/anland-mobilegl/display.sock
+   Environment=QT_LOGGING_RULES=kwin_*.info=true
+   ```
+   `mobilegl-startup.sh plasma` writes the user drop-in for `plasma-kwin_wayland.service` (the `/opt/mobilegl/kwin` binary, `ANLAND_MOBILEGL=1`, `MOBILEGL_IPC_SURFACE=server`, `GBM_BACKEND=mobilegl`) and runs `startplasma-wayland`.
+
+## 7. Bring it up
+
+Nothing has to be started by hand, and the order no longer matters: **open the app** (its launcher
+icon, or `am start -n com.anland.consumer.mobilegl/com.anland.consumer.MainActivity` with no extras).
+That is the single trigger. The app starts its foreground service and runs its bundled
+`mobilegl-desktop.sh up` as root, which:
+1. starts the display daemon if it is not running;
+2. publishes the backend (the app setting "Renderer backend", default DirectGLES) to
+   `/data/local/tmp/anland-mobilegl/backend`;
+3. starts the container if it is stopped;
+4. starts `desktop-session`, which waits for the server and the daemon socket, copies the backend into
+   `/etc/mobilegl/backend` and runs Plasma.
+
+Cold start to a desktop takes about 8-12 s. `scripts/device/bringup.sh` wakes the phone, dismisses the
+keyguard, opens the app and prints the timeline. To change the backend, use the app setting and then
+Stop desktop and reopen, or run `scripts/device/switch.sh DirectGLES|DirectVulkan`. Do not use
+`setprop debug.mobilegl.backend`: it overrides the setting until the next reboot. **Stop desktop**
+(the notification action) stops the session and the container.
+
+**After a phone reboot:** wireless adb does not come back by itself. Re-enable it on the phone
+(Developer options > Wireless debugging), then `adb connect <ip>:<port>`; the port changes. After
+that, open the app as above. See SKILL.md "Starting, idling, stopping".
+
+## 8. Verify
+
+- **Plasma:** it appears in the app. The tray must NOT show the "software renderer" icon; if it does, delete `SceneGraphBackend=software` from the user's `kdeglobals` and restart.
+- **Plain shell:** in a plain container shell as the desktop user, `eglinfo -B` names MobileGL, and `glxinfo -B` from a login shell does too.
+- **Wayland client:** `timeout 25 mgrun offscreen glmark2-es2-wayland -s 1280x720`. The log must say `presented through linux-dmabuf shared images`.
+- **Chrome:** `bash scripts/chrome-launch.sh t1` gives windows with correct icons.
+- **Resize and lock:** BACK toggles the extra-keys bar (resize); lock/unlock resumes the desktop. `rtest.sh` in the scratch dir checks both.
+- **Unit tests:** `bash scripts/wsl-all.sh` in WSL. NEVER run the DirectVulkan integration tests.
+
+## 9. Machine-specific values to change
+
+| Value | Original | Where |
+|---|---|---|
+| MobileGL worktree | this skill's worktree | `MGL_WORKTREE` (auto) |
+| scratch dir | `%LOCALAPPDATA%/Temp/anl` | `ANL_DIR` |
+| NDK | `<SDK>/ndk/27.2.12479018` | `ANDROID_NDK` |
+| anland worktree | `%USERPROFILE%/AndroidStudioProjects/anland/.claude/worktrees/mgl-unified` | `ANLAND_WORKTREE` (`kwin-stage.sh`) |
+| APK id | `com.anland.consumer.mobilegl` | `MGL_PKG` |
+| container | `arch-kde-mgl` | `MGL_CONTAINER` |
+| desktop user | `swung0x48` | `MGL_DESKTOP_USER` + `desktop-session.service` |
+| WSL distro | `archlinux` | `WSL_DISTRO` (`xbuild.sh`, `wsl-*.sh`) |
+| cross build dir | WSL `~/mgl-xbuild-a64` | `MGL_XBUILD_DIR` |
+| phone | Lenovo TB321FU, Adreno 750 | nothing: no serial is pinned; with several devices set `ANDROID_SERIAL` |
+
+**Thermals:** a phone vendor's thermal daemon may throttle charging while the screen is on during long sessions. That is not part of this setup.
@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+# build-android.sh - build the Android SERVER libMobileGL.so (NDK, arm64-v8a) that the anland APK embeds.
+# Output: $MGL_WORKTREE/build-android/libMobileGL.so (RelWithDebInfo, unstripped, ~320 MB; keep it for
+# llvm-addr2line, it is what cyc.sh pushes).  Stamp = worktree HEAD, so the container client
+# (xbuild.sh, same stamp) and the server agree on the build ("compatible wire with different build"
+# WARN otherwise).  Env: MGL_WORKTREE, ANDROID_NDK, BUILD_DIR (default build-android), JOBS (12).
+# Log level is compiled to INFO.  android-30 = the APK's minSdk: thread_local is then the platform's
+# own TLS rather than the emulated one (which cost 7-12% of an apply thread on the API 26 build).
+set -euo pipefail
+. "$(dirname "$0")/env.sh"
+W=$MGL_WORKTREE
+B=$W/${BUILD_DIR:-build-android}
+STAMP=$(git -C "$W" rev-parse HEAD)
+if [ ! -f "$B/build.ninja" ]; then
+cmake -S "$W" -B "$B" -G Ninja -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
+  -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-30 -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+  -DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_BUILD_DISAGGREGATED=ON -DMOBILEGL_BUILD_DISAGGREGATED_INPROC=ON \
+  -DMOBILEGL_BUILD_TEST=OFF -DMOBILEGL_BUILD_BENCHMARK=OFF \
+  -DCMAKE_CXX_FLAGS=-DMOBILEGL_LOG_ACTIVE_LEVEL=MOBILEGL_LOG_LEVEL_INFO \
+  -DMOBILEGL_BUILD_STAMP="$STAMP"
+else
+cmake "$B" -DMOBILEGL_BUILD_STAMP="$STAMP" >/dev/null
+fi
+cmake --build "$B" --target MobileGL -j"${JOBS:-12}"
+ls -la "$B/libMobileGL.so"
@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+# chrome-launch.sh <tag> [extra chrome flags...] - (re)start Chrome inside the running Plasma session
+# through mobilegl-startup.sh's `chrome` mode (ANGLE GLES over MobileGL EGL), client log
+# /tmp/mgl-chrome-<tag>.client.log, then print its "presented through" line and save a screenshot
+# to $ANL_DIR/ch-<tag>.png.  Kills the running chrome first (pkill -x, never -f).  MobileGL is the
+# container's system-wide vendor (10_mobilegl.json + /etc/mobilegl/client.conf), so only the log is set here.
+# Env: WAIT (s before checking, default 20), CHANNEL (ct.sh channel, default chrome),
+#      CHROME_GPU (process, the default: the GPU in its own process, which presents
+#      through GBM dma-bufs and is restarted by Chrome after a device loss; in-process opts out -
+#      needs the anland startup script that knows MOBILEGL_CHROME_GPU; in process mode the GPU process and the browser
+#      write the one client log, the GPU process last); CHROME_GPU_SANDBOX=on keeps its GPU sandbox.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+TAG=${1:?usage: chrome-launch.sh <tag> [chrome flags...]}; shift
+bash "$here/ct.sh" "${CHANNEL:-chrome}" <<IN
+pkill -x chrome; sleep 2
+U=\$(id -u $MGL_DESKTOP_USER)
+rm -f /tmp/mgl-chrome-$TAG.client.log /tmp/chrome-$TAG-out.log
+sudo -u $MGL_DESKTOP_USER env XDG_RUNTIME_DIR=/run/user/\$U WAYLAND_DISPLAY=wayland-0 DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/\$U/bus MOBILEGL_LOG_FILE_PATH=/tmp/mgl-chrome-$TAG.log MOBILEGL_CHROME_GPU=${CHROME_GPU:-process} MOBILEGL_CHROME_GPU_SANDBOX=${CHROME_GPU_SANDBOX:-off} bash -c 'nohup /opt/mobilegl/bin/mobilegl-startup.sh chrome --restore-last-session $* > /tmp/chrome-$TAG-out.log 2>&1 &'
+sleep ${WAIT:-20}
+grep -ah 'presented through' /tmp/mgl-chrome-$TAG.client.log | head -3
+IN
+adb_ shell "su -c 'screencap -p $DEV_SHARE/ch-$TAG.png'"
+adb_ pull "$DEV_SHARE/ch-$TAG.png" "$ANL_DIR/ch-$TAG.png" >/dev/null && echo "screenshot: $ANL_DIR/ch-$TAG.png"
@@ -0,0 +1,146 @@
+#!/usr/bin/env bash
+# chrome-recover.sh <in-process|process> [rounds] - does Chrome repaint after a GPU device loss,
+# without a restart?  Starts Chrome (chrome-launch.sh, CHROME_GPU=<mode>) on an animated WebGL page,
+# then per round: screenshots (the page must animate), loses the session of Chrome's GPU client with
+# the server's pid-targeted debug knob (debug.mobilegl.inject_device_lost_pid; the pid as Chrome sees
+# itself in the container), waits, and screenshots again.  A round PASSES when the server latched
+# that pid's session, the browser process is the same one (no restart), a GPU client is drawing on a
+# NEW server session, and the screen still animates.  Stops early below MIN_BATTERY percent.
+#
+#   in-process  the GPU in the browser process (the startup script's default today): the target is
+#               the browser itself; recovery needs ANGLE to make a new native context.
+#   process     the GPU in its own process (MOBILEGL_CHROME_GPU=process): the target is the GPU
+#               process; Chrome restarts it after the loss, so the next round's target is a new pid.
+#
+# Deploy first (same MobileGL commit on both sides; see SKILL.md):
+#   bash scripts/build-android.sh && bash scripts/cyc.sh pre      # server -> APK, restarts the session
+#   bash scripts/xbuild.sh && bash scripts/xdeploy.sh             # client -> container
+#   STARTUP=<anland worktree>/producers/kde/Arch_v5/mobilegl-startup.sh bash scripts/chrome-recover.sh deploy-startup
+# Env: ROUNDS (default 2), SETTLE (s after the loss, default 8), MIN_BATTERY (default 40),
+#      CHANNEL (ct.sh channel, default crec), TAG (log tag, default crec-<mode>).
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+MODE=${1:?usage: chrome-recover.sh <in-process|process|deploy-startup> [rounds]}
+ROUNDS=${2:-${ROUNDS:-2}}
+SETTLE=${SETTLE:-8}
+MIN_BATTERY=${MIN_BATTERY:-40}
+CH=${CHANNEL:-crec}
+TAG=${TAG:-crec-$MODE}
+PAGE=/tmp/mgl-spin.html
+
+ct() { bash "$here/ct.sh" "$CH"; }
+
+battery() { adb_ shell dumpsys battery | tr -d '\r' | awk '/ level:/ {print $2}'; }
+check_battery() {
+    local level; level=$(battery)
+    echo "battery: $level%"
+    if [ -n "$level" ] && [ "$level" -lt "$MIN_BATTERY" ]; then
+        echo "STOP: battery below $MIN_BATTERY%"; cleanup; exit 3
+    fi
+}
+cleanup() { adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid 0'" || true; }
+
+if [ "$MODE" = deploy-startup ]; then
+    # The anland startup script that knows MOBILEGL_CHROME_GPU, installed by rename (a running
+    # session keeps its open copy).
+    src=${STARTUP:?STARTUP=<path to the anland mobilegl-startup.sh>}
+    tr -d '\r' < "$src" > "$ANL_DIR/mobilegl-startup.sh"
+    adb_ push "$ANL_DIR/mobilegl-startup.sh" "$DEV_SHARE/mobilegl-startup.sh.new" >/dev/null
+    ct <<'IN'
+install -m 755 /run/anland-mobilegl/mobilegl-startup.sh.new /opt/mobilegl/bin/mobilegl-startup.sh.new
+mv -f /opt/mobilegl/bin/mobilegl-startup.sh.new /opt/mobilegl/bin/mobilegl-startup.sh
+grep -c MOBILEGL_CHROME_GPU /opt/mobilegl/bin/mobilegl-startup.sh
+IN
+    exit 0
+fi
+case "$MODE" in in-process|process) ;; *) echo "mode must be in-process or process" >&2; exit 2;; esac
+trap cleanup EXIT
+
+check_battery
+adb_ shell "su -c 'sh $DEV_TOOLS/status.sh'" || true
+cleanup
+
+# The page: WebGL clears that change colour every frame plus a moving block, so a frame that
+# reaches the screen changes the screenshot.
+ct <<IN
+cat > $PAGE <<'HTML'
+<!doctype html><html><body style="margin:0;background:#000">
+<canvas id=c width=640 height=360 style="width:100vw;height:100vh"></canvas>
+<script>
+const gl = document.getElementById('c').getContext('webgl');
+let t = 0, lost = 0;
+document.getElementById('c').addEventListener('webglcontextlost', e => { e.preventDefault(); lost++; });
+function frame() {
+  t++;
+  if (gl && !gl.isContextLost()) {
+    gl.clearColor((t % 60) / 60, ((t + 20) % 60) / 60, ((t + 40) % 60) / 60, 1);
+    gl.clear(gl.COLOR_BUFFER_BIT);
+    gl.enable(gl.SCISSOR_TEST); gl.scissor((t * 7) % 600, 100, 40, 160);
+    gl.clearColor(1, 1, 1, 1); gl.clear(gl.COLOR_BUFFER_BIT); gl.disable(gl.SCISSOR_TEST);
+  }
+  document.title = 'spin ' + t + ' lost ' + lost;
+  requestAnimationFrame(frame);
+}
+frame();
+</script></body></html>
+HTML
+chmod 644 $PAGE
+IN
+
+WAIT=${WAIT:-20} CHROME_GPU=$MODE CHANNEL=$CH bash "$here/chrome-launch.sh" "$TAG" --new-window "file://$PAGE"
+
+# The target pid inside the container: the browser (in-process) or the GPU process.
+target_pid() {
+    ct <<'IN' | tr -d '\r' | tail -1
+for p in $(pgrep -x chrome); do
+  args=$(tr '\0' ' ' < /proc/$p/cmdline)
+  case "$args" in
+    *--type=gpu-process*) echo "gpu $p" ;;
+    *--type=*) ;;
+    *) echo "browser $p" ;;
+  esac
+done | sort > /tmp/mgl-crec-pids
+cat /tmp/mgl-crec-pids >&2
+echo "$(awk '/^browser/ {print $2; exit}' /tmp/mgl-crec-pids) $(awk '/^gpu/ {print $2; exit}' /tmp/mgl-crec-pids)"
+IN
+}
+
+pass=0
+for round in $(seq 1 "$ROUNDS"); do
+    echo "=== round $round/$ROUNDS ($MODE)"
+    check_battery
+    read -r browser gpu < <(target_pid)
+    if [ "$MODE" = process ]; then target=$gpu; else target=$browser; fi
+    echo "browser=$browser gpu=${gpu:-none} target=$target"
+    if [ -z "$target" ]; then echo "FAIL: no target process"; break; fi
+    bash "$here/shot.sh" "$TAG-r$round-pre" 2 1
+    adb_ logcat -c
+    adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid $target'"
+    sleep "$SETTLE"
+    adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid 0'"
+    bash "$here/shot.sh" "$TAG-r$round-post" 3 1 | tee "$ANL_DIR/$TAG-r$round-post.md5"
+    latched=$(adb_ logcat -d | grep -c "injected:pid=$target" || true)
+    started=$(adb_ logcat -d | grep -c "session #[0-9]* started" || true)
+    adb_ logcat -d | grep -E "inject_device_lost_pid|SessionLatch|session #[0-9]+ (started|pid)" | cut -c1-220 | head -12 || true
+    read -r browser2 gpu2 < <(target_pid)
+    echo "after: browser=$browser2 gpu=${gpu2:-none}"
+    ct <<IN || true
+echo "--- client log ($TAG):"; grep -aE "DEVICE LOST|RECOVERED|GL_UNKNOWN_CONTEXT_RESET|failed to make|session #" /tmp/mgl-chrome-$TAG.client.log | tail -8
+echo "--- chrome stderr:"; grep -aiE "context lost|GPU process|make.*current|lost" /tmp/chrome-$TAG-out.log | tail -8
+IN
+    adb_ logcat -b crash -d | tail -5
+    hashes=$(awk '{print $1}' "$ANL_DIR/$TAG-r$round-post.md5" | sort -u | wc -l)
+    verdict=PASS
+    [ "$latched" -ge 1 ] || { echo "  - the server never latched pid $target"; verdict=FAIL; }
+    [ "$browser2" = "$browser" ] || { echo "  - the browser process changed ($browser -> $browser2): Chrome restarted"; verdict=FAIL; }
+    [ "$started" -ge 1 ] || { echo "  - no new server session started after the loss"; verdict=FAIL; }
+    if [ "$MODE" = process ]; then
+        [ -n "$gpu2" ] && [ "$gpu2" != "$gpu" ] || { echo "  - no new GPU process"; verdict=FAIL; }
+    fi
+    [ "$hashes" -ge 2 ] || { echo "  - the screen does not change after the loss (frozen)"; verdict=FAIL; }
+    echo "round $round: $verdict"
+    [ "$verdict" = PASS ] && pass=$((pass + 1))
+    [ "$verdict" = PASS ] || break
+done
+echo "=== $pass/$ROUNDS rounds passed ($MODE); screenshots in $ANL_DIR/$TAG-r*-*.png"
@@ -0,0 +1,29 @@
+#!/bin/bash
+# kwprep.sh - (container, as root) bring the live KWin tree /root/mobilegl-build/kwin-6.7.4 in line with
+# anland's kwin.patch + anland backend sources, touching ONLY files whose content changed (keeps the
+# ninja build incremental).  It re-applies the whole patch to a pristine copy of KWin's src/
+# (/root/gbm-dev/kwin-a/src) in /root/kp and copies every patched file that differs into the live tree,
+# then copies changed backend files.  Build + install afterwards with anland's sync-build-kwin.sh.
+#
+# Input: /root/kwsync.tgz, a tar made in the anland worktree's producers/kde directory:
+#   tar czf kwsync.tgz Arch_v5/kwin.patch Arch_v5/mobilegl-startup.sh anland_backend_Arch_v5/src/backends/anland/
+# and pushed to the container's /root (e.g. via /mnt/Droidspaces/arch-kde-mgl/root/ on the Android side).
+# The pristine tree /root/gbm-dev/kwin-a/src was set up by hand earlier (unverified that it is exactly
+# upstream 6.7.4); patch hunks that fail are reported by `patch`.
+set -e
+rm -rf /root/kwsync && mkdir -p /root/kwsync && tar xzf /root/kwsync.tgz -C /root/kwsync
+P=/root/kwsync/Arch_v5/kwin.patch
+rm -rf /root/kp && mkdir -p /root/kp && cp -a /root/gbm-dev/kwin-a/src /root/kp/
+cd /root/kp && patch -p1 -s < $P
+LIVE=/root/mobilegl-build/kwin-6.7.4
+FILES=$(grep -E '^\+\+\+ b/' $P | sed 's|^+++ b/||' | awk '{print $1}' | sort -u)
+for f in $FILES; do
+  if [ -f "/root/kp/$f" ] && ! cmp -s "/root/kp/$f" "$LIVE/$f"; then cp "/root/kp/$f" "$LIVE/$f"; echo "updated $f"; fi
+done
+for f in /root/kwsync/anland_backend_Arch_v5/src/backends/anland/*; do
+  # libdisplay_producer is a symlink in git; a Windows checkout turns it into a small text file. The live
+  # tree has the real directory (repo-root libdisplay_producer/), which neither script syncs.
+  [ -f "$f" ] && [ "$(basename $f)" != libdisplay_producer ] || continue
+  b=$(basename $f); if ! cmp -s "$f" "$LIVE/src/backends/anland/$b"; then cp "$f" "$LIVE/src/backends/anland/$b"; echo "updated anland/$b"; fi
+done
+echo PREP_DONE
@@ -0,0 +1,21 @@
+#!/bin/bash
+# mgrun <surface: offscreen|server> <command...> - (container, as root) run one MobileGL client as the
+# desktop user inside the running Plasma session.  Installed as /usr/local/bin/mgrun (push-tools.sh).
+#   mgrun offscreen glmark2-es2-wayland -s 1280x720     # Wayland client, presents through KWin
+#   mgrun offscreen eglinfo -B ; mgrun offscreen glxinfo -B   # glxinfo goes through Xwayland + GLX vendor
+# "server" is only for a client that owns the Android Surface itself (KWin does); everything else is
+# offscreen.  MobileGL is the container's system-wide vendor, so nothing selects it here: EGL comes from
+# /usr/share/glvnd/egl_vendor.d/10_mobilegl.json, transport/endpoint from /etc/mobilegl/client.conf,
+# the backend from /etc/mobilegl/backend (must match the server), and the GLX vendor name / GBM backend
+# from /etc/profile.d/mobilegl.sh (sourced by the login shell below).  Client log:
+# ${MGLOG:-/tmp/mgl-run.log} -> the library writes /tmp/mgl-run.client.log (truncated per run; give
+# concurrent clients different MGLOG values).  Env: MGLOG, DESKTOP_USER (swung0x48).
+S=$1; shift
+DU=${DESKTOP_USER:-swung0x48}
+U=$(id -u "$DU")
+cd /tmp
+exec sudo -u "$DU" env XDG_RUNTIME_DIR=/run/user/$U WAYLAND_DISPLAY=wayland-0 DISPLAY=:0 \
+  XAUTHORITY=$(ls /run/user/$U/xauth_* 2>/dev/null | head -1) \
+  DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/$U/bus QT_QPA_PLATFORM=wayland \
+  MOBILEGL_IPC_SURFACE=$S MOBILEGL_LOG_FILE_PATH=${MGLOG:-/tmp/mgl-run.log} \
+  bash -lc 'exec "$@"' mgrun "$@"
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# ct.sh <channel> [script-file] - run a bash script as ROOT inside the container (login shell),
+# reading the script from the file or from stdin.  Generic form of the old c.sh/c2.sh/c3.sh.
+#
+#   printf '%s\n' 'ls -la /tmp/*.log' | bash ct.sh mychan
+#   bash ct.sh mychan ./something.cmd
+#
+# How: the script is written to $ANL_DIR/<channel>.cmd, pushed to $DEV_TOOLS/<channel>.cmd together with
+# a one-line <channel>.sh that does `droidspaces --name=<container> run bash -lc "$(cat <channel>.cmd)"`,
+# and that runs under `su`.  Quoting through adb + su + droidspaces is thereby avoided entirely.
+#
+# Pick a channel name of your own: other sessions on the same phone use c, c2, c3, xd, xr, gbm...;
+# writing to their channel files races with them.  CR characters are stripped (Windows checkouts).
+# Inside, run user-side commands with `sudo -u <desktop user> env XDG_RUNTIME_DIR=/run/user/<uid> ...`.
+# Never `pkill -f <word>` in a channel script: the `bash -lc "<script>"` process carries the whole
+# script on its command line and kills itself.  Use `pkill -x <name>`.
+set -euo pipefail
+. "$(dirname "$0")/env.sh"
+ch=${1:?usage: ct.sh <channel> [script-file]}
+src=${2:--}
+if [ "$src" = - ]; then tr -d '\r' > "$ANL_DIR/$ch.cmd"; else tr -d '\r' < "$src" > "$ANL_DIR/$ch.cmd"; fi
+printf '%s\n' "$DS_BIN --name=$MGL_CONTAINER run bash -lc \"\$(cat $DEV_TOOLS/$ch.cmd)\"" > "$ANL_DIR/$ch.sh"
+adb_ shell "mkdir -p $DEV_TOOLS"
+adb_ push "$ANL_DIR/$ch.cmd" "$DEV_TOOLS/$ch.cmd" >/dev/null
+adb_ push "$ANL_DIR/$ch.sh" "$DEV_TOOLS/$ch.sh" >/dev/null
+adb_ shell "su -c 'sh $DEV_TOOLS/$ch.sh'"
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+# cyc.sh <shot-name> - one server iteration: clear Plasma's sticky software-renderer key, push the
+# NDK-built server lib, swap it into the installed APK (device/swap.sh), restart app + Plasma session
+# (device/run-plasma.sh, ~90 s), screenshot to $ANL_DIR/<shot-name>.png, print fatal logcat lines.
+# Env: SERVER_LIB (default $MGL_WORKTREE/build-android/libMobileGL.so), WAIT (s after the session
+# restart, default 15), N (logcat lines, default 20), CHANNEL (ct.sh channel, default cyc).
+# Needs device/swap.sh + device/run-plasma.sh in $DEV_TOOLS (push-tools.sh).  RESTARTS THE SESSION.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+shot=${1:?usage: cyc.sh <shot-name>}
+LIB=${SERVER_LIB:-$MGL_WORKTREE/build-android/libMobileGL.so}
+printf '%s\n' "sudo -u $MGL_DESKTOP_USER kwriteconfig6 --file kdeglobals --group QtQuickRendererSettings --key SceneGraphBackend --delete; rm -f /tmp/mgl-plasmashell.client.log" \
+    | bash "$here/ct.sh" "${CHANNEL:-cyc}"
+adb_ push "$LIB" "$DEV_SHARE/libMobileGL.new.so" >/dev/null
+adb_ logcat -c
+adb_ shell "su -c 'sh $DEV_TOOLS/swap.sh >/dev/null && sh $DEV_TOOLS/run-plasma.sh >/dev/null 2>&1; sleep ${WAIT:-15}; screencap -p $DEV_SHARE/$shot.png'"
+adb_ pull "$DEV_SHARE/$shot.png" "$ANL_DIR/$shot.png" >/dev/null && echo "screenshot: $ANL_DIR/$shot.png"
+adb_ logcat -d | grep -E "Fatal signal|FATAL|Abort message|Magma wire buffer|E MobileGL" | cut -c1-260 | head -"${N:-20}" || true
@@ -0,0 +1,94 @@
+#!/usr/bin/env bash
+# desktop-verify.sh <phase>... - check that the MobileGL desktop starts, idles and stops by itself
+# (anland app: foreground service + mobilegl-desktop.sh; container: desktop-session.service).
+# Phases run in the order given; screenshots go to $ANL_DIR/dv-<phase>.png.
+#
+#   push     install device/*.sh into $DEV_TOOLS (push-tools.sh's device half)
+#   reset    everything down: app force-stopped (its service with it), container stopped,
+#            experiment display daemon killed - the state after a phone reboot
+#   cold     reset, then open the app like its icon does and time each piece (desktop-timeline.sh)
+#   open     open the app without resetting (after a reboot: the reboot was the reset)
+#   chrome   (re)start Chrome in the session on an animated WebGL page (kills a running Chrome)
+#   idle     CPU/GPU for visible, hidden (HOME), locked (screen off), visible again (measure.sh)
+#   reopen   swipe the app away (remove its task), check KWin survives, open it again, time it
+#   stop     Stop desktop from the notification (the service's STOP action) and check it all ends
+#   rtest    lock/unlock + window resize regression ($ANL_DIR/rtest.sh)
+# Env: SECS (idle window, default 20), URL (chrome page).
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+SECS=${SECS:-20}
+URL=${URL:-https://webglsamples.org/aquarium/aquarium.html}
+SOCK=$DEV_SHARE/display.sock
+su_() { adb_ shell "su -c '$*'"; }
+shot() {
+    su_ "screencap -p $DEV_SHARE/dv-$1.png"
+    adb_ pull "$DEV_SHARE/dv-$1.png" "$ANL_DIR/dv-$1.png" > /dev/null && echo "screenshot $ANL_DIR/dv-$1.png"
+}
+kwin_log() {   # the anland backend's lines since <time>
+    printf '%s\n' "journalctl --no-pager --since '$1' 2>/dev/null | grep -a -E 'viewer (gone|back)|consumer (reconnected|disconnected)|failed to connect' | tail -${2:-8}" |
+        bash "$here/ct.sh" dv 2>/dev/null | grep -v pushed || true
+}
+task_id() {
+    su_ "am stack list" | tr -d '\r' | grep -m1 "$MGL_PKG/" | sed -n 's/.*taskId=\([0-9]*\).*/\1/p'
+}
+wake() {
+    su_ "input keyevent KEYCODE_WAKEUP"; sleep 1; su_ "wm dismiss-keyguard"; sleep 2
+}
+
+for phase in "$@"; do
+    echo "=== $phase"
+    case $phase in
+    push)
+        for f in "$here"/device/*.sh; do
+            tr -d '\r' < "$f" > "$ANL_DIR/push.tmp"
+            adb_ push "$ANL_DIR/push.tmp" "$DEV_TOOLS/$(basename "$f")" > /dev/null
+        done
+        rm -f "$ANL_DIR/push.tmp"; echo "device scripts in $DEV_TOOLS" ;;
+    reset)
+        su_ "sh $DEV_TOOLS/desktop-reset.sh" ;;
+    cold)
+        bash "$0" reset
+        wake
+        su_ "sh $DEV_TOOLS/desktop-timeline.sh 240"
+        sleep 3; shot cold ;;
+    open)
+        wake
+        su_ "sh $DEV_TOOLS/desktop-timeline.sh 300"
+        sleep 3; shot open ;;
+    chrome)
+        WAIT=15 bash "$here/chrome-launch.sh" dv "$URL" ;;
+    idle)
+        su_ "sh $DEV_TOOLS/measure.sh visible $SECS"
+        t=$(date -u '+%Y-%m-%d %H:%M:%S UTC')
+        su_ "input keyevent KEYCODE_HOME"; sleep 4
+        su_ "sh $DEV_TOOLS/measure.sh hidden $SECS"
+        su_ "am start -n $MGL_PKG/com.anland.consumer.MainActivity >/dev/null"; sleep 5
+        su_ "input keyevent KEYCODE_SLEEP"; sleep 4
+        su_ "sh $DEV_TOOLS/measure.sh locked $SECS"
+        wake
+        su_ "am start -n $MGL_PKG/com.anland.consumer.MainActivity >/dev/null"; sleep 5
+        su_ "sh $DEV_TOOLS/measure.sh visible2 $SECS"
+        shot idle
+        kwin_log "$t" 12 ;;
+    reopen)
+        t=$(date -u '+%Y-%m-%d %H:%M:%S UTC')
+        id=$(task_id); echo "removing task $id"
+        [ -n "$id" ] && su_ "am stack remove $id"
+        sleep 8
+        su_ "echo server: \$(pidof $MGL_PKG:mobilegl) ui: \$(pidof $MGL_PKG) kwin: \$(pidof kwin_wayland) plasmashell: \$(pidof plasmashell)"
+        su_ "sh $DEV_TOOLS/measure.sh closed 10"
+        su_ "sh $DEV_TOOLS/desktop-timeline.sh 120"
+        sleep 3; shot reopen
+        kwin_log "$t" ;;
+    stop)
+        su_ "am startservice -n $MGL_PKG/com.anland.consumer.MobileGLWorker -a com.anland.consumer.mobilegl.STOP" || true
+        sleep 25
+        su_ "echo server: \$(pidof $MGL_PKG:mobilegl) ui: \$(pidof $MGL_PKG); $DS_BIN show"
+        su_ "tail -3 $DEV_SHARE/desktop.log" ;;
+    rtest)
+        bash "$ANL_DIR/rtest.sh" dv ;;
+    *)
+        echo "unknown phase $phase" >&2; exit 2 ;;
+    esac
+done
@@ -0,0 +1,7 @@
+# bringup.sh - (Android root shell) bring the MobileGL desktop up after a phone reboot. There is
+# nothing left to do by hand: opening the anland app (its launcher icon, or this am start) starts
+# the display daemon, the MobileGL server, the container and its Plasma session. Prints the timeline.
+# Run: su -c "sh /data/local/tmp/anl/bringup.sh"      Env: as desktop-timeline.sh.
+input keyevent KEYCODE_WAKEUP
+wm dismiss-keyguard
+sh "$(dirname "$0")/desktop-timeline.sh" 300
@@ -0,0 +1,17 @@
+# desktop-reset.sh - (Android root shell) take the MobileGL desktop fully down, as after a phone
+# reboot: the app force-stopped (its foreground service and server with it), the container
+# stopped, the experiment display daemon killed. Leaves the original anland daemon alone.
+# Run: su -c "sh /data/local/tmp/anl/desktop-reset.sh"      Env: PKG, CONTAINER, SOCK, DS.
+P=${PKG:-com.anland.consumer.mobilegl}
+C=${CONTAINER:-arch-kde-mgl}
+SOCK=${SOCK:-/data/local/tmp/anland-mobilegl/display.sock}
+D=${DS:-/data/local/Droidspaces/bin/droidspaces}
+am force-stop $P
+$D --name=$C stop > /dev/null 2>&1
+for p in $(pidof display_daemon); do
+    tr '\0' ' ' < /proc/$p/cmdline | grep -qF " $SOCK " && kill $p
+done
+rm -f $SOCK
+sleep 2
+echo "app: $(pidof $P) server: $(pidof $P:mobilegl) daemon: $(ls $SOCK 2>/dev/null)"
+$D show
@@ -0,0 +1,30 @@
+# desktop-timeline.sh [timeout-s] - (Android root shell) open the anland app the way its launcher
+# icon does (am start, no extras) and print when each piece of the MobileGL desktop comes up,
+# in seconds after the open: display daemon socket, the app's MobileGL server listening, the
+# container, kwin_wayland, plasmashell, and the splash (ksplashqml) gone = desktop on screen.
+# Pieces that already run show 0.0. Read-only apart from the am start.
+# Run: su -c "sh /data/local/tmp/anl/desktop-timeline.sh 180"      Env: PKG, CONTAINER, SOCK, DS.
+T=${1:-180}
+P=${PKG:-com.anland.consumer.mobilegl}
+C=${CONTAINER:-arch-kde-mgl}
+SOCK=${SOCK:-/data/local/tmp/anland-mobilegl/display.sock}
+D=${DS:-/data/local/Droidspaces/bin/droidspaces}
+now() { awk '{print $1}' /proc/uptime; }
+since() { awk -v a="$1" -v b="$(now)" 'BEGIN{printf "%.1f", b - a}'; }
+listening() { grep -q '00010000 0001 01 [0-9]* @anland-mobilegl$' /proc/net/unix; }
+t0=$(now)
+am start -n $P/com.anland.consumer.MainActivity > /dev/null
+for k in daemon server container kwin shell splash_seen shown; do eval "t_$k="; done
+while :; do
+    [ -z "$t_daemon" ] && [ -S "$SOCK" ] && t_daemon=$(since $t0) && echo "daemon socket     +${t_daemon}s"
+    [ -z "$t_server" ] && listening && t_server=$(since $t0) && echo "server listening  +${t_server}s"
+    [ -z "$t_container" ] && [ -n "$($D --name=$C pid 2>/dev/null | tr -dc 0-9)" ] && t_container=$(since $t0) && echo "container up      +${t_container}s"
+    [ -z "$t_kwin" ] && pidof kwin_wayland > /dev/null && t_kwin=$(since $t0) && echo "kwin_wayland      +${t_kwin}s"
+    [ -z "$t_shell" ] && pidof plasmashell > /dev/null && t_shell=$(since $t0) && echo "plasmashell       +${t_shell}s"
+    [ -z "$t_splash_seen" ] && pidof ksplashqml > /dev/null && t_splash_seen=1
+    if [ -n "$t_shell" ] && [ -z "$t_shown" ] && ! pidof ksplashqml > /dev/null; then
+        t_shown=$(since $t0); echo "splash gone       +${t_shown}s (desktop on screen)"; break
+    fi
+    [ "$(since $t0 | cut -d. -f1)" -ge "$T" ] && { echo "TIMEOUT after ${T}s"; break; }
+    sleep 0.3
+done
@@ -0,0 +1,47 @@
+# measure.sh <label> [seconds] - (Android root shell) CPU and GPU load of the MobileGL desktop over a
+# window: % of ONE core (so 800 = all 8 cores) for the whole phone, KWin (+Xwayland), plasmashell,
+# Chrome (all its processes), the app's :mobilegl server and its UI process, plus the mean of the
+# kgsl gpu_busy_percentage samples (1/s). The container's processes are visible from Android's
+# /proc (Droidspaces only gives them their own pid namespace).
+# Run: su -c "sh /data/local/tmp/anl/measure.sh visible 20"      Env: PKG.
+L=${1:-sample}
+N=${2:-20}
+P=${PKG:-com.anland.consumer.mobilegl}
+GPU=/sys/class/kgsl/kgsl-3d0/gpu_busy_percentage
+
+pids_named() {   # exact process names (comm or argv0 basename)
+    for n in "$@"; do pidof "$n"; done
+}
+pids_matching() {   # argv contains the word
+    for d in /proc/[0-9]*; do
+        { tr '\0' ' ' < $d/cmdline; } 2>/dev/null | grep -q "$1" && echo ${d#/proc/}
+    done
+}
+ticks() {   # utime+stime of the pids, in clock ticks
+    t=0
+    for p in "$@"; do
+        [ -r /proc/$p/stat ] || continue
+        v=$(sed 's/^.*) //' /proc/$p/stat | awk '{print $12 + $13}')
+        t=$((t + v))
+    done
+    echo $t
+}
+total() { awk '/^cpu /{print $2+$3+$4+$6+$7+$8}' /proc/stat; }
+
+KW="$(pids_named kwin_wayland Xwayland)"
+SH="$(pids_named plasmashell)"
+CH="$(pids_matching /opt/google/chrome/)"
+SV="$(pidof $P:mobilegl)"
+UI="$(pidof $P)"
+
+a0=$(total); k0=$(ticks $KW); s0=$(ticks $SH); c0=$(ticks $CH); v0=$(ticks $SV); u0=$(ticks $UI)
+g=0; i=0
+while [ $i -lt $N ]; do
+    sleep 1
+    b=$(cat $GPU 2>/dev/null | tr -dc '0-9'); g=$((g + ${b:-0})); i=$((i + 1))
+done
+a1=$(total); k1=$(ticks $KW); s1=$(ticks $SH); c1=$(ticks $CH); v1=$(ticks $SV); u1=$(ticks $UI)
+# CLK_TCK is 100: ticks per second == percent of one core.
+pct() { echo $(( ($2 - $1) / N )); }
+printf '%-10s %3ss  phone %4s%%  kwin %3s%%  plasmashell %3s%%  chrome %3s%%  server %3s%%  app-ui %3s%%  gpu %3s%%\n' \
+    "$L" "$N" "$(pct $a0 $a1)" "$(pct $k0 $k1)" "$(pct $s0 $s1)" "$(pct $c0 $c1)" "$(pct $v0 $v1)" "$(pct $u0 $u1)" "$((g / N))"
@@ -0,0 +1,20 @@
+# restart-daemon.sh - (Android root shell) (re)start the EXPERIMENT's anland display daemon on $SOCK.
+# It is not started at boot (the anland-daemon module's service.sh only starts the original daemon on
+# /data/local/tmp/display_daemon.sock, which must be left alone).  Without it KWin logs
+# "failed to connect to display daemon at /run/anland-mobilegl/display.sock" and the session dies in ~3 s.
+# The container sees $SOCK as /run/anland-mobilegl/display.sock (bind mount in container.config);
+# uid 1000 in the container must connect, hence 777 dir / 666 socket.
+# Env: SOCK, DAEMON (/data/adb/modules/anland-daemon/display_daemon).
+SOCK=${SOCK:-/data/local/tmp/anland-mobilegl/display.sock}
+DAEMON=${DAEMON:-/data/adb/modules/anland-daemon/display_daemon}
+DIR=$(dirname $SOCK)
+for p in $(ps -A -o PID,ARGS | grep "display_daemon $SOCK" | grep -v grep | awk '{print $1}'); do kill $p; done
+sleep 1
+rm -f $SOCK
+mkdir -p $DIR
+nohup $DAEMON $SOCK > $DIR/display-daemon.log 2>&1 &
+sleep 1
+chmod 777 $DIR
+chmod 666 $SOCK
+echo "daemon restarted:"
+ps -A -o PID,ARGS | grep "display_daemon $SOCK" | grep -v grep
@@ -0,0 +1,27 @@
+# run-plasma.sh - (Android root shell) DEVELOPER restart of the whole MobileGL desktop, e.g. after a
+# server, client or KWin rebuild: ends the Plasma session (and the user's logind session, which takes
+# a while to wind down), then restarts the anland app; opening it brings up the server, starts
+# desktop-session.service again and the session waits for the server. Normal use needs none of this:
+# opening the app is enough. ~60 s. Prints the unit status, plasma user units and recent coredumps.
+# Run: su -c "sh /data/local/tmp/anl/run-plasma.sh"
+# Env: PKG, CONTAINER, DESKTOP_USER (swung0x48), DS (droidspaces binary).
+D=${DS:-/data/local/Droidspaces/bin/droidspaces}
+P=${PKG:-com.anland.consumer.mobilegl}
+C=${CONTAINER:-arch-kde-mgl}
+U=${DESKTOP_USER:-swung0x48}
+$D --name=$C run bash -lc "
+systemctl stop desktop-session.service 2>/dev/null
+loginctl terminate-user $U 2>/dev/null; sleep 30
+systemctl reset-failed desktop-session.service
+rm -f /tmp/mobilegl-compositor.log
+date +%s > /tmp/plasma-start
+"
+am force-stop $P
+am start -n $P/com.anland.consumer.MainActivity > /dev/null
+sleep 30
+$D --name=$C run bash -lc "
+systemctl --no-pager status desktop-session.service | head -8
+UID_=\$(id -u $U)
+sudo -u $U XDG_RUNTIME_DIR=/run/user/\$UID_ systemctl --user --no-pager list-units --all 'plasma*' 2>&1 | head -30
+coredumpctl --no-pager list --since '-2min' 2>&1 | tail -5
+"
@@ -0,0 +1,13 @@
+# status.sh - (Android root shell) READ-ONLY health check of the whole stack.
+# Run: su -c "sh /data/local/tmp/anl/status.sh"      Env: PKG, CONTAINER, DS, SOCK.
+D=${DS:-/data/local/Droidspaces/bin/droidspaces}
+P=${PKG:-com.anland.consumer.mobilegl}
+C=${CONTAINER:-arch-kde-mgl}
+SOCK=${SOCK:-/data/local/tmp/anland-mobilegl/display.sock}
+echo "battery: $(dumpsys battery | grep ' level:' | tr -d ' ')  $(dumpsys power | grep -o 'mStayOn=[a-z]*')"
+echo "backend: app setting=$(cat /data/data/$P/files/mobilegl-backend 2>/dev/null || echo 'DirectGLES (default)')  published=$(cat ${SOCK%/*}/backend 2>/dev/null)  container /etc/mobilegl/backend=$(cat /mnt/Droidspaces/$C/etc/mobilegl/backend 2>/dev/null)  prop override=$(getprop debug.mobilegl.backend)"
+echo "desktop service log: $(tail -1 ${SOCK%/*}/desktop.log 2>/dev/null)"
+echo "app pid: $(pidof $P)  server (:mobilegl) pid: $(pidof $P:mobilegl)"
+ps -A -o PID,ARGS | grep "display_daemon $SOCK" | grep -v grep || echo "experiment display daemon NOT running"
+$D show
+$D --name=$C run bash -lc 'echo "desktop-session: $(systemctl is-active desktop-session.service)"; pgrep -a -x kwin_wayland; pgrep -a -x plasmashell; grep -A1 QtQuickRendererSettings /home/*/.config/kdeglobals 2>/dev/null || echo "kdeglobals: no sticky software renderer"'
@@ -0,0 +1,17 @@
+# swap.sh - (Android root shell) replace the server libMobileGL.so inside the INSTALLED anland APK with
+# $SHARE/libMobileGL.new.so, keeping owner, mode and SELinux context; force-stops the app.  The first
+# run backs the APK's original lib up to $SHARE/libMobileGL.so.orig.  No re-signing/reinstall (that
+# would change the app UID and lose its root grant).  Run: su -c "sh /data/local/tmp/anl/swap.sh"
+# Env: PKG (com.anland.consumer.mobilegl), SHARE (/data/local/tmp/anland-mobilegl).
+set -e
+P=${PKG:-com.anland.consumer.mobilegl}
+SHARE=${SHARE:-/data/local/tmp/anland-mobilegl}
+D=$(dirname $(pm path $P | sed 's/package://'))/lib/arm64
+ls -laZ $D
+[ -f $SHARE/libMobileGL.so.orig ] || cp $D/libMobileGL.so $SHARE/libMobileGL.so.orig
+am force-stop $P
+CTX=$(ls -Z $D/libMobileGL.so | awk '{print $1}')
+OWN=$(stat -c %u:%g $D/libMobileGL.so)
+cp $SHARE/libMobileGL.new.so $D/libMobileGL.so
+chown $OWN $D/libMobileGL.so; chmod 755 $D/libMobileGL.so; chcon $CTX $D/libMobileGL.so
+ls -laZ $D/libMobileGL.so
@@ -0,0 +1,17 @@
+# switch.sh <DirectGLES|DirectVulkan> - (Android root shell) change the MobileGL backend the anland
+# app's server runs (Espryt = DirectGLES, Magma = DirectVulkan). The choice is the app's setting
+# (Settings > Connection > MobileGL desktop, saved in the app's files/mobilegl-backend); this does
+# what a user does there: save it, Stop desktop, open the app again. The new server publishes the
+# backend and the container's session copies it to /etc/mobilegl/backend before it starts, so
+# server and clients agree. ~30-60 s. No setprop: debug.mobilegl.backend, if set, still overrides
+# the setting until the next reboot - this clears it.
+# Env: PKG, SOCK.
+B=${1:?usage: switch.sh DirectGLES|DirectVulkan}
+P=${PKG:-com.anland.consumer.mobilegl}
+[ -n "$(getprop debug.mobilegl.backend)" ] && setprop debug.mobilegl.backend ""
+if pidof $P:mobilegl > /dev/null; then
+    am startservice -n $P/com.anland.consumer.MobileGLWorker -a com.anland.consumer.mobilegl.STOP > /dev/null
+    i=0; while pidof $P:mobilegl > /dev/null && [ $i -lt 60 ]; do sleep 1; i=$((i + 1)); done
+fi
+am start -n $P/com.anland.consumer.MainActivity --es mobilegl_backend $B > /dev/null
+echo "backend $B saved; the desktop is starting (watch: sh $(dirname $0)/status.sh)"
@@ -0,0 +1,8 @@
+# unlock.sh - (Android root shell) ask logind to unlock the desktop user's session.  UNVERIFIED that this
+# helps here: the handoff notes this logind session type does not support lock-session (a real lock
+# screen then needs the password).  Prevention: [Daemon] Autolock=false in ~/.config/kscreenlockerrc.
+# Env: DS, CONTAINER, DESKTOP_USER.
+D=${DS:-/data/local/Droidspaces/bin/droidspaces}
+C=${CONTAINER:-arch-kde-mgl}
+U=${DESKTOP_USER:-swung0x48}
+$D --name=$C run bash -lc "S=\$(loginctl list-sessions --no-legend | awk '\$3==\"$U\"{print \$1}' | head -1); loginctl unlock-session \$S; echo unlocked \$S"
@@ -0,0 +1,47 @@
+# x11-measure.sh <label> <app-process-name> [seconds] - (Android root shell) CPU of the X11 path while
+# an X11 app runs: % of ONE core (800 = all 8 cores) for the whole phone, the app, Xwayland,
+# kwin_wayland, the app's :mobilegl server; mean kgsl gpu_busy_percentage; and the number of dma-buf
+# descriptors Xwayland and KWin hold (window pixmaps / client buffers as dma-bufs). Refuses to run
+# below 40% battery: the phone drains faster under load than USB charges it.
+# Run: su -c "sh /data/local/tmp/anl/x11-measure.sh glamor glxgears 10"      Env: PKG, MIN_BATTERY.
+L=${1:-sample}
+APP=${2:-glxgears}
+N=${3:-10}
+P=${PKG:-com.anland.consumer.mobilegl}
+GPU=/sys/class/kgsl/kgsl-3d0/gpu_busy_percentage
+level=$(dumpsys battery | sed -n 's/^ *level: *//p')
+if [ "${level:-0}" -lt "${MIN_BATTERY:-40}" ]; then
+    echo "battery ${level}% < ${MIN_BATTERY:-40}%: not loading the phone; charge first" >&2
+    exit 3
+fi
+ticks() {
+    t=0
+    for p in "$@"; do
+        [ -r /proc/$p/stat ] || continue
+        v=$(sed 's/^.*) //' /proc/$p/stat | awk '{print $12 + $13}')
+        t=$((t + v))
+    done
+    echo $t
+}
+total() { awk '/^cpu /{print $2+$3+$4+$6+$7+$8}' /proc/stat; }
+dmabufs() {   # dma-buf descriptors held by the pids (Android kernels link them to /dmabuf:...)
+    n=0
+    for p in "$@"; do n=$((n + $(ls -l /proc/$p/fd 2>/dev/null | grep -c dmabuf))); done
+    echo $n
+}
+AP="$(pidof $APP)"
+XW="$(pidof Xwayland)"
+KW="$(pidof kwin_wayland)"
+SV="$(pidof $P:mobilegl)"
+[ -n "$AP" ] || { echo "no running process named $APP" >&2; exit 2; }
+a0=$(total); p0=$(ticks $AP); x0=$(ticks $XW); k0=$(ticks $KW); v0=$(ticks $SV)
+g=0; i=0
+while [ $i -lt $N ]; do
+    sleep 1
+    b=$(cat $GPU 2>/dev/null | tr -dc '0-9'); g=$((g + ${b:-0})); i=$((i + 1))
+done
+a1=$(total); p1=$(ticks $AP); x1=$(ticks $XW); k1=$(ticks $KW); v1=$(ticks $SV)
+pct() { echo $(( ($2 - $1) / N )); }
+printf '%-12s %3ss  phone %4s%%  %s %3s%%  Xwayland %3s%%  kwin %3s%%  server %3s%%  gpu %3s%%  dmabufs Xwayland %s kwin %s  battery %s%%\n' \
+    "$L" "$N" "$(pct $a0 $a1)" "$APP" "$(pct $p0 $p1)" "$(pct $x0 $x1)" "$(pct $k0 $k1)" "$(pct $v0 $v1)" "$((g / N))" \
+    "$(dmabufs $XW)" "$(dmabufs $KW)" "$level"
@@ -0,0 +1,41 @@
+# env.sh - shared parameters for the host-side (Git Bash) scripts of the anland-mobilegl-plasma skill.
+# Sourced, never run. Every value can be overridden from the environment before calling a script.
+#
+#   MGL_WORKTREE    MobileGL worktree (Windows form C:/...). Default: the worktree this skill lives in.
+#   ANL_DIR         Local scratch dir: channel files, screenshots, xdist/. Default: %LOCALAPPDATA%/Temp/anl
+#   ANDROID_NDK     NDK used for the Android server lib. Default: <SDK>/ndk/27.2.12479018
+#   MGL_PKG         Anland APK that embeds the server. Default: com.anland.consumer.mobilegl
+#   MGL_CONTAINER   Droidspaces container. Default: arch-kde-mgl
+#   MGL_DESKTOP_USER  Desktop user inside the container. Default: swung0x48
+#   DS_BIN          droidspaces binary on the phone. Default: /data/local/Droidspaces/bin/droidspaces
+#   DEV_TOOLS       Phone dir for device/*.sh and channel files. Default: /data/local/tmp/anl
+#   DEV_SHARE       Phone dir bind-mounted into the container as /run/anland-mobilegl (screenshots,
+#                   server lib staging, display daemon socket). Default: /data/local/tmp/anland-mobilegl
+#   ANDROID_SERIAL  Honoured by adb itself; leave unset when only one device is attached
+#                   (USB serial or a wireless "ip:port").
+
+_skill_scripts=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+if [ -z "${MGL_WORKTREE:-}" ]; then
+    MGL_WORKTREE=$(cd "$_skill_scripts/../../../.." && pwd -W 2>/dev/null || pwd)
+fi
+: "${ANL_DIR:=$(cygpath -m "${LOCALAPPDATA:-C:/Users/$USERNAME/AppData/Local}" 2>/dev/null || echo /tmp)/Temp/anl}"
+: "${ANDROID_NDK:=$(cygpath -m "${LOCALAPPDATA:-C:/Users/$USERNAME/AppData/Local}" 2>/dev/null)/Android/Sdk/ndk/27.2.12479018}"
+: "${MGL_PKG:=com.anland.consumer.mobilegl}"
+: "${MGL_CONTAINER:=arch-kde-mgl}"
+: "${MGL_DESKTOP_USER:=swung0x48}"
+: "${DS_BIN:=/data/local/Droidspaces/bin/droidspaces}"
+: "${DEV_TOOLS:=/data/local/tmp/anl}"
+: "${DEV_SHARE:=/data/local/tmp/anland-mobilegl}"
+# The container rootfs as Android root sees it (the container's /tmp is NOT visible here).
+CT_ROOT=/mnt/Droidspaces/$MGL_CONTAINER
+export MGL_WORKTREE ANL_DIR ANDROID_NDK MGL_PKG MGL_CONTAINER MGL_DESKTOP_USER DS_BIN DEV_TOOLS DEV_SHARE CT_ROOT
+
+# Git Bash would rewrite /data/... arguments into C:/Program Files/Git/data/...; keep them literal.
+# Consequence: pass LOCAL paths to Windows tools (adb, cmake, git) in C:/ form, never /c/.
+export MSYS_NO_PATHCONV=1
+adb_() { MSYS2_ARG_CONV_EXCL='*' adb "$@"; }
+
+# C:/x/y -> /mnt/c/x/y (for WSL).
+win2wsl() { local p=$1 d; d=$(printf '%s' "${p:0:1}" | tr 'A-Z' 'a-z'); printf '/mnt/%s%s' "$d" "${p:2}"; }
+
+mkdir -p "$ANL_DIR"
@@ -0,0 +1,146 @@
+#!/usr/bin/env bash
+# kwin-recover.sh [rounds] - does the desktop keep working when KWin's OWN MobileGL session is lost?
+# Starts animated clients (glmark2-es2-wayland: Wayland, zero-copy shared images; glxgears: X11 through
+# Xwayland's GLX; with CHROME=1 also Chrome on an animated WebGL page), then per round loses the server
+# session of KWin with the pid-targeted debug knob (debug.mobilegl.inject_device_lost_pid, KWin's pid as
+# it sees itself in the container), waits, and checks.  A round PASSES when:
+#   - the server latched KWin's session (the server log file: "injected:pid=<kwin>"; logcat is too small);
+#   - KWin, plasmashell and every test client are the SAME processes (nothing restarted);
+#   - KWin restarted compositing on a fresh session (its journal: "compositing restarted on a fresh
+#     MobileGL session"; its client log: one more "RECOVERED");
+#   - the screen still animates (shot.sh hashes differ) and the clients keep presenting;
+#   - no native crash (logcat -b crash).
+# Between rounds the knob is reset to 0 while the clients run, which re-arms it for the same pid.
+# Stops early below MIN_BATTERY percent (the phone drains faster under load than USB charges it).
+#
+# Deploy first (same MobileGL commit on both sides; see SKILL.md):
+#   bash scripts/build-android.sh && bash scripts/cyc.sh pre      # server -> APK, restarts the session
+#   bash scripts/xbuild.sh && bash scripts/xdeploy.sh             # client -> container
+#   ANLAND_WORKTREE=<anland worktree> bash scripts/kwin-stage.sh build   # kwin.patch + anland backend
+#   (wait for "== ALL DONE ==" in /root/kwin-build.log), then: adb shell 'su -c "sh /data/local/tmp/anl/run-plasma.sh"'
+# Env: ROUNDS (default 2), SETTLE (s after the loss, default 10), REARM (s with the knob at 0 between
+#      rounds, default 4), MIN_BATTERY (default 40), CHROME (1: add Chrome), CHANNEL (ct.sh channel,
+#      default krec), TAG (screenshot/log tag, default krec).
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+ROUNDS=${1:-${ROUNDS:-2}}
+SETTLE=${SETTLE:-10}
+REARM=${REARM:-4}
+MIN_BATTERY=${MIN_BATTERY:-40}
+CH=${CHANNEL:-krec}
+TAG=${TAG:-krec}
+COMPOSITOR_LOG=/tmp/mobilegl-compositor.client.log
+SERVER_LOG=/data/data/$MGL_PKG/files/mobilegl-server.server.log
+
+ct() { bash "$here/ct.sh" "$CH"; }
+
+battery() { adb_ shell dumpsys battery | tr -d '\r' | awk '/ level:/ {print $2}'; }
+stop_clients() {
+    ct <<'IN' || true
+pkill -x glmark2-es2-way 2>/dev/null; pkill -x glmark2-es2-wayland 2>/dev/null; pkill -x glxgears 2>/dev/null
+true
+IN
+}
+cleanup() {
+    adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid 0'" || true
+    stop_clients
+}
+check_battery() {
+    local level; level=$(battery)
+    echo "battery: $level%"
+    if [ -n "$level" ] && [ "$level" -lt "$MIN_BATTERY" ]; then
+        echo "STOP: battery below $MIN_BATTERY%"; exit 3
+    fi
+}
+trap cleanup EXIT
+
+check_battery
+adb_ shell "su -c 'sh $DEV_TOOLS/status.sh'" || true
+adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid 0'"
+
+# The clients, each with its own log (mgrun's MGLOG), in the background for the whole run.
+ct <<'IN'
+rm -f /tmp/mgl-krec-gm.client.log /tmp/mgl-krec-gx.client.log /tmp/krec-gm.out /tmp/krec-gx.out
+MGLOG=/tmp/mgl-krec-gm.log nohup timeout 900 mgrun offscreen glmark2-es2-wayland -s 960x540 --run-forever \
+  > /tmp/krec-gm.out 2>&1 < /dev/null &
+MGLOG=/tmp/mgl-krec-gx.log nohup timeout 900 mgrun offscreen glxgears -info > /tmp/krec-gx.out 2>&1 < /dev/null &
+sleep 8
+echo "glmark2: $(pgrep -x glmark2-es2-way || pgrep -x glmark2-es2-wayland)  glxgears: $(pgrep -x glxgears)"
+grep -aE "presented through|could not be presented" /tmp/mgl-krec-gm.client.log | head -2
+grep -aE "presented through|readback" /tmp/mgl-krec-gx.client.log | head -2
+IN
+if [ "${CHROME:-0}" = 1 ]; then
+    ct <<'IN'
+cat > /tmp/mgl-krec-spin.html <<'HTML'
+<!doctype html><html><body style="margin:0;background:#000">
+<canvas id=c width=640 height=360 style="width:100vw;height:100vh"></canvas>
+<script>
+const gl = document.getElementById('c').getContext('webgl'); let t = 0;
+function frame() { t++; if (gl && !gl.isContextLost()) { gl.clearColor((t % 60) / 60, 0.2, 0.6, 1); gl.clear(gl.COLOR_BUFFER_BIT); }
+  requestAnimationFrame(frame); }
+frame();
+</script></body></html>
+HTML
+chmod 644 /tmp/mgl-krec-spin.html
+IN
+    WAIT=${WAIT:-20} CHANNEL=$CH bash "$here/chrome-launch.sh" "$TAG" --new-window "file:///tmp/mgl-krec-spin.html"
+fi
+
+# KWin, plasmashell and the clients, as pids inside the container; how often KWin recovered so far.
+snapshot() {
+    ct <<IN | tr -d '\r' | tail -1
+k=\$(pgrep -x kwin_wayland | head -1); p=\$(pgrep -x plasmashell | head -1)
+g=\$(pgrep -x glmark2-es2-way || pgrep -x glmark2-es2-wayland); x=\$(pgrep -x glxgears | head -1)
+c=\$(pgrep -x chrome | head -1)
+r=\$(grep -ac "RECOVERED" $COMPOSITOR_LOG 2>/dev/null || true)
+f=\$(grep -ac "FPS" /tmp/krec-gm.out 2>/dev/null || true)
+echo "\${k:-none} \${p:-none} \${g:-none} \${x:-none} \${c:-none} \${r:-0} \${f:-0}"
+IN
+}
+
+pass=0
+for round in $(seq 1 "$ROUNDS"); do
+    echo "=== round $round/$ROUNDS"
+    check_battery
+    read -r kwin plasma gm gx chrome recovered fps < <(snapshot)
+    echo "kwin=$kwin plasmashell=$plasma glmark2=$gm glxgears=$gx chrome=$chrome recoveries=$recovered fps-lines=$fps"
+    if [ "$kwin" = none ]; then echo "FAIL: KWin is not running"; break; fi
+    bash "$here/shot.sh" "$TAG-r$round-pre" 2 1
+    srvlines=$(adb_ shell "su -c 'wc -l < $SERVER_LOG'" | tr -dc '0-9')
+    adb_ logcat -b crash -c || true
+    adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid $kwin'"
+    sleep "$SETTLE"
+    adb_ shell "su -c 'setprop debug.mobilegl.inject_device_lost_pid 0'"
+    bash "$here/shot.sh" "$TAG-r$round-post" 3 1 | tee "$ANL_DIR/$TAG-r$round-post.md5"
+    read -r kwin2 plasma2 gm2 gx2 chrome2 recovered2 fps2 < <(snapshot)
+    echo "after: kwin=$kwin2 plasmashell=$plasma2 glmark2=$gm2 glxgears=$gx2 chrome=$chrome2 recoveries=$recovered2 fps-lines=$fps2"
+    adb_ shell "su -c 'tail -n +$((srvlines + 1)) $SERVER_LOG'" > "$ANL_DIR/$TAG-r$round-server.log"
+    latched=$(grep -c "injected:pid=$kwin" "$ANL_DIR/$TAG-r$round-server.log" || true)
+    grep -aE "inject_device_lost_pid|SessionLatch|session #[0-9]+ (started|pid)|owner=server|lease" \
+        "$ANL_DIR/$TAG-r$round-server.log" | cut -c1-220 | head -14 || true
+    ct <<IN || true
+echo "--- KWin's MobileGL client log:"; grep -aE "DEVICE LOST|RECOVERED|GL_UNKNOWN_CONTEXT_RESET|re-created on the fresh|Refuse" $COMPOSITOR_LOG | tail -6
+echo "--- KWin journal:"; journalctl --no-pager --since -40s 2>/dev/null | grep -aE "kwin_wayland" | grep -aiE "reset|restarted|MobileGL swap|context|fatal|abort" | tail -10
+echo "--- clients:"; grep -aE "DEVICE LOST|lost|Fatal" /tmp/mgl-krec-gm.client.log /tmp/mgl-krec-gx.client.log | tail -4
+IN
+    crashes=$(adb_ logcat -b crash -d | grep -c "Fatal signal\|Abort message" || true)
+    hashes=$(awk '{print $1}' "$ANL_DIR/$TAG-r$round-post.md5" | sort -u | wc -l)
+    verdict=PASS
+    [ "$latched" -ge 1 ] || { echo "  - the server never latched KWin's session (pid $kwin)"; verdict=FAIL; }
+    [ "$kwin2" = "$kwin" ] || { echo "  - KWin is a different process ($kwin -> $kwin2): it restarted"; verdict=FAIL; }
+    [ "$plasma2" = "$plasma" ] || { echo "  - plasmashell restarted ($plasma -> $plasma2)"; verdict=FAIL; }
+    [ "$gm2" = "$gm" ] && [ "$gx2" = "$gx" ] || { echo "  - a test client restarted or died"; verdict=FAIL; }
+    [ "${CHROME:-0}" != 1 ] || [ "$chrome2" = "$chrome" ] || { echo "  - Chrome restarted"; verdict=FAIL; }
+    [ "$recovered2" -gt "$recovered" ] || { echo "  - KWin's client log shows no new RECOVERED session"; verdict=FAIL; }
+    [ "$fps2" -gt "$fps" ] || { echo "  - glmark2 stopped reporting frames"; verdict=FAIL; }
+    [ "$hashes" -ge 2 ] || { echo "  - the screen does not change after the loss (frozen)"; verdict=FAIL; }
+    [ "$crashes" -eq 0 ] || { echo "  - native crash(es):"; adb_ logcat -b crash -d | tail -8; verdict=FAIL; }
+    echo "round $round: $verdict"
+    [ "$verdict" = PASS ] && pass=$((pass + 1))
+    [ "$verdict" = PASS ] || break
+    # Re-arm the knob for the same pid: a session must observe 0 before the next value counts.
+    sleep "$REARM"
+done
+echo "=== $pass/$ROUNDS rounds passed; screenshots in $ANL_DIR/$TAG-r*-*.png"
+[ "$pass" = "$ROUNDS" ]
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# kwin-stage.sh [build] - stage anland's KWin changes (kwin.patch, anland backend sources,
+# mobilegl-startup.sh, sync-build-kwin.sh) for the in-container KWin rebuild; with "build", also run
+# kwprep.sh (apply changed patch hunks / backend files to the live tree) and start sync-build-kwin.sh
+# (ninja -j4 + atomic install into /opt/mobilegl/kwin, screencast plugin, startup script) in the
+# background, log /root/kwin-build.log.  KWin is NOT restarted: run device/run-plasma.sh afterwards.
+# Poll:  printf 'tail -5 /root/kwin-build.log\n' | bash ct.sh kw
+#
+# Two staging places, because the two container scripts read different ones:
+#   /root/kwsync.tgz                            <- kwprep.sh (installed by push-tools.sh)
+#   $DEV_SHARE/kwin-sync = /run/anland-mobilegl/kwin-sync  <- sync-build-kwin.sh (as kwin-sync/sync-build.sh)
+# All files are CR-stripped: the Windows checkout of anland has CRLF backend sources and a CRLF
+# sync-build-kwin.sh (bash then fails on $'\r').
+# Assembled from kwprep.sh + sync-build-kwin.sh; not exercised as one unit yet.
+# Env: ANLAND_WORKTREE (anland checkout on branch legacy-mobilegl-unified, Windows form),
+#      CHANNEL (ct.sh channel, default kw).
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+: "${ANLAND_WORKTREE:=$(cygpath -m "$USERPROFILE" 2>/dev/null)/AndroidStudioProjects/anland/.claude/worktrees/mgl-unified}"
+K=$ANLAND_WORKTREE/producers/kde
+st=$ANL_DIR/kwstage
+rm -rf "$st"; mkdir -p "$st/kwsync/Arch_v5" "$st/kwsync/anland_backend_Arch_v5/src/backends/anland" "$st/kwin-sync/anland" "$st/kwin-sync/misc"
+lf() { tr -d '\r' < "$1" > "$2"; }
+for f in "$K"/anland_backend_Arch_v5/src/backends/anland/*; do
+    b=$(basename "$f"); [ "$b" = libdisplay_producer ] && continue
+    lf "$f" "$st/kwsync/anland_backend_Arch_v5/src/backends/anland/$b"
+    lf "$f" "$st/kwin-sync/anland/$b"
+done
+for f in kwin.patch mobilegl-startup.sh desktop-session-mobilegl.conf; do
+    lf "$K/Arch_v5/$f" "$st/kwsync/Arch_v5/$f"
+    lf "$K/Arch_v5/$f" "$st/kwin-sync/misc/$f"
+done
+lf "$K/Arch_v5/sync-build-kwin.sh" "$st/kwin-sync/sync-build.sh"
+(cd "$st/kwsync" && tar --force-local -czf ../kwsync.tgz Arch_v5 anland_backend_Arch_v5)
+tar --force-local -C "$st" -czf "$st/kwin-sync.tgz" kwin-sync
+adb_ push "$st/kwsync.tgz" "$DEV_TOOLS/kwsync.tgz" >/dev/null
+adb_ push "$st/kwin-sync.tgz" "$DEV_TOOLS/kwin-sync.tgz" >/dev/null
+adb_ shell "su -c 'cp $DEV_TOOLS/kwsync.tgz $CT_ROOT/root/kwsync.tgz && tar xzf $DEV_TOOLS/kwin-sync.tgz -C $DEV_SHARE && chmod -R a+rX $DEV_SHARE/kwin-sync && rm -f $DEV_TOOLS/kwsync.tgz $DEV_TOOLS/kwin-sync.tgz && echo staged'"
+if [ "${1:-}" = build ]; then
+    bash "$here/ct.sh" "${CHANNEL:-kw}" <<'IN'
+bash /root/kwprep.sh || exit 1
+nohup bash /run/anland-mobilegl/kwin-sync/sync-build.sh > /root/kwin-build.log 2>&1 < /dev/null &
+echo "kwin build started (pid $!), log /root/kwin-build.log; it ends with '== ALL DONE =='"
+IN
+fi
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+# cdpmedia.py [port] [seconds] - ask a Chrome started with --remote-debugging-port=<port> (and a
+# non-default --user-data-dir) which video decoder its first page's players use, and evaluate
+# `window.__mglStats` / the element `#t` text of the page. Standard library only (no websocket
+# module in the container): a minimal RFC 6455 client.
+#   python3 cdpmedia.py 9333 6
+# Prints kVideoDecoderName / kIsPlatformVideoDecoder / kVideoTracks properties and media errors as
+# they arrive, then the page's stats text.
+import base64, json, os, socket, struct, sys, time, urllib.request
+
+port = int(sys.argv[1]) if len(sys.argv) > 1 else 9333
+seconds = float(sys.argv[2]) if len(sys.argv) > 2 else 6.0
+targets = json.load(urllib.request.urlopen(f"http://127.0.0.1:{port}/json"))
+page = next(t for t in targets if t.get("type") == "page")
+url = page["webSocketDebuggerUrl"]
+path = url.split(f"127.0.0.1:{port}", 1)[1]
+sock = socket.create_connection(("127.0.0.1", port))
+key = base64.b64encode(os.urandom(16)).decode()
+sock.sendall((f"GET {path} HTTP/1.1\r\nHost: 127.0.0.1:{port}\r\nUpgrade: websocket\r\nConnection: Upgrade\r\n"
+              f"Sec-WebSocket-Key: {key}\r\nSec-WebSocket-Version: 13\r\n\r\n").encode())
+response = b""
+while b"\r\n\r\n" not in response:
+    response += sock.recv(4096)
+buffer = response.split(b"\r\n\r\n", 1)[1]
+
+
+def send(message):
+    data = json.dumps(message).encode()
+    header = bytes([0x81])
+    mask = os.urandom(4)
+    if len(data) < 126:
+        header += bytes([0x80 | len(data)])
+    elif len(data) < 65536:
+        header += bytes([0x80 | 126]) + struct.pack(">H", len(data))
+    else:
+        header += bytes([0x80 | 127]) + struct.pack(">Q", len(data))
+    sock.sendall(header + mask + bytes(b ^ mask[i % 4] for i, b in enumerate(data)))
+
+
+def recv_exact(n):
+    global buffer
+    while len(buffer) < n:
+        chunk = sock.recv(65536)
+        if not chunk:
+            raise EOFError
+        buffer += chunk
+    out, buffer = buffer[:n], buffer[n:]
+    return out
+
+
+def recv():
+    first, second = recv_exact(2)
+    length = second & 0x7F
+    if length == 126:
+        length = struct.unpack(">H", recv_exact(2))[0]
+    elif length == 127:
+        length = struct.unpack(">Q", recv_exact(8))[0]
+    payload = recv_exact(length)
+    return json.loads(payload) if first & 0x0F == 1 else None
+
+
+send({"id": 1, "method": "Media.enable"})
+# Players created before the domain was enabled report nothing: reload so the page makes new ones.
+if os.environ.get("CDP_NAVIGATE"):
+    send({"id": 3, "method": "Page.navigate", "params": {"url": os.environ["CDP_NAVIGATE"]}})
+elif os.environ.get("CDP_RELOAD", "1") == "1":
+    send({"id": 3, "method": "Page.reload"})
+deadline = time.time() + seconds
+sock.settimeout(0.5)
+seen = set()
+while time.time() < deadline:
+    try:
+        message = recv()
+    except socket.timeout:
+        continue
+    if not message or "method" not in message:
+        continue
+    if message["method"] == "Media.playerPropertiesChanged":
+        for prop in message["params"]["properties"]:
+            if prop["name"] in ("kVideoDecoderName", "kIsPlatformVideoDecoder", "kVideoTracks", "kResolution"):
+                line = f"{prop['name']} = {prop['value']}"
+                if line not in seen:
+                    seen.add(line)
+                    print(line)
+    elif message["method"] in ("Media.playerErrorsRaised",):
+        print("media error:", json.dumps(message["params"])[:300])
+send({"id": 2, "method": "Runtime.evaluate",
+      "params": {"expression": os.environ.get("CDP_EXPR", "(document.getElementById('t')||{}).textContent"), "returnByValue": True}})
+sock.settimeout(3)
+while True:
+    message = recv()
+    if message and message.get("id") == 2:
+        print("page:", message.get("result", {}).get("result", {}).get("value"))
+        break
@@ -0,0 +1,34 @@
+// dmahash <pid> : for each dma-buf fd of <pid> at least 1 MB, print inode, size and a sampled checksum.
+#include <dirent.h>
+#include <fcntl.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+int main(int argc, char** argv) {
+    char dir[64];
+    snprintf(dir, sizeof dir, "/proc/%s/fd", argv[1]);
+    DIR* d = opendir(dir);
+    if (!d) { perror("opendir"); return 1; }
+    struct dirent* e;
+    while ((e = readdir(d))) {
+        char p[128], l[128] = {0};
+        snprintf(p, sizeof p, "%s/%s", dir, e->d_name);
+        if (readlink(p, l, sizeof l - 1) <= 0 || strncmp(l, "/dmabuf", 7) != 0) continue;
+        int fd = open(p, O_RDONLY);
+        if (fd < 0) continue;
+        struct stat st; fstat(fd, &st);
+        if (st.st_size < (1 << 20)) { close(fd); continue; }
+        const uint8_t* m = mmap(NULL, st.st_size, PROT_READ, MAP_SHARED, fd, 0);
+        if (m == MAP_FAILED) { printf("fd %s ino %lu mmap failed\n", e->d_name, (unsigned long)st.st_ino); close(fd); continue; }
+        uint64_t h = 1469598103934665603ull; unsigned nz = 0;
+        for (off_t i = 0; i < st.st_size; i += 64) { h = (h ^ m[i]) * 1099511628211ull; nz += m[i] != 0; }
+        printf("fd %s ino %lu size %ld hash %016llx nz %u\n", e->d_name, (unsigned long)st.st_ino, (long)st.st_size,
+               (unsigned long long)h, nz);
+        munmap((void*)m, st.st_size); close(fd);
+    }
+    return 0;
+}
@@ -0,0 +1,55 @@
+// gpuprobe - a plain Android process (outside MobileGL) that measures whether the GPU keeps executing
+// other contexts' work: a pbuffer GLES context clears + draws and glFinish()es every ~10 ms, printing any
+// round trip over 50 ms with a CLOCK_REALTIME timestamp.  Usage: gpuprobe <seconds> [priority: low|med|high]
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+#include <GLES3/gl3.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <unistd.h>
+
+#ifndef EGL_CONTEXT_PRIORITY_LEVEL_IMG
+#define EGL_CONTEXT_PRIORITY_LEVEL_IMG 0x3100
+#define EGL_CONTEXT_PRIORITY_HIGH_IMG 0x3101
+#define EGL_CONTEXT_PRIORITY_MEDIUM_IMG 0x3102
+#define EGL_CONTEXT_PRIORITY_LOW_IMG 0x3103
+#endif
+
+static double now_ms(void) { struct timespec ts; clock_gettime(CLOCK_REALTIME, &ts); return ts.tv_sec * 1e3 + ts.tv_nsec / 1e6; }
+
+int main(int argc, char** argv) {
+    int seconds = argc > 1 ? atoi(argv[1]) : 20;
+    EGLint prio = EGL_CONTEXT_PRIORITY_MEDIUM_IMG;
+    if (argc > 2 && !strcmp(argv[2], "high")) prio = EGL_CONTEXT_PRIORITY_HIGH_IMG;
+    if (argc > 2 && !strcmp(argv[2], "low")) prio = EGL_CONTEXT_PRIORITY_LOW_IMG;
+    setvbuf(stdout, NULL, _IOLBF, 0);
+    EGLDisplay d = eglGetDisplay(EGL_DEFAULT_DISPLAY);
+    eglInitialize(d, NULL, NULL);
+    const EGLint ca[] = {EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT, EGL_RED_SIZE, 8, EGL_NONE};
+    EGLConfig c; EGLint n;
+    eglChooseConfig(d, ca, &c, 1, &n);
+    const EGLint pa[] = {EGL_WIDTH, 256, EGL_HEIGHT, 256, EGL_NONE};
+    EGLSurface s = eglCreatePbufferSurface(d, c, pa);
+    const EGLint xa[] = {EGL_CONTEXT_CLIENT_VERSION, 3, EGL_CONTEXT_PRIORITY_LEVEL_IMG, prio, EGL_NONE};
+    EGLContext x = eglCreateContext(d, c, EGL_NO_CONTEXT, xa);
+    if (x == EGL_NO_CONTEXT) { printf("no ctx 0x%x\n", eglGetError()); return 1; }
+    eglMakeCurrent(d, s, s, x);
+    EGLint got = 0; eglQueryContext(d, x, EGL_CONTEXT_PRIORITY_LEVEL_IMG, &got);
+    const double t0 = now_ms(); double maxRt = 0; long iters = 0, slow = 0;
+    printf("[probe] start %.0f prio=0x%x\n", t0, got);
+    while (now_ms() - t0 < seconds * 1000.0) {
+        const double a = now_ms();
+        glClearColor((iters & 1) ? 1.f : 0.f, 0.f, 0.f, 1.f);
+        glClear(GL_COLOR_BUFFER_BIT);
+        glFinish();
+        const double rt = now_ms() - a;
+        if (rt > maxRt) maxRt = rt;
+        if (rt > 50) { ++slow; printf("[probe] glFinish %.0f ms ending at %.0f\n", rt, now_ms()); }
+        ++iters;
+        usleep(10000);
+    }
+    printf("[probe] END iters=%ld max=%.0f ms slow=%ld\n", iters, maxRt, slow);
+    return 0;
+}
@@ -0,0 +1,150 @@
+// hangclient - a Wayland xdg-toplevel, EGL + desktop GL 3.3 core client that animates and, on request,
+// submits fragment work that never ends (the driver has to reset it).  Prints per-frame gaps so the same
+// binary measures "other app" frame pacing (frame callbacks come from the compositor) when run with -H -1.
+//
+//   hangclient -T tag [-H hang_at_frame] [-N hang_repeat] [-I iters] [-S seconds] [-W w] [-X h] [-F]
+//     -H n   frame at which the hanging draw is submitted (-1: never; default -1)
+//     -N k   how many consecutive hanging frames (default 1)
+//     -E e   re-arm a hang every e frames after the first (0: once)
+//     -I i   loop count of the hanging shader (default 2^31-1)
+//     -S s   run time in seconds (default 30)
+//     -F     glFinish after every frame (puts the wait inside the client's own session)
+#define GL_GLEXT_PROTOTYPES 1
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+#include <GL/glcorearb.h>
+#include <wayland-client.h>
+#include <wayland-egl.h>
+#include "xdg-shell-client-protocol.h"
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <unistd.h>
+
+static struct wl_compositor* g_comp;
+static struct xdg_wm_base* g_wm;
+static int g_configured, g_closed;
+static int g_w = 480, g_h = 360;
+
+static void reg_global(void* d, struct wl_registry* r, uint32_t name, const char* iface, uint32_t ver) {
+    (void)d;
+    if (!strcmp(iface, "wl_compositor")) g_comp = wl_registry_bind(r, name, &wl_compositor_interface, 4);
+    else if (!strcmp(iface, "xdg_wm_base")) g_wm = wl_registry_bind(r, name, &xdg_wm_base_interface, 1);
+}
+static void reg_remove(void* d, struct wl_registry* r, uint32_t n) { (void)d; (void)r; (void)n; }
+static const struct wl_registry_listener reg_l = {reg_global, reg_remove};
+static void wm_ping(void* d, struct xdg_wm_base* wm, uint32_t s) { (void)d; xdg_wm_base_pong(wm, s); }
+static const struct xdg_wm_base_listener wm_l = {wm_ping};
+static void xs_conf(void* d, struct xdg_surface* s, uint32_t serial) { (void)d; xdg_surface_ack_configure(s, serial); g_configured = 1; }
+static const struct xdg_surface_listener xs_l = {xs_conf};
+static void tl_conf(void* d, struct xdg_toplevel* t, int32_t w, int32_t h, struct wl_array* st) { (void)d; (void)t; (void)st; (void)w; (void)h; }
+static void tl_close(void* d, struct xdg_toplevel* t) { (void)d; (void)t; g_closed = 1; }
+static const struct xdg_toplevel_listener tl_l = {tl_conf, tl_close};
+
+static double now_ms(void) { struct timespec ts; clock_gettime(CLOCK_REALTIME, &ts); return ts.tv_sec * 1e3 + ts.tv_nsec / 1e6; }
+
+static GLuint sh(GLenum type, const char* src) {
+    GLuint s = glCreateShader(type);
+    glShaderSource(s, 1, &src, NULL);
+    glCompileShader(s);
+    GLint ok = 0; glGetShaderiv(s, GL_COMPILE_STATUS, &ok);
+    if (!ok) { char log[2048]; glGetShaderInfoLog(s, sizeof log, NULL, log); fprintf(stderr, "shader: %s\n", log); exit(2); }
+    return s;
+}
+
+int main(int argc, char** argv) {
+    const char* tag = "hc"; int hangAt = -1, hangN = 1, every = 0, seconds = 30, finish = 0; long iters = 2147483647L;
+    for (int i = 1; i < argc; ++i) {
+        if (!strcmp(argv[i], "-T") && i + 1 < argc) tag = argv[++i];
+        else if (!strcmp(argv[i], "-H") && i + 1 < argc) hangAt = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-N") && i + 1 < argc) hangN = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-E") && i + 1 < argc) every = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-I") && i + 1 < argc) iters = atol(argv[++i]);
+        else if (!strcmp(argv[i], "-S") && i + 1 < argc) seconds = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-W") && i + 1 < argc) g_w = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-X") && i + 1 < argc) g_h = atoi(argv[++i]);
+        else if (!strcmp(argv[i], "-F")) finish = 1;
+    }
+    setvbuf(stdout, NULL, _IOLBF, 0);
+    struct wl_display* dpy = wl_display_connect(NULL);
+    if (!dpy) { fprintf(stderr, "no wayland\n"); return 1; }
+    struct wl_registry* reg = wl_display_get_registry(dpy);
+    wl_registry_add_listener(reg, &reg_l, NULL);
+    wl_display_roundtrip(dpy);
+    if (!g_comp || !g_wm) { fprintf(stderr, "no compositor/xdg_wm_base\n"); return 1; }
+    xdg_wm_base_add_listener(g_wm, &wm_l, NULL);
+    struct wl_surface* surf = wl_compositor_create_surface(g_comp);
+    struct xdg_surface* xs = xdg_wm_base_get_xdg_surface(g_wm, surf);
+    xdg_surface_add_listener(xs, &xs_l, NULL);
+    struct xdg_toplevel* tl = xdg_surface_get_toplevel(xs);
+    xdg_toplevel_add_listener(tl, &tl_l, NULL);
+    xdg_toplevel_set_title(tl, tag);
+    wl_surface_commit(surf);
+    while (!g_configured) wl_display_dispatch(dpy);
+
+    EGLDisplay ed = eglGetPlatformDisplay(EGL_PLATFORM_WAYLAND_KHR, dpy, NULL);
+    if (!eglInitialize(ed, NULL, NULL)) { fprintf(stderr, "eglInitialize\n"); return 1; }
+    eglBindAPI(EGL_OPENGL_API);
+    const EGLint ca[] = {EGL_SURFACE_TYPE, EGL_WINDOW_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT, EGL_RED_SIZE, 8, EGL_GREEN_SIZE, 8, EGL_BLUE_SIZE, 8, EGL_NONE};
+    EGLConfig cfg; EGLint n = 0;
+    if (!eglChooseConfig(ed, ca, &cfg, 1, &n) || n < 1) { fprintf(stderr, "no config\n"); return 1; }
+    const EGLint xa[] = {EGL_CONTEXT_MAJOR_VERSION, 3, EGL_CONTEXT_MINOR_VERSION, 3, EGL_CONTEXT_OPENGL_PROFILE_MASK, EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT, EGL_NONE};
+    EGLContext ctx = eglCreateContext(ed, cfg, EGL_NO_CONTEXT, xa);
+    if (ctx == EGL_NO_CONTEXT) { fprintf(stderr, "no context 0x%x\n", eglGetError()); return 1; }
+    struct wl_egl_window* ew = wl_egl_window_create(surf, g_w, g_h);
+    EGLSurface es = eglCreatePlatformWindowSurface(ed, cfg, ew, NULL);
+    eglMakeCurrent(ed, es, es, ctx);
+    eglSwapInterval(ed, 1);
+
+    const char* vs = "#version 330 core\nlayout(location=0) in vec2 p; uniform float t; out vec2 uv;\n"
+                     "void main(){ uv=p*0.5+0.5; gl_Position=vec4(p*0.4+vec2(0.5*sin(t),0.0),0.0,1.0);}";
+    const char* fs = "#version 330 core\nin vec2 uv; uniform int iters; uniform float t; out vec4 c;\n"
+                     "void main(){ float a=uv.x; for(int i=0;i<iters;++i){ a=fract(a*1.0001+0.37+uv.y); }\n"
+                     " c=vec4(uv, 0.5+0.5*sin(t)+a*1e-6, 1.0);}";
+    GLuint prog = glCreateProgram();
+    glAttachShader(prog, sh(GL_VERTEX_SHADER, vs));
+    glAttachShader(prog, sh(GL_FRAGMENT_SHADER, fs));
+    glLinkProgram(prog);
+    GLint ok = 0; glGetProgramiv(prog, GL_LINK_STATUS, &ok);
+    if (!ok) { fprintf(stderr, "link failed\n"); return 2; }
+    GLint uIt = glGetUniformLocation(prog, "iters"), uT = glGetUniformLocation(prog, "t");
+    GLuint vao, vbo; glGenVertexArrays(1, &vao); glBindVertexArray(vao);
+    const float q[] = {-1, -1, 1, -1, -1, 1, 1, 1};
+    glGenBuffers(1, &vbo); glBindBuffer(GL_ARRAY_BUFFER, vbo); glBufferData(GL_ARRAY_BUFFER, sizeof q, q, GL_STATIC_DRAW);
+    glEnableVertexAttribArray(0); glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 0, 0);
+
+    const double t0 = now_ms(); double last = t0, maxGap = 0; long frames = 0, gaps100 = 0;
+    printf("[%s] start %.0f renderer=%s hangAt=%d N=%d every=%d iters=%ld\n", tag, t0, glGetString(GL_RENDERER), hangAt, hangN, every, iters);
+    while (!g_closed && now_ms() - t0 < seconds * 1000.0) {
+        wl_display_dispatch_pending(dpy);
+        const double t = (now_ms() - t0) / 1000.0;
+        int hang = 0;
+        if (hangAt >= 0 && frames >= hangAt) {
+            long rel = frames - hangAt;
+            if (every > 0) rel %= every;
+            hang = rel < hangN;
+        }
+        glViewport(0, 0, g_w, g_h);
+        glClearColor(0.5f + 0.5f * (float)sin(t * 3.0), 0.2f, 0.4f, 1.0f);
+        glClear(GL_COLOR_BUFFER_BIT);
+        glUseProgram(prog);
+        glUniform1i(uIt, hang ? (int)iters : 0);
+        glUniform1f(uT, (float)t);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        if (hang) printf("[%s] HANG submitted frame %ld at %.0f\n", tag, frames, now_ms());
+        if (finish) glFinish();
+        if (!eglSwapBuffers(ed, es)) printf("[%s] swap failed 0x%x at %.0f\n", tag, eglGetError(), now_ms());
+        const GLenum rs = glGetError();
+        if (rs != GL_NO_ERROR) printf("[%s] glGetError 0x%x at %.0f\n", tag, rs, now_ms());
+        const double nowv = now_ms(), gap = nowv - last;
+        last = nowv; ++frames;
+        if (gap > maxGap) maxGap = gap;
+        if (gap > 100.0) { ++gaps100; printf("[%s] GAP %.0f ms ending at %.0f (frame %ld)\n", tag, gap, nowv, frames); }
+        if (frames % 120 == 0) printf("[%s] frame %ld t=%.1fs fps=%.1f\n", tag, frames, (nowv - t0) / 1e3, frames * 1e3 / (nowv - t0));
+    }
+    const double total = now_ms() - t0;
+    printf("[%s] END frames=%ld fps=%.1f maxGap=%.0f ms gaps>100ms=%ld\n", tag, frames, frames * 1e3 / total, maxGap, gaps100);
+    return 0;
+}
@@ -0,0 +1,95 @@
+/* texfmt: does a texture keep its alpha? One opaque red texel and one transparent texel per
+ * format, sampled across a 64x64 ES3 pbuffer cleared white with SRC_ALPHA blending. Expect the
+ * left half red and the right half WHITE; a black right half means the alpha was lost. */
+#include <EGL/egl.h>
+#include <GLES3/gl3.h>
+#include <GLES2/gl2ext.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#define P(T, n) T n = (T)eglGetProcAddress(#n)
+static const char* vs = "#version 300 es\nlayout(location=0) in vec2 p; out vec2 uv; void main(){uv=p*0.5+0.5; gl_Position=vec4(p,0,1);}";
+static const char* fs = "#version 300 es\nprecision mediump float; uniform sampler2D t; uniform int mode; in vec2 uv; out vec4 o;"
+                        "void main(){vec4 c=texture(t,uv); o = mode==1 ? vec4(1.0,0.0,0.0,c.a) : (mode==2 ? vec4(c.r,0.0,0.0,c.r) : c);}";
+int main(void) {
+    EGLDisplay d = eglGetDisplay(EGL_DEFAULT_DISPLAY); eglInitialize(d, 0, 0); eglBindAPI(EGL_OPENGL_ES_API);
+    EGLint ca[] = {EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT, EGL_RED_SIZE, 8, EGL_GREEN_SIZE, 8, EGL_BLUE_SIZE, 8, EGL_ALPHA_SIZE, 8, EGL_NONE};
+    EGLConfig c; EGLint n; eglChooseConfig(d, ca, &c, 1, &n);
+    EGLint xa[] = {EGL_CONTEXT_CLIENT_VERSION, 3, EGL_NONE};
+    EGLContext x = eglCreateContext(d, c, EGL_NO_CONTEXT, xa);
+    EGLint pa[] = {EGL_WIDTH, 64, EGL_HEIGHT, 64, EGL_NONE};
+    EGLSurface s = eglCreatePbufferSurface(d, c, pa); eglMakeCurrent(d, s, s, x);
+    P(PFNGLCREATESHADERPROC, glCreateShader); P(PFNGLSHADERSOURCEPROC, glShaderSource); P(PFNGLCOMPILESHADERPROC, glCompileShader);
+    P(PFNGLCREATEPROGRAMPROC, glCreateProgram); P(PFNGLATTACHSHADERPROC, glAttachShader); P(PFNGLLINKPROGRAMPROC, glLinkProgram);
+    P(PFNGLUSEPROGRAMPROC, glUseProgram); P(PFNGLGETUNIFORMLOCATIONPROC, glGetUniformLocation); P(PFNGLUNIFORM1IPROC, glUniform1i);
+    P(PFNGLGENBUFFERSPROC, glGenBuffers); P(PFNGLBINDBUFFERPROC, glBindBuffer); P(PFNGLBUFFERDATAPROC, glBufferData);
+    P(PFNGLVERTEXATTRIBPOINTERPROC, glVertexAttribPointer); P(PFNGLENABLEVERTEXATTRIBARRAYPROC, glEnableVertexAttribArray);
+    P(PFNGLDRAWARRAYSPROC, glDrawArrays); P(PFNGLCLEARCOLORPROC, glClearColor); P(PFNGLCLEARPROC, glClear);
+    P(PFNGLREADPIXELSPROC, glReadPixels); P(PFNGLVIEWPORTPROC, glViewport); P(PFNGLGENTEXTURESPROC, glGenTextures);
+    P(PFNGLBINDTEXTUREPROC, glBindTexture); P(PFNGLTEXIMAGE2DPROC, glTexImage2D); P(PFNGLTEXPARAMETERIPROC, glTexParameteri);
+    P(PFNGLENABLEPROC, glEnable); P(PFNGLBLENDFUNCPROC, glBlendFunc); P(PFNGLGETERRORPROC, glGetError);
+    P(PFNGLTEXSTORAGE2DPROC, glTexStorage2D); P(PFNGLTEXSUBIMAGE2DPROC, glTexSubImage2D); P(PFNGLGETSTRINGPROC, glGetString);
+    printf("GL_VERSION %s\n", glGetString(GL_VERSION));
+    printf("BGRA ext: %s\n", strstr((const char*)glGetString(GL_EXTENSIONS), "GL_EXT_texture_format_BGRA8888") ? "yes" : "no");
+    GLuint v = glCreateShader(GL_VERTEX_SHADER), f = glCreateShader(GL_FRAGMENT_SHADER);
+    glShaderSource(v, 1, &vs, 0); glCompileShader(v); glShaderSource(f, 1, &fs, 0); glCompileShader(f);
+    GLuint pr = glCreateProgram(); glAttachShader(pr, v); glAttachShader(pr, f); glLinkProgram(pr); glUseProgram(pr);
+    float q[] = {-1,-1, 1,-1, -1,1, 1,1};
+    GLuint vbo; glGenBuffers(1, &vbo); glBindBuffer(GL_ARRAY_BUFFER, vbo);
+    glBufferData(GL_ARRAY_BUFFER, sizeof q, q, GL_STATIC_DRAW); glVertexAttribPointer(0, 2, GL_FLOAT, 0, 0, 0); glEnableVertexAttribArray(0);
+    GLint lm = glGetUniformLocation(pr, "mode");
+    glEnable(GL_BLEND); glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
+    struct { const char* name; GLint ifmt; GLenum fmt; int bpp; int mode; int storage; GLenum sized; unsigned char px[8]; } cases[] = {
+        {"RGBA/TexImage", GL_RGBA, GL_RGBA, 4, 0, 0, 0, {255,0,0,255, 0,0,0,0}},
+        {"BGRA_EXT/TexImage", GL_BGRA_EXT, GL_BGRA_EXT, 4, 0, 0, 0, {0,0,255,255, 0,0,0,0}},
+        {"BGRA8_EXT/TexStorage+Sub", 0, GL_BGRA_EXT, 4, 0, 1, GL_BGRA8_EXT, {0,0,255,255, 0,0,0,0}},
+        {"RGBA8/TexStorage+Sub", 0, GL_RGBA, 4, 0, 1, GL_RGBA8, {255,0,0,255, 0,0,0,0}},
+        {"ALPHA", GL_ALPHA, GL_ALPHA, 1, 1, 0, 0, {255, 0}},
+        {"LUMINANCE_ALPHA", GL_LUMINANCE_ALPHA, GL_LUMINANCE_ALPHA, 2, 1, 0, 0, {255,255, 0,0}},
+        {"R8/TexStorage+Sub", 0, GL_RED, 1, 2, 1, GL_R8, {255, 0}},
+    };
+    for (unsigned i = 0; i < sizeof cases / sizeof cases[0]; ++i) {
+        GLuint t; glGenTextures(1, &t); glBindTexture(GL_TEXTURE_2D, t);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        while (glGetError()) {}
+        if (cases[i].storage) {
+            glTexStorage2D(GL_TEXTURE_2D, 1, cases[i].sized, 2, 1);
+            glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 2, 1, cases[i].fmt, GL_UNSIGNED_BYTE, cases[i].px);
+        } else {
+            glTexImage2D(GL_TEXTURE_2D, 0, cases[i].ifmt, 2, 1, 0, cases[i].fmt, GL_UNSIGNED_BYTE, cases[i].px);
+        }
+        GLenum err = glGetError();
+        glUniform1i(lm, cases[i].mode);
+        glViewport(0, 0, 64, 64); glClearColor(1, 1, 1, 1); glClear(GL_COLOR_BUFFER_BIT);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        unsigned char l[4], r[4];
+        glReadPixels(16, 32, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, l); glReadPixels(48, 32, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, r);
+        const int ok = l[0] > 200 && l[1] < 60 && l[2] < 60 && r[0] > 200 && r[1] > 200 && r[2] > 200;
+        printf("%-26s err=0x%04x left=%3d,%3d,%3d right=%3d,%3d,%3d %s\n", cases[i].name, err, l[0], l[1], l[2], r[0], r[1], r[2], ok ? "ok" : "BAD");
+    }
+    /* ANGLE's emulation of ALPHA8 / LUMINANCE_ALPHA8 on a desktop-GL driver: R8 / RG8 + swizzle. */
+    P(PFNGLTEXPARAMETERIPROC, glTexParameteri2);
+    struct { const char* name; GLenum sized, fmt; int bpp; GLint sw[4]; unsigned char px[4]; unsigned char wantL[3], wantR[3]; } sz[] = {
+        {"R8 swizzled as ALPHA8", GL_R8, GL_RED, 1, {GL_ZERO, GL_ZERO, GL_ZERO, GL_RED}, {255, 0}, {0,0,0}, {255,255,255}},
+        {"RG8 swizzled as LUM_ALPHA", GL_RG8, GL_RG, 2, {GL_RED, GL_RED, GL_RED, GL_GREEN}, {255,255, 0,0}, {255,255,255}, {255,255,255}},
+        {"RG8 as LUM_ALPHA (lum 0)", GL_RG8, GL_RG, 2, {GL_RED, GL_RED, GL_RED, GL_GREEN}, {0,255, 0,0}, {0,0,0}, {255,255,255}},
+    };
+    glUniform1i(lm, 0);
+    for (unsigned i = 0; i < sizeof sz / sizeof sz[0]; ++i) {
+        GLuint t; glGenTextures(1, &t); glBindTexture(GL_TEXTURE_2D, t);
+        glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glTexStorage2D(GL_TEXTURE_2D, 1, sz[i].sized, 2, 1);
+        glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 2, 1, sz[i].fmt, GL_UNSIGNED_BYTE, sz[i].px);
+        glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_SWIZZLE_R, sz[i].sw[0]); glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_SWIZZLE_G, sz[i].sw[1]);
+        glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_SWIZZLE_B, sz[i].sw[2]); glTexParameteri2(GL_TEXTURE_2D, GL_TEXTURE_SWIZZLE_A, sz[i].sw[3]);
+        GLenum err = glGetError();
+        glClearColor(1, 1, 1, 1); glClear(GL_COLOR_BUFFER_BIT); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        unsigned char l[4], r[4];
+        glReadPixels(16, 32, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, l); glReadPixels(48, 32, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, r);
+        int ok = 1;
+        for (int k = 0; k < 3; ++k) ok = ok && abs(l[k] - sz[i].wantL[k]) < 40 && abs(r[k] - sz[i].wantR[k]) < 40;
+        printf("%-26s err=0x%04x left=%3d,%3d,%3d right=%3d,%3d,%3d %s\n", sz[i].name, err, l[0], l[1], l[2], r[0], r[1], r[2], ok ? "ok" : "BAD");
+    }
+    return 0;
+}
@@ -0,0 +1,410 @@
+// yuvprobe.c - pixel check of YUV dma-buf sampling through MobileGL (GL_OES_EGL_image_external).
+//
+//   gcc -O1 yuvprobe.c -o /usr/local/bin/mgl-yuvprobe -lEGL -lGLESv2 -lgbm
+//   mgrun offscreen mgl-yuvprobe            # exit 0 = every check passed
+//
+// What it does, in an ES 3.0 context on a pbuffer:
+//   1. FOREIGN NV12 (and P010): a dma-buf from /dev/dma_heap/system - a buffer the MobileGL server
+//      did not allocate, as a video decoder's would be - with padded pitches and a gap before the
+//      chroma plane, filled with four vertical bars (red, green, blue, white) in BT.709 narrow
+//      range. Imported with eglCreateImage(EGL_LINUX_DMA_BUF_EXT) and its colour hints, bound to
+//      GL_TEXTURE_EXTERNAL_OES, drawn through a samplerExternalOES shader into an RGBA8 FBO and read
+//      back: each bar must come out its colour. The buffer is then rewritten with the bars in
+//      reverse order and, after a native fence (what ends a frame for an offscreen producer), drawn
+//      again: the new order must show (the server refreshes a foreign buffer at each frame).
+//   2. GBM NV12: gbm_bo_create(NV12) through the mobilegl GBM backend (a server-allocated YUV
+//      buffer, zero-copy), its planes printed, imported and sampled once (content undefined; the
+//      check is that every call succeeds).
+// Also prints the dma-buf formats/modifiers EGL lists for NV12/P010.
+#define _GNU_SOURCE
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+#include <GLES3/gl3.h>
+#include <GLES2/gl2ext.h>
+#include <fcntl.h>
+#include <gbm.h>
+#include <linux/dma-buf.h>
+#include <linux/dma-heap.h>
+#include <math.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#ifndef DRM_FORMAT_NV12
+#define DRM_FORMAT_NV12 0x3231564E
+#endif
+#ifndef DRM_FORMAT_P010
+#define DRM_FORMAT_P010 0x30313050
+#endif
+
+static int g_failures = 0;
+#define CHECK(cond, ...)                                                                                              \
+    do {                                                                                                               \
+        if (!(cond)) {                                                                                                 \
+            printf("FAIL: " __VA_ARGS__);                                                                              \
+            printf("\n");                                                                                              \
+            ++g_failures;                                                                                              \
+        }                                                                                                              \
+    } while (0)
+
+static PFNEGLCREATEIMAGEKHRPROC pCreateImage;
+static PFNEGLDESTROYIMAGEKHRPROC pDestroyImage;
+static PFNGLEGLIMAGETARGETTEXTURE2DOESPROC pTargetTexture;
+static PFNEGLQUERYDMABUFFORMATSEXTPROC pQueryFormats;
+static PFNEGLQUERYDMABUFMODIFIERSEXTPROC pQueryModifiers;
+static PFNEGLCREATESYNCKHRPROC pCreateSync;
+static PFNEGLDESTROYSYNCKHRPROC pDestroySync;
+static EGLDisplay g_display;
+
+enum { W = 256, H = 128, BARS = 4 };
+static const unsigned char kBars[BARS][3] = {{255, 0, 0}, {0, 255, 0}, {0, 0, 255}, {255, 255, 255}};
+
+// RGB -> BT.709 narrow-range codes at `bits` (8 or 10).
+static void Encode709(const unsigned char rgb[3], int bits, int out[3]) {
+    const double kr = 0.2126, kb = 0.0722, kg = 1.0 - kr - kb;
+    const double r = rgb[0] / 255.0, g = rgb[1] / 255.0, b = rgb[2] / 255.0;
+    const double y = kr * r + kg * g + kb * b;
+    const double cb = (b - y) / (2.0 * (1.0 - kb)), cr = (r - y) / (2.0 * (1.0 - kr));
+    const double scale = (double)(1 << (bits - 8));
+    out[0] = (int)lround((16.0 + 219.0 * y) * scale);
+    out[1] = (int)lround((128.0 + 224.0 * cb) * scale);
+    out[2] = (int)lround((128.0 + 224.0 * cr) * scale);
+}
+
+// Writes the bars (`reverse`d or not) into a semi-planar buffer: 1 byte samples, or P010's 16-bit
+// samples with the 10 significant bits high.
+static void FillBars(uint8_t* base, uint32_t pitchY, uint32_t offsetC, uint32_t pitchC, int bytes, int reverse) {
+    const int bits = bytes == 1 ? 8 : 10;
+    for (int y = 0; y < H; ++y) {
+        for (int x = 0; x < W; ++x) {
+            int bar = x * BARS / W;
+            if (reverse) bar = BARS - 1 - bar;
+            int code[3];
+            Encode709(kBars[bar], bits, code);
+            uint8_t* py = base + (size_t)y * pitchY + (size_t)x * bytes;
+            if (bytes == 1) py[0] = (uint8_t)code[0];
+            else *(uint16_t*)py = (uint16_t)(code[0] << 6);
+            if ((x & 1) == 0 && (y & 1) == 0) {
+                uint8_t* pc = base + offsetC + (size_t)(y / 2) * pitchC + (size_t)(x / 2) * 2 * bytes;
+                if (bytes == 1) {
+                    pc[0] = (uint8_t)code[1];
+                    pc[1] = (uint8_t)code[2];
+                } else {
+                    ((uint16_t*)pc)[0] = (uint16_t)(code[1] << 6);
+                    ((uint16_t*)pc)[1] = (uint16_t)(code[2] << 6);
+                }
+            }
+        }
+    }
+}
+
+static GLuint Compile(GLenum type, const char* source) {
+    GLuint shader = glCreateShader(type);
+    glShaderSource(shader, 1, &source, NULL);
+    glCompileShader(shader);
+    GLint ok = 0;
+    glGetShaderiv(shader, GL_COMPILE_STATUS, &ok);
+    if (!ok) {
+        char log[2048];
+        glGetShaderInfoLog(shader, sizeof(log), NULL, log);
+        printf("shader: %s\n", log);
+    }
+    return shader;
+}
+
+static GLuint g_program, g_fbo, g_rgba;
+
+static void SetUpDraw(void) {
+    const char* vs = "#version 300 es\n"
+                     "out vec2 uv;\n"
+                     "void main() {\n"
+                     "  vec2 p = vec2(float((gl_VertexID << 1) & 2), float(gl_VertexID & 2));\n"
+                     "  gl_Position = vec4(p * 2.0 - 1.0, 0.0, 1.0);\n"
+                     "  uv = p;\n"
+                     "}\n";
+    const char* fs = "#version 300 es\n"
+                     "#extension GL_OES_EGL_image_external_essl3 : require\n"
+                     "precision mediump float;\n"
+                     "uniform samplerExternalOES tex;\n"
+                     "in vec2 uv;\n"
+                     "out vec4 color;\n"
+                     "void main() { color = texture(tex, uv); }\n";
+    g_program = glCreateProgram();
+    glAttachShader(g_program, Compile(GL_VERTEX_SHADER, vs));
+    glAttachShader(g_program, Compile(GL_FRAGMENT_SHADER, fs));
+    glLinkProgram(g_program);
+    GLint linked = 0;
+    glGetProgramiv(g_program, GL_LINK_STATUS, &linked);
+    CHECK(linked, "the samplerExternalOES program did not link");
+    GLint type = 0, size = 0;
+    char name[64];
+    glGetActiveUniform(g_program, 0, sizeof(name), NULL, &size, (GLenum*)&type, name);
+    printf("uniform 0: %s type 0x%04x\n", name, type);
+    CHECK(type == GL_SAMPLER_EXTERNAL_OES, "the sampler reports type 0x%04x, not GL_SAMPLER_EXTERNAL_OES", type);
+    glGenTextures(1, &g_rgba);
+    glBindTexture(GL_TEXTURE_2D, g_rgba);
+    glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, W, H);
+    glGenFramebuffers(1, &g_fbo);
+    glBindFramebuffer(GL_FRAMEBUFFER, g_fbo);
+    glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, g_rgba, 0);
+    CHECK(glCheckFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE, "the readback FBO is incomplete");
+    GLuint vao;
+    glGenVertexArrays(1, &vao);
+    glBindVertexArray(vao);
+}
+
+// Draws the external texture `texture` and checks each bar's center (row `y` of the image).
+static void DrawAndCheck(GLuint texture, const char* what, int reverse, int tolerance) {
+    glBindFramebuffer(GL_FRAMEBUFFER, g_fbo);
+    glViewport(0, 0, W, H);
+    glUseProgram(g_program);
+    glActiveTexture(GL_TEXTURE0);
+    glBindTexture(GL_TEXTURE_EXTERNAL_OES, texture);
+    glUniform1i(glGetUniformLocation(g_program, "tex"), 0);
+    glDrawArrays(GL_TRIANGLES, 0, 3);
+    static unsigned char pixels[W * H * 4];
+    glReadPixels(0, 0, W, H, GL_RGBA, GL_UNSIGNED_BYTE, pixels);
+    GLenum error = glGetError();
+    CHECK(error == GL_NO_ERROR, "%s: GL error 0x%04x", what, error);
+    if (tolerance < 0) {
+        printf("%s: center pixel %d %d %d %d (content undefined)\n", what, pixels[(H / 2 * W + W / 2) * 4],
+               pixels[(H / 2 * W + W / 2) * 4 + 1], pixels[(H / 2 * W + W / 2) * 4 + 2],
+               pixels[(H / 2 * W + W / 2) * 4 + 3]);
+        return;
+    }
+    // Rows: GL row 0 is the image's first row (the uv mapping above), and the bars are vertical,
+    // so any row does; take two to catch a broken chroma row pitch.
+    const int rows[2] = {H / 4, (3 * H) / 4};
+    for (int r = 0; r < 2; ++r) {
+        for (int bar = 0; bar < BARS; ++bar) {
+            const int x = bar * (W / BARS) + W / (2 * BARS);
+            const unsigned char* p = &pixels[(rows[r] * W + x) * 4];
+            const unsigned char* want = kBars[reverse ? BARS - 1 - bar : bar];
+            int worst = 0;
+            for (int c = 0; c < 3; ++c) {
+                int d = abs((int)p[c] - (int)want[c]);
+                if (d > worst) worst = d;
+            }
+            printf("%s: row %d bar %d -> %3d %3d %3d %3d (want %3d %3d %3d)%s\n", what, rows[r], bar, p[0], p[1], p[2],
+                   p[3], want[0], want[1], want[2], worst > tolerance ? "  <-- off" : "");
+            CHECK(worst <= tolerance, "%s: bar %d is off by %d", what, bar, worst);
+        }
+    }
+}
+
+static int AllocateDmaHeap(size_t size) {
+    int heap = open("/dev/dma_heap/system", O_RDWR | O_CLOEXEC);
+    if (heap < 0) {
+        perror("open /dev/dma_heap/system");
+        return -1;
+    }
+    struct dma_heap_allocation_data data = {.len = size, .fd_flags = O_RDWR | O_CLOEXEC};
+    if (ioctl(heap, DMA_HEAP_IOCTL_ALLOC, &data) != 0) {
+        perror("DMA_HEAP_IOCTL_ALLOC");
+        close(heap);
+        return -1;
+    }
+    close(heap);
+    return (int)data.fd;
+}
+
+static void Sync(int fd, int start) {
+    struct dma_buf_sync sync = {.flags = DMA_BUF_SYNC_WRITE | (start ? DMA_BUF_SYNC_START : DMA_BUF_SYNC_END)};
+    ioctl(fd, DMA_BUF_IOCTL_SYNC, &sync);
+}
+
+static EGLImageKHR Import(int fd, uint32_t fourcc, uint32_t pitchY, uint32_t offsetC, uint32_t pitchC, int modifier) {
+    EGLint attribs[64];
+    int n = 0;
+#define A(k, v) (attribs[n++] = (k), attribs[n++] = (EGLint)(v))
+    A(EGL_WIDTH, W);
+    A(EGL_HEIGHT, H);
+    A(EGL_LINUX_DRM_FOURCC_EXT, fourcc);
+    A(EGL_DMA_BUF_PLANE0_FD_EXT, fd);
+    A(EGL_DMA_BUF_PLANE0_OFFSET_EXT, 0);
+    A(EGL_DMA_BUF_PLANE0_PITCH_EXT, pitchY);
+    A(EGL_DMA_BUF_PLANE1_FD_EXT, fd);
+    A(EGL_DMA_BUF_PLANE1_OFFSET_EXT, offsetC);
+    A(EGL_DMA_BUF_PLANE1_PITCH_EXT, pitchC);
+    if (modifier) {
+        A(EGL_DMA_BUF_PLANE0_MODIFIER_LO_EXT, 0);
+        A(EGL_DMA_BUF_PLANE0_MODIFIER_HI_EXT, 0);
+        A(EGL_DMA_BUF_PLANE1_MODIFIER_LO_EXT, 0);
+        A(EGL_DMA_BUF_PLANE1_MODIFIER_HI_EXT, 0);
+    }
+    A(EGL_YUV_COLOR_SPACE_HINT_EXT, EGL_ITU_REC709_EXT);
+    A(EGL_SAMPLE_RANGE_HINT_EXT, EGL_YUV_NARROW_RANGE_EXT);
+    attribs[n] = EGL_NONE;
+#undef A
+    EGLImageKHR image = pCreateImage(g_display, EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, attribs);
+    if (image == EGL_NO_IMAGE_KHR) printf("eglCreateImage failed: 0x%04x\n", eglGetError());
+    return image;
+}
+
+static GLuint ExternalTexture(EGLImageKHR image) {
+    GLuint texture;
+    glGenTextures(1, &texture);
+    glBindTexture(GL_TEXTURE_EXTERNAL_OES, texture);
+    pTargetTexture(GL_TEXTURE_EXTERNAL_OES, image);
+    GLenum error = glGetError();
+    CHECK(error == GL_NO_ERROR, "glEGLImageTargetTexture2DOES(GL_TEXTURE_EXTERNAL_OES) -> 0x%04x", error);
+    GLint bound = 0;
+    glGetIntegerv(GL_TEXTURE_BINDING_EXTERNAL_OES, &bound);
+    CHECK((GLuint)bound == texture, "GL_TEXTURE_BINDING_EXTERNAL_OES reads %d", bound);
+    return texture;
+}
+
+// A frame boundary as an offscreen producer makes one: a native fence.
+static void EndFrame(void) {
+    EGLint attribs[] = {EGL_NONE};
+    EGLSyncKHR sync = pCreateSync(g_display, EGL_SYNC_NATIVE_FENCE_ANDROID, attribs);
+    CHECK(sync != EGL_NO_SYNC_KHR, "eglCreateSync(native fence) failed");
+    if (sync != EGL_NO_SYNC_KHR) pDestroySync(g_display, sync);
+}
+
+static void ForeignTest(uint32_t fourcc, int bytes, int tolerance) {
+    const char* what = fourcc == DRM_FORMAT_NV12 ? "foreign NV12" : "foreign P010";
+    // Padded pitches and a gap before the chroma plane, as decoders lay buffers out.
+    const uint32_t pitchY = W * bytes + 64, pitchC = W * bytes + 128;
+    const uint32_t offsetC = pitchY * (H + 16);
+    const size_t size = offsetC + (size_t)pitchC * (H / 2) + 4096;
+    int fd = AllocateDmaHeap(size);
+    if (fd < 0) {
+        CHECK(0, "%s: no dma-buf to test with", what);
+        return;
+    }
+    uint8_t* map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+    CHECK(map != MAP_FAILED, "%s: mmap failed", what);
+    if (map == MAP_FAILED) return;
+    Sync(fd, 1);
+    FillBars(map, pitchY, offsetC, pitchC, bytes, 0);
+    Sync(fd, 0);
+    EGLImageKHR image = Import(fd, fourcc, pitchY, offsetC, pitchC, fourcc == DRM_FORMAT_NV12);
+    CHECK(image != EGL_NO_IMAGE_KHR, "%s: import refused", what);
+    if (image == EGL_NO_IMAGE_KHR) return;
+    GLuint texture = ExternalTexture(image);
+    DrawAndCheck(texture, what, 0, tolerance);
+    EndFrame();
+    // New content, same buffer: shows after the frame boundary.
+    Sync(fd, 1);
+    FillBars(map, pitchY, offsetC, pitchC, bytes, 1);
+    Sync(fd, 0);
+    EndFrame();
+    DrawAndCheck(texture, what, 1, tolerance);
+    glDeleteTextures(1, &texture);
+    pDestroyImage(g_display, image);
+    munmap(map, size);
+    close(fd);
+}
+
+static void GbmTest(void) {
+    int node = open("/dev/dri/renderD128", O_RDWR | O_CLOEXEC);
+    if (node < 0) node = open("/dev/null", O_RDWR | O_CLOEXEC);
+    struct gbm_device* gbm = gbm_create_device(node);
+    if (gbm == NULL) {
+        CHECK(0, "gbm_create_device failed");
+        return;
+    }
+    printf("gbm backend: %s\n", gbm_device_get_backend_name(gbm));
+    const int supported = gbm_device_is_format_supported(gbm, DRM_FORMAT_NV12, 0);
+    printf("gbm NV12 supported: %d, P010: %d\n", supported,
+           gbm_device_is_format_supported(gbm, DRM_FORMAT_P010, 0));
+    CHECK(supported == 1, "gbm does not support NV12");
+    struct gbm_bo* bo = gbm_bo_create(gbm, W, H, DRM_FORMAT_NV12, 0);
+    CHECK(bo != NULL, "gbm_bo_create(NV12) failed");
+    if (bo != NULL) {
+        const int planes = gbm_bo_get_plane_count(bo);
+        printf("gbm NV12 bo: planes %d, stride %u/%u, offset %u/%u, modifier 0x%llx\n", planes,
+               gbm_bo_get_stride_for_plane(bo, 0), gbm_bo_get_stride_for_plane(bo, 1), gbm_bo_get_offset(bo, 0),
+               gbm_bo_get_offset(bo, 1), (unsigned long long)gbm_bo_get_modifier(bo));
+        CHECK(planes == 2, "the NV12 bo has %d planes", planes);
+        int fd = gbm_bo_get_fd(bo);
+        EGLImageKHR image =
+            Import(fd, DRM_FORMAT_NV12, gbm_bo_get_stride_for_plane(bo, 0), gbm_bo_get_offset(bo, 1),
+                   gbm_bo_get_stride_for_plane(bo, 1), 0);
+        CHECK(image != EGL_NO_IMAGE_KHR, "the server's own NV12 buffer was not imported");
+        if (image != EGL_NO_IMAGE_KHR) {
+            GLuint texture = ExternalTexture(image);
+            DrawAndCheck(texture, "gbm NV12", 0, -1);
+            glDeleteTextures(1, &texture);
+            pDestroyImage(g_display, image);
+        }
+        close(fd);
+        gbm_bo_destroy(bo);
+    }
+    gbm_device_destroy(gbm);
+    close(node);
+}
+
+int main(void) {
+    setvbuf(stdout, NULL, _IONBF, 0);
+    g_display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
+    EGLint major, minor;
+    if (!eglInitialize(g_display, &major, &minor)) {
+        printf("FAIL: eglInitialize\n");
+        return 2;
+    }
+    const char* eglExtensions = eglQueryString(g_display, EGL_EXTENSIONS);
+    printf("EGL %d.%d dma_buf_import: %s\n", major, minor,
+           strstr(eglExtensions, "EGL_EXT_image_dma_buf_import") ? "yes" : "no");
+    eglBindAPI(EGL_OPENGL_ES_API);
+    const EGLint configAttribs[] = {EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT,
+                                    EGL_RED_SIZE, 8, EGL_GREEN_SIZE, 8, EGL_BLUE_SIZE, 8, EGL_NONE};
+    EGLConfig config;
+    EGLint count = 0;
+    eglChooseConfig(g_display, configAttribs, &config, 1, &count);
+    if (count < 1) {
+        printf("FAIL: no config\n");
+        return 2;
+    }
+    const EGLint pbufferAttribs[] = {EGL_WIDTH, 16, EGL_HEIGHT, 16, EGL_NONE};
+    EGLSurface surface = eglCreatePbufferSurface(g_display, config, pbufferAttribs);
+    const EGLint contextAttribs[] = {EGL_CONTEXT_MAJOR_VERSION, 3, EGL_NONE};
+    EGLContext context = eglCreateContext(g_display, config, EGL_NO_CONTEXT, contextAttribs);
+    if (context == EGL_NO_CONTEXT || !eglMakeCurrent(g_display, surface, surface, context)) {
+        printf("FAIL: no ES 3 context\n");
+        return 2;
+    }
+    pCreateImage = (PFNEGLCREATEIMAGEKHRPROC)eglGetProcAddress("eglCreateImageKHR");
+    pDestroyImage = (PFNEGLDESTROYIMAGEKHRPROC)eglGetProcAddress("eglDestroyImageKHR");
+    pTargetTexture = (PFNGLEGLIMAGETARGETTEXTURE2DOESPROC)eglGetProcAddress("glEGLImageTargetTexture2DOES");
+    pQueryFormats = (PFNEGLQUERYDMABUFFORMATSEXTPROC)eglGetProcAddress("eglQueryDmaBufFormatsEXT");
+    pQueryModifiers = (PFNEGLQUERYDMABUFMODIFIERSEXTPROC)eglGetProcAddress("eglQueryDmaBufModifiersEXT");
+    pCreateSync = (PFNEGLCREATESYNCKHRPROC)eglGetProcAddress("eglCreateSyncKHR");
+    pDestroySync = (PFNEGLDESTROYSYNCKHRPROC)eglGetProcAddress("eglDestroySyncKHR");
+    printf("GL_RENDERER: %s\n", glGetString(GL_RENDERER));
+    const char* glExtensions = (const char*)glGetString(GL_EXTENSIONS);
+    const int external = strstr(glExtensions, "GL_OES_EGL_image_external_essl3") != NULL;
+    printf("GL_OES_EGL_image_external_essl3: %s\n", external ? "yes" : "no");
+    CHECK(external, "GL_OES_EGL_image_external_essl3 is not advertised");
+
+    EGLint formats[32], formatCount = 0;
+    pQueryFormats(g_display, 32, formats, &formatCount);
+    int nv12 = 0;
+    for (int i = 0; i < formatCount; ++i) {
+        printf("dma-buf format %.4s", (const char*)&formats[i]);
+        EGLuint64KHR modifiers[4];
+        EGLBoolean externalOnly[4];
+        EGLint modifierCount = 0;
+        pQueryModifiers(g_display, formats[i], 4, modifiers, externalOnly, &modifierCount);
+        for (int m = 0; m < modifierCount; ++m)
+            printf(" [modifier 0x%llx%s]", (unsigned long long)modifiers[m], externalOnly[m] ? " external-only" : "");
+        printf("\n");
+        nv12 |= formats[i] == (EGLint)DRM_FORMAT_NV12;
+    }
+    CHECK(nv12, "NV12 is not a listed dma-buf format");
+
+    SetUpDraw();
+    ForeignTest(DRM_FORMAT_NV12, 1, 12);
+    ForeignTest(DRM_FORMAT_P010, 2, 16);
+    GbmTest();
+
+    printf("%s (%d failure(s))\n", g_failures ? "FAILED" : "PASSED", g_failures);
+    eglMakeCurrent(g_display, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
+    eglTerminate(g_display);
+    return g_failures ? 1 : 0;
+}
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+# push-tools.sh - install this skill's device- and container-side helpers on the phone.
+#   device/*.sh      -> $DEV_TOOLS            (run as root: adb shell 'su -c "sh $DEV_TOOLS/<x>.sh"')
+#   container/mgrun  -> <container>/usr/local/bin/mgrun
+#   container/kwprep.sh -> <container>/root/kwprep.sh
+# Needs the container running (its rootfs is mounted at $CT_ROOT only then).
+# Overwrites same-named files other sessions may be using in $DEV_TOOLS; check first on a shared phone.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+adb_ shell "mkdir -p $DEV_TOOLS"
+for f in "$here"/device/*.sh; do
+    tr -d '\r' < "$f" > "$ANL_DIR/push.tmp"
+    adb_ push "$ANL_DIR/push.tmp" "$DEV_TOOLS/$(basename "$f")" >/dev/null
+    echo "pushed $DEV_TOOLS/$(basename "$f")"
+done
+for f in mgrun kwprep.sh; do
+    tr -d '\r' < "$here/container/$f" > "$ANL_DIR/push.tmp"
+    adb_ push "$ANL_DIR/push.tmp" "$DEV_TOOLS/$f" >/dev/null
+done
+rm -f "$ANL_DIR/push.tmp"
+adb_ shell "su -c 'cp $DEV_TOOLS/mgrun $CT_ROOT/usr/local/bin/mgrun && chmod 755 $CT_ROOT/usr/local/bin/mgrun && cp $DEV_TOOLS/kwprep.sh $CT_ROOT/root/kwprep.sh && echo installed mgrun kwprep.sh'"
@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+# rec.sh <name> - record 3 s of the screen while opening Kickoff, extract frames inside the container
+# with ffmpeg, pull them to $ANL_DIR/af_<name>/ and build a 3x4 contact sheet $ANL_DIR/grid_<name>.png
+# (needs Python + Pillow on Windows).  Use it to catch flicker/black frames a single screenshot misses.
+# Env: TAP="x y" Kickoff button position (default "69 1349" = bottom-left on the 2560x1600 tablet;
+# device specific).  Note: anland grabs the touchscreen (EVIOCGRAB); if `input tap` does nothing, use
+# `input mouse tap x y`.  Env CHANNEL (default rec).
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+N=${1:?usage: rec.sh <name>}; TAP=${TAP:-69 1349}
+adb_ shell "su -c 'input keyevent 111; sleep 1.5; (screenrecord --time-limit 3 --bit-rate 20000000 $DEV_SHARE/$N.mp4 &); sleep 0.7; input tap $TAP; sleep 3; cp $DEV_SHARE/$N.mp4 $CT_ROOT/root/$N.mp4'"
+bash "$here/ct.sh" "${CHANNEL:-rec}" <<IN
+rm -rf /root/af_$N; mkdir -p /root/af_$N; ffmpeg -loglevel quiet -i /root/$N.mp4 -vf scale=1000:-1 -fps_mode passthrough /root/af_$N/f%03d.png; ls /root/af_$N | wc -l
+IN
+adb_ shell "su -c 'rm -rf $DEV_SHARE/af_$N; cp -r $CT_ROOT/root/af_$N $DEV_SHARE/ && chmod -R 777 $DEV_SHARE/af_$N'"
+rm -rf "$ANL_DIR/af_$N"; adb_ pull "$DEV_SHARE/af_$N" "$ANL_DIR/" >/dev/null
+python - "$ANL_DIR/af_$N" "$ANL_DIR/grid_$N.png" <<'P'
+import sys, os
+from PIL import Image
+d, out = sys.argv[1], sys.argv[2]
+fs = sorted(os.listdir(d)); n = len(fs)
+pick = [fs[min(n-1, int(i * n / 12))] for i in range(12)]
+ims = [Image.open(os.path.join(d, f)) for f in pick]
+w, h = ims[0].size; tw, th = w // 2, h // 2
+g = Image.new('RGB', (tw * 3, th * 4))
+for i, im in enumerate(ims): g.paste(im.resize((tw, th)), ((i % 3) * tw, (i // 3) * th))
+g.save(out); print(n, pick)
+P
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# shot.sh <name> [count] [interval-s] - take <count> (default 2) phone screenshots <interval> s apart
+# (default 1), pull them to $ANL_DIR/<name>-<i>.png and print their md5.  Identical hashes while an
+# animated client (glmark2, a video) runs mean the frames are NOT reaching the screen.
+set -euo pipefail
+. "$(dirname "$0")/env.sh"
+name=${1:?usage: shot.sh <name> [count] [interval]}; n=${2:-2}; dt=${3:-1}
+cmd=""; i=1
+while [ "$i" -le "$n" ]; do cmd="$cmd screencap -p $DEV_SHARE/$name-$i.png; sleep $dt;"; i=$((i+1)); done
+adb_ shell "su -c '$cmd'"
+i=1
+while [ "$i" -le "$n" ]; do
+    adb_ pull "$DEV_SHARE/$name-$i.png" "$ANL_DIR/$name-$i.png" >/dev/null
+    md5sum "$ANL_DIR/$name-$i.png"; i=$((i+1))
+done
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+# sync-src.sh - FALLBACK path only (the WSL cross build, xbuild.sh, replaced it).  Pushes the files the
+# worktree changed since the container's source stamp into the container's MobileGL source tree
+# (/root/mobilegl-unified-src), keeping mtimes so the in-container ninja build stays incremental;
+# then build/install inside the container with anland's anland-build-client.sh
+# (producers/kde/Arch_v5/mobilegl-tools/, gcc, -j4, slow on the phone).
+# A full re-sync of the tree is anland's anland-sync-source.sh (MOBILEGL_REPO=<worktree>).
+set -euo pipefail
+. "$(dirname "$0")/env.sh"
+SRC=$CT_ROOT/root/mobilegl-unified-src
+STAMP=$(adb_ shell "su -c 'cat $SRC/.anland-build-stamp'" | tr -d '\r\n')
+cd "$MGL_WORKTREE"
+FILES=$( (git diff --name-only --diff-filter=d "$STAMP"; git ls-files --others --exclude-standard) \
+    | grep -v -E '^(3rdparty/|build-android|\.anland-build-stamp|\.claude/)' | sort -u)
+echo "$(echo "$FILES" | wc -l) files since $STAMP"
+tar --force-local -czf "$ANL_DIR/src-delta.tgz" $FILES
+adb_ push "$ANL_DIR/src-delta.tgz" /data/local/tmp/src-delta.tgz >/dev/null
+HEAD=$(git rev-parse HEAD)
+adb_ shell "su -c 'tar xzf /data/local/tmp/src-delta.tgz -C $SRC && echo $HEAD > $SRC/.anland-build-stamp && echo synced'"
@@ -0,0 +1,9 @@
+#!/bin/bash
+# wsl-all.sh - run INSIDE WSL archlinux after wsl-test.sh configured $B: build every *Test target except
+# DirectVulkan*/Benchmark* and run the unit label.  NEVER run the DirectVulkan integration tests
+# (they need a live device + manual intervention); never run two ctest suites at once.
+B=${B:-$HOME/mgl-anl-bld}
+cd $B
+T=$(ninja -t targets all 2>/dev/null | grep -oE "^[A-Za-z]+Test:" | tr -d : | sort -u | grep -v -E "^(DirectVulkan|Benchmark)" | tr "\n" " ")
+ninja -j14 MobileGLServer $T 2>&1 | grep -E "error:|FAILED" | head -10
+ctest -L unit -j10 --timeout 300 -E "DirectVulkan" 2>&1 | tail -25
@@ -0,0 +1,13 @@
+#!/bin/bash
+# wsl-test.sh <targets...> - run INSIDE WSL archlinux: configure (once) and build MobileGL unit-test
+# targets for the host (clang, disaggregated + in-process server ON).  From Git Bash:
+#   MSYS2_ARG_CONV_EXCL='*' wsl -d archlinux -- bash /mnt/c/<worktree>/.claude/skills/anland-mobilegl-plasma/scripts/wsl-test.sh MultiSessionTest
+# then ctest in $B (see wsl-all.sh).  Env: W (worktree, WSL path; default: derived from this file),
+# B (build dir, default ~/mgl-anl-bld).  The Windows MSVC build of feat/disaggregated is broken; use this.
+set -e
+W=${W:-$(cd "$(dirname "$0")/../../../.." && pwd)}
+B=${B:-$HOME/mgl-anl-bld}
+if [ ! -f $B/build.ninja ]; then
+cmake -S $W -B $B -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DMOBILEGL_BUILD_TEST=ON -DMOBILEGL_BUILD_DISAGGREGATED=ON -DMOBILEGL_BUILD_DISAGGREGATED_INPROC=ON -DMOBILEGL_BUILD_STAMP=wsl
+fi
+cmake --build $B --target "$@" -j16
@@ -0,0 +1,135 @@
+#!/usr/bin/env bash
+# x11-verify.sh <step> [...] - device verification of X11 on the GPU (docs/Disaggregated/notes/anland/
+# plan-x11-gpu.md): Xwayland drawing with glamor on MobileGL, and GLX windows presented through
+# DRI3 + Present from shared images. Steps, in the order a full run takes them:
+#
+#   preflight             read-only: battery, stack status, Xwayland package, render node, xcb libs,
+#                         test apps, Xwayland's extensions and its GLX vendor
+#   measure <label> [s]   glxgears 1280x720 (GEARS_ARGS; vblank_mode=0) for s seconds (10): its fps, then CPU/GPU/dma-buf
+#                         counts (device/x11-measure.sh); also with MOBILEGL_GLX_PRESENT=readback
+#   deploy-client         xbuild.sh + xdeploy.sh of this worktree's client (MGL_XBUILD_DIR)
+#   deploy-kwin           kwin-stage.sh build from ANLAND_WORKTREE; waits for "== ALL DONE =="
+#   deploy-server         build-android.sh + cyc.sh x11-server (swaps the APK lib, RESTARTS the session)
+#   check                 the evidence: Xwayland took glamor (journal), its MobileGL log, GLX's path
+#                         line, dma-buf counts in Xwayland/KWin, the server's import lines
+#   apps                  X11 clients on screen: konsole and kate on Qt's xcb platform (glamor-drawn
+#                         core X), glxgears (GLX), xterm/xeyes when installed; two screenshots
+#   stop                  kill the test apps
+#
+# Typical run: preflight; measure before; deploy-client; deploy-kwin; deploy-server; check; measure
+# after; apps; stop. Battery: every load step refuses below MIN_BATTERY (40) %.
+# Env: CHANNEL (ct.sh channel, default x11), ANLAND_WORKTREE (anland x11-gpu worktree, Windows
+# form), MGL_XBUILD_DIR (default /home/swung/mgl-xbuild-x11), MIN_BATTERY.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+CH=${CHANNEL:-x11}
+: "${MGL_XBUILD_DIR:=/home/swung/mgl-xbuild-x11}"
+: "${ANLAND_WORKTREE:=$(cygpath -m "$USERPROFILE" 2>/dev/null)/AndroidStudioProjects/anland/.claude/worktrees/x11-gpu}"
+export MGL_XBUILD_DIR ANLAND_WORKTREE
+
+battery_ok() {
+    local level
+    level=$(adb_ shell "dumpsys battery" | sed -n 's/^ *level: *//p' | tr -d '\r')
+    echo "battery: ${level}%"
+    if [ "${level:-0}" -lt "${MIN_BATTERY:-40}" ]; then
+        echo "below ${MIN_BATTERY:-40}%: stopping GPU-heavy work; charge the phone first" >&2
+        return 1
+    fi
+}
+
+push_measure() {
+    tr -d '\r' < "$here/device/x11-measure.sh" > "$ANL_DIR/x11-measure.sh"
+    adb_ push "$ANL_DIR/x11-measure.sh" "$DEV_TOOLS/x11-measure.sh" >/dev/null
+}
+
+step=${1:?usage: x11-verify.sh <preflight|measure|deploy-client|deploy-kwin|deploy-server|check|apps|stop>}
+shift
+case "$step" in
+preflight)
+    battery_ok || true
+    adb_ shell "su -c 'sh $DEV_TOOLS/status.sh'" || true
+    bash "$here/ct.sh" "$CH" <<'IN'
+echo "== packages"; pacman -Q xorg-xwayland libxcb mesa-utils xorg-xdpyinfo qt6-base 2>&1 | sed 's/^/  /'
+echo "== render nodes"; ls -l /dev/dri/ 2>&1 | sed 's/^/  /'
+U=swung0x48; for n in /dev/dri/renderD*; do sudo -u $U test -r "$n" -a -w "$n" && echo "  $U can open $n"; done
+echo "== KWin's thin node"; tr '\0' '\n' < /proc/$(pgrep -x kwin_wayland)/environ 2>/dev/null | grep -E "MOBILEGL_GBM_NODE|ANLAND_XWAYLAND_GLAMOR|GBM_BACKEND" | sed 's/^/  /'
+echo "== xcb libraries"; for l in libxcb-dri3.so.0 libxcb-present.so.0 libxcb-xfixes.so.0 libxcb-shm.so.0; do ls /usr/lib/$l >/dev/null 2>&1 && echo "  $l ok" || echo "  $l MISSING"; done
+echo "== Xwayland"; pgrep -a -x Xwayland | sed 's/^/  /'
+mgrun offscreen xdpyinfo 2>/dev/null | grep -E "^ +(DRI3|Present|MIT-SHM|XFIXES|GLX)$" | sed 's/^/  ext/'
+mgrun offscreen glxinfo -B 2>&1 | grep -E "vendor|renderer|direct rendering" | sed 's/^/  /'
+IN
+    ;;
+measure)
+    label=${1:?usage: x11-verify.sh measure <label> [seconds]}
+    secs=${2:-10}
+    battery_ok
+    push_measure
+    for mode in auto readback; do
+        printf '%s\n' \
+            "pkill -x glxgears; rm -f /tmp/mgl-x11-gears.client.log /tmp/x11-gears.out" \
+            "MGLOG=/tmp/mgl-x11-gears.log timeout $((secs + 8)) mgrun offscreen env vblank_mode=${VBLANK_MODE:-0} MOBILEGL_GLX_PRESENT=$mode glxgears ${GEARS_ARGS:--geometry 1280x720} > /tmp/x11-gears.out 2>&1 &" \
+            "sleep 3" | bash "$here/ct.sh" "$CH" >/dev/null
+        adb_ shell "su -c 'sh $DEV_TOOLS/x11-measure.sh $label-$mode glxgears $secs'" || true
+        sleep 6
+        printf '%s\n' "grep -a 'frames in' /tmp/x11-gears.out | tail -3; grep -a 'presented through' /tmp/mgl-x11-gears.client.log | head -1; pkill -x glxgears; true" \
+            | bash "$here/ct.sh" "$CH"
+    done
+    ;;
+deploy-client)
+    bash "$here/xbuild.sh"
+    CHANNEL=$CH bash "$here/xdeploy.sh"
+    ;;
+deploy-kwin)
+    CHANNEL=$CH bash "$here/kwin-stage.sh" build
+    for _ in $(seq 1 120); do
+        out=$(printf 'tail -3 /root/kwin-build.log\n' | bash "$here/ct.sh" "$CH" 2>/dev/null || true)
+        echo "$out" | tail -1
+        echo "$out" | grep -q "== ALL DONE ==" && exit 0
+        echo "$out" | grep -qiE "error|failed" && { echo "KWin build failed" >&2; exit 1; }
+        sleep 30
+    done
+    echo "KWin build did not finish in an hour" >&2; exit 1
+    ;;
+deploy-server)
+    battery_ok
+    bash "$here/build-android.sh"
+    CHANNEL=$CH bash "$here/cyc.sh" x11-server
+    ;;
+check)
+    bash "$here/ct.sh" "$CH" <<'IN'
+echo "== Xwayland's command line (no -shm with glamor)"; pgrep -a -x Xwayland | sed 's/^/  /'
+echo "== Xwayland's glamor (journal)"
+journalctl --no-pager --since -30min 2>/dev/null | grep -aiE "glamor|Xwayland.*(EGL|gbm|dri3|falling back)" | tail -10 | sed 's/^/  /'
+echo "== Xwayland's MobileGL client log"; grep -aE "presented|shared image|declin|Fatal|ERROR" /tmp/mgl-xwayland.client.log /tmp/mgl-xwayland.client 2>/dev/null | tail -8 | sed 's/^/  /' || echo "  (none)"
+X=$(pgrep -x Xwayland); K=$(pgrep -x kwin_wayland)
+echo "== dma-bufs held: Xwayland $(ls -l /proc/$X/fd 2>/dev/null | grep -c dmabuf), kwin $(ls -l /proc/$K/fd 2>/dev/null | grep -c dmabuf)"
+echo "== Xwayland's GLX vendor and extensions"
+mgrun offscreen xdpyinfo 2>/dev/null | grep -E "^ +(DRI3|Present|MIT-SHM)$" | sed 's/^/  ext/'
+echo "== a GLX window's path"
+rm -f /tmp/mgl-x11-check.client.log
+MGLOG=/tmp/mgl-x11-check.log timeout 6 mgrun offscreen glxgears >/dev/null 2>&1
+grep -a "presented through\|leaves DRI3" /tmp/mgl-x11-check.client.log | sed 's/^/  /'
+IN
+    adb_ logcat -d | grep -aE "MobileGL.*(shared image|dma-buf import|refused)" | tail -8 || true
+    adb_ logcat -b crash -d | tail -5 || true
+    ;;
+apps)
+    battery_ok
+    printf '%s\n' \
+        "pkill -x xterm; pkill -x xeyes; pkill -x glxgears; pkill -x konsole; pkill -x kate" \
+        "MGLOG=/tmp/mgl-x11-konsole.log mgrun offscreen env QT_QPA_PLATFORM=xcb konsole >/dev/null 2>&1 &" \
+        "MGLOG=/tmp/mgl-x11-kate.log mgrun offscreen env QT_QPA_PLATFORM=xcb kate >/dev/null 2>&1 &" \
+        "command -v xterm >/dev/null && MGLOG=/tmp/mgl-x11-xterm.log mgrun offscreen xterm >/dev/null 2>&1 &" \
+        "command -v xeyes >/dev/null && MGLOG=/tmp/mgl-x11-xeyes.log mgrun offscreen xeyes >/dev/null 2>&1 &" \
+        "MGLOG=/tmp/mgl-x11-gears.log mgrun offscreen glxgears >/dev/null 2>&1 &" \
+        "sleep 6; pgrep -a -x konsole; pgrep -a -x kate; pgrep -a -x xterm; pgrep -a -x xeyes; pgrep -a -x glxgears; true" | bash "$here/ct.sh" "$CH"
+    bash "$here/shot.sh" x11-apps 2 1
+    ;;
+stop)
+    printf '%s\n' "pkill -x xterm; pkill -x xeyes; pkill -x glxgears; pkill -x konsole; pkill -x kate; true" | bash "$here/ct.sh" "$CH"
+    ;;
+*)
+    echo "unknown step $step" >&2; exit 2
+    ;;
+esac
@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+# xbuild.sh [--reconfigure] [extra ninja targets...] - cross-build the container CLIENT: libMobileGL.so
+# (glibc aarch64, glvnd EGL/GLX vendor) + mobilegl_gbm.so (GBM backend), in WSL archlinux with
+# clang + lld against a sysroot exported from the container (xsysroot/refresh-sysroot.sh).
+# Full build ~150 s, incremental seconds.  Deploy with xdeploy.sh.
+# Env: MGL_WORKTREE (source, Windows form), MGL_XBUILD_DIR (WSL build dir, default ~/mgl-xbuild-a64),
+#      MGL_JOBS (16), WSL_DISTRO (archlinux).
+# Prereq in WSL: clang, lld, llvm (lld major MUST equal clang/llvm-libs major - lld 23 on llvm 22
+# broke ld.lld), cmake, ninja; ~/sysroots/aarch64-arch.cmake + ~/sysroots/arch-kde-mgl.
+set -euo pipefail
+. "$(dirname "$0")/env.sh"
+STAMP=$(git -C "$MGL_WORKTREE" rev-parse HEAD)
+RECONF=0
+if [ "${1:-}" = "--reconfigure" ]; then RECONF=1; shift; fi
+EXTRA="$*"
+SRC_WSL=$(win2wsl "$MGL_WORKTREE")
+wsl -d "${WSL_DISTRO:-archlinux}" -- env SRC="$SRC_WSL" STAMP="$STAMP" RECONF="$RECONF" EXTRA="$EXTRA" \
+    OUT="${MGL_XBUILD_DIR:-}" JOBS="${MGL_JOBS:-16}" bash -s <<'IN' 2>&1 | tr -d '\0' | grep -av '^wsl:'
+set -euo pipefail
+OUT=${OUT:-$HOME/mgl-xbuild-a64}
+TC=$HOME/sysroots/aarch64-arch.cmake
+[ -f "$TC" ] || { echo "missing toolchain $TC (run xsysroot/refresh-sysroot.sh)"; exit 1; }
+ld.lld --version >/dev/null || { echo "ld.lld is not runnable in WSL (lld/llvm version mismatch?)"; exit 1; }
+t0=$(date +%s)
+if [ "$RECONF" = 1 ] || [ ! -f "$OUT/build.ninja" ]; then
+    rm -rf "$OUT/CMakeCache.txt" "$OUT/CMakeFiles"
+    cmake -S "$SRC" -B "$OUT" -G Ninja --toolchain "$TC" -DMOBILEGL_BUILD_STAMP="$STAMP" \
+        -DCMAKE_BUILD_TYPE=Release -DMOBILEGL_BUILD_TEST=OFF -DMOBILEGL_BUILD_BENCHMARK=OFF \
+        -DMOBILEGL_BUILD_DISAGGREGATED=ON -DMOBILEGL_ENABLE_LTO=OFF
+elif ! grep -qx "MOBILEGL_BUILD_STAMP:STRING=$STAMP" "$OUT/CMakeCache.txt"; then
+    cmake -S "$SRC" -B "$OUT" -DMOBILEGL_BUILD_STAMP="$STAMP" >/dev/null
+fi
+t1=$(date +%s)
+cmake --build "$OUT" --target MobileGL mobilegl_gbm $EXTRA -j"$JOBS"
+t2=$(date +%s)
+echo "stamp:     $STAMP"
+echo "configure: $((t1-t0))s   build: $((t2-t1))s"
+for f in "$OUT/libMobileGL.so" "$OUT/MobileGL/MG_Gbm/mobilegl_gbm.so"; do
+    ls -la "$f"; echo "  windows: $(wslpath -w "$f")"
+done
+IN
+exit "${PIPESTATUS[0]}"
@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+# xdeploy.sh - install the WSL cross-built libMobileGL.so + mobilegl_gbm.so into the container, exactly
+# like anland's anland-build-client.sh: copy to a NEW inode and rename into place (overwriting a mapped
+# .so in place crashes every running client, KWin included), refresh the libEGL/libGL aliases, the
+# glvnd GLX vendor link, the EGL vendor JSON (system-wide 10_mobilegl.json + the /opt copy), the client
+# config /etc/mobilegl/client.conf and the GLX/GBM selection drop-ins.  Does not restart anything: already-running clients
+# keep the old inode; restart the client (or the session) to pick up the new library.
+# Env: MGL_XBUILD_DIR (WSL build dir), CHANNEL (ct.sh channel, default xd), WSL_DISTRO.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/env.sh"
+mkdir -p "$ANL_DIR/xdist"
+# 1. Copy the build outputs out of WSL.
+wsl -d "${WSL_DISTRO:-archlinux}" -- env OUT="${MGL_XBUILD_DIR:-}" D="$(win2wsl "$ANL_DIR")/xdist" bash -s <<'IN' 2>&1 | tr -d '\0' | grep -av '^wsl:'
+set -e; OUT=${OUT:-$HOME/mgl-xbuild-a64}
+cp -f "$OUT/libMobileGL.so" "$D/libMobileGL.so"; cp -f "$OUT/MobileGL/MG_Gbm/mobilegl_gbm.so" "$D/mobilegl_gbm.so"
+sha256sum "$D/libMobileGL.so" "$D/mobilegl_gbm.so"
+IN
+# 2. Push to the phone and stage them in the container's /root/xdeploy (the container's /tmp and the
+#    phone's /data/local/tmp are different places; the rootfs is reachable at $CT_ROOT).
+adb_ shell "mkdir -p $DEV_TOOLS/xdist"
+adb_ push "$ANL_DIR/xdist/libMobileGL.so" "$DEV_TOOLS/xdist/libMobileGL.so"
+adb_ push "$ANL_DIR/xdist/mobilegl_gbm.so" "$DEV_TOOLS/xdist/mobilegl_gbm.so"
+adb_ shell "su -c 'mkdir -p $CT_ROOT/root/xdeploy && cp $DEV_TOOLS/xdist/libMobileGL.so $DEV_TOOLS/xdist/mobilegl_gbm.so $CT_ROOT/root/xdeploy/ && rm -f $DEV_TOOLS/xdist/*.so'"
+# 3. Install inside the container.
+bash "$here/ct.sh" "${CHANNEL:-xd}" <<'IN'
+set -euo pipefail
+in=/root/xdeploy
+mkdir -p /opt/mobilegl/lib /opt/mobilegl/share/glvnd/egl_vendor.d
+cp "$in/libMobileGL.so" /opt/mobilegl/lib/libMobileGL.so.new
+mv -f /opt/mobilegl/lib/libMobileGL.so.new /opt/mobilegl/lib/libMobileGL.so
+for lib in libEGL.so libEGL.so.1 libGL.so libGL.so.1; do ln -sf libMobileGL.so "/opt/mobilegl/lib/$lib"; done
+ln -sf /opt/mobilegl/lib/libMobileGL.so /usr/lib/libGLX_mobilegl.so.0
+vendor='{"file_format_version":"1.0.0","ICD":{"library_path":"/opt/mobilegl/lib/libMobileGL.so"}}'
+# Per-process forcing (__EGL_VENDOR_LIBRARY_FILENAMES=<this file>) keeps working off the /opt copy.
+printf '%s\n' "$vendor" >/opt/mobilegl/share/glvnd/egl_vendor.d/50_mobilegl.json
+# System-wide vendor: 10_ sorts ahead of the system's 50_mesa.json, so stock libglvnd asks MobileGL
+# first and moves on to the other vendor whenever MobileGL declines (no server reachable).
+mkdir -p /usr/share/glvnd/egl_vendor.d
+printf '%s\n' "$vendor" >/usr/share/glvnd/egl_vendor.d/10_mobilegl.json.new
+mv -f /usr/share/glvnd/egl_vendor.d/10_mobilegl.json.new /usr/share/glvnd/egl_vendor.d/10_mobilegl.json
+# Client settings every process reads when its environment does not say otherwise (the backend
+# comes from /etc/mobilegl/backend, which switch.sh rewrites).
+mkdir -p /etc/mobilegl
+cat >/etc/mobilegl/client.conf <<'CONF'
+# MobileGL client configuration, read by every process that loads libMobileGL.so.
+# KEY=value lines; a MOBILEGL_* variable in a process's environment overrides the line here.
+# The backend is /etc/mobilegl/backend. Written by xdeploy.sh / anland-build-client.sh.
+MOBILEGL_TRANSPORT=spawn
+MOBILEGL_IPC_DATA=shm
+MOBILEGL_IPC_CONTROL=unix:@anland-mobilegl
+CONF
+# What glvnd's libGLX and libgbm cannot infer: environment.d reaches the systemd user manager (every
+# Plasma service and what it launches, terminals included), profile.d reaches login shells.
+mkdir -p /etc/environment.d /etc/profile.d
+printf '%s\n' '__GLX_VENDOR_LIBRARY_NAME=mobilegl' 'GBM_BACKEND=mobilegl' >/etc/environment.d/10-mobilegl.conf
+printf '%s\n' 'export __GLX_VENDOR_LIBRARY_NAME=mobilegl' 'export GBM_BACKEND=mobilegl' >/etc/profile.d/mobilegl.sh
+mkdir -p /usr/lib/gbm
+cp "$in/mobilegl_gbm.so" /usr/lib/gbm/mobilegl_gbm.so.new
+mv -f /usr/lib/gbm/mobilegl_gbm.so.new /usr/lib/gbm/mobilegl_gbm.so
+rm -f "$in/libMobileGL.so" "$in/mobilegl_gbm.so"
+sha256sum /opt/mobilegl/lib/libMobileGL.so /usr/lib/gbm/mobilegl_gbm.so
+ls /usr/share/glvnd/egl_vendor.d/ /etc/mobilegl/ /etc/environment.d/
+echo DEPLOYED
+IN
@@ -0,0 +1,47 @@
+# Cross toolchain: WSL clang -> aarch64 Arch Linux ARM (glibc) container "arch-kde-mgl".
+# Sysroot exported from the container by refresh-sysroot.sh (default $HOME/sysroots/arch-kde-mgl).
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR aarch64)
+
+set(_triple aarch64-unknown-linux-gnu)
+if (DEFINED ENV{MGL_SYSROOT})
+    set(_sysroot "$ENV{MGL_SYSROOT}")
+else()
+    set(_sysroot "$ENV{HOME}/sysroots/arch-kde-mgl")
+endif()
+# The container's GCC install dir (crtbegin/crtend, libgcc, libstdc++ via /usr/include/c++/<ver>).
+file(GLOB _gcc_dirs LIST_DIRECTORIES true "${_sysroot}/usr/lib/gcc/${_triple}/*")
+list(SORT _gcc_dirs COMPARE NATURAL ORDER DESCENDING)
+list(GET _gcc_dirs 0 _gcc_install_dir)
+
+set(CMAKE_SYSROOT "${_sysroot}")
+set(CMAKE_C_COMPILER clang)
+set(CMAKE_CXX_COMPILER clang++)
+set(CMAKE_ASM_COMPILER clang)
+set(CMAKE_C_COMPILER_TARGET ${_triple})
+set(CMAKE_CXX_COMPILER_TARGET ${_triple})
+set(CMAKE_ASM_COMPILER_TARGET ${_triple})
+set(CMAKE_AR llvm-ar)
+set(CMAKE_RANLIB llvm-ranlib)
+set(CMAKE_NM llvm-nm)
+set(CMAKE_OBJCOPY llvm-objcopy)
+set(CMAKE_STRIP llvm-strip)
+
+set(_xflags "--gcc-install-dir=${_gcc_install_dir}")
+set(CMAKE_C_FLAGS_INIT "${_xflags}")
+set(CMAKE_CXX_FLAGS_INIT "${_xflags}")
+set(CMAKE_ASM_FLAGS_INIT "${_xflags}")
+set(CMAKE_EXE_LINKER_FLAGS_INIT "-fuse-ld=lld")
+set(CMAKE_SHARED_LINKER_FLAGS_INIT "-fuse-ld=lld")
+set(CMAKE_MODULE_LINKER_FLAGS_INIT "-fuse-ld=lld")
+set(CMAKE_LINKER_TYPE LLD)
+
+set(CMAKE_FIND_ROOT_PATH "${_sysroot}")
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
+
+set(ENV{PKG_CONFIG_SYSROOT_DIR} "${_sysroot}")
+set(ENV{PKG_CONFIG_LIBDIR} "${_sysroot}/usr/lib/pkgconfig:${_sysroot}/usr/share/pkgconfig")
+set(ENV{PKG_CONFIG_PATH} "")
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# xsysroot/refresh-sysroot.sh - (re)export the cross sysroot from the container into WSL
+# ~/sysroots/arch-kde-mgl and install the toolchain file as ~/sysroots/aarch64-arch.cmake.
+# Read-only on the container side (it only tars headers/libs into /root and moves the tarball out).
+# Run once on a new machine, and again after the container upgrades glibc/gcc/vulkan/libdrm/gbm;
+# then `bash xbuild.sh --reconfigure`.  Env: CHANNEL (default xr), WSL_DISTRO.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+. "$here/../env.sh"
+cp -f "$here/aarch64-arch.cmake" "$ANL_DIR/aarch64-arch.cmake"
+bash "$here/../ct.sh" "${CHANNEL:-xr}" <<'IN'
+set -e; cd /
+L=$(ls -d usr/lib/*crt*.o usr/lib/lib{c,m,mvec,pthread,dl,rt,util,resolv,anl,stdc++,stdc++fs,stdc++exp,supc++,gcc_s,atomic,vulkan,drm,drm_*,gbm,z}.* usr/lib/lib{c,m}_nonshared* usr/lib/ld-linux* usr/lib/pkgconfig usr/share/pkgconfig usr/share/cmake/VulkanHeaders usr/lib/cmake/Vulkan* 2>/dev/null)
+tar czf /root/xsysroot.tar.gz --exclude='usr/lib/gcc/*/*/cc1' --exclude='usr/lib/gcc/*/*/cc1plus' --exclude='usr/lib/gcc/*/*/lto1' --exclude='usr/lib/gcc/*/*/lto-wrapper' --exclude='usr/lib/gcc/*/*/collect2' --exclude='usr/lib/gcc/*/*/plugin' --exclude='usr/lib/gcc/*/*/install-tools' --exclude='usr/lib/gcc/*/*/liblto_plugin.so' lib usr/include usr/lib/gcc $L
+echo SYSROOT_TAR_DONE
+IN
+adb_ shell "su -c 'mv $CT_ROOT/root/xsysroot.tar.gz $DEV_TOOLS/xsysroot.tar.gz && chmod 644 $DEV_TOOLS/xsysroot.tar.gz'"
+adb_ pull "$DEV_TOOLS/xsysroot.tar.gz" "$ANL_DIR/xsysroot.tar.gz"
+adb_ shell "su -c 'rm -f $DEV_TOOLS/xsysroot.tar.gz'"
+wsl -d "${WSL_DISTRO:-archlinux}" -- env A="$(win2wsl "$ANL_DIR")" bash -s <<'IN' 2>&1 | tr -d '\0' | grep -av '^wsl:'
+set -e; S=$HOME/sysroots/arch-kde-mgl; rm -rf "$S.new"; mkdir -p "$S.new"
+tar xzf "$A/xsysroot.tar.gz" -C "$S.new"
+rm -rf "$S"; mv "$S.new" "$S"; cp -f "$A/aarch64-arch.cmake" "$HOME/sysroots/aarch64-arch.cmake"
+du -sh "$S"; ls "$S/usr/lib/gcc/aarch64-unknown-linux-gnu/"
+IN
+rm -f "$ANL_DIR/xsysroot.tar.gz"
@@ -61,15 +61,40 @@ if ! grep -Eq '^lib/[^/]+/libMobileGL\.so$' <<<"$plugin_contents"; then
  exit 1
 fi

+# The render server is not trace-only: both flavors carry the control screen, the offscreen
+# service, the on-screen display activity, and (split native builds) the supervisor executable.
+# The plugin APK ships to users, so its build keeps the three components private (manifest
+# placeholder serverComponentsExported=false); only the debuggable trace APK exports them for adb.
+for component in ServerControlActivity MobileGLServerService MobileGLDisplayActivity; do
+  require "$component" "$plugin_manifest" "plugin render server component ($component)"
+  require "$component" "$trace_manifest" "trace render server component ($component)"
+done
+if [[ "${MOBILEGL_CI_DISAGGREGATED:-OFF}" == "ON" ]]; then
+  trace_contents=$(unzip -Z1 "$trace_apk")
+  for entry in "plugin:${plugin_contents}" "trace:${trace_contents}"; do
+    if ! grep -Eq '^lib/[^/]+/libMobileGLServer\.so$' <<<"${entry#*:}"; then
+      echo "::error::${entry%%:*} APK does not contain libMobileGLServer.so (both flavors ship the render server)" >&2
+      exit 1
+    fi
+  done
+fi
+
 require 'top.mobilegl.plugin.trace' "$trace_manifest" 'trace package name'
 require 'top.mobilegl.plugin.TRACE_REPLAY' "$trace_manifest" 'trace replay action'
 if grep -Fq 'fclPlugin' <<<"$trace_manifest"; then
  echo '::error::Trace APK must not advertise renderer-plugin metadata' >&2
  exit 1
 fi
-if grep -Fq 'android.intent.action.MAIN' <<<"$trace_manifest"; then
-  echo '::error::Trace APK must not expose a launcher activity' >&2
+# The trace APK IS a launchable app now: PostActivity (the POST self-test screen) is its
+# entry icon and the bottom navigation bar switches to the render server screen. What is
+# left of the old "no launcher" contract is the count - exactly ONE MAIN activity, so a
+# screen that accidentally grows a launcher entry (ServerControlActivity was one for a day)
+# reds here instead of shipping two icons.
+main_count=$(grep -Fc 'android.intent.action.MAIN' <<<"$trace_manifest")
+if [[ "$main_count" -ne 1 ]]; then
+  echo "::error::Trace APK must expose exactly one launcher activity (PostActivity); found ${main_count} android.intent.action.MAIN" >&2
  exit 1
 fi
+require 'PostActivity' "$trace_manifest" 'trace launcher activity (PostActivity)'

 echo 'Validated unified MobileGL plugin APK and isolated trace APK.'
@@ -18,6 +18,10 @@ on:
      - '**/*.md'
  workflow_dispatch:

+env:
+  # Software-rendered CI submissions may exceed the interactive GPU hang budget.
+  MOBILEGL_GPU_HANG_BUDGET_MS: "0"
+
 # ONE LIVE RUN PER BRANCH, as in test.yml: a superseded push is cancelled rather than left queued
 # behind its successor; a manual dispatch gets a group of its own; dev queues rather than cancels.
 concurrency:
@@ -477,7 +481,7 @@ jobs:
          extra_retrace_args=()
          if [ "${MOBILEGL_CI_TRANSPORT}" = "inproc" ]; then
            extra_retrace_args+=(
-              --env "MOBILEGL_TRANSPORT=inproc;MOBILEGL_IPC_ROLE_SPLIT_STATE=1;MOBILEGL_IPC_STRICT_ERRORS=1;MOBILEGL_IPC_RUN_AHEAD=1"
+              --env "MOBILEGL_TRANSPORT=inproc;MOBILEGL_IPC_ROLE_SPLIT_STATE=1;MOBILEGL_IPC_STRICT_ERRORS=1;MOBILEGL_IPC_RUN_AHEAD=1;MOBILEGL_GPU_HANG_BUDGET_MS=${MOBILEGL_GPU_HANG_BUDGET_MS}"
              --require-inproc
            )
          elif [ "${MOBILEGL_CI_TRANSPORT}" = "spawn" ]; then
@@ -493,12 +497,12 @@ jobs:
            # into the CLIENT's process and SetWindowHandle is refused on the way across with
            # Fatal{UnmigratedSurface, "AndroidNativeWindow@P12"} (Rule H).
            extra_retrace_args+=(
-              --env "MOBILEGL_TRANSPORT=spawn;MOBILEGL_IPC_ROLE_SPLIT_STATE=1;MOBILEGL_IPC_STRICT_ERRORS=1;MOBILEGL_IPC_RUN_AHEAD=1"
+              --env "MOBILEGL_TRANSPORT=spawn;MOBILEGL_IPC_ROLE_SPLIT_STATE=1;MOBILEGL_IPC_STRICT_ERRORS=1;MOBILEGL_IPC_RUN_AHEAD=1;MOBILEGL_GPU_HANG_BUDGET_MS=${MOBILEGL_GPU_HANG_BUDGET_MS}"
              --use-pbuffer
              --require-spawn
            )
          else
-            extra_retrace_args+=(--env "MOBILEGL_TRANSPORT=monolith")
+            extra_retrace_args+=(--env "MOBILEGL_TRANSPORT=monolith;MOBILEGL_GPU_HANG_BUDGET_MS=${MOBILEGL_GPU_HANG_BUDGET_MS}")
          fi
          # Bliss needs the newer signed ANGLE variant plus sampler mipmap
          # min-filter downgrading on ANGLE llvmpipe.
@@ -21,6 +21,10 @@ on:
        required: false
        default: "087685d1"

+env:
+  # Software-rendered CI submissions may exceed the interactive GPU hang budget.
+  MOBILEGL_GPU_HANG_BUDGET_MS: "0"
+
 # ONE LIVE RUN PER BRANCH. The disaggregation branch pushes (often docs-only) commits minutes apart
 # and each push fans out to ~330 retrace legs here and ~275 jobs in apk.yml; with no group, runs
 # sat queued for 10+ hours behind their own successors. A manual dispatch gets a group of its own,
@@ -413,6 +417,12 @@ jobs:
            --discovery "${RUNNER_TEMP}/primary-inproc-lane.json" \
            --junit "${RUNNER_TEMP}/primary-inproc.xml" --output "${RUNNER_TEMP}/primary-inproc-proof.json"

+      # The run log needs a signed-in reader; annotations do not. Name every failed ctest entry,
+      # with the lines of its output that say why, where anyone reading the run can see them.
+      - name: Annotate failed entries
+        if: failure()
+        run: python3 scripts/ci/junit_annotate.py "${RUNNER_TEMP}"
+
      - name: Upload integration lane evidence
        if: always()
        uses: actions/upload-artifact@v7
@@ -1169,6 +1179,12 @@ jobs:
      # arm proof read; verify-split-logs/ is whatever the control steps left behind, which for
      # twelve entries is a deliberately red re-run under the same name. Uploading only the
      # second - which is what this step did before the copy above existed - loses the first.
+      # The run log needs a signed-in reader; annotations do not. Name every failed ctest entry,
+      # with the lines of its output that say why, where anyone reading the run can see them.
+      - name: Annotate failed entries
+        if: failure()
+        run: python3 scripts/ci/junit_annotate.py "${RUNNER_TEMP}"
+
      - name: Upload verify-split lane logs
        if: always()
        uses: actions/upload-artifact@v7
@@ -1461,11 +1477,27 @@ jobs:
      - name: Spawn/split lane parity (9.2)
        run: python3 scripts/ci/spawn_lane_parity.py build-split

+      # P8-A (ID-P8-2): every monolith case, tails normalised, is on a gated split arm or is named
+      # with a reason class in scripts/data/split_coverage_exemptions.txt. The same `ctest -N`
+      # cost as the parity step above, and the question it cannot answer: parity compares the
+      # split arms with EACH OTHER, so a case missing from all three passes it.
+      - name: Split coverage against monolith (P8-A)
+        run: |
+          python3 scripts/ci/split_coverage.py --self-test
+          python3 scripts/ci/split_coverage.py build-split
+
+      # P8-SE: PipeCatalogueTest's unmigrated-emulation list equals the MGPipeUnmigratedEmulation
+      # call sites in the tree, one site per name. A source lint: no build needed.
+      - name: Unmigrated-emulation list against its call sites (P8-SE)
+        run: |
+          python3 scripts/ci/unmigrated_emulation_sites.py --self-test
+          python3 scripts/ci/unmigrated_emulation_sites.py
+
      - name: Unit tests on the split runtime
        working-directory: build-split
        env:
          MOBILEGL_ITEST_REQUIRE_GPU: "1"
-        run: ctest --output-on-failure -L unit --no-tests=error -j "$(nproc)"
+        run: ctest --output-on-failure -L unit --no-tests=error -j "$(nproc)" --output-junit "${RUNNER_TEMP}/unit-split.xml"

      # P7 wave 2-F, PH-7 (1). THE TCP SUPERVISOR LANE NEEDS flatc, AND CI HAD NONE.
      #
@@ -1632,6 +1664,12 @@ jobs:
          CONTROL_TMPDIR: ${{ runner.temp }}
        run: bash "${GITHUB_WORKSPACE}/scripts/ci/split_negative_controls.sh"

+      # The run log needs a signed-in reader; annotations do not. Name every failed ctest entry,
+      # with the lines of its output that say why, where anyone reading the run can see them.
+      - name: Annotate failed entries
+        if: failure()
+        run: python3 scripts/ci/junit_annotate.py "${RUNNER_TEMP}"
+
      - name: Upload split lane logs
        if: always()
        uses: actions/upload-artifact@v7
@@ -1721,7 +1759,7 @@ jobs:
        env:
          MOBILEGL_ITEST_REQUIRE_GPU: "1"
          MOBILEGL_IPC_STRICT_ERRORS: "1"
-        run: ctest --output-on-failure -L unit --no-tests=error -j "$(nproc)"
+        run: ctest --output-on-failure -L unit --no-tests=error -j "$(nproc)" --output-junit "${RUNNER_TEMP}/unit-strict.xml"

      # ---- P5e (ra), MG_Remote/CONTRACT-P5E.md §7: ONE STEP, TWO ARMS, AND THE TREE PICKS ----
      #
@@ -2042,6 +2080,13 @@ jobs:
              "PersistentCoherentMapScenario.TwoWritesThroughTheCoherentPointerEachReachTheirOwnDraw",
              "PersistentCoherentMapScenario.AWriteAfterAFrameBoundaryReachesTheNextFramesDraw",
              "StorageBufferRegrowScenario.AGrownStoreIsVisibleThroughItsExistingIndexedBinding",
+              # P11 A2: the >= 16 MiB arena and its tier assertion (emulated).
+              "LargeArenaAdoptionScenario.SubDataAfterAnInFlightDrawReachesTheNextDraw",
+              "LargeArenaAdoptionScenario.RespecifiedVertexArenaKeepsVaoBindings",
+              "LargeArenaAdoptionScenario.RespecifiedIndexArenaKeepsVaoBinding",
+              "LargeArenaAdoptionScenario.ReadbackSeesTheLatestCpuWrite",
+              "LargeArenaAdoptionScenario.GpuWriteIntoTheArenaIsReadBack",
+              "LargeArenaAdoptionScenario.TheArenaLandsInTheTierItsLaneDeclares",
          ]
          slp.require_green(json.load(open(sys.argv[2])), sys.argv[3],
                            ["DirectVulkan.Split.Buffers." + name for name in cases])
@@ -2104,6 +2149,12 @@ jobs:
              "PersistentCoherentMapScenario.TwoWritesThroughTheCoherentPointerEachReachTheirOwnDraw",
              "PersistentCoherentMapScenario.AWriteAfterAFrameBoundaryReachesTheNextFramesDraw",
              "StorageBufferRegrowScenario.AGrownStoreIsVisibleThroughItsExistingIndexedBinding",
+              "LargeArenaAdoptionScenario.SubDataAfterAnInFlightDrawReachesTheNextDraw",
+              "LargeArenaAdoptionScenario.RespecifiedVertexArenaKeepsVaoBindings",
+              "LargeArenaAdoptionScenario.RespecifiedIndexArenaKeepsVaoBinding",
+              "LargeArenaAdoptionScenario.ReadbackSeesTheLatestCpuWrite",
+              "LargeArenaAdoptionScenario.GpuWriteIntoTheArenaIsReadBack",
+              "LargeArenaAdoptionScenario.TheArenaLandsInTheTierItsLaneDeclares",
          ]
          required = [sys.argv[4] + name for name in cases]
          if sys.argv[5] == "tcp":
@@ -2217,6 +2268,12 @@ jobs:
          slp.require_green(lane, sys.argv[3], names)
          PY

+      # The run log needs a signed-in reader; annotations do not. Name every failed ctest entry,
+      # with the lines of its output that say why, where anyone reading the run can see them.
+      - name: Annotate failed entries
+        if: failure()
+        run: python3 scripts/ci/junit_annotate.py "${RUNNER_TEMP}"
+
      - name: Upload strict lane logs
        if: always()
        uses: actions/upload-artifact@v7
@@ -2722,7 +2779,11 @@ jobs:
             && [ '${{ matrix.case }}' = 'improved-transparency-minecraft-26.3' ]; then
            export MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE=1
          fi
-          ctest -V --no-tests=error -R '^MobileGLTraceReplay\.${{ matrix.case }}\.${{ matrix.backend }}$'
+          ctest -V --no-tests=error -R '^MobileGLTraceReplay\.${{ matrix.case }}\.${{ matrix.backend }}$'             --output-junit "${RUNNER_TEMP}/retrace.xml"
+
+      - name: Annotate failed entries
+        if: failure()
+        run: python3 scripts/ci/junit_annotate.py --excerpts 1 "${RUNNER_TEMP}"

      - name: Upload core dumps
        if: failure()
@@ -3477,27 +3538,10 @@ jobs:
    runs-on: ubuntu-latest
    # Deliberately independent of build-linux: these are source-level gates, they take
    # seconds, and a broken build must not hide a drifted interface.
-    env:
-      # THE CURRENT PHASE's base ref, for the two G5 region gates below. It is 37da3c3a - P4a's
-      # base ref, INTEGRATOR-DECISIONS ID-1 - and NOT the workflow's baseline_sha input: that input
-      # is the SYMBOL baseline (087685d1, P1's G1 reading) and it is empty on a push, whereas these
-      # gates ask "did the do-not-touch list move since the phase started".
-      #
-      # IT MOVED FROM P3a's 44c2b5cf TO P4a's 37da3c3a WITH THE PHASE, and that is a deliberate
-      # narrowing rather than a loss: P3a's eleven functions were compared against 44c2b5cf at P3a's
-      # own exit and were byte-identical there, so 37da3c3a carries the same bodies (measured: the
-      # eleven shas at 37da3c3a are the eleven shas at 44c2b5cf, and FlushPendingRangesFrom's is
-      # still the sha pinned in the script at 3e298c9a). What the two gates now both answer is "did
-      # anything on the list move during P4a", which is the question this phase can act on.
-      BASELINE: "37da3c3a"

    steps:
      - name: Checkout repo
        uses: actions/checkout@v6
-        with:
-          # The G5 gate reads Managers.cpp at BASELINE with `git show`, which a depth-1 checkout
-          # does not have. Nothing else in this job needs history.
-          fetch-depth: 0

      # The seven generators all read MG_Pipe/*.def, so regenerating and diffing is what
      # keeps the two interface tables, the wire records, the verify comparators, the
@@ -3626,76 +3670,6 @@ jobs:
          python3 tools/trace_replay/test_run_tcp_matrix.py
          python3 tools/trace_replay/test_compare_actuals.py

-      # A GATE as of P3a (G5). "pool 与延迟释放原样搬" (ROADMAP.md:19) is meant literally: the
-      # buffer pool, the deferred-release drain and the three persistently mapped rings move
-      # VERBATIM, and ARCHITECTURE.md:515 says why - their retire happens only inside Present, so
-      # a batching or ordering change there starves them, and nothing else in this workflow can
-      # see it. P3a rewrites the rest of Managers.cpp by design, so a file diff says nothing; the
-      # script extracts the ELEVEN named bodies and compares their hashes on their own.
-      #
-      # Eleven and not ten (ID-15): Managers.cpp carries the three-tier flush drain TWICE, once
-      # per preprocessor arm, and a push build compiles only FlushPendingRangesFrom while the
-      # untouched FlushPendingRangesNow lives in the `#else`. Hashing the pull name alone would
-      # protect text no shipping build compiles, so both are hashed - the pull ladder against
-      # BASELINE, the push ladder against a sha pinned in the script at 3dadd4c1 (ID-41), because
-      # that one was born in P3a and is compared against the reviewed body rather than against the
-      # base ref.
-      #
-      # Scoped to the disaggregation branch and to a manual dispatch, deliberately: the question
-      # is "did these eleven move since P3a started", and BASELINE is P3a's base ref. On dev,
-      # where unrelated buffer fixes land on their own schedule, the same comparison would be
-      # asking a question nobody posed - it belongs with the TEMPORARY trigger lines at the top
-      # of this file and retires with them.
-      #
-      # --self-test is the half that keeps it honest, and it is not optional: a comparison that
-      # silently stopped comparing produces exactly the same green as eleven untouched bodies. It
-      # runs eight canned controls - eleven bodies extracted, an untouched copy compared equal, an
-      # edit OUTSIDE them ignored, each of the three perturbation targets (ClearBufferPool and BOTH
-      # flush ladders) reported BY NAME, plus the pin-precedence control and a one-token edit to the
-      # pinned ladder compared against the pin - and fails if any of them does not answer.
-      # Same shape as gen_pipe.py --self-test above.
-      - name: P3a regions unchanged (G5)
-        if: ${{ github.ref == 'refs/heads/feat/disaggregated' || github.event_name == 'workflow_dispatch' }}
-        run: bash scripts/p3a_untouched_regions.sh "${BASELINE}" HEAD
-
-      - name: P3a untouched-region self-test (G5)
-        if: ${{ github.ref == 'refs/heads/feat/disaggregated' || github.event_name == 'workflow_dispatch' }}
-        run: bash scripts/p3a_untouched_regions.sh --self-test
-
-      # A GATE AS OF P4a (G5), and a SECOND script rather than an edit to the one above. P4a extends
-      # the same claim to the rest of ARCHITECTURE.md:318's do-not-touch list - the unpack PBO ring's
-      # staging repack and its two helpers, the attachment permutation, the D24S8 sampling-emulation
-      # core and the format-caveat handler - which is SEVENTEEN regions across THREE files
-      # (BRIEF-P4A.md D-N: P3a's eleven, which P4a must not touch either, plus P4a's six). The
-      # parent's SOURCE_PATH is a single file, so the extension needed a per-region source path and
-      # a region KIND (DepthStencilSamplingReadImpl is a namespace, not a function); everything else
-      # about the extraction is its parent's, verbatim.
-      #
-      # Both scripts run. The parent keeps answering its own question against its own eleven, so a
-      # regression in either half names itself, and neither gate can be silenced by editing the
-      # other's list.
-      #
-      # Same feat/disaggregated-or-dispatch guard as the P3a step, for the same reason: the question
-      # is "did these move since the phase started", and on dev - where unrelated buffer and texture
-      # fixes land on their own schedule - it would be a question nobody posed. It belongs with the
-      # TEMPORARY trigger lines at the top of this file and retires with them.
-      #
-      # --self-test is the half that keeps it honest and is not optional: a comparison that silently
-      # stopped comparing produces exactly the same green as seventeen untouched regions. It runs
-      # three positive controls (seventeen regions extracted, an untouched copy compared equal, an
-      # edit OUTSIDE them invisible in all three files) and NINE negative ones - ClearBufferPool,
-      # FlushPendingRangesNow, RecomputeBackendColorSlots and StageBlocksIntoUnpackRing, each
-      # perturbed at its HEAD and at its TAIL and each required to be named BY NAME, plus a
-      # one-token edit to the pinned FlushPendingRangesFrom body compared against the pin - and
-      # fails if any of them does not answer.
-      - name: P4a regions unchanged (G5)
-        if: ${{ github.ref == 'refs/heads/feat/disaggregated' || github.event_name == 'workflow_dispatch' }}
-        run: bash scripts/p4a_untouched_regions.sh "${BASELINE}" HEAD
-
-      - name: P4a untouched-region self-test (G5)
-        if: ${{ github.ref == 'refs/heads/feat/disaggregated' || github.event_name == 'workflow_dispatch' }}
-        run: bash scripts/p4a_untouched_regions.sh --self-test
-
      # Warning only for now: the disaggregation documents are still being written, and a
      # lint that fails a rewrite in progress teaches people to ignore it. It becomes
      # --strict when the documents settle.
@@ -0,0 +1,31 @@
+name: Arch WSL GPU smoke
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - ci/arch-wsl-gpu-runner
+    paths:
+      - .github/workflows/wsl-gpu-smoke.yml
+      - docs/ci/arch-wsl-gpu-runner.md
+
+permissions: {}
+
+concurrency:
+  group: arch-wsl-gpu-smoke
+  cancel-in-progress: false
+
+jobs:
+  gpu-smoke:
+    runs-on: [self-hosted, Linux, X64, arch, wsl, gpu, gles, vulkan]
+    timeout-minutes: 5
+    steps:
+      - name: Report runner identity
+        shell: bash
+        run: |
+          printf 'Runner: %s\n' "$RUNNER_NAME"
+          uname -n
+          uname -a
+      - name: Verify hardware GLES and Vulkan
+        shell: bash
+        run: /usr/local/libexec/mobilegl-runner/gpu-smoke
@@ -28,3 +28,4 @@ tools/trace_replay/work/
 __pycache__/
 *.py[cod]
 /.gradle
+/.trace-work/
@@ -0,0 +1,11 @@
+# glvnd vendor ABI headers
+
+'libeglabi.h' and 'GLdispatchABI.h' from libglvnd (MIT), copied verbatim with their own licence
+headers intact.
+
+They are here because MobileGL implements the *vendor* half of glvnd's EGL ABI - the side a library
+exposes so the system's 'libEGL.so.1', which IS glvnd, can dispatch to it.  The distribution ships
+the runtime but not these headers in the base image, and a vendored copy of two MIT headers is a
+smaller dependency than a build-time download of a source tree.
+
+Upstream ABI version implemented: 'EGL_VENDOR_ABI_MAJOR_VERSION' 0 (the header pins it).
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2013, NVIDIA CORPORATION.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and/or associated documentation files (the
+ * "Materials"), to deal in the Materials without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sublicense, and/or sell copies of the Materials, and to
+ * permit persons to whom the Materials are furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * unaltered in all copies or substantial portions of the Materials.
+ * Any additions, deletions, or changes to the original source files
+ * must be clearly indicated in accompanying documentation.
+ *
+ * THE MATERIALS ARE 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
+ * MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS.
+ */
+
+#include <GL/gl.h>
+
+#if !defined(__GL_DISPATCH_ABI_H)
+#define __GL_DISPATCH_ABI_H
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/*!
+ * \defgroup gldispatchabi GL dispatching ABI
+ *
+ * This is not a complete ABI, but rather a fragment common to the libEGL and
+ * libGLX ABIs.  Changes to this file should be accompanied by a version bump to
+ * these client ABIs.
+ */
+
+/*!
+ * Thread-local implementation used by libglvnd. This is passed into the patch
+ * function callback via the type parameter.
+ *
+ * For most architectures, the vendor library can ignore this parameter, since
+ * it will always be the same value. It's used for systems like ARM, where the
+ * stubs might be use the ARM or Thumb instruction sets.
+ *
+ * The stub type does not make any distinction between TLS and TSD stubs. The
+ * entire purpose of entrypoint rewriting is to skip the dispatch table in
+ * libGLdispatch.so, so it doesn't matter how that dispatch table is stored.
+ */
+enum {
+    /*!
+     * Indicates that the stubs aren't defined in assembly. For example, if the
+     * dispatch stubs are written in C. Vendor libraries generally won't see
+     * this value.
+     */
+    __GLDISPATCH_STUB_UNKNOWN,
+
+    /*!
+     * Used for stubs on x86 systems.
+     */
+    __GLDISPATCH_STUB_X86,
+
+    /*!
+     * Used for stubs on x86-64 systems.
+     */
+    __GLDISPATCH_STUB_X86_64,
+
+    /*!
+     * Used for stubs on ARMv7, using the Thumb instruction set.
+     */
+    __GLDISPATCH_STUB_ARMV7_THUMB,
+
+    /*!
+     * Used for stubs on ARMv7, using the normal ARM instruction set.
+     */
+    __GLDISPATCH_STUB_ARMV7_ARM,
+
+    /*!
+     * Used for stubs on ARMv8/aarch64.
+     */
+    __GLDISPATCH_STUB_AARCH64,
+
+    /*!
+     * Used for stubs on x32 builds (x86-64 with 32-bit pointers).
+     */
+    __GLDISPATCH_STUB_X32,
+
+    /*!
+     * Used for stubs on PPC64 systems.
+     */
+    __GLDISPATCH_STUB_PPC64,
+
+    /*!
+     * Used for stubs on PPC64LE systems. Same as PPC64, for compatibility.
+     */
+    __GLDISPATCH_STUB_PPC64LE = __GLDISPATCH_STUB_PPC64,
+
+    /*!
+     * Used for stubs on LOONGARCH/loongarch64.
+     */
+    __GLDISPATCH_STUB_LOONGARCH64,
+};
+
+/*!
+ * A callback function called by the vendor library to fetch the address of an
+ * entrypoint.
+ *
+ * The function returns two pointers, one writable and one executable. The two
+ * pointers may or may not be the same virtual address, but they will both be
+ * mappings of the same physical memory.
+ *
+ * The vendor library should write its entrypoint to the address returned by
+ * \p writePtr, but should use the address from \p execPtr for things like
+ * calculating PC-relative offsets.
+ *
+ * Note that if this function fails, then the vendor library can still try to
+ * patch other entrypoints.
+ *
+ * Note that on ARM, the low-order bit of both \c execPtr and \p writePtr will
+ * be zero, even if the stub uses the thumb instruction set. The vendor library
+ * should use the \c type parameter of \c initiatePatch to determine which
+ * instruction set to use.
+ *
+ * \param funcName The function name.
+ * \param[out] writePtr The pointer that the vendor library can write to.
+ * \param[out] execPtr The pointer to the executable code.
+ * \return GL_TRUE if the entrypoint exists, or GL_FALSE if it doesn't.
+ */
+typedef GLboolean (*DispatchPatchLookupStubOffset)(const char *funcName,
+        void **writePtr, const void **execPtr);
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif // __GL_DISPATCH_ABI_H
@@ -0,0 +1,464 @@
+/*
+ * Copyright (c) 2013, NVIDIA CORPORATION.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and/or associated documentation files (the
+ * "Materials"), to deal in the Materials without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sublicense, and/or sell copies of the Materials, and to
+ * permit persons to whom the Materials are furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * unaltered in all copies or substantial portions of the Materials.
+ * Any additions, deletions, or changes to the original source files
+ * must be clearly indicated in accompanying documentation.
+ *
+ * THE MATERIALS ARE 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
+ * MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS.
+ */
+
+#if !defined(__LIB_EGL_ABI_H)
+#define __LIB_EGL_ABI_H
+
+#include <stdint.h>
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+
+#include "glvnd/GLdispatchABI.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/*!
+ * \defgroup eglvendorabi EGL Vendor ABI
+ *
+ * Definition of ABI exported by libEGL.so to libEGL_VENDOR.so libraries.
+ *
+ * Each vendor is associated with three distinct dispatch table types:
+ *
+ * - static EGL dispatch table: this is the fixed list of EGL 1.5 entrypoints
+ *   provided by the vendor at load time during the initial handshake.
+ * - dynamic EGL dispatch table: this is a structure allocated by the API
+ *   library at runtime used to manage EGL extension functions which are not
+ *   present in the static table.
+ * - core GL dispatch table: this is a structure maintained by the API library
+ *   which contains both GL core (static) and GL extension (dynamic) functions.
+ *
+ * @{
+ */
+
+/*
+ * Rendering API handling:
+ *
+ * libEGL only supports OpenGL and OpenGL ES, not OpenVG. If OpenVG or any
+ * other API is added, then the major version number will be incremented.
+ *
+ * When the application calls eglBindAPI, libEGL will forward the call to every
+ * vendor library. In addition, a vendor library can query the current API from
+ * libEGL using the getCurrentApi callback.
+ *
+ * Vendor libraries are not required to support both GL and GLES, but they must
+ * be able to deal with either one as the current rendering API. If a vendor
+ * doesn't support the current API, then it should return an error from
+ * eglCreateContext.
+ */
+
+/*!
+ * Current version of the ABI.
+ *
+ * This version number contains a major number in the high-order 16 bits, and
+ * a minor version number in the low-order 16 bits.
+ *
+ * The major version number is incremented when an interface change will break
+ * backwards compatibility with existing vendor libraries. The minor version
+ * number is incremented when there's a change but existing vendor libraries
+ * will still work.
+ */
+#define EGL_VENDOR_ABI_MAJOR_VERSION ((uint32_t) 0)
+#define EGL_VENDOR_ABI_MINOR_VERSION ((uint32_t) 2)
+#define EGL_VENDOR_ABI_VERSION ((EGL_VENDOR_ABI_MAJOR_VERSION << 16) | EGL_VENDOR_ABI_MINOR_VERSION)
+static inline uint32_t EGL_VENDOR_ABI_GET_MAJOR_VERSION(uint32_t version)
+{
+    return version >> 16;
+}
+static inline uint32_t EGL_VENDOR_ABI_GET_MINOR_VERSION(uint32_t version)
+{
+    return version & 0xFFFF;
+}
+
+
+/*!
+ * This opaque structure stores function pointers for EGL extension functions.
+ * It is allocated at runtime by the API library. Vendor-provided dispatch
+ * functions retrieve and operate on this structure using the API below.
+ */
+typedef struct __EGLvendorInfoRec __EGLvendorInfo;
+
+/****************************************************************************
+ * API library exports                                                      *
+ ****************************************************************************/
+
+typedef struct __EGLapiExportsRec {
+    /************************************************************************
+     * The following routines are used by vendor-implemented EGL dispatch
+     * functions to lookup and call into the right vendor.
+     ************************************************************************/
+
+    /*!
+     * This function must be called at the start of every EGL dispatch stub. It
+     * performs any necessary per-call bookkeeping.
+     */
+    void (* threadInit) (void);
+
+    EGLenum (* getCurrentApi)(void);
+
+    /*!
+     * This function retrieves the appropriate current dynamic dispatch table,
+     * if a GL context is current. Otherwise, this returns NULL.
+     */
+    __EGLvendorInfo *(*getCurrentVendor)(void);
+
+    /*!
+     * This retrieves the current context for this thread.
+     */
+    EGLContext (*getCurrentContext)(void);
+
+    /*!
+     * Returns the current display for this thread.
+     */
+    EGLDisplay (*getCurrentDisplay)(void);
+
+    /*!
+     * Returns the current drawable for this thread.
+     */
+    EGLSurface (*getCurrentSurface)(EGLint readDraw);
+
+    /*!
+     * This function retrieves an entry point from the dynamic dispatch table
+     * given an index into the table.
+     */
+    __eglMustCastToProperFunctionPointerType (*fetchDispatchEntry) (__EGLvendorInfo *dynDispatch, int index);
+
+    /*!
+     * Sets the last error for the current thread. The error will be returned
+     * the next time the app calls eglGetError().
+     *
+     * This function will override a call to \c setLastVendor, and vice-versa.
+     */
+    void (* setEGLError) (EGLint errorCode);
+
+    /*!
+     * Notifies libEGL about the vendor library that an EGL function is
+     * dispatched to. This is used to look up the last EGL error code from a
+     * vendor.
+     *
+     * A vendor library may generate an EGL error asynchronously, if it
+     * offloads some EGL functions onto a worker thread. In that case, calling
+     * \c setEGLError to set the last error would not work.
+     *
+     * When an EGL dispatch stubs finds the vendor library, it must call this
+     * function before it dispatches the function call.
+     *
+     * This function will override a call to \c setEGLError, and vice-versa.
+     *
+     * \param vendor The vendor library of the current EGL function call.
+     * \return EGL_TRUE on success, or EGL_FALSE on failure.
+     */
+    EGLBoolean (* setLastVendor) (__EGLvendorInfo *vendor);
+
+    /************************************************************************
+     * These routines are used by vendor dispatch functions to look up
+     * and add mappings between various objects and vendors.
+     ************************************************************************/
+
+    // As far as I can tell, all EGL functions that take a context or drawable
+    // also take an EGLDisplay. So, we don't need any extra functions for
+    // mapping a context/surface/whatever to a EGLDisplay or vendor.
+
+    /*!
+     * Returns the EGL vendor for an EGLDisplay handle.
+     */
+    __EGLvendorInfo *(*getVendorFromDisplay)(EGLDisplay dpy);
+
+    /*!
+     * Returns the EGL vendor for an EGLDeviceEXT handle.
+     */
+    __EGLvendorInfo *(*getVendorFromDevice)(EGLDeviceEXT dev);
+
+    /**
+     * Sets the EGL vendor for an EGLDeviceEXT handle. The dispatch stub for
+     * any function that returns an EGLDeviceEXT handle should call this.
+     *
+     * Supported since ABI version 1.2.
+     */
+    EGLBoolean (* setVendorForDevice)(EGLDeviceEXT dev, __EGLvendorInfo *vendor);
+} __EGLapiExports;
+
+/*****************************************************************************
+ * API library imports                                                       *
+ *****************************************************************************/
+
+/*!
+ * The enum values accepted by \c __EGLapiImports::getVendorString.
+ */
+enum
+{
+    /*!
+     * The set of platform extensions that the vendor supports.
+     *
+     * Platform extensions typically just add enums to the existing
+     * eglGetPlatformDisplay function, so we don't need any additional logic
+     * or special handling from libEGL itself.
+     *
+     * As a result, vendor libraries can add these extensions without having
+     * to modify libEGL to match.
+     */
+    __EGL_VENDOR_STRING_PLATFORM_EXTENSIONS,
+};
+
+/*!
+ * This structure stores required and optional vendor library callbacks.
+ */
+typedef struct __EGLapiImportsRec {
+    /**
+     * Creates an EGLDisplay. This function is used to handle both
+     * \c eglGetDisplay and \c eglGetPlatformDisplay.
+     *
+     * For some platform types, libEGL may be able to identify a specific
+     * vendor to use.
+     *
+     * If libEGL can't figure out which vendor to use on its own, then it will
+     * go through the list of available vendor libraries and call this function
+     * until one succeeds.
+     *
+     * If the application calls eglGetPlatformDisplay, then the parameters are
+     * passed through from the application.
+     *
+     * If the application calls eglGetDisplay, and the native display handle is
+     * not \c EGL_DEFAULT_DISPLAY, then libEGL will try to guess a platform
+     * type, and will pass that platform to the vendor library.
+     *
+     * Lastly, if eglGetDisplay is called with \c EGL_DEFAULT_DISPLAY, then
+     * libEGL will call into each vendor library with \p platform set to
+     * \c EGL_NONE. The vendor library can then select a default display to
+     * return.
+     *
+     * In all cases, if the vendor library can't find a matching EGLDisplay,
+     * then it should return \c EGL_NO_DISPLAY. Any errors should be reported
+     * through the vendor's \c eglGetError function.
+     */
+    EGLDisplay (* getPlatformDisplay) (EGLenum platform, void *nativeDisplay,
+            const EGLAttrib *attrib_list);
+
+    /*!
+     * Checks if the vendor library supports a given client API (that is, the
+     * API value passed to \c eglBindAPI).
+     */
+    EGLBoolean (* getSupportsAPI) (EGLenum api);
+
+    /*!
+     * (OPTIONAL) Returns a string from the vendor library. This is used for
+     * anything that isn't available from eglQueryString.
+     */
+    const char * (* getVendorString) (int name);
+
+    /*
+     * NOTE: libEGL will load all of the vendor libraries up front, so it
+     * handles extension functions slightly differently than EGL.
+     */
+
+    /*!
+     * This retrieves the pointer to the real EGL or core GL function.
+     *
+     * \param procName The name of the function.
+     * \return A pointer to a function, or \c NULL if the vendor does not
+     * support the function.
+     */
+    void * (* getProcAddress) (const char *procName);
+
+    /*!
+     * This retrieves vendor-neutral functions which use the
+     * __EGLdispatchTableDynamic API above to dispatch to the correct vendor.
+     *
+     * A vendor library must provide a dispatch function for all EGL display
+     * extension functions that it supports.
+     *
+     * Client extension functions cannot be dispatched based on an EGDisplay,
+     * so they must be handled in libEGL itself.
+     *
+     * \param procName The name of the function.
+     * \return A pointer to a function, or \c NULL if the vendor does not
+     * support the function or \p procName is not a EGL display function.
+     */
+    void * (*getDispatchAddress) (const char *procName);
+
+    /*!
+     * This notifies the vendor library which dispatch table index is
+     * assigned to a particular EGL extension function.
+     */
+    void (*setDispatchIndex) (const char *procName, int index);
+
+    /*
+     * The vendor library may use the isPatchSupported, initiatePatch,
+     * releasePatch, and patchThreadAttach callbacks to re-write libglvnd's
+     * entrypoints at make current time, provided no other contexts are current
+     * and the TLS model supports this functionality. This is a performance
+     * optimization that may not be available at runtime; the vendor library
+     * must not depend on this functionality for correctness.
+     *
+     * To use this optimization, the vendor library must provide at least the
+     * isPatchSupported and initiatePatch entrypoints.
+     *
+     * Note that these functions are identical to the ones used in GLX.
+     */
+
+    /*!
+     * (OPTIONAL) Checks to see if the vendor library supports patching the
+     * given stub type and size.
+     *
+     * \param type The type of entrypoints. This will be a one of the
+     * __GLDISPATCH_STUB_* values.
+     * \param stubSize The maximum size of the stub that the vendor library can
+     * write, in bytes.
+     * \param lookupStubOffset A callback into libglvnd to look up the address
+     * of each entrypoint.
+     */
+    GLboolean (* isPatchSupported)(int type, int stubSize);
+
+    /*!
+     * (OPTIONAL) Called by libglvnd to request that a vendor library patch its
+     * top-level entrypoints.
+     *
+     * The vendor library should use the \p lookupStubOffset callback to find
+     * the addresses of each entrypoint.
+     *
+     * This function may be called more than once to patch multiple sets of
+     * entrypoints. For example, depending on how they're built, libOpenGL.so
+     * or libGL.so may have their own entrypoints that are separate functions
+     * from the ones in libGLdispatch.
+     *
+     * Note that during this call is the only time that the entrypoints can be
+     * modified. After the call to \c initiatePatch returns, the vendor library
+     * should treat the entrypoints as read-only.
+     *
+     * \param type The type of entrypoints. This will be a one of the
+     * __GLDISPATCH_STUB_* values.
+     * \param stubSize The maximum size of the stub that the vendor library can
+     * write, in bytes.
+     * \param lookupStubOffset A callback into libglvnd to look up the address
+     * of each entrypoint.
+     *
+     * \return GL_TRUE if the vendor library supports patching with this type
+     * and size.
+     */
+    GLboolean (*initiatePatch)(int type,
+                               int stubSize,
+                               DispatchPatchLookupStubOffset lookupStubOffset);
+
+    /*!
+     * (OPTIONAL) Called by libglvnd to notify the current vendor that it no
+     * longer owns the top-level entrypoints.
+     *
+     * Libglvnd will take care of the restoring the entrypoints back to their
+     * original state. The vendor library must not try to modify them.
+     */
+    void (*releasePatch)(void);
+
+    /*!
+     * (OPTIONAL) Called at the start of window-system functions (GLX and EGL).
+     * This callback allows vendor libraries to perform any per-thread
+     * initialization.
+     *
+     * This is basically a workaround for broken applications. A lot of apps
+     * will make one or more invalid GLX/EGL calls on a thread (often including
+     * a MakeCurrent with invalid parameters), and then will try to call an
+     * OpenGL function.
+     *
+     * A non-libglvnd-based driver would be able to initialize any thread state
+     * even on a bogus GLX call, but with libglvnd, those calls wouldn't get
+     * past libGLX.
+     *
+     * This function is optional. If it's \c NULL, then libGLdispatch will
+     * simply ignore it.
+     *
+     * \note This function may be called concurrently from multiple threads.
+     */
+    void (*patchThreadAttach)(void);
+
+    /*!
+     * (OPTIONAL) Tries to determine the platform type for a native display.
+     *
+     * If the vendor library provides this function, then libglvnd will call it
+     * to determine which platform to use for a native display handle in
+     * eglGetDisplay.
+     *
+     * If no vendor library identifies the platform, then libglvnd will fall
+     * back to its own platform detection logic.
+     *
+     * Libglvnd can call this function for any native display handle except
+     * \c EGL_DEFAULT_DISPLAY.
+     *
+     * No matter what the value of \p native_display, the vendor library must
+     * not crash, and must not return a false match. If the vendor library
+     * can't identify the display, then it must return \c EGL_NONE.
+     *
+     * In particular, that means that a vendor library must not return any sort
+     * of default or fallback platform.
+     *
+     * \param native_display The native display handle passed to eglGetDisplay.
+     * \return Either a platform type enum or EGL_NONE.
+     */
+    EGLenum (* findNativeDisplayPlatform) (void *native_display);
+} __EGLapiImports;
+
+/*****************************************************************************/
+
+#define __EGL_MAIN_PROTO_NAME "__egl_Main"
+
+typedef EGLBoolean (* __PFNEGLMAINPROC) (uint32_t version, const __EGLapiExports *exports,
+        __EGLvendorInfo *vendor, __EGLapiImports *imports);
+
+/*!
+ * Vendor libraries must export a function called __egl_Main() with the
+ * following prototype.
+ *
+ * This function also performs a handshake based on the ABI version number.
+ * Vendor libraries can optionally use the version number to support older
+ * versions of the ABI.
+ *
+ * \param[in] version The ABI version. The upper 16 bits contains the major version
+ * number, and the lower 16 bits contains the minor version number.
+ *
+ * \param[in] exports The table of functions provided by libEGL. This pointer will
+ * remain valid for as long as the vendor is loaded.
+ *
+ * \param[in] vendor The opaque pointer used to identify this vendor library. This
+ * may be used in future versions to provide additional per-vendor information.
+ *
+ * \param[out] imports The function table that the vendor library should fill
+ * in. The vendor library must assign every non-optional function in the
+ * struct.
+ *
+ * \return True on success. If the vendor library does not support the
+ * requested ABI version or if some other error occurs, then it should return
+ * False.
+ */
+EGLBoolean __egl_Main(uint32_t version, const __EGLapiExports *exports,
+        __EGLvendorInfo *vendor, __EGLapiImports *imports);
+
+/*!
+ * @}
+ */
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* __LIB_EGL_ABI_H */
@@ -29,6 +29,8 @@ option(MOBILEGL_BUILD_DISAGGREGATED "Build the MG_Remote transport layer (two-pr
 # only decides whether the shims are compiled in.
 option(MOBILEGL_BUILD_DISAGGREGATED_INPROC "Compile the in-process (one-process, two-role) split shims; implies MOBILEGL_BUILD_DISAGGREGATED" OFF)
 option(MOBILEGL_BUILD_SERVER_SPIKE "Build the P0 spike-A MobileGLServer delivery-chain executable (Android only)" OFF)
+option(MOBILEGL_BUILD_EXTMEM_PROBE "Package tools/spikes/extmem_probe as libMobileGLExtmemProbe.so for the trace APK's exec hook (P11 A3, Android only)" OFF)
+option(MOBILEGL_EXTMEM_PROBE_AS_SERVER "Spike only (P11 B2 step 1): also package the probe as libMobileGLServer.so, the real server as libMobileGLServerReal.so" OFF)
 # The PipeInputs strangler (ARCHITECTURE.md 9.2). OFF is the pull build and must stay
 # byte-identical to a tree without either option: MGB_CTX is the live GLContext, no
 # MGPipe/PipeInputs source is compiled, every MGP_FILL is ((void)0).
@@ -302,6 +304,7 @@ set(SOURCE_FILES
    MobileGL/ConfigLoader.cpp

    MobileGL/MG_Util/Debug/Log.cpp
+    MobileGL/MG_Util/X11/DisplayGuard.cpp

    MobileGL/MG_Util/Async/JobNode.cpp
    MobileGL/MG_Util/Async/ShaderCompilePool.cpp
@@ -369,6 +372,7 @@ set(SOURCE_FILES
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FixIterationRPSubgroupScratchPass.cpp
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateSubgroupsPass.cpp
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/NormalizeRectCoordinatesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeduplicateImageTypesPass.cpp
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/Lower1DArrayImagesPass.cpp
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/Lower1DSampledImagesPass.cpp
    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/BakeImageFormatsPass.cpp
@@ -397,11 +401,17 @@ set(SOURCE_FILES
    MobileGL/MG_Util/Texture/TextureFormatProcessor.cpp

    MobileGL/MG_Impl/GLXImpl/Exporting/Definitions.cpp
+    MobileGL/MG_Impl/GLXImpl/Exporting/GlvndVendor.cpp
    MobileGL/MG_Impl/GLXImpl/GLXImpl.cpp
+    MobileGL/MG_Impl/GLXImpl/X11Present.cpp
+    MobileGL/MG_Impl/GLXImpl/X11PresentXcb.cpp
    MobileGL/MG_Impl/GLXImpl/LookUp/LookUp.cpp

    MobileGL/MG_Impl/EGLImpl/Exporting/Definitions.cpp
+
+    MobileGL/MG_Impl/EGLImpl/Exporting/GlvndVendor.cpp
    MobileGL/MG_Impl/EGLImpl/EGLImpl.cpp
+    MobileGL/MG_Impl/EGLImpl/WaylandWindow.cpp
    
    # @INSERTION_POINT:SOURCE_FILE_GLIMPL@ #
    MobileGL/MG_Impl/GLImpl/Sampler/Validators.cpp
@@ -431,6 +441,7 @@ set(SOURCE_FILES

    MobileGL/MG_Backend/Init.cpp
    MobileGL/MG_Backend/BackendObject.cpp
+    MobileGL/MG_Backend/GpuHangWatch.cpp

    MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp
    MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp
@@ -440,6 +451,7 @@ set(SOURCE_FILES

    MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp
    MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp
+    MobileGL/MG_Backend/DirectVulkan/MagmaSession.cpp
    MobileGL/MG_Backend/DirectVulkan/VmaImpl.cpp
    MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp
    MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp
@@ -457,8 +469,10 @@ set(SOURCE_FILES
    MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp
    MobileGL/MG_Backend/DirectVulkan/Renderer/VkClearManager.cpp
    MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp
+    MobileGL/MG_Backend/DirectVulkan/Renderer/GpuProgressMarkers.cpp

    MobileGL/MG_State/GLState/Core.cpp
+    MobileGL/MG_State/GLState/ShareGroupState.cpp
    MobileGL/MG_State/EGLState/Core.cpp
    MobileGL/MG_State/GLState/ErrorState/Error.cpp
    MobileGL/MG_State/GLState/BufferState/BufferState.cpp
@@ -607,9 +621,17 @@ if (MOBILEGL_BUILD_DISAGGREGATED)
        MobileGL/MG_Remote/Transport/InProcessTransport.cpp
        MobileGL/MG_Remote/Transport/SocketTransport.cpp
        MobileGL/MG_Remote/Transport/ControlInbox.cpp
+        # P9 W3: the server's log forward - a queue and a sender thread, graded by severity
+        # (<= WARN lossy, >= ERROR lossless and paced), so no logging thread writes the socket.
+        MobileGL/MG_Remote/Transport/LogForward.cpp
        # Keeps MG_Util/Debug/Log.h - and through it the GL frontend's
        # umbrella header - out of the header-only wire code (WireLog.h).
        MobileGL/MG_Remote/Transport/WireLog.cpp
+        # P11 A1: the adopt tier settled at the handshake from the knob and the data plane, by
+        # both roles (so SERVER-partition code in the link ratchet, not the client's tracker).
+        MobileGL/MG_Remote/Transport/AdoptTier.cpp
+        # P11 B2: T0 - the client's AHardwareBuffer store and its hop descriptor (both roles).
+        MobileGL/MG_Remote/Transport/AdoptT0.cpp
        # ---- P5: the three new directories ------------------------------
        # Wire/   the G3 codec: MGPWireRec_* in and out of SEG_CMD, blobs and
        #         var-tails in and out of SEG_STAGE.                    [w1]
@@ -630,6 +652,9 @@ if (MOBILEGL_BUILD_DISAGGREGATED)
        MobileGL/MG_Remote/FatalFunnel.cpp
        MobileGL/MG_Remote/Wire/PipeWireCodec.cpp
        MobileGL/MG_Remote/Client/ClientSession.cpp
+        # Whether the configured server is there at all, asked before a display is offered so the
+        # system's EGL/GLX/GBM loaders can move on to another implementation when it is not.
+        MobileGL/MG_Remote/Client/ServerProbe.cpp
        MobileGL/MG_Remote/Client/EmitTables.cpp
        MobileGL/MG_Remote/Client/CapsMirror.cpp
        # P5 c1: the client role's BackendObject. pActiveBackendObject holds one of these
@@ -649,7 +674,19 @@ if (MOBILEGL_BUILD_DISAGGREGATED)
        MobileGL/MG_Remote/Server/ServerSession.cpp
        MobileGL/MG_Remote/Server/PipeApplier.cpp
        MobileGL/MG_Remote/Server/ServerLoop.cpp
+        # P14 S2: the per-session runtime and the thread routing that lets one process serve
+        # several sessions (the real session a singleton accessor answers with on that
+        # session's own threads).
+        MobileGL/MG_Remote/Server/SessionRuntime.cpp
        MobileGL/MG_Remote/Server/ServerMain.cpp
+        # P11 PAIR: a socket client's control and aux connections paired by the nonce each
+        # presents first (PairBind), not by arrival order - both unix supervisors take pairs here.
+        MobileGL/MG_Remote/Server/PairAcceptor.cpp
+        # P11 B2: T0's Offers on the aux socket, matched to their map_persistent records by seq.
+        MobileGL/MG_Remote/Server/AdoptInbox.cpp
+        # Shared images: server-allocated AHardwareBuffers exported to clients as dma-buf descriptors
+        # (docs/Disaggregated/notes/anland/plan-ahb-dmabuf.md).
+        MobileGL/MG_Remote/Server/SharedImageRegistry.cpp
        MobileGL/MG_Remote/Server/ServerSpawn.cpp
        # P5f fc: the value frame that replaced the control mailbox's function-pointer payload,
        # and its codec against protocol.fbs's SurfaceOp/SurfaceReply (the spawn arm's seam).
@@ -659,6 +696,8 @@ if (MOBILEGL_BUILD_DISAGGREGATED)
        # (the display Activity's SurfaceView on Android; nobody installs one on the host, so a
        # ServerOwned request is refused by name there).
        MobileGL/MG_Remote/Server/ServerDisplay.cpp
+        # Package-neutral C ABI for an APK that owns the server and its ANativeWindow.
+        MobileGL/MG_Remote/Server/EmbeddedServer.cpp
        # P7 gate 5 (g5-msprobe): the Magma wire arm's two multisample depth/stencil resolve
        # arms and the probe that measures the first against the second, shared by the
        # renderer's bring-up and the Vulkan POST. Wire-only - the monolith records neither arm -
@@ -682,7 +721,7 @@ if (ANDROID)
        MobileGL/MG_Util/SelfTest/DriverBenchJni.cpp
    )
    if (MOBILEGL_BUILD_DISAGGREGATED)
-        # P12 (on-screen server window): the trace APK's MobileGLDisplayActivity - the in-process
+        # P12 (on-screen server window): the APK's MobileGLDisplayActivity - the in-process
        # display server's thread, its SurfaceView's ANativeWindow in and out of ServerDisplay, and
        # the geometry up-call. Split build only: nothing else has an in-process server to drive.
        list(APPEND SOURCE_FILES
@@ -763,6 +802,9 @@ if (MOBILEGL_BUILD_DISAGGREGATED)
    # target, and above all no flatc in the build graph. protocol_generated.h
    # is committed and regenerated by scripts/gen_protocol.py.
    list(APPEND MOBILEGL_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/3rdparty/flatbuffers/include)
+    # The vendored glvnd vendor-ABI headers (3rdparty/glvnd): the entry point a vendor library
+    # exposes so the system libEGL - which IS glvnd - can dispatch to it.
+    list(APPEND MOBILEGL_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/3rdparty/glvnd/include)
 endif()

 add_library(${CMAKE_PROJECT_NAME} SHARED
@@ -809,7 +851,7 @@ target_compile_definitions(${CMAKE_PROJECT_NAME}

 if(UNIX AND NOT APPLE AND NOT ANDROID)
    foreach(MOBILEGL_LOADER_ALIAS
-            libEGL.so libEGL.so.1)
+            libEGL.so libEGL.so.1 libGL.so libGL.so.1 libGLX_mobilegl.so.0)
        add_custom_command(TARGET ${CMAKE_PROJECT_NAME} POST_BUILD
            COMMAND ${CMAKE_COMMAND} -E create_symlink
                    "$<TARGET_FILE_NAME:${CMAKE_PROJECT_NAME}>"
@@ -873,6 +915,28 @@ if (TRACY_ENABLE)
    target_compile_definitions(${CMAKE_PROJECT_NAME}_s PUBLIC -DTRACY_ENABLE)
 endif ()

+# GLVND, AND WHY THIS IS A LINKER FLAG RATHER THAN A VISIBILITY ONE.
+#
+# MobileGL can serve as a glvnd EGL vendor - the library the system's libEGL.so.1 dispatches to -
+# and a vendor must hand glvnd ITS OWN entry points: glvnd calls whatever address it is given, so
+# an address that resolves to another libEGL in the process is a call back into the dispatcher.
+# With default visibility every reference inside this library resolves against the whole process,
+# and a process that loaded the system libEGL first wins.  Marking the entry points protected says
+# the same thing in the source, except the platform's own GL/EGL headers declare them default and
+# the two visibilities do not mix; this flag is that promise made by the linker instead.
+if (UNIX AND NOT APPLE)
+    target_link_options(${CMAKE_PROJECT_NAME} PRIVATE "LINKER:-Bsymbolic-functions")
+endif()
+
+# ---- Shared images outside GL: the C ABI and the GBM backend (Linux client, split builds) ----
+# docs/Disaggregated/notes/anland/plan-ahb-dmabuf.md. libMobileGL.so exports the shared-image C ABI
+# (MG_Impl/SharedImageApi.h); mobilegl_gbm.so (MG_Gbm/) is a separate GBM loader module that reaches
+# it through dlsym. Self-contained: nothing outside this block refers to either.
+if (MOBILEGL_BUILD_DISAGGREGATED AND UNIX AND NOT APPLE AND NOT ANDROID)
+    target_sources(${CMAKE_PROJECT_NAME} PRIVATE MobileGL/MG_Impl/SharedImageApi.cpp)
+    add_subdirectory(MobileGL/MG_Gbm)
+endif()
+
 if (ANDROID)
    target_link_libraries(${CMAKE_PROJECT_NAME} PUBLIC
        android
@@ -946,6 +1010,17 @@ if (NOT ANDROID AND NOT MOBILEGL_IOS)
    endif ()
 endif ()

+# Discover gtest cases when ctest starts, one binary after another, instead of in a post-build
+# step. The post-build discovery of the CMake releases that never pass the target name down
+# writes every binary's `--gtest_output=json` result to ONE per-directory file
+# (cmake_test_discovery_<hash of an empty name>.json), so binaries linked in parallel in the same
+# directory overwrite and delete each other's list: the registered count then differs between
+# runs of the same tree and a binary ends up with none or only some of its cases. The pre-test
+# mode reads the list from stdout, serially, and also works when the binary is cross-built.
+if (NOT DEFINED CMAKE_GTEST_DISCOVER_TESTS_DISCOVERY_MODE)
+    set(CMAKE_GTEST_DISCOVER_TESTS_DISCOVERY_MODE PRE_TEST)
+endif()
+
 if (NOT ANDROID)
    # Enable testing in the top-level scope so a CTestTestfile.cmake is emitted
    # at the build-tree root. This lets `ctest` be invoked from the top-level
@@ -982,6 +1057,12 @@ if (ANDROID AND MOBILEGL_BUILD_INTEGRATION_TEST)
    add_subdirectory(MobileGL/MG_IntegrationTest)
 endif()

+# The replay runner also builds for Android: an adb-shell executable that can
+# dial a render server on the device (unix or TCP) without any Activity.
+if (ANDROID AND MOBILEGL_BUILD_TRACE_REPLAY)
+    add_subdirectory(tools/trace_replay)
+endif()
+
 # ---------------------------------------------------------------------------
 # P0 spike A: the Android delivery chain for a second native executable.
 #
@@ -1090,3 +1171,48 @@ if (ANDROID AND MOBILEGL_BUILD_SERVER_SPIKE)
            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_LIBRARY_OUTPUT_DIRECTORY}")
    endif()
 endif()
+
+if (ANDROID AND MOBILEGL_BUILD_EXTMEM_PROBE)
+    # P11 A3 (open question 3): the spike-B external-memory probe, exec'd from the trace
+    # app's own process by the spike hook (`--es mobilegl_spike_spawn
+    # libMobileGLExtmemProbe.so`), so its rows answer for the untrusted_app domain. The
+    # same PIE-executable-named-like-a-library shape as MobileGLServerSpike above; it
+    # links nothing from MobileGL.
+    add_executable(MobileGLExtmemProbe
+        ${CMAKE_CURRENT_SOURCE_DIR}/tools/spikes/extmem_probe/extmem_probe.cpp)
+    set_target_properties(MobileGLExtmemProbe PROPERTIES
+        PREFIX "lib"
+        SUFFIX ".so"
+        OUTPUT_NAME "MobileGLExtmemProbe"
+        POSITION_INDEPENDENT_CODE ON)
+    target_compile_options(MobileGLExtmemProbe PRIVATE -fPIE)
+    target_link_options(MobileGLExtmemProbe PRIVATE -pie)
+    target_link_libraries(MobileGLExtmemProbe PRIVATE vulkan EGL GLESv3 android log)
+    if (CMAKE_LIBRARY_OUTPUT_DIRECTORY)
+        set_target_properties(MobileGLExtmemProbe PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_LIBRARY_OUTPUT_DIRECTORY}")
+    endif()
+    if (MOBILEGL_EXTMEM_PROBE_AS_SERVER)
+        # P11 B2 step 1, spike APKs only: MobileGLServerService execs `libMobileGLServer.so
+        # <endpoint> --serve` and B1's broker connects to whatever that listens on, so the probe
+        # takes the server's file name and answers "does a client's AHardwareBuffer cross this
+        # route into the server app" on the real route (tools/spikes/extmem_probe, route mode).
+        if (TARGET MobileGLServer)
+            set_target_properties(MobileGLServer PROPERTIES OUTPUT_NAME "MobileGLServerReal")
+        endif()
+        add_executable(MobileGLExtmemProbeAsServer
+            ${CMAKE_CURRENT_SOURCE_DIR}/tools/spikes/extmem_probe/extmem_probe.cpp)
+        set_target_properties(MobileGLExtmemProbeAsServer PROPERTIES
+            PREFIX "lib"
+            SUFFIX ".so"
+            OUTPUT_NAME "MobileGLServer"
+            POSITION_INDEPENDENT_CODE ON)
+        target_compile_options(MobileGLExtmemProbeAsServer PRIVATE -fPIE)
+        target_link_options(MobileGLExtmemProbeAsServer PRIVATE -pie)
+        target_link_libraries(MobileGLExtmemProbeAsServer PRIVATE vulkan EGL GLESv3 android log)
+        if (CMAKE_LIBRARY_OUTPUT_DIRECTORY)
+            set_target_properties(MobileGLExtmemProbeAsServer PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_LIBRARY_OUTPUT_DIRECTORY}")
+        endif()
+    endif()
+endif()
@@ -14,7 +14,7 @@ namespace MobileGL::MG_Config {
    inline const String ProjectName = "MobileGL";
    inline const String CoreName = "MobileGL Core";
    inline const String CoreVendor = "MobileGL-Dev (BZLZHH, Swung0x48, Tungsten)";
-    inline const Version CoreVersion = {26, 9, 0, "-dev", VersionType::Development};
+    inline const Version CoreVersion = {26, 10, 0, "-dev", VersionType::Development};
    inline const VersionStringFormatAttrib DefaultVersionStringFormatAttrib = {2, 2, 0, true, true};
    inline const Uint64 CacheVersion = 0;

@@ -130,8 +130,25 @@ namespace MobileGL::MG_Config {
        Bool AdvertiseFp64 = false;
        // MOBILEGL_MAGMA_R11G11B10F_FALLBACK: use fallback format for R11G11B10F on Vulkan.
        Bool MagmaR11G11B10FFallback = false;
+        // MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV: hand a shader module that fails SPIR-V validation to
+        // the driver anyway. Off by default: Magma declines such a program (no shader module, its
+        // draws refused by name), because a driver that accepts invalid SPIR-V may fault the GPU
+        // into device loss. For diagnosing a driver that copes with a specific module only.
+        Bool MagmaAllowInvalidSpirv = false;
        // MOBILEGL_MAGMA_FRAMESINFLIGHT: requested Magma frames in flight, defaulting to 3.
        Uint32 MagmaFramesInFlight = 3;
+        // MOBILEGL_MAGMA_MAX_DRAWS_PER_COMMAND_BUFFER: GL draws and dispatches Magma records into
+        // one command buffer before it submits it and continues on a fresh one. 0 = unbounded.
+        // Drivers back a command buffer with GPU memory that is only returned when the buffer is
+        // freed; Adreno maps it into the process in 16 KiB chunks, so one ~1M-draw loading frame
+        // exhausts vm.max_map_count. See VulkanRenderer::SplitOversizedRecording.
+        Uint32 MagmaMaxDrawsPerCommandBuffer = 16384;
+        // MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES: how many consecutive frames (strictly, BeginFrame
+        // epochs) a frame slot's descriptor pools must stay under a quarter full before Magma
+        // hands the slot's whole set cache back and lets it re-grow. 0 = never trim. A loading
+        // frame otherwise keeps its burst of sets (rd12: ~15k, where steady frames use a few dozen)
+        // for the life of the process. See UniformManager::TrimFrameDescriptorPools.
+        Uint32 MagmaDescriptorTrimFrames = 120;
        // MOBILEGL_ESPRYT_AVOID_SAMPLER_MIPMAP_MIN_FILTER: avoid mipmap min filters in samplers,
        // resolves certain rendering bugs on ANGLE + llvmpipe.
        Bool EsprytAvoidSamplerMipmapMinFilter = false;
@@ -202,9 +219,12 @@ namespace MobileGL::MG_Config {
        // persistently+coherently mapped storage at definition time (negative control /
        // escape hatch). Frontend-scoped: it engages only where the active backend
        // provides AcquirePersistentMap. With adoption on, an app SubData into a busy
-        // 128MB arena is a plain memcpy into GPU-visible memory; every driver-mediated
-        // route for the same write stalls the thread or ghost-copies the whole arena on
-        // this class of Mali driver, and the arena stops costing its size again in RAM.
+        // 128MB arena lands GPU-ordered through the backend's resident-subdata op
+        // (BufferObject::LandBytesIntoResidentStore, since 0ee3384b: an in-place write
+        // tore the frames still reading the old bytes) - still no driver-mediated write
+        // into the busy mutable store, which on this class of Mali driver stalls the
+        // thread or ghost-copies the whole arena, and the arena stops costing its size
+        // again in RAM.
        Bool DisableLargeBufferAdoption = false;
        // MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION: make DirectGLES skip the native ES
        // depth/stencil reads and always go through the shader-sampling emulation. Core GL
@@ -408,8 +428,9 @@ namespace MobileGL::MG_Config {
        Uint32 PipeIndexMirrorMb = 64;
        // MOBILEGL_PIPE_STATS_PERIOD: frames per boundary-counter summary line. 120 is the
        // steady-state cadence; the device retrace harness never reaches the teardown dump
-        // and a trimmed fixture (create-indirect) is shorter than 120 frames, so a run that
-        // needs its numbers at all sets this low enough to land at least one window.
+        // and a trimmed fixture (minecraft-1.21.1-neoforge-create-indirect-in-world-align1024)
+        // is shorter than 120 frames, so a run that needs its numbers at all sets this low
+        // enough to land at least one window.
        Uint32 PipeStatsPeriod = 120;
        // MOBILEGL_PIPE_STATS_FILE: where the boundary counters' teardown JSON dump goes.
        // Empty (the default) means no dump; the per-120-frame summary line still goes to
@@ -485,6 +506,9 @@ namespace MobileGL::MG_Config {
        Offscreen = 0, // a window surface is the client's own window, as before (the default)
        Server = 1,    // a window surface is the SERVER's window: WindowKind::ServerOwned
    };
+    // IpcTable::AdoptTier when MOBILEGL_IPC_ADOPT_TIER is unset (or unparseable): T0 by default,
+    // with QUIET fallbacks. Not a tier a user can type - the knob admits 0..2.
+    inline constexpr Uint32 kAdoptTierUnset = 3;
    struct IpcTable {
        // MOBILEGL_IPC_SERVER_PATH: where to find libMobileGLServer. P6 consumes it; P5
        // lands the parse because t1's ctest ENVIRONMENT blocks and add_trace_replay_test's
@@ -595,11 +619,40 @@ namespace MobileGL::MG_Config {
        // flag keyed on wantReply would disable the very mechanism it exists to enable
        // (docs/Disaggregated/notes/p12/CREATE-WINDOW-MEASURED.md sections 7-11).
        Uint32 CreateWindow = 10;
-        // MOBILEGL_IPC_ADOPT_TIER: 2 = emulate (client keeps the shadow and pushes), which
-        // is the only tier P5 implements and the reason persistent-map-push can be non-zero
-        // at all (R-6). 0 and 1 parse and are Fatal at use with "P11"; they exist now so the
-        // negative control has a spelling the day P11 writes it.
-        Uint32 AdoptTier = 2;
+        // MOBILEGL_IPC_PBO_READBACK_SYNC (P9 W1, MG_Remote/CONTRACT-P9.md §1): 1 = a glReadPixels
+        // or glGetTexImage into a bound GL_PIXEL_PACK_BUFFER takes the reply-slot form again -
+        // the client waits for the pixels and uploads them into the buffer itself, as before P9.
+        // 0 (default) sends read_pixels_to_buffer / get_texture_image_to_buffer, waits for
+        // nothing, and marks the buffer GPU-written. The 1 arm is the A/B control and the
+        // red-once of the wait-count gate (PackBufferReadbackScenario): it must show a reply
+        // wait per read.
+        Uint32 PboReadbackSync = 0;
+        // MOBILEGL_IPC_POLL_ESCALATE (P10, MG_Remote/CONTRACT-P10.md §1): a glClientWaitSync with a
+        // zero timeout or a glGetSynciv(GL_SYNC_STATUS) is answered by the client from what the
+        // server reported (kEventFenceSignaled) - "signaled" when it has been, "not yet" otherwise
+        // - and only the Nth "not yet" in a row on one fence crosses as a real round trip, so a
+        // poll loop cannot outlive a late report. 0 = every poll crosses, as before P10: the A/B
+        // arm and the red-once of the no-reply gate (FencePollScenario).
+        Uint32 PollEscalate = 64;
+        // MOBILEGL_IPC_RECOVER (client): 1 (default) = after a device loss ends this client's
+        // session, the application's next eglCreateContext / eglCreate*Surface / eglInitialize
+        // dials the server for a FRESH session, and contexts created from then on work. Contexts
+        // (and their objects) from before the loss stay lost, as EGL's power-management event
+        // has them; surfaces survive and are re-created on the new session when next used.
+        // 0 = the loss is final for the process (every verb DECLINED for the rest of its life).
+        Uint32 RecoverAfterLoss = 1;
+        // MOBILEGL_IPC_ADOPT_TIER, THE ZERO-COPY SWITCH (ruling ID-P11-14, CONTRACT-P11 B2):
+        // UNSET (kAdoptTierUnset, the default since 2026-09-29, the user's decision) = T0 where the
+        // session can run it, T2 QUIETLY where it cannot (MGLOG_D only, still counted);
+        // 0 = T0 asked by name - a T0 this session cannot use (a stream plane, the server's allow
+        // switch, no AHardwareBuffer, a failed server POST) is T2 plus one named line;
+        // 1 = T1, closed - refused by name, then T2; 2 = emulate (T2: the client keeps the shadow
+        // and pushes). Never a Fatal (MG_Remote/Transport/AdoptTier.h).
+        Uint32 AdoptTier = kAdoptTierUnset;
+        // MOBILEGL_IPC_ALLOW_ADOPT_T0 (server): 1 (default) = grant a client's T0 ask when this
+        // platform and the backend's POST self-test allow it; 0 = never grant it (the session
+        // runs T2 and says so). Read by the server role only; kept through ServerSpawn.
+        Uint32 AllowAdoptT0 = 1;
        // MOBILEGL_IPC_VERB_BARRIER: 1 = the client blocks at every verb boundary until
        // appliedSeq reaches its emitSeq (R-1). 0 is the negative control: it is EXPECTED to
        // be red, because 31 of the 63 PipeInputs fields are still pulled from a live
@@ -6,6 +6,9 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header

+#if defined(__ANDROID__)
+#include <sys/system_properties.h>
+#endif
 #include "Config.h"
 #if MOBILEGL_PIPE_PUSH
 // For kMGPipeSubsystemsMigratedAtP5e, the push build's PipePush default (the P2, P3a and P4a
@@ -15,12 +18,10 @@
 #endif

 #include <cerrno>
+#include <cstdio>
 #include <cstdlib>
-#if MOBILEGL_BUILD_DISAGGREGATED
-// std::strcmp, for InitIpc's server-role check (P12). Split-only, so the pull build's TU is
-// unchanged (G1).
+// std::strcmp: InitIpc's server-role check (P12) and the system configuration's.
 #include <cstring>
-#endif

 #ifndef _WIN32
 extern char** environ;
@@ -49,6 +50,115 @@ namespace MobileGL::MG_ConfigLoader {
        return (key.compare(0, 6, "LIBGL_") == 0 || key.compare(0, 9, "MOBILEGL_") == 0);
    }

+    // ---- the system configuration -----------------------------------------------------------
+    //
+    // A library that is every process's GL implementation (installed as the system's EGL/GLX
+    // vendor) cannot rely on each process having been started with MOBILEGL_* in its environment:
+    // most are started by something that has never heard of it. So the settings a whole machine
+    // shares live in files, and the environment only overrides them:
+    //
+    //     environment  >  /etc/mobilegl/backend (MOBILEGL_BACKEND_TYPE only)  >  client.conf  >  built-in
+    //
+    // client.conf is `KEY=value` lines; blank lines and lines starting with '#' are skipped, the
+    // key and the value are trimmed, and one pair of matching quotes around the value is removed.
+    // Only MOBILEGL_* / LIBGL_* keys are taken, exactly as from the environment. The backend file
+    // holds just the backend's name (the file the backend switch rewrites), and names it over a
+    // MOBILEGL_BACKEND_TYPE line in client.conf so that switch cannot be silently shadowed.
+    //
+    // MOBILEGL_CONFIG_FILE / MOBILEGL_BACKEND_FILE name other files; set to empty they turn the
+    // file off. The defaults exist only on desktop Linux: an Android app configures its process
+    // itself, and a server process (MOBILEGL_IPC_ROLE=server) never reads the clients' file.
+    namespace {
+#if defined(__linux__) && !defined(__ANDROID__)
+        constexpr const char* kDefaultConfigFile = "/etc/mobilegl/client.conf";
+        constexpr const char* kDefaultBackendFile = "/etc/mobilegl/backend";
+#else
+        constexpr const char* kDefaultConfigFile = "";
+        constexpr const char* kDefaultBackendFile = "";
+#endif
+
+        String FilePathSetting(const char* variable, const char* fallback) {
+            const char* value = std::getenv(variable);
+            return value != nullptr ? String(value) : String(fallback);
+        }
+
+        String Trimmed(const String& text) {
+            const auto first = text.find_first_not_of(" \t\r\n");
+            if (first == String::npos) return String();
+            const auto last = text.find_last_not_of(" \t\r\n");
+            return text.substr(first, last - first + 1);
+        }
+
+        Bool IsServerProcess() {
+            const char* role = std::getenv("MOBILEGL_IPC_ROLE");
+            return role != nullptr && std::strcmp(role, "server") == 0;
+        }
+
+        // Reads `path` into `out` without replacing what `out` already holds. A missing file is
+        // not an error: the defaults are the absence of one.
+        void ReadConfigFile(const String& path, UnorderedMap<String, String>& out, Bool log) {
+            if (path.empty()) return;
+            std::FILE* file = std::fopen(path.c_str(), "r");
+            if (file == nullptr) return;
+            char buffer[4096];
+            Uint32 lineNumber = 0;
+            while (std::fgets(buffer, sizeof(buffer), file) != nullptr) {
+                ++lineNumber;
+                const String line = Trimmed(buffer);
+                if (line.empty() || line[0] == '#') continue;
+                const auto equals = line.find('=');
+                const String key = equals == String::npos ? String() : Trimmed(line.substr(0, equals));
+                if (key.empty() || !IsAcceptedPrefix(key)) {
+                    if (log) {
+                        MGLOG_W("Config: %s:%u is not a MOBILEGL_*/LIBGL_* KEY=value line; ignored", path.c_str(),
+                                lineNumber);
+                    }
+                    continue;
+                }
+                String value = Trimmed(line.substr(equals + 1));
+                if (value.size() >= 2 && (value.front() == '"' || value.front() == '\'') && value.back() == value.front()) {
+                    value = value.substr(1, value.size() - 2);
+                }
+                out.emplace(key, value);
+            }
+            std::fclose(file);
+        }
+
+        void ReadBackendFile(const String& path, UnorderedMap<String, String>& out) {
+            if (path.empty()) return;
+            std::FILE* file = std::fopen(path.c_str(), "r");
+            if (file == nullptr) return;
+            char buffer[256];
+            const Bool read = std::fgets(buffer, sizeof(buffer), file) != nullptr;
+            std::fclose(file);
+            if (!read) return;
+            const String name = Trimmed(buffer);
+            if (!name.empty()) out.emplace("MOBILEGL_BACKEND_TYPE", name);
+        }
+
+        // Everything the system files say, the backend file ahead of client.conf (first emplace
+        // wins). Re-read on every call: callers are few (Init, and a handful of lookups before it),
+        // and a file that changed between two processes' starts must not be served stale.
+        void ReadSystemConfig(UnorderedMap<String, String>& out, Bool log) {
+            if (IsServerProcess()) return;
+            ReadBackendFile(FilePathSetting("MOBILEGL_BACKEND_FILE", kDefaultBackendFile), out);
+            ReadConfigFile(FilePathSetting("MOBILEGL_CONFIG_FILE", kDefaultConfigFile), out, log);
+        }
+    } // namespace
+
+    Bool LookupSetting(const char* key, String& outValue) {
+        if (const char* value = std::getenv(key)) {
+            outValue = value;
+            return true;
+        }
+        UnorderedMap<String, String> system;
+        ReadSystemConfig(system, false);
+        const auto it = system.find(key);
+        if (it == system.end()) return false;
+        outValue = it->second;
+        return true;
+    }
+
    inline void InitializeAcceptedEnvVariables() {
        if (!acceptedEnvVariablesMap) {
            acceptedEnvVariablesMap = MakeUnique<UnorderedMap<String, String>>();
@@ -79,6 +189,15 @@ namespace MobileGL::MG_ConfigLoader {
                }
            }
        }
+
+        // Below the environment: a key the environment set is never replaced.
+        UnorderedMap<String, String> system;
+        ReadSystemConfig(system, true);
+        for (const auto& [key, value] : system) {
+            if (acceptedEnvVariablesMap->emplace(key, value).second) {
+                MGLOG_D("Config: Accepted system setting: %s=%s", key.c_str(), value.c_str());
+            }
+        }
    }

    inline void QueryEnvVariable(const String& key, String& outValue, const String& defaultValue) {
@@ -229,7 +348,11 @@ namespace MobileGL::MG_ConfigLoader {
        features.MagmaDeriveNumSubgroups = QueryEnvQuirkOverride("MOBILEGL_MAGMA_DERIVE_NUM_SUBGROUPS");
        features.AdvertiseFp64 = QueryEnvFlag("MOBILEGL_ADVERTISE_FP64");
        features.MagmaR11G11B10FFallback = QueryEnvFlag("MOBILEGL_MAGMA_R11G11B10F_FALLBACK");
+        features.MagmaAllowInvalidSpirv = QueryEnvFlag("MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV");
        features.MagmaFramesInFlight = QueryEnvUint32("MOBILEGL_MAGMA_FRAMESINFLIGHT", 3, 1, 64);
+        features.MagmaMaxDrawsPerCommandBuffer =
+            QueryEnvUint32("MOBILEGL_MAGMA_MAX_DRAWS_PER_COMMAND_BUFFER", 16384, 0, 1u << 24);
+        features.MagmaDescriptorTrimFrames = QueryEnvUint32("MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES", 120, 0, 1u << 20);
        features.EsprytAvoidSamplerMipmapMinFilter =
            QueryEnvFlag("MOBILEGL_ESPRYT_AVOID_SAMPLER_MIPMAP_MIN_FILTER");
        features.EsprytAvoidExplicitLodBias = QueryEnvFlag("MOBILEGL_ESPRYT_AVOID_EXPLICIT_LOD_BIAS");
@@ -299,6 +422,17 @@ namespace MobileGL::MG_ConfigLoader {
    inline void InitBackendType() {
        String backendTypeStr;
        QueryEnvVariable("MOBILEGL_BACKEND_TYPE", backendTypeStr, "DirectGLES");
+#if defined(__ANDROID__)
+        // An Android app that hosts the library sets the environment itself, before the library
+        // loads; the debug property is how a developer picks the backend for such a process
+        // without rebuilding it (`adb shell setprop debug.mobilegl.backend DirectVulkan`).
+        char property[PROP_VALUE_MAX] = {};
+        if (__system_property_get("debug.mobilegl.backend", property) > 0) {
+            MGLOG_I("Config: debug.mobilegl.backend=%s overrides MOBILEGL_BACKEND_TYPE=%s", property,
+                    backendTypeStr.c_str());
+            backendTypeStr = property;
+        }
+#endif
 #define ENTRY(backendType)                                                                                             \
    if (backendTypeStr == #backendType) {                                                                              \
        MG_Config::ActiveBackendType = BackendType::backendType;                                                       \
@@ -421,14 +555,23 @@ namespace MobileGL::MG_ConfigLoader {
        // cap of 4 (see WireTables.cpp's kCreateWindowMax). WireTables.cpp's static_assert and
        // StreamLink's crossing log are the checks on this number.
        ipc.CreateWindow = QueryEnvUint32("MOBILEGL_IPC_CREATE_WINDOW", 10, 1, 12);
+        // P9 W1: the pack-buffer readback's A/B and red-once control (Config.h says both).
+        ipc.PboReadbackSync = QueryEnvUint32("MOBILEGL_IPC_PBO_READBACK_SYNC", 0, 0, 1);
+        // P10: fence polls answered locally; 0 is the A/B arm (Config.h).
+        ipc.PollEscalate = QueryEnvUint32("MOBILEGL_IPC_POLL_ESCALATE", 64, 0, 1000000);
+        // A lost session is replaced by a fresh one on the application's next create (Config.h).
+        ipc.RecoverAfterLoss = QueryEnvUint32("MOBILEGL_IPC_RECOVER", 1, 0, 1);
        // The verify harness compares the pushed block against the applier per verb; a
        // batched queue lets the comparer read a supplied field mid-apply, which is a
        // torn read rather than a divergence. The batch is therefore off whenever the
        // shadow comparer is armed.
        if (MG_Config::Features.PipeVerify) ipc.BatchWaits = 0;
-        // 2 is the only tier P5 implements (R-6). 0 and 1 parse here and are refused at the
-        // point of use, which is where the "P11" in the message belongs.
-        ipc.AdoptTier = QueryEnvUint32("MOBILEGL_IPC_ADOPT_TIER", 2, 0, 2);
+        // UNSET is T0 with quiet fallbacks (the default since 2026-09-29); 0 asks for T0 by name, 1 is
+        // refused by name, 2 is T2 - settled at the handshake from the data plane (P11 A1/B2,
+        // MG_Remote/Transport/AdoptTier.h). An unparseable value is warned about and reads as unset.
+        ipc.AdoptTier = QueryEnvUint32("MOBILEGL_IPC_ADOPT_TIER", MG_Config::kAdoptTierUnset, 0, 2);
+        // P11 B2: the server's own T0 allow switch (Config.h).
+        ipc.AllowAdoptT0 = QueryEnvUint32("MOBILEGL_IPC_ALLOW_ADOPT_T0", 1, 0, 1);
        ipc.VerbBarrier = QueryEnvUint32("MOBILEGL_IPC_VERB_BARRIER", 1, 0, 1);
        // P5e (MG_Remote/CONTRACT-P5E.md §1). The wait rule's A/B, parsed here like every
        // other IPC knob and armed only where the server publishes kCapRunAheadApply.
@@ -522,11 +665,13 @@ namespace MobileGL::MG_ConfigLoader {
        // changes the wall clock by seconds (the P12 sweep: 28.3 s at 1 down to 10.8 s at 6) a
        // log that does not name it makes every run ambiguous about its own configuration.
        MGLOG_I("Config: IPC ring=%uMiB stage=%uMiB wire-deferred=%uMiB spin=%uus event-wait=%ums "
-                "persistent-block=%uKiB create-window=%u "
-                "adopt-tier=%u verb-barrier=%u run-ahead=%u present-credit=%u control-timeout=%ums "
+                "persistent-block=%uKiB create-window=%u pbo-readback-sync=%u poll-escalate=%u "
+                "adopt-tier=%s allow-t0=%u verb-barrier=%u run-ahead=%u present-credit=%u control-timeout=%ums "
                "cold-start=%ums strict=%d audit=%d role-split-state=%d affinity='%s' surface=%s",
                ipc.RingMb, ipc.StageMb, ipc.WireDeferredMb, ipc.SpinUs, ipc.EventWaitMs,
-                ipc.PersistentBlockKb, ipc.CreateWindow, ipc.AdoptTier,
+                ipc.PersistentBlockKb, ipc.CreateWindow, ipc.PboReadbackSync, ipc.PollEscalate,
+                ipc.AdoptTier == 0 ? "0" : ipc.AdoptTier == 1 ? "1" : ipc.AdoptTier == 2 ? "2" : "unset(T0)",
+                ipc.AllowAdoptT0,
                ipc.VerbBarrier, ipc.RunAhead, ipc.PresentCredit, ipc.ControlTimeoutMs, ipc.ColdStartMs,
                static_cast<int>(ipc.StrictErrors), static_cast<int>(ipc.Audit),
                static_cast<int>(ipc.RoleSplitState), ipc.ServerAffinity.c_str(),
@@ -34,6 +34,11 @@
 #ifdef _WIN32
 #define MOBILEGL_EXPORT extern "C" __declspec(dllexport)
 #else
+// Default visibility, which is what the platform's own GL and EGL headers declare their entry
+// points with.  The one place where that is not enough - a vendor library handing glvnd its own
+// entry points, where a reference resolving against the process instead of this library would be a
+// call back into the dispatcher - is answered at link time with -Bsymbolic-functions rather than
+// here, because a second visibility on the same declaration does not compile.
 #define MOBILEGL_EXPORT extern "C" __attribute__((visibility("default")))
 #endif

@@ -24,6 +24,9 @@
 #if MOBILEGL_PIPE_VERIFY
 #include <MG_Backend/MGPipe/PipeInputs.h> // MGPipeVerifyFlushSummary, before Debug::Close in DestroyImpl
 #endif
+#if MOBILEGL_BUILD_DISAGGREGATED
+#include <MG_Remote/Client/ServerProbe.h>
+#endif

 #include <atomic>
 #include <mutex>
@@ -102,6 +105,10 @@ namespace MobileGL {
            MG_State::GLState::ClearProgramTranslationCache();
            MG_Backend::gBackendFunctionsTable = {};
            g_isInitialized = false;
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // A later re-initialization asks whether the server is there again.
+            MG_Remote::Client::ResetServerProbeCache();
+#endif
 #if MOBILEGL_PIPE_VERIFY
            // BEFORE Close, and from here rather than from a static destructor (V1 fix round 2):
            // Close() nulls the role's sink and Log.cpp's next write reopens it with "w", so the
@@ -220,6 +227,25 @@ namespace MobileGL {
        DestroyImpl(true);
    }

+    Bool ImplementationAvailable() {
+        if (g_isInitialized.load(std::memory_order_acquire)) {
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // Initialized: under a split transport the session either came up (a backend object
+            // exists) or it did not, and a display on a session that never came up cannot draw.
+            if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
+                return MG_Backend::pActiveBackendObject != nullptr;
+            }
+#endif
+            return true;
+        }
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // Not yet: the bring-up would dial the configured server and wait for it. Ask first.
+        return MG_Remote::Client::ConfiguredServerAvailable();
+#else
+        return true;
+#endif
+    }
+
    // MobileGL's lifecycle is owned entirely by the host-API layers
    // (EGL/WGL/CGL): initialization happens lazily on the first entry point
    // via EnsureInitialized(), and full teardown happens deterministically
@@ -26,13 +26,29 @@ namespace MobileGL {

    namespace MG_ConfigLoader {
        void Init();
+        // One setting resolved the way Init() resolves it - the environment, then the system
+        // configuration files (ConfigLoader.cpp) - for the few readers that must decide something
+        // before MobileGL initializes. False when neither names it.
+        Bool LookupSetting(const char* key, String& outValue);
    } // namespace MG_ConfigLoader

+    // Whether this process can be served right now. False when MobileGL is configured to reach a
+    // server that is not there (or a split session already failed to come up), or when the process
+    // asked for software rendering (LIBGL_ALWAYS_SOFTWARE, desktop Linux), so a
+    // caller that can choose another implementation - the system's EGL/GLX/GBM loaders - should
+    // decline rather than hand out something that cannot draw. Cheap: once true it stays true
+    // until the library is torn down, and a probe is one non-blocking connect with a short bound.
+    Bool ImplementationAvailable();
+
    namespace MG_Backend {
        void Init();
    } // namespace MG_Backend

    namespace MG_Impl {
        void Init();
+        // GL framebuffer 0, built into the calling thread's GLContext (the process default at
+        // Init) and installed as the one every later context starts with. Init's step, and the
+        // rebuild after a recovered device loss, whose fresh session holds none of the old one's.
+        void BuildDefaultFramebuffer();
    } // namespace MG_Impl
 } // namespace MobileGL
@@ -97,6 +97,8 @@ namespace MobileGL::MG_Backend {
            static_cast<SizeT>(target) >= kFormatCapabilityTextureTargetCount) {
            return kFormatCapabilityTargetCount;
        }
+        // An external texture is a 2D one wherever a format can go.
+        if (target == TextureTarget::External) return static_cast<SizeT>(TextureTarget::Texture2D);
        return static_cast<SizeT>(target);
    }

@@ -172,11 +174,15 @@ namespace MobileGL::MG_Backend {
            return false;
        }

-        if (m_eglDisplayInitialized && m_eglDisplay != dpy) {
-            MGLOG_E("InitializeEGLDisplay failed: backend already bound to a different EGLDisplay");
-            return false;
-        }
-
+        // A DISPLAY IS A CONNECTION, NOT A RESOURCE THIS BACKEND OWNS, and refusing the second one
+        // refused a caller for a reason that is not there: the surface and context tables are keyed
+        // by the display each was created on, every operation carries its own display handle, and
+        // this field has no reader anywhere - it was the whole of the rule.
+        //
+        // What the rule cost was a compositor.  One asks for the display it will draw into AFTER
+        // probing another, so the second eglInitialize is the one that matters, and a refusal there
+        // is reported one layer up as "no usable DRM render device" - the probe having succeeded
+        // and the real display having failed, with nothing in between that says which was refused.
        m_eglDisplay = dpy;
        m_eglDisplayInitialized = true;
        if (major) {
@@ -196,12 +202,15 @@ namespace MobileGL::MG_Backend {
    Bool BackendObject::ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height) {
        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
        auto surfaceIt = m_eglSurfaces.find(surface);
-        if (surfaceIt == m_eglSurfaces.end() || surfaceIt->second.Kind != SurfaceKind::Window) {
+        if (surfaceIt == m_eglSurfaces.end() || surfaceIt->second.Kind == SurfaceKind::None) {
            MGLOG_E("ResizeEGLWindowSurface failed: no window surface is initialized");
            return false;
        }
+        // A pbuffer-backed window (a Wayland window presented through wl_shm) resizes too.
        surfaceIt->second.Window.Width = width;
        surfaceIt->second.Window.Height = height;
+        surfaceIt->second.Width = static_cast<EGLint>(std::max<Uint32>(width, 1));
+        surfaceIt->second.Height = static_cast<EGLint>(std::max<Uint32>(height, 1));
        if (m_eglSurface == surface) {
            m_windowHandle.Width = width;
            m_windowHandle.Height = height;
@@ -279,6 +288,7 @@ namespace MobileGL::MG_Backend {
            return true;
        }

+        m_activatingSurface = surface;
        if (surfaceState->Kind == SurfaceKind::Window) {
            SetWindowHandle(surfaceState->Window);
            if (!InitWindowSurface()) {
@@ -298,8 +308,14 @@ namespace MobileGL::MG_Backend {
        m_eglSurface = surface;
        m_eglSurfaceInitialized = true;
        m_eglSurfaceKind = surfaceState->Kind;
-        m_eglCurrentThreads.clear();
-        m_backendCapabilitiesInitialized = false;
+        // NOT m_eglCurrentThreads.clear(): another thread's current context and surface are its
+        // own (a multi-window client renders each window on its own thread), and wiping them made
+        // that thread's next eglSwapBuffers fail with "no current context attached".
+        auto activatedIt = m_eglSurfaces.find(surface);
+        if (activatedIt != m_eglSurfaces.end() && !activatedIt->second.Activated) {
+            activatedIt->second.Activated = true;
+            m_backendCapabilitiesInitialized = false;
+        }
        return true;
    }

@@ -366,6 +382,24 @@ namespace MobileGL::MG_Backend {
        m_eglCurrentThreads.clear();
    }

+    void BackendObject::ForgetActiveEGLSurface() {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        m_eglSurfaceInitialized = false;
+        m_eglSurfaceKind = SurfaceKind::None;
+        m_eglSurface = EGL_NO_SURFACE;
+        m_windowHandle = {};
+    }
+
+    namespace {
+        MG_Util::Damage::Region& PresentDamageSlot() {
+            thread_local MG_Util::Damage::Region damage = MG_Util::Damage::Region::Full();
+            return damage;
+        }
+    } // namespace
+
+    const MG_Util::Damage::Region& CurrentPresentDamage() { return PresentDamageSlot(); }
+    void SetCurrentPresentDamage(const MG_Util::Damage::Region& damage) { PresentDamageSlot() = damage; }
+
    Bool BackendObject::SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw) {
        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
        if (!m_eglDisplayInitialized || m_eglDisplay != dpy) {
@@ -382,10 +416,16 @@ namespace MobileGL::MG_Backend {
            MGLOG_E("SwapEGLBuffers failed: draw surface is not current on this thread");
            return false;
        }
-        if (!m_eglSurfaceInitialized || draw == EGL_NO_SURFACE || draw != m_eglSurface) {
+        if (draw == EGL_NO_SURFACE || !GetRegisteredEGLSurface(draw)) {
            MGLOG_E("SwapEGLBuffers failed: invalid draw surface");
            return false;
        }
+        // The calling thread's surface, which another thread's make-current may have moved the
+        // backend off: bring it back before presenting it.
+        if ((!m_eglSurfaceInitialized || draw != m_eglSurface) && !ActivateEGLSurface(draw)) {
+            MGLOG_E("SwapEGLBuffers failed: the draw surface could not be re-activated");
+            return false;
+        }

        const auto& backendFunctions = GetBackendFunctions();
        if (!backendFunctions.Present) {
@@ -426,8 +466,9 @@ namespace MobileGL::MG_Backend {
        m_eglSurfaces.erase(surfaceIt);
        if (m_eglSurface == surface) {
            OnEGLSurfaceReleased(surface);
-            ResetEGLRuntimeState();
+            ForgetActiveEGLSurface();
        }
+        OnEGLSurfaceForgotten(surface);
    }

    void BackendObject::ReleaseEGLCurrentThread(const std::thread::id& threadKey) {
@@ -458,8 +499,9 @@ namespace MobileGL::MG_Backend {
        m_eglSurfaces.erase(surfaceIt);
        if (m_eglSurface == surface) {
            OnEGLSurfaceReleased(surface);
-            ResetEGLRuntimeState();
+            ForgetActiveEGLSurface();
        }
+        OnEGLSurfaceForgotten(surface);
    }

    void BackendObject::ReleaseEGLResources() {
@@ -492,4 +534,8 @@ namespace MobileGL::MG_Backend {
    void BackendObject::OnEGLSurfaceReleased(EGLSurface surface) {
        (void)surface;
    }
+
+    void BackendObject::OnEGLSurfaceForgotten(EGLSurface surface) {
+        (void)surface;
+    }
 } // namespace MobileGL::MG_Backend
@@ -9,6 +9,7 @@
 #pragma once
 #include <Includes.h>
 #include "MG_State/GLState/TextureState/TextureEnum.h"
+#include <MG_Util/Damage/Damage.h>
 #if MOBILEGL_BUILD_DISAGGREGATED
 #include <MG_Pipe/MGPipeHandles.h>
 #endif
@@ -583,26 +584,256 @@ namespace MobileGL {
            Uint32 Height = 0;
        };

+        // SHARED IMAGES (docs/Disaggregated/notes/anland/plan-ahb-dmabuf.md): server-allocated colour
+        // buffers exported to clients as dma-buf descriptors.
+        //
+        // The server-side view of one image, as a backend sees it: `NativeBuffer` is the image's
+        // AHardwareBuffer (null where there is none), `Fourcc` a DRM fourcc - ABGR8888 (R, G, B, A in
+        // memory, GL RGBA8) or XBGR8888 (alpha ignored); ARGB8888 / XRGB8888 images are stored the
+        // same way (SharedImageRegistry.h says why), so a backend binds every one as RGBA8.
+        struct SharedImageView {
+            Uint64 Id = 0;
+            Uint32 Width = 0;
+            Uint32 Height = 0;
+            Uint32 Fourcc = 0;
+            void* NativeBuffer = nullptr;
+        };
+
+        // The client-side answer of an allocation: the image's id and its exported descriptor
+        // (`Fd`, owned by the caller), with the plane layout a linux-dmabuf wl_buffer is created with.
+        struct SharedImageExport {
+            Uint64 Id = 0;
+            int Fd = -1;
+            Uint32 Width = 0;
+            Uint32 Height = 0;
+            Uint32 Fourcc = 0;
+            Uint32 Stride = 0;
+            Uint32 Offset = 0;
+            Uint64 Modifier = 0;
+            // A YUV image's Cb Cr plane: in the same descriptor, at this offset and pitch.
+            Uint32 Plane1Stride = 0;
+            Uint32 Plane1Offset = 0;
+        };
+
+        // An import of a two-plane (YUV) dma-buf: where its planes are in the one descriptor, and
+        // EGL_EXT_image_dma_buf_import's colour hints (0 = not given).
+        struct SharedImageImportLayout {
+            Uint32 PlaneCount = 1;
+            Uint32 Offset[2] = {0, 0};
+            Uint32 Pitch[2] = {0, 0};
+            Uint32 ColorSpace = 0;
+            Uint32 Range = 0;
+            Uint32 SitingX = 0;
+            Uint32 SitingY = 0;
+            // The DRM format modifier the import named; HasModifier false = none (the implicit one).
+            Bool HasModifier = false;
+            Uint64 Modifier = 0;
+        };
+
+        // THE DAMAGE OF THE PRESENT BEING APPLIED on this thread: the server's apply thread sets it
+        // around the backend's Present (from the present record), and a backend that can hand damage
+        // to its native present reads it there. Full whenever nothing set it.
+        const MG_Util::Damage::Region& CurrentPresentDamage();
+        void SetCurrentPresentDamage(const MG_Util::Damage::Region& damage);
+
        class BackendObject {
        public:
            virtual ~BackendObject() = default;

+            // ---- shared images, SERVER side (a backend implements these) ---------------------
+            //
+            // Copies the default framebuffer of the surface current on this (the apply) thread into
+            // `image`, rows TOP FIRST - GL's top row (y = height-1) lands in the image's first memory
+            // row, which is what a wl_buffer's first row is - scaled to the image's extent if the two
+            // differ. The image is ORDERED THROUGH ITS SYNC STATE (SharedImageRegistry.h): the copy
+            // waits for the image's pending read fences before it writes, and its own completion
+            // is published with SharedImages::PublishWrite as a sync_file - the call returns once
+            // the copy is submitted, never waiting for it on the CPU (a backend with no fence export
+            // waits and publishes -1). A reader binding the image waits for that fence on its GPU
+            // when the write generation moved, and at its frame boundary publishes its frame's fence
+            // with PublishRead / ReadTracker. The application's GL state is untouched. False = not
+            // supported / failed (logged).
+            //
+            // `region` (GL window coordinates, origin bottom-left) is the part of the frame to copy;
+            // Full copies all of it. The rest of the image is left as it is - the client sends a
+            // region only for an image that already holds the frame outside it - and a backend that
+            // cannot be sure the image's contents survived (a target it has never written) copies
+            // all of it regardless.
+            virtual Bool BlitDefaultFramebufferToSharedImage(const SharedImageView& image,
+                                                             const MG_Util::Damage::Region& region) {
+                (void)image, (void)region;
+                return false;
+            }
+
+            // EGL_BUFFER_AGE_EXT of the draw surface of the context current on this (the apply)
+            // thread: how many presents ago the buffer the next frame draws into held the frame then
+            // presented; 0 = unknown (the application repaints everything). Asking says the client
+            // relies on the answer, so from then on the backend keeps what it presents.
+            //
+            // `damageRegionFollows`: the client declares the region it draws (SetCurrentDamageRegion)
+            // before drawing. A driver whose only buffer age is EGL_KHR_partial_update's keeps a
+            // buffer's content only outside that region, so without one coming it is not asked.
+            virtual Int32 QueryCurrentBufferAge(Bool damageRegionFollows) {
+                (void)damageRegionFollows;
+                return 0;
+            }
+            // eglSetDamageRegionKHR for that draw surface: the frame draws `region` (GL window
+            // coordinates; Full = all of it) and nothing outside it. A backend without the notion
+            // ignores it.
+            virtual Bool SetCurrentDamageRegion(const MG_Util::Damage::Region& region) {
+                (void)region;
+                return true;
+            }
+            // IMPLICIT SYNC BY FLUSH (docs/Disaggregated/notes/anland/plan-x11-gpu.md): the session
+            // on this (the apply) thread ends a frame at a glFlush. Every shared image it used since
+            // its last boundary is published as written and read by one fence of everything it
+            // submitted (SharedImages::ReadTracker::PublishFrameAsWrite), and from now on its first
+            // use of an image in a frame waits for that image's pending reads (it may be about to
+            // overwrite what another session still samples). A backend without shared images has
+            // nothing to publish.
+            virtual Bool PublishSharedImageAccesses() { return true; }
+            // EGL_ANDROID_native_fence_sync's fence command, on the server: in `*fence`, a sync_file
+            // (the caller's to close) that signals once everything the session submitted so far has
+            // completed - and, for a session that writes images, what PublishSharedImageAccesses
+            // does. -1 when there is none to give: the backend cannot export one, or the work has
+            // already completed - in both cases it HAS completed when this returns.
+            virtual Bool ExportNativeFence(int* fence) {
+                *fence = -1;
+                return PublishSharedImageAccesses();
+            }
+
+            // ---- shared images, CLIENT side (BackendObject_Remote implements these) ----------
+            //
+            // A new image, allocated by the server; `out->Fd` is the caller's to close.
+            virtual Bool AllocateSharedImage(Uint32 width, Uint32 height, Uint32 fourcc, SharedImageExport* out) {
+                (void)width, (void)height, (void)fourcc, (void)out;
+                return false;
+            }
+            // Names the image `fd` was exported from (the caller keeps `fd`). False: not one of this
+            // server's images, or the size/format disagree.
+            virtual Bool ImportSharedImage(int fd, Uint32 width, Uint32 height, Uint32 fourcc, Uint64* outId) {
+                (void)fd, (void)width, (void)height, (void)fourcc, (void)outId;
+                return false;
+            }
+            // The same for a YUV fourcc, whose planes and hints `layout` gives. A YUV buffer the
+            // server did not export is taken too (its labelled CPU-copy fallback).
+            virtual Bool ImportSharedImagePlanes(int fd, Uint32 width, Uint32 height, Uint32 fourcc,
+                                                 const SharedImageImportLayout& layout, Uint64* outId) {
+                (void)fd, (void)width, (void)height, (void)fourcc, (void)layout, (void)outId;
+                return false;
+            }
+            // Drops this client's reference (an allocation's or an import's).
+            virtual Bool ReleaseSharedImage(Uint64 id) {
+                (void)id;
+                return false;
+            }
+            // In the GL stream: the current context's default framebuffer, copied top row first into
+            // the image (BlitDefaultFramebufferToSharedImage on the server) - `region` of it, Full for
+            // all. Returns when it completed.
+            virtual Bool PresentToSharedImage(Uint64 id, const MG_Util::Damage::Region& region) {
+                (void)id, (void)region;
+                return false;
+            }
+            // In the GL stream: EGL_BUFFER_AGE_EXT of the current draw surface (QueryCurrentBufferAge
+            // on the server). False: the backend has no answer (the caller reports 0).
+            virtual Bool QueryBufferAge(Bool damageRegionFollows, EGLint* age) {
+                (void)damageRegionFollows, (void)age;
+                return false;
+            }
+            // In the GL stream: eglSetDamageRegionKHR for the current draw surface
+            // (SetCurrentDamageRegion on the server).
+            virtual Bool SetDamageRegion(const MG_Util::Damage::Region& region) {
+                (void)region;
+                return false;
+            }
+            // In the GL stream: the texture whose lifetime id is `textureLifetimeId` takes the image as
+            // its level-0 storage (glEGLImageTargetTexture2DOES), until its next storage definition.
+            // The caller has already defined level 0 as Width x Height RGBA8 with no data.
+            virtual Bool AttachSharedImageToTexture(Uint64 textureLifetimeId, Uint64 id) {
+                (void)textureLifetimeId, (void)id;
+                return false;
+            }
+            // In the GL stream, from glFlush/glFinish: PublishSharedImageAccesses on the server,
+            // returning once the fence is published. Skipped (true) when nothing reached the stream
+            // since the last one.
+            virtual Bool FlushSharedImageAccesses() { return false; }
+            // In the GL stream, from eglCreateSync(EGL_SYNC_NATIVE_FENCE_ANDROID) with no descriptor:
+            // ExportNativeFence on the server. `*fence` is the sync_file it handed back (the
+            // caller's to close), or -1 when it had none - the fenced work has then completed.
+            virtual Bool CreateNativeFence(int* fence) {
+                *fence = -1;
+                return false;
+            }
+
            virtual void Initialize() = 0;
            virtual Bool InitCapabilities() = 0;
            virtual Bool InitWindowSurface() = 0;

+            // A served session's apply thread - the one thread that runs this backend from here on -
+            // calls this first, in a process that may serve several sessions at once. A backend whose
+            // native state is process-global by construction gives this object its own copy and binds
+            // it to the calling thread. The default has nothing per process to separate.
+            virtual void BindSessionStateToThisThread() {}
+            // A served session's apply thread asks this each time it wakes, BEFORE it applies another
+            // record: true when this session's own GPU work faulted or hung the GPU and the backend has
+            // latched the session for it (MGPipeSessionLatch, family DeviceLost), so the next frame -
+            // which would very likely stop the GPU again - is never applied. Cheap by contract: no
+            // wait, at most a status read. The default knows of no fault.
+            virtual Bool LatchIfGpuFaulted() { return false; }
+            // A served session's ring moved to another of its client's contexts (bind_context). A
+            // backend whose surface follows the native context it binds needs nothing; one that keeps
+            // a single native surface binding re-binds the surface that context was last made current
+            // with. `token` is the client's context token (the EGLContext MakeEGLCurrent was given).
+            virtual void OnClientContextBound(Uint64 token) { (void)token; }
+
            virtual Bool InitializeEGLDisplay(EGLDisplay dpy, EGLint* major, EGLint* minor);
            virtual Bool CreateEGLWindowSurface(EGLSurface surface, const WindowHandle& handle);
            virtual Bool ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height);
            virtual Bool CreateEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height);
            virtual Bool MakeEGLCurrent(EGLDisplay dpy, EGLSurface draw, EGLSurface read, EGLContext ctx);
            virtual Bool SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw);
+            // eglSwapBuffersWithDamage: `damage` (GL window coordinates) is a hint for the native
+            // present. A backend with nothing to hand it to swaps as SwapEGLBuffers does.
+            virtual Bool SwapEGLBuffersWithDamage(EGLDisplay dpy, EGLSurface draw, const MG_Util::Damage::Region& damage) {
+                (void)damage;
+                return SwapEGLBuffers(dpy, draw);
+            }
            // Forwards the app-requested eglSwapInterval to the backend's native
            // presentation path (no-op for backends without a SetSwapInterval hook).
            virtual void SetEGLSwapInterval(Int interval);
            virtual void ReleaseEGLSurface(EGLSurface surface);
            virtual void ReleaseEGLResources();

+            // ---- the server's own display window going away and coming back -------------------
+            //
+            // `window` (a window surface's native window) is about to be destroyed - the display's
+            // screen went off. Release ONLY what is bound to it: every window surface on it keeps its
+            // client handle and draws into a placeholder from here on, and every context, object and
+            // device stays. Called on the thread that runs this backend; nothing may reference the
+            // window once it returns true. False: not supported (the caller ends the session instead).
+            virtual Bool SuspendServerWindow(void* window) {
+                (void)window;
+                return false;
+            }
+            // The surfaces SuspendServerWindow parked are rebuilt on `window` (a new native window,
+            // `width`x`height`), and their new extent is published as a creation publishes it. False:
+            // they could not be (the caller ends the session).
+            virtual Bool ResumeServerWindow(void* window, Uint32 width, Uint32 height) {
+                (void)window, (void)width, (void)height;
+                return false;
+            }
+            // `window` changed size to `width`x`height` (the display's report - the one size to publish).
+            // Its surfaces publish that extent to the client now, without waiting for a present; a
+            // swapchain that has to follow is rebuilt when no frame is half-recorded, else at the
+            // present. False: nothing was published (no surface on it, or not supported).
+            virtual Bool ServerWindowResized(void* window, Uint32 width, Uint32 height) {
+                (void)window, (void)width, (void)height;
+                return false;
+            }
+            // CLIENT side, from eglQuerySurface(EGL_WIDTH/EGL_HEIGHT): a backend whose surface extents
+            // change under the application (a remote server's window) applies what it has received.
+            virtual void RefreshSurfaceExtent(EGLSurface surface) { (void)surface; }
+
            void SetWindowHandle(const WindowHandle& handle);

            virtual const RendererInfo& GetRendererInfo() const = 0;
@@ -629,18 +860,28 @@ namespace MobileGL {
            struct EGLSurfaceState {
                SurfaceKind Kind = SurfaceKind::None;
                Bool DestroyPending = false;
+                // Activated at least once: a re-activation (a multi-window client switching back)
+                // is a re-bind and does not re-run the backend's capability probe.
+                Bool Activated = false;
                WindowHandle Window;
                EGLint Width = 1;
                EGLint Height = 1;
            };

            void ResetEGLRuntimeState();
+            // ResetEGLRuntimeState for a released ACTIVE surface: the backend has no active surface
+            // any more, but every OTHER thread's current context and surface stay as they were (a
+            // multi-window client closing one popup must not unbind its other windows' threads).
+            void ForgetActiveEGLSurface();
            Bool RegisterEGLWindowSurface(EGLSurface surface, const WindowHandle& handle);
            Bool RegisterEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height);
            const EGLSurfaceState* GetRegisteredEGLSurface(EGLSurface surface) const;
            Bool ActivateEGLSurface(EGLSurface surface);
            virtual Bool InitPbufferSurface(EGLint width, EGLint height);
            virtual void OnEGLSurfaceReleased(EGLSurface surface);
+            // Every registered surface the base class drops, current or not, after it is gone from
+            // m_eglSurfaces. A backend that keeps a native surface per client surface frees it here.
+            virtual void OnEGLSurfaceForgotten(EGLSurface surface);
            FormatCapabilityCache& MutableFormatCapabilities();

            mutable std::recursive_mutex m_eglStateMutex;
@@ -648,6 +889,10 @@ namespace MobileGL {
            WindowHandle m_windowHandle;
            EGLDisplay m_eglDisplay = EGL_NO_DISPLAY;
            EGLSurface m_eglSurface = EGL_NO_SURFACE;
+            // The client surface ActivateEGLSurface is bringing up while it calls InitWindowSurface /
+            // InitPbufferSurface, which take no surface argument: a backend that maps client surfaces
+            // to native ones reads it to re-bind an existing native surface instead of making another.
+            EGLSurface m_activatingSurface = EGL_NO_SURFACE;
            Bool m_eglDisplayInitialized = false;
            Bool m_eglSurfaceInitialized = false;
            Bool m_backendCapabilitiesInitialized = false;
@@ -906,16 +906,69 @@ namespace MobileGL::MG_Backend::DirectGLES {
 #endif
    }

+    Bool BackendObject_DirectGLES::RebindNativeSurfaceFor(EGLSurface surface) {
+        const auto it = m_nativeSurfaces.find(surface);
+        if (it == m_nativeSurfaces.end()) return false;
+        if (DirectGLES::BindSessionSurface(it->second)) return true;
+        // The session no longer owns it (destroyed with a window loss): forget and recreate.
+        m_nativeSurfaces.erase(it);
+        return false;
+    }
+
+    void BackendObject_DirectGLES::RestoreSurfaceAfterCreation(EGLSurface previousNativeDraw) {
+        // The ACTIVE CONTEXT's own surface, exactly: the creation's bind attached the new surface to
+        // whichever context the session had bound (often another window's thread), and putting back
+        // "the surface activated last" instead re-attached THAT context to a third window's surface -
+        // a desktop drawing into and reading back from the panel's pbuffer.
+        if (previousNativeDraw == EGL_NO_SURFACE) return;
+        if (!DirectGLES::BindSessionSurface(previousNativeDraw)) return;
+        for (const auto& [client, native] : m_nativeSurfaces) {
+            if (native != previousNativeDraw) continue;
+            const auto state = m_eglSurfaces.find(client);
+            if (state == m_eglSurfaces.end()) break;
+            m_eglSurface = client;
+            m_eglSurfaceInitialized = true;
+            m_eglSurfaceKind = state->second.Kind;
+            if (state->second.Kind == SurfaceKind::Window) m_windowHandle = state->second.Window;
+            break;
+        }
+    }
+
    Bool BackendObject_DirectGLES::InitWindowSurface() {
+        if (RebindNativeSurfaceFor(m_activatingSurface)) return true;
+        // A suspended surface whose placeholder went: another placeholder - never the window, which
+        // is gone.
+        if (std::find(m_suspendedWindowSurfaces.begin(), m_suspendedWindowSurfaces.end(), m_activatingSurface) !=
+            m_suspendedWindowSurfaces.end()) {
+            const EGLSurface placeholder = DirectGLES::CreateSessionPbuffer(1, 1);
+            if (placeholder == EGL_NO_SURFACE || !DirectGLES::BindSessionSurface(placeholder)) {
+                MGLOG_E("DirectGLES: no placeholder for the suspended window surface %p",
+                        reinterpret_cast<void*>(m_activatingSurface));
+                return false;
+            }
+            m_nativeSurfaces[m_activatingSurface] = placeholder;
+            return true;
+        }
        // Only use EGL for now
        auto nativeWindow = reinterpret_cast<NativeWindowType>(m_windowHandle.Handle);
        if (!DirectGLES::InitWindowSurface(nativeWindow)) {
            MGLOG_E("Failed to initialize window surface for DirectGLES backend");
            return false;
        }
+        if (m_activatingSurface != EGL_NO_SURFACE) {
+            m_nativeSurfaces[m_activatingSurface] = DirectGLES::CurrentSessionDrawSurface();
+        }
        return true;
    }

+    Bool BackendObject_DirectGLES::LatchIfGpuFaulted() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return LatchIfGuiltyBeforeApply();
+#else
+        return false;
+#endif
+    }
+
    void BackendObject_DirectGLES::Initialize() {
        MG_Util::BackendLoader::AcquireEGLFunctions(m_EGLFunctions);
        MG_Util::BackendLoader::AcquireGLESFunctions(m_GLESFunctions, m_EGLFunctions.eglGetProcAddress);
@@ -976,8 +1029,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
            return true;
        }

+        // ---- P14 S4: A SECOND SURFACE NO LONGER TEARS THE CONTEXT DOWN. ----
+        //
+        // This used to be `DestroyEGLContext(); ResetEGLRuntimeState();` whenever a surface was
+        // created while another was active, i.e. every pbuffer -> window -> pbuffer switch destroyed
+        // the native context - and with it every buffer, texture and program the application had
+        // built. The native context now belongs to the (session, context token) tuple the backend
+        // resolves, and creating a surface for it only adds a surface: the tuple keeps its context
+        // and its objects. ResetEGLRuntimeState still runs because the base class's per-surface
+        // bookkeeping (which surface is active, which threads made it current) does describe the
+        // surface that is being replaced.
+        const EGLSurface previous = m_eglSurfaceInitialized ? m_eglSurface : EGL_NO_SURFACE;
+        const EGLSurface previousNativeDraw = DirectGLES::CurrentContextDrawSurface();
        if (m_eglSurfaceInitialized) {
-            DestroyEGLContext();
            ResetEGLRuntimeState();
        }

@@ -986,7 +1050,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // the first virtual tuple adopts that bind. On failure nothing is known to be bound. The
        // pull arm below is the original statement, byte for byte (G1).
        const Bool created = BackendObject::CreateEGLWindowSurface(surface, handle);
-        if (created) {
+        if (created && previous != EGL_NO_SURFACE && previous != surface) {
+            // Another surface was current: the creation must not leave the context that held it
+            // drawing to the new one (P14 multi-window clients).
+            RestoreSurfaceAfterCreation(previousNativeDraw);
+            NoteNativeContextGone();
+        } else if (created) {
            NoteNativeContextFreshFromSurfaceCreation();
        } else {
            NoteNativeContextGone();
@@ -1008,8 +1077,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
            return true;
        }

+        // P14 S4: as for the window surface above - the native context is the (session, context
+        // token) tuple's and survives a surface switch. This used to be DestroyEGLContext().
+        const EGLSurface previous = m_eglSurfaceInitialized ? m_eglSurface : EGL_NO_SURFACE;
+        const EGLSurface previousNativeDraw = DirectGLES::CurrentContextDrawSurface();
        if (m_eglSurfaceInitialized) {
-            DestroyEGLContext();
            ResetEGLRuntimeState();
        }

@@ -1017,7 +1089,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // ID-54 / ID-67: as for the window surface - InitPbufferSurface bound natively. The pull
        // arm below is the original statement, byte for byte (G1).
        const Bool created = BackendObject::CreateEGLPbufferSurface(surface, width, height);
-        if (created) {
+        if (created && previous != EGL_NO_SURFACE && previous != surface) {
+            RestoreSurfaceAfterCreation(previousNativeDraw);
+            NoteNativeContextGone();
+        } else if (created) {
            NoteNativeContextFreshFromSurfaceCreation();
        } else {
            NoteNativeContextGone();
@@ -1028,8 +1103,34 @@ namespace MobileGL::MG_Backend::DirectGLES {
 #endif
    }

+    Bool BackendObject_DirectGLES::ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        const auto state = m_eglSurfaces.find(surface);
+        const auto native = m_nativeSurfaces.find(surface);
+        if (state == m_eglSurfaces.end() || state->second.Kind != SurfaceKind::Pbuffer || native == m_nativeSurfaces.end()) {
+            return BackendObject::ResizeEGLWindowSurface(surface, width, height);
+        }
+        // A pbuffer has no resize: the client surface gets a new native pbuffer at the new size,
+        // and every context that drew to the old one draws to the new one.
+        const EGLint w = static_cast<EGLint>(std::max<Uint32>(width, 1));
+        const EGLint h = static_cast<EGLint>(std::max<Uint32>(height, 1));
+        const EGLSurface replacement = DirectGLES::CreateSessionPbuffer(w, h);
+        if (replacement == EGL_NO_SURFACE) {
+            MGLOG_E("ResizeEGLWindowSurface: no %dx%d pbuffer could replace the old drawable", w, h);
+            return false;
+        }
+        DirectGLES::ReplaceSessionSurface(native->second, replacement);
+        native->second = replacement;
+        return BackendObject::ResizeEGLWindowSurface(surface, width, height);
+    }
+
    Bool BackendObject_DirectGLES::InitPbufferSurface(EGLint width, EGLint height) {
-        return DirectGLES::InitPbufferSurface(width, height);
+        if (RebindNativeSurfaceFor(m_activatingSurface)) return true;
+        if (!DirectGLES::InitPbufferSurface(width, height)) return false;
+        if (m_activatingSurface != EGL_NO_SURFACE) {
+            m_nativeSurfaces[m_activatingSurface] = DirectGLES::CurrentSessionDrawSurface();
+        }
+        return true;
    }

    Bool BackendObject_DirectGLES::MakeEGLCurrent(EGLDisplay dpy, EGLSurface draw, EGLSurface read, EGLContext ctx) {
@@ -1089,7 +1190,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                            m_eglSurfaceInitialized && m_eglSurface == draw &&
                                            DirectGLES::IsBackendContextCurrentOnThisThread() &&
                                            NativeBindCanBeSkippedFor(dpy, draw, read, ctx);
-        if (!nativelyCurrentAlready && !DirectGLES::MakeCurrent()) {
+        // The draw surface's own native surface (P14 multi-window): a known one is re-bound, so the
+        // bind goes to it and not to whichever surface the session drew to last.
+        // The real bind still runs whenever the skip above says so (ID-67: a different tuple onto
+        // the same surface needs MakeCurrent's invalidations even though the driver triple is equal).
+        const auto nativeDraw = m_nativeSurfaces.find(draw);
+        if (!nativelyCurrentAlready && nativeDraw != m_nativeSurfaces.end()) {
+            (void)DirectGLES::SelectSessionSurface(nativeDraw->second);
+        }
+        const Bool bound = nativelyCurrentAlready || DirectGLES::MakeCurrent();
+        if (!bound) {
            NoteNativeContextGone();
            return false;
        }
@@ -1110,6 +1220,41 @@ namespace MobileGL::MG_Backend::DirectGLES {
        return true;
    }

+    Bool BackendObject_DirectGLES::BlitDefaultFramebufferToSharedImage(const SharedImageView& image,
+                                                                       const MG_Util::Damage::Region& region) {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return DirectGLES::SharedImageImpl::BlitDefaultFramebufferTo(image, region);
+#else
+        (void)image, (void)region;
+        return false;
+#endif
+    }
+
+    Int32 BackendObject_DirectGLES::QueryCurrentBufferAge(Bool damageRegionFollows) {
+        return DirectGLES::CurrentDrawBufferAge(damageRegionFollows);
+    }
+
+    Bool BackendObject_DirectGLES::SetCurrentDamageRegion(const MG_Util::Damage::Region& region) {
+        return DirectGLES::SetCurrentDrawDamageRegion(region);
+    }
+
+    Bool BackendObject_DirectGLES::PublishSharedImageAccesses() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return DirectGLES::SharedImageImpl::PublishPendingAccesses();
+#else
+        return true;
+#endif
+    }
+
+    Bool BackendObject_DirectGLES::ExportNativeFence(int* fence) {
+        *fence = -1;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return DirectGLES::SharedImageImpl::ExportNativeFence(fence);
+#else
+        return true;
+#endif
+    }
+
    Bool BackendObject_DirectGLES::SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw) {
        return BackendObject::SwapEGLBuffers(dpy, draw);
    }
@@ -1125,15 +1270,116 @@ namespace MobileGL::MG_Backend::DirectGLES {
 #if MOBILEGL_BUILD_DISAGGREGATED
        NoteNativeContextGone();
 #endif
+        m_suspendedWindowSurfaces.clear();
        BackendObject::ReleaseEGLResources();
    }

-    void BackendObject_DirectGLES::OnEGLSurfaceReleased(EGLSurface surface) {
-        (void)surface;
-        DestroyEGLContext();
+    Bool BackendObject_DirectGLES::SuspendServerWindow(void* window) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (window == nullptr) return false;
+        Vector<EGLSurface> onWindow;
+        for (const auto& [client, state] : m_eglSurfaces) {
+            if (state.Kind != SurfaceKind::Window || state.Window.Handle != window) continue;
+            if (std::find(m_suspendedWindowSurfaces.begin(), m_suspendedWindowSurfaces.end(), client) !=
+                m_suspendedWindowSurfaces.end())
+                continue;
+            onWindow.push_back(client);
+        }
+        for (const EGLSurface client : onWindow) {
+            // A surface never activated has no native surface: only its record names the window.
+            const auto native = m_nativeSurfaces.find(client);
+            if (native != m_nativeSurfaces.end()) {
+                const EGLSurface placeholder = DirectGLES::SuspendSessionWindowSurface(native->second);
+                if (placeholder == EGL_NO_SURFACE) return false;
+                native->second = placeholder;
+            }
+            m_suspendedWindowSurfaces.push_back(client);
+        }
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // The native binding moved onto the placeholder: the next make-current binds for real.
+        NoteNativeContextGone();
+#endif
+        MGLOG_I("DirectGLES: %zu window surface(s) on server window %p now draw into placeholder pbuffers; the "
+                "contexts and their objects stay",
+                onWindow.size(), window);
+        return true;
+    }
+
+    Bool BackendObject_DirectGLES::ResumeServerWindow(void* window, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (window == nullptr) return false;
+        Vector<EGLSurface> suspended;
+        suspended.swap(m_suspendedWindowSurfaces);
+        for (const EGLSurface client : suspended) {
+            const auto state = m_eglSurfaces.find(client);
+            if (state == m_eglSurfaces.end()) continue; // released while suspended
+            state->second.Window.Handle = window;
+            state->second.Window.Width = width;
+            state->second.Window.Height = height;
+            state->second.Width = static_cast<EGLint>(std::max<Uint32>(width, 1));
+            state->second.Height = static_cast<EGLint>(std::max<Uint32>(height, 1));
+            if (m_eglSurface == client) m_windowHandle = state->second.Window;
+            // Never activated: its first activation creates it on the new window.
+            const auto native = m_nativeSurfaces.find(client);
+            if (native == m_nativeSurfaces.end()) continue;
+            const EGLSurface surface =
+                DirectGLES::ResumeSessionWindowSurface(native->second, reinterpret_cast<NativeWindowType>(window));
+            if (surface == EGL_NO_SURFACE) return false;
+            native->second = surface;
+        }
 #if MOBILEGL_BUILD_DISAGGREGATED
        NoteNativeContextGone();
 #endif
+        MGLOG_I("DirectGLES: %zu suspended window surface(s) rebuilt on server window %p (%ux%u)", suspended.size(),
+                window, width, height);
+        return true;
+    }
+
+    Bool BackendObject_DirectGLES::ServerWindowResized(void* window, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (window == nullptr) return false;
+        Bool published = false;
+        for (auto& [client, state] : m_eglSurfaces) {
+            if (state.Kind != SurfaceKind::Window || state.Window.Handle != window) continue;
+            if (std::find(m_suspendedWindowSurfaces.begin(), m_suspendedWindowSurfaces.end(), client) !=
+                m_suspendedWindowSurfaces.end())
+                continue;
+            state.Window.Width = width;
+            state.Window.Height = height;
+            state.Width = static_cast<EGLint>(std::max<Uint32>(width, 1));
+            state.Height = static_cast<EGLint>(std::max<Uint32>(height, 1));
+            if (m_eglSurface == client) m_windowHandle = state.Window;
+            // The window surface's buffers follow the window by themselves (from the next one the
+            // driver dequeues); only the client has to be told.
+            const auto native = m_nativeSurfaces.find(client);
+            if (native == m_nativeSurfaces.end()) continue;
+            published = DirectGLES::RepublishWindowSurfaceShape(native->second,
+                                                                reinterpret_cast<NativeWindowType>(window)) ||
+                        published;
+        }
+        return published;
+    }
+
+    void BackendObject_DirectGLES::OnEGLSurfaceReleased(EGLSurface surface) {
+        // P14: releasing a surface releases THAT SURFACE (OnEGLSurfaceForgotten), never the
+        // contexts. This used to be DestroyEGLContext() - every native context of the session, with
+        // every object the client had built - which a multi-window client hit on its first closed
+        // popup.
+        (void)surface;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        NoteNativeContextGone();
+#endif
+    }
+
+    void BackendObject_DirectGLES::OnEGLSurfaceForgotten(EGLSurface surface) {
+        m_suspendedWindowSurfaces.erase(
+            std::remove(m_suspendedWindowSurfaces.begin(), m_suspendedWindowSurfaces.end(), surface),
+            m_suspendedWindowSurfaces.end());
+        const auto it = m_nativeSurfaces.find(surface);
+        if (it == m_nativeSurfaces.end()) return;
+        const EGLSurface native = it->second;
+        m_nativeSurfaces.erase(it);
+        DirectGLES::DestroySessionSurface(native);
    }

    const RendererInfo& BackendObject_DirectGLES::GetRendererInfo() const {
@@ -1165,6 +1411,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
            V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, V_OpenGL41, V_OpenGL42, V_OpenGL43,
            V_OpenGL44, V_OpenGL45, V_OpenGL46,
            E_GL_ARB_draw_buffers_blend,
+            // The GLSL 1.00/1.10-era ARB trio plus NPOT textures: all core since GL 2.0 and
+            // trivially true for every context this backend claims (4.6). KWin 6's
+            // EglContext::checkSupported() gates desktop-GL compositing on exactly these four
+            // strings, so withholding them fails its renderer bring-up on an otherwise capable
+            // implementation.
+            E_GL_ARB_shader_objects, E_GL_ARB_vertex_shader, E_GL_ARB_fragment_shader,
+            E_GL_ARB_texture_non_power_of_two,
            E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
            E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage, E_GL_ARB_texture_storage,
@@ -32,16 +32,38 @@ namespace MobileGL::MG_Backend::DirectGLES {
    public:
        ~BackendObject_DirectGLES() override;

+        // The driver's verdict on this session's context (DirectGLES::LatchIfDeviceLost).
+        Bool LatchIfGpuFaulted() override;
        void Initialize() override;
        Bool InitCapabilities() override;
        Bool InitWindowSurface() override;
        Bool InitializeEGLDisplay(EGLDisplay dpy, EGLint* major, EGLint* minor) override;
        Bool CreateEGLWindowSurface(EGLSurface surface, const WindowHandle& handle) override;
        Bool CreateEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height) override;
+        Bool ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height) override;
        Bool MakeEGLCurrent(EGLDisplay dpy, EGLSurface draw, EGLSurface read, EGLContext ctx) override;
        Bool SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw) override;
+        // Shared images: a GL blit of default framebuffer 0 into a renderbuffer whose storage is
+        // the image's EGLImage, ordered after the image's pending reads and published as a native
+        // fence without a CPU wait (DirectGLES.h, SharedImageImpl).
+        Bool BlitDefaultFramebufferToSharedImage(const SharedImageView& image,
+                                                 const MG_Util::Damage::Region& region) override;
+        // A window surface's age is the driver's (EGL_EXT_buffer_age); a pbuffer keeps its content,
+        // so it is 1 once it has been presented (DirectGLES::CurrentDrawBufferAge).
+        Int32 QueryCurrentBufferAge(Bool damageRegionFollows) override;
+        // eglSetDamageRegionKHR on a window surface whose driver has EGL_KHR_partial_update.
+        Bool SetCurrentDamageRegion(const MG_Util::Damage::Region& region) override;
+        // A glFlush that ends a shared-image producer's frame (SharedImageImpl::PublishPendingAccesses).
+        Bool PublishSharedImageAccesses() override;
+        // An EGL native fence of the session's work (SharedImageImpl::ExportNativeFence).
+        Bool ExportNativeFence(int* fence) override;
        void ReleaseEGLSurface(EGLSurface surface) override;
        void ReleaseEGLResources() override;
+        // The server window going and coming back: each window surface on it swaps its native EGL
+        // window surface for a placeholder pbuffer and back (DirectGLES::Suspend/ResumeSessionWindowSurface).
+        Bool SuspendServerWindow(void* window) override;
+        Bool ResumeServerWindow(void* window, Uint32 width, Uint32 height) override;
+        Bool ServerWindowResized(void* window, Uint32 width, Uint32 height) override;

        const RendererInfo& GetRendererInfo() const override;
        String GetBackendAPIVersionString() const override;
@@ -57,8 +79,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
        void UpdateDynamicBackendParameters();
        Bool InitPbufferSurface(EGLint width, EGLint height) override;
        void OnEGLSurfaceReleased(EGLSurface surface) override;
+        void OnEGLSurfaceForgotten(EGLSurface surface) override;
+        // An existing native surface for the client surface being activated, re-bound; false when
+        // there is none yet (the caller creates it).
+        Bool RebindNativeSurfaceFor(EGLSurface surface);
+        // EGL: creating a surface does not change what is current. The creation bound the new
+        // surface natively; put the previous one back for the context that held it.
+        void RestoreSurfaceAfterCreation(EGLSurface previousNativeDraw);

        Bool m_initialized = false;
+        // P14: ONE NATIVE SURFACE PER CLIENT SURFACE, created on its first activation and re-bound
+        // on every later one. A multi-window client switches surfaces every frame; making a new
+        // native surface each time leaked them and lost what the window had drawn.
+        UnorderedMap<EGLSurface, EGLSurface> m_nativeSurfaces;
+        // Window surfaces whose window is gone (SuspendServerWindow): their native surface is a
+        // placeholder pbuffer until ResumeServerWindow.
+        Vector<EGLSurface> m_suspendedWindowSurfaces;
        MG_External::EGLFunctionsTable m_EGLFunctions;
        MG_External::GLESFunctionsTable m_GLESFunctions;
        MG_External::GLESCapabilities m_GLESCapabilities;
@@ -23,15 +23,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
    // Content uploads use scratch bindings, so draws and dispatches call this after
    // texture synchronization.
    void BindCurrentTextures();
-#if MOBILEGL_BUILD_DISAGGREGATED
-    Bool ReadTextureImageWire(const MG_Pipe::MGPReadbackInfo& info, Vector<Uint8>& bytes);
-    // Native/server-only color level readback. Returns tightly packed, owned
+    // Driver-side color/depth level readback. Returns tightly packed, owned
    // bytes in the requested pair; never consults a frontend texture or PACK/PBO.
-    // The caller resolves identity/extent from its resource record and supplies
-    // whether the SOURCE allocation already uses an image carrier.
+    // The caller supplies identity and extent, and whether the SOURCE allocation
+    // already uses an image carrier.
    Bool ReadTextureLevelTight(GLuint texture, TextureTarget target, TextureUploadTarget uploadTarget,
                               TextureInternalFormat logicalFormat, GLint level, const IntVec3& logicalExtent,
                               Bool sourceUsesImageCarrier, GLenum format, GLenum type, Vector<Uint8>& bytes);
+#if MOBILEGL_BUILD_DISAGGREGATED
+    Bool ReadTextureImageWire(const MG_Pipe::MGPReadbackInfo& info, Vector<Uint8>& bytes);
+    // Whether the NATIVE texture currently holds `level` of `uploadTarget` at `logicalExtent`
+    // (glGetTexLevelParameteriv). False for a level the driver never allocated - one defined after
+    // the texture's last sync - so a re-mint does not read what is not there. A query the driver
+    // refuses answers true, keeping the caller on its readback path.
+    Bool NativeTextureLevelHasExtent(GLuint texture, TextureTarget target, TextureUploadTarget uploadTarget,
+                                     GLint level, const IntVec3& logicalExtent);
 #endif
    void ClearBufferfi(GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil);
    void ClearBufferfv(GLenum buffer, GLint drawbuffer, const GLfloat* value);
@@ -105,11 +111,72 @@ namespace MobileGL::MG_Backend::DirectGLES {
    void ShaderStorageBlockBinding(GLuint program, const GLchar* storageBlockName, GLuint storageBlockBinding);
    Bool InitWindowSurface(NativeWindowType window);
    Bool InitPbufferSurface(EGLint width, EGLint height);
+    // The calling thread's session's current native draw surface (the one the last Init*Surface
+    // created, or the last BindSessionSurface named).
+    EGLSurface CurrentSessionDrawSurface();
+    // The surface the calling thread's ACTIVE context last drew to (EGL_NO_SURFACE for none).
+    EGLSurface CurrentContextDrawSurface();
+    // EGL_BUFFER_AGE_EXT of the session's current draw surface for the next frame: the driver's
+    // answer for a window surface, 1 for a pbuffer that has been presented (a pbuffer's content stays
+    // across swaps), 0 otherwise. A driver whose buffer age is only EGL_KHR_partial_update's keeps
+    // nothing inside the frame's damage region - the whole surface unless one is set - so it is asked
+    // only when `damageRegionFollows` (SetCurrentDrawDamageRegion before the first draw).
+    Int32 CurrentDrawBufferAge(Bool damageRegionFollows);
+    // eglSetDamageRegionKHR(region) on the session's current window surface (a pbuffer, or a driver
+    // without EGL_KHR_partial_update, ignores it). The driver wants the age asked first in the frame;
+    // when the client was answered without asking it, it is asked here.
+    Bool SetCurrentDrawDamageRegion(const MG_Util::Damage::Region& region);
+    // Makes an EXISTING native surface of the calling thread's session the one it draws to, and
+    // binds it to the active context. False for a surface the session does not own.
+    Bool BindSessionSurface(EGLSurface surface);
+    // The same selection without the bind (the caller binds - MakeCurrent - unconditionally).
+    Bool SelectSessionSurface(EGLSurface surface);
+    // Destroys one native surface of the calling thread's session - and only it; the session's
+    // contexts and other surfaces are untouched.
+    void DestroySessionSurface(EGLSurface surface);
+    // A new pbuffer of the calling thread's session, created with the active context's config and
+    // bound to nothing (EGL: creating a surface changes no binding).
+    EGLSurface CreateSessionPbuffer(EGLint width, EGLint height);
+    // Every context of the session that drew to `previous` draws to `replacement` instead, and
+    // `previous` is destroyed (a pbuffer-backed window's resize).
+    void ReplaceSessionSurface(EGLSurface previous, EGLSurface replacement);
+    // The server window going away (screen off): every context of the session that drew to the
+    // window surface `windowSurface` draws to a new 1x1 placeholder pbuffer instead, and the window
+    // surface is destroyed - no longer current anywhere once this returns. The contexts and their
+    // objects stay. Returns the placeholder, or EGL_NO_SURFACE (nothing changed) when none could be
+    // made.
+    EGLSurface SuspendSessionWindowSurface(EGLSurface windowSurface);
+    // The way back: a window surface on `window` replaces `placeholder` (EGL_NO_SURFACE: none to
+    // replace) for every context that drew to it, is bound, and its extent is published as a
+    // creation publishes it. Returns the new surface, or EGL_NO_SURFACE.
+    EGLSurface ResumeSessionWindowSurface(EGLSurface placeholder, NativeWindowType window);
+    // The window under `surface` changed size: its extent (the display's) is published again, at once.
+    Bool RepublishWindowSurfaceShape(EGLSurface surface, NativeWindowType window);
    Bool MakeCurrent();
    Bool ReleaseCurrent();
    // True when the backend ES context is current on the calling thread, i.e.
    // immediate buffer ops may issue GL calls right now.
    Bool IsBackendContextCurrentOnThisThread();
+#if MOBILEGL_BUILD_DISAGGREGATED
+    // DEVICE LOSS, Espryt's half (Magma's is VulkanRenderer::LatchWireDeviceLoss). Asks the
+    // driver whether the calling thread's context was reset (glGetGraphicsResetStatus); a reset -
+    // or the debug knob MGPipeDebugDeviceLossDue - latches THIS session through
+    // MGPipeSessionLatch and answers true, and the caller returns from its verb without issuing
+    // more GL. Asked at the frame boundaries (Present, a shared-image present), before a readback,
+    // and by the apply loop each time it wakes (LatchIfGpuFaulted), never per draw. The contexts ask
+    // for reset notification (NativeContextsNotifyResets) so the driver can name the GUILTY one;
+    // only GUILTY (or a reset of unknown cause) ends the session - an INNOCENT report, a context
+    // whose work merely waited behind another's hang (the compositor's among them), is only noted.
+    // `debugKnob` false: the knob is not consulted here (the on-screen Present, so an injected loss
+    // lands on an offscreen client's shared-image present or readback, never on the compositor's).
+    Bool LatchIfDeviceLost(const char* site, Bool debugKnob = true);
+    // The apply loop's form of it (BackendObject::LatchIfGpuFaulted), asked each time the apply thread
+    // wakes: the driver's report only, no debug knob, and no EGL ground-truth verification spent.
+    Bool LatchIfGuiltyBeforeApply();
+#endif
+    // True once the process's native contexts were created with EGL_LOSE_CONTEXT_ON_RESET (decided
+    // by the first one; the driver must offer EGL_EXT_create_context_robustness).
+    Bool NativeContextsNotifyResets();
    // GL fence sync objects, backed by native ES fences. FenceSync returns null
    // (the frontend then falls back to an always-signaled sync) when the calling
    // thread does not own the ES context. Waits/queries degrade to "signaled" in
@@ -198,6 +265,200 @@ namespace MobileGL::MG_Backend::DirectGLES {
    void SetEGLFuncsTable(const MG_External::EGLFunctionsTable& eglFuncs);
    void SetGLESFuncsTable(const MG_External::GLESFunctionsTable& glesFuncs);
    void SetGLESCapabilities(const MG_External::GLESCapabilities& capabilities);
+
+    // ---------------------------------------------------------------------------------
+    // P14 S4 (docs/Disaggregated/design/11-state-ownership.md): THE NATIVE TUPLE IS PER CONTEXT
+    // ---------------------------------------------------------------------------------
+    //
+    // The display is process state and stays one object (eglInitialize once per process, terminated
+    // when the last native context leaves). The native EGLContext, the config it was created with
+    // and the native surfaces are keyed by (session, client context token) - the identity S1 put on
+    // the wire and S2 gave a per-thread owner. Creating a second context therefore no longer tears
+    // the first one's native context down, which is what lets two contexts (and two share groups)
+    // be live at once in one process.
+
+    // Answers, for the CALLING THREAD, which session it belongs to and which client context token
+    // that session has bound. Installed once by the server layer; with no resolver installed -
+    // monolith, a frontend-only process, a unit case - the answer is the single key {0, 0}, which is
+    // exactly the process singleton this replaces. `false` also means {0,0}.
+    using NativeContextKeyResolver = Bool (*)(Uint64* outSessionKey, Uint64* outContextToken);
+    void SetNativeContextKeyResolver(NativeContextKeyResolver resolver);
+
+#if MOBILEGL_PIPE_PUSH
+    // ---------------------------------------------------------------------------------
+    // P14 S6 (docs/Disaggregated/design/11-state-ownership.md): THE TWIN TABLES' KEY, AND ITS
+    // SHAPE IS S4's, ONE LEVEL DOWN. A twin is a DRIVER object, so it lives in one native
+    // context's object namespace; the identity a twin table must be keyed on is therefore
+    // {session, share group} - the SAME two words the applier's OBJECT RECORDS are filed under
+    // (PipeApply.h's MGPipeApplierKey, derived by the same probe and the same `0 means its own
+    // group` rule, so the two agree by construction rather than by two conventions).
+    //
+    // Answers, for the CALLING THREAD, which session it belongs to and which share group that
+    // session's current context is in. Installed once by the server layer beside the native and
+    // applier probes; with no probe installed - or on a thread that is no session's own, which
+    // is the client process, the in-process client+server shape and every unit case - the
+    // answer is the single key {0, 0}, exactly the process-wide table this replaces.
+    struct TwinKey {
+        Uint64 SessionKey = 0;
+        Uint64 ShareGroupKey = 0;
+        Bool operator==(const TwinKey& other) const {
+            return SessionKey == other.SessionKey && ShareGroupKey == other.ShareGroupKey;
+        }
+        Bool operator!=(const TwinKey& other) const { return !(*this == other); }
+    };
+    struct TwinKeyHash {
+        SizeT operator()(const TwinKey& key) const {
+            // The 64-bit mix SlotAllocator and the applier's registry already use for their own
+            // pair keys, so every keyed table in this tree hashes a session the same way.
+            Uint64 mixed = key.SessionKey ^ (key.ShareGroupKey + 0x9E3779B97F4A7C15ull +
+                                             (key.SessionKey << 6) + (key.SessionKey >> 2));
+            mixed ^= mixed >> 33;
+            mixed *= 0xFF51AFD7ED558CCDull;
+            mixed ^= mixed >> 33;
+            return static_cast<SizeT>(mixed);
+        }
+    };
+    using TwinKeyResolver = Bool (*)(TwinKey* outKey);
+    void SetTwinKeyResolver(TwinKeyResolver resolver);
+    // The calling thread's key. Answers false for {0, 0} - the process-wide group - which is
+    // also the answer when no probe has ever been installed.
+    Bool CurrentTwinKey(TwinKey* outKey);
+#endif
+
+    // Builds the native context for `contextToken` under the calling thread's session. A
+    // `shareGroupToken` whose group already has a live native context in this session hands that
+    // context to eglCreateContext as the share argument - which is what makes two contexts see one
+    // set of buffer/texture objects. Idempotent: a token that already has one answers true.
+    Bool CreateNativeContextFor(Uint64 contextToken, Uint64 shareGroupToken);
+    // Destroys ONE context's native tuple (the DestroyContext control frame). The session's other
+    // contexts are untouched; the display is terminated only when the last context in the process
+    // has gone.
+    void DestroyNativeContextFor(Uint64 contextToken);
+
+    // ---- objects that belong to ONE native context ----
+    //
+    // Framebuffers, vertex arrays, transform feedbacks and queries are container objects: GL never
+    // shares them, not even between contexts of one share group, and every context numbers its own
+    // from 1. A process-wide scratch FBO id minted in one context and bound in another therefore
+    // names THAT context's own framebuffer - often an application's - and the scratch path's
+    // attach/scrub then rewires it onto staging textures. With one process serving several
+    // sessions (each with its own native contexts) that silently corrupted other clients' frames.
+    //
+    // The calling thread's current native context as a lifetime serial: assigned when the context
+    // is created and never reused (a recycled EGLContext address is a new serial). 0 when none was
+    // made current through MakeCurrent.
+    Uint64 CurrentNativeContextSerial();
+
+    class PerNativeContextBase {
+    public:
+        PerNativeContextBase();
+        PerNativeContextBase(const PerNativeContextBase&) = delete;
+        PerNativeContextBase& operator=(const PerNativeContextBase&) = delete;
+        // The context died, and its names with it: forget them (no GL call - there is no context
+        // left to make one in).
+        virtual void ForgetContext(Uint64 serial) = 0;
+
+    protected:
+        ~PerNativeContextBase();
+    };
+
+    // Called where a native context is destroyed.
+    void ForgetNativeContextObjects(Uint64 serial);
+    // Moves whenever any per-context entry is dropped; a thread's cached entry is valid while
+    // it has not moved.
+    Uint64 PerNativeContextEpoch();
+    void BumpPerNativeContextEpoch();
+
+    // One T per native context, created on first use from that context.
+    template <typename T>
+    class PerNativeContext final : public PerNativeContextBase {
+    public:
+        T& Current() {
+            const Uint64 serial = CurrentNativeContextSerial();
+            // The draw path asks for the same entry over and over from one thread: a one-entry
+            // thread-local cache, invalidated by any drop anywhere (rare).
+            thread_local CurrentCache cache;
+            const Uint64 epoch = PerNativeContextEpoch();
+            if (cache.Owner == this && cache.Serial == serial && cache.Epoch == epoch) return *cache.Entry;
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            T& entry = m_bySerial[serial]; // node-based: the reference survives other contexts' inserts
+            cache = {this, serial, epoch, &entry};
+            return entry;
+        }
+        template <typename F>
+        void ForEach(F&& visit) {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            for (auto& entry : m_bySerial) visit(entry.second);
+        }
+        void Clear() {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            BumpPerNativeContextEpoch();
+            m_bySerial.clear();
+        }
+        void ForgetContext(Uint64 serial) override {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            BumpPerNativeContextEpoch();
+            m_bySerial.erase(serial);
+        }
+
+    private:
+        struct CurrentCache {
+            const PerNativeContext* Owner = nullptr;
+            Uint64 Serial = 0;
+            Uint64 Epoch = 0;
+            T* Entry = nullptr;
+        };
+        std::mutex m_mutex;
+        std::unordered_map<Uint64, T> m_bySerial;
+    };
+
+    // The GL name of one per-context object, spelled like the plain `GLuint` it replaces: it reads
+    // as the CURRENT context's name, `= 0` resets that one, and `&name` is where glGen* writes it.
+    class NativeContextName {
+    public:
+        operator GLuint() const { return m_names.Current(); }
+        NativeContextName& operator=(GLuint name) {
+            m_names.Current() = name;
+            return *this;
+        }
+        GLuint* operator&() { return &m_names.Current(); }
+
+    private:
+        mutable PerNativeContext<GLuint> m_names;
+    };
+
+    // P14 S4. THE NATIVE HALF OF A CONTEXT SWITCH. `bind_context` moves which context the session's
+    // records belong to; this makes the calling thread's native tuple follow it, binding the
+    // session's current surface to the newly named context's native EGLContext. A no-op when the
+    // named context has no native context yet (the session never sent CreateContext) or when the
+    // native context already current on this thread IS that one - so the single-context world pays
+    // nothing for it. Called from the bind_context path on the apply thread.
+    Bool MakeNativeContextCurrentForBoundToken();
+#if MOBILEGL_BUILD_DISAGGREGATED
+    // P14: the server's backend turn moved to the calling apply thread (ServerLoop.cpp's
+    // BackendTurnLock). Takes the backend's GL ownership for this thread and drops every
+    // binding shadow, which still describes the previous holder's native context.
+    void OnBackendTurnHandoff();
+#endif
+    // How many native contexts this process holds, and how many of them belong to one session. The
+    // test observables for "a second context did not tear the first one down".
+    Uint64 NativeContextCount();
+    Uint64 NativeContextCountForSession(Uint64 sessionKey);
+
+    // The calling thread's session, as the resolver above reports it (0 with no resolver).
+    Uint64 CurrentNativeSessionKey();
+
+    // ---- Test observables (P14 S4). EGL exposes no "which EGLContext does this client context
+    // hold" query and no way to ask a context for its share group, so a case that must tell "the
+    // same native context still serves" from "a fresh one was built", or a sharing pair from two
+    // unrelated contexts, reads it here. These are pure reads plus one key override; neither is on
+    // a production path and neither changes behaviour.
+    EGLContext NativeContextHandleFor(Uint64 sessionKey, Uint64 contextToken);
+    // Overrides the key the calling thread resolves to, for the life of the call. The backend's
+    // ordinary answer comes from the server layer's probe; this is how a case drives exactly the
+    // tuple it means when no session runtime is installed.
+    void ForceNativeContextKeyForTesting(Uint64 sessionKey, Uint64 contextToken);
+
    void DestroyEGLContext();

    // Transform feedback capture spans, performed by the real ES driver. The
@@ -241,8 +502,77 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // answers from the frontend state, which this function only reads.
        void SyncRenderState(Bool forColorClear = false);
        void InvalidateSyncedRenderState();
+
+        // GL_SCISSOR_TEST as the driver has it in the native context current on this thread -
+        // what a guard that turns the test off around its own blit, and puts it back, needs to
+        // know. The shadow answers only while it describes this context (synced since the last
+        // invalidation, in this context epoch); otherwise the driver is asked. The shadow is one
+        // per process: once another session's context synced into it, or a context switch
+        // invalidated it, its bits are some other context's.
+        Bool DriverScissorTestEnabled();
+
+        // TESTING: the shadow as SyncRenderState leaves it in the current context epoch, with
+        // GL_SCISSOR_TEST `enabled`.
+        void SetSyncedScissorTestForTesting(Bool enabled);
+        // TESTING: `inside` runs within the guard the shared-image present and the readbacks
+        // copy under (ScopedScissorDisable).
+        void RunScissorDisabledForTesting(void (*inside)(void*), void* data);
    } // namespace RenderStateImpl

+#if MOBILEGL_BUILD_DISAGGREGATED
+    // ---- SHARED IMAGES (docs/Disaggregated/notes/anland/plan-ahb-dmabuf.md) ----
+    //
+    // A shared image reaches the driver only through the Android import path: its AHardwareBuffer
+    // becomes an EGLImage (EGL_ANDROID_get_native_client_buffer + EGL_ANDROID_image_native_buffer)
+    // and that EGLImage the storage of a texture or renderbuffer (GL_OES_EGL_image). On a host
+    // build there is no such path and every entry below declines.
+    namespace SharedImageImpl {
+        // One EGLImage over one shared image, holding the image's registry reference for as long
+        // as the EGLImage lives (SharedImageRegistry.h's lifetime rule). The last reference
+        // destroys the EGLImage, unless the display that made it has been terminated since (which
+        // destroyed it already). Opaque outside DirectGLES.cpp.
+        struct EglImage;
+        using EglImageRef = SharedPtr<EglImage>;
+
+        // An EGLImage for the live shared image `id`, or null with `why`. Needs a current context
+        // (the first call resolves the extension entry points).
+        EglImageRef CreateEglImage(Uint64 id, String& why);
+        // glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, image) on what the active unit has bound to
+        // GL_TEXTURE_2D. False when the driver refused it (its GL error is consumed).
+        Bool TargetBoundTexture2D(const EglImageRef& image);
+        // BackendObject::BlitDefaultFramebufferToSharedImage, for the calling thread's context.
+        Bool BlitDefaultFramebufferTo(const SharedImageView& image, const MG_Util::Damage::Region& region);
+
+        // THE READER'S HALF of the image's sync state (SharedImageRegistry.h). Called on every use
+        // of a name whose level 0 is `image`, before the command that samples it: when a write
+        // landed since this context last waited, the context's later commands wait for its fence
+        // on the GPU; and the image is noted as read by the calling session's current frame.
+        void AcquireForSampling(const EglImageRef& image);
+        // YUV IMAGES (NV12/P010, MG_Remote/Server/SharedImageYuv.h) are never a name's level 0: the
+        // name keeps its own RGBA8 storage `target` (width x height), into which the image is
+        // converted at its first use in each of the calling session's frames - the image bound as
+        // an external texture, one triangle drawn - after the registry's CPU copy for a foreign
+        // buffer. Every touched binding is put back. Called after AcquireForSampling.
+        Bool IsYuv(const EglImageRef& image);
+        void ConvertYuvForSampling(const EglImageRef& image, GLuint target, Uint32 width, Uint32 height);
+        // The calling session's frame boundary for what it has read: hands a fence for everything
+        // the current context has submitted to every image noted since the last boundary. Called
+        // at Present and wherever the current native context stops being current (its reads can
+        // only be fenced from it). Nothing noted = nothing done.
+        void PublishPendingReads();
+        // IMPLICIT SYNC BY FLUSH (BackendObject::PublishSharedImageAccesses): everything the
+        // calling session used since its last boundary is published as written and read by one
+        // fence, and the session's first use of an image in each frame from now on also waits for
+        // the image's pending reads.
+        Bool PublishPendingAccesses();
+        // EGL_ANDROID_native_fence_sync's fence command (BackendObject::ExportNativeFence): a
+        // sync_file of everything the current context submitted in `*fence` (-1: none, and the work
+        // is waited out here), and - for a session that writes images - the boundary
+        // PublishPendingAccesses makes.
+        Bool ExportNativeFence(int* fence);
+    } // namespace SharedImageImpl
+#endif
+
    extern MG_External::EGLFunctionsTable g_EGLFuncs;
    extern MG_External::GLESFunctionsTable g_GLESFuncs;
    extern MG_External::GLESCapabilities g_GLESCapabilities;
@@ -74,6 +74,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
    // flag while the process is very much alive.
    Bool InProcessTeardown();
    void EnsureProcessTeardownSentinel();
+    // The exit() half of InProcessTeardown alone, for an object whose release needs only the
+    // display, not a current context (an EGLImage): a server session's twin teardown has no
+    // context current but a live display, and skipping the release there would leak the image.
+    Bool InProcessExit();

    // Generation of the backend ES context that owns the driver ids currently handed
    // out. Bumped exactly once per DestroyEGLContext. Every backend twin that owns a
@@ -332,6 +336,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
        using const_iterator = typename BackendMap::const_iterator;
 #if MOBILEGL_PIPE_PUSH
        using SlotTable = BackendSlotTable<StateObject, BackendObject, kKind>;
+        // P14 S6: the ONE bucket of this kind the calling thread is in. Every handle-keyed entry
+        // below resolves through it, so the six twin registries answer per {session, share
+        // group} without a single call site naming a bucket.
+        using SlotTables = SlotTableRegistry<StateObject, BackendObject, kKind>;
+        using TwinKey = DirectGLES::TwinKey;
 #endif

        BackendPtr& GetOrCreate(const StatePtr& stateObj) {
@@ -348,7 +357,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                // insertion (D13; SlotTables.h) - so a table used outside a registry arms
                // it too, which is right: it is the twin, not the registry, that owns the
                // driver id a guarded destructor exists for.
-                return m_slotTable.GetOrCreate(stateObj);
+                return m_slotTables.ForCallingThread().GetOrCreate(stateObj);
            }
 #endif
            // Twin creation is the moment a driver-owned id starts needing a guarded
@@ -407,7 +416,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
 #endif
 #if MOBILEGL_PIPE_PUSH
            if (EsprytSlotTablesEnabled()) {
-                return m_slotTable.Find(stateObj);
+                return m_slotTables.ForCallingThread().Find(stateObj);
            }
 #endif
            const auto entryIt = m_entries.find(stateObj);
@@ -445,7 +454,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // backend memo stores instead of a raw pointer, a GL name or a bare lifetime id.
        MG_Pipe::MGPipeHandle HandleOf(const StateObject* stateObj) const {
            if (EsprytSlotTablesEnabled()) {
-                return m_slotTable.HandleOf(stateObj);
+                return m_slotTables.HandleOf(stateObj);
            }
            return MG_Pipe::kMGPipeNullHandle;
        }
@@ -454,7 +463,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // the lookup a backend memo that already holds a handle wants: no lifetime-id probe.
        BackendPtr* FindByHandle(MG_Pipe::MGPipeHandle handle) {
            if (EsprytSlotTablesEnabled()) {
-                return m_slotTable.FindByHandle(handle);
+                return m_slotTables.FindByHandle(handle);
            }
            return nullptr;
        }
@@ -490,7 +499,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                return nullptr;
 #endif
            }
-            const Uint32 liveGen = m_slotTable.LiveGenAt(handle.Slot);
+            const Uint32 liveGen = m_slotTables.LiveGenAt(handle.Slot);
            if (liveGen != 0 && liveGen > handle.Gen) {
 #if MOBILEGL_BUILD_DISAGGREGATED
                MG_Pipe::MGPipeSessionFail( // @Ph-declined (ID-P7-1): PH-2 stays Fatal, CONTRACT-P7 §12
@@ -502,7 +511,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                return nullptr;
 #endif
            }
-            return &m_slotTable.GetOrCreate(handle);
+            return &m_slotTables.GetOrCreate(handle);
        }

        // The generation of the LIVE entry at this slot, or 0. It exists so a caller can
@@ -510,7 +519,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // performs silently.
        Uint32 LiveGenAt(Uint32 slot) const {
            if (!EsprytSlotTablesEnabled()) return 0;
-            return m_slotTable.LiveGenAt(slot);
+            return m_slotTables.LiveGenAt(slot);
        }

        // P5c (hd): the two halves of SlotTables.h's state note, forwarded. A caller holding
@@ -524,12 +533,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // which is the answer the legacy walk's own weak_ptr test already gives.
        void NoteStateForHandle(MG_Pipe::MGPipeHandle handle, const StatePtr& stateObj) {
            if (EsprytSlotTablesEnabled()) {
-                m_slotTable.NoteStateForHandle(handle, stateObj);
+                m_slotTables.NoteStateForHandle(handle, stateObj);
            }
        }
        StatePtr StateForHandle(MG_Pipe::MGPipeHandle handle) const {
            if (EsprytSlotTablesEnabled()) {
-                return m_slotTable.StateForHandle(handle);
+                return m_slotTables.StateForHandle(handle);
            }
            return nullptr;
        }
@@ -591,9 +600,52 @@ namespace MobileGL::MG_Backend::DirectGLES {
        template <typename Fn>
        void ForEachLive(Fn&& fn) const {
            if (EsprytSlotTablesEnabled()) {
-                m_slotTable.ForEachLive(fn);
+                m_slotTables.ForEachLive(fn);
            }
        }
+
+        // ---- P14 S6: the KEYED shape's own controls ----------------------------------------
+        //
+        // Everything above is the shape a CALL SITE sees, and it is deliberately the shape it
+        // saw before the rekey. These four are the shape only a caller that has to reason about
+        // buckets needs: the teardown of a session, a test asking whether two contexts landed in
+        // two namespaces, and a diagnostic.
+        //
+        // THE ONE BUCKET THIS THREAD RESOLVES TO. A by-value copy of it is exactly what a
+        // fixture's saved registry was, and it is what makes `= {}` and "put it back" both mean
+        // something under a key.
+        SlotTable& BucketForCallingThread() {
+            if (EsprytSlotTablesEnabled()) return m_slotTables.ForCallingThread();
+            // The legacy arm has no buckets; the empty table's Reset() is the only honest
+            // answer, and no shipping path asks (the arm is compiled out or the bit is clear).
+            return m_legacyEmptyTable;
+        }
+
+        Vector<SharedPtr<SlotTable>> BucketSnapshot() const {
+            if (EsprytSlotTablesEnabled()) return m_slotTables.Snapshot();
+            return {};
+        }
+        SizeT BucketCount() const {
+            return EsprytSlotTablesEnabled() ? m_slotTables.BucketCount() : 0;
+        }
+        // Drop one {session, share group}. Answers whether such a bucket existed.
+        Bool DropBucket(const TwinKey& key) {
+            return EsprytSlotTablesEnabled() && m_slotTables.Drop(key);
+        }
+        // Drop every bucket of one session - the difference between "this session's twins go"
+        // and "the process's do" - and answer how many groups that took with it.
+        Uint32 DropBucketsOfSession(Uint64 sessionKey) {
+            return EsprytSlotTablesEnabled() ? m_slotTables.DropSession(sessionKey) : 0;
+        }
+        // Empty every bucket. The teardown of a process-shaped world, where the key is {0, 0}.
+        void ResetBuckets() {
+            if (EsprytSlotTablesEnabled()) m_slotTables.Reset();
+        }
+        // Every bucket, walked with its key.
+        template <typename Fn>
+        void ForEachBucket(Fn&& fn) const {
+            if (EsprytSlotTablesEnabled()) m_slotTables.ForEachBucket(fn);
+        }
 #endif

        // The seven DirectGLES.cpp call sites drive the LEGACY arm and nothing else. On the
@@ -664,7 +716,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
        Uint32 m_creationTick = 0;
        Bool m_isCollecting = false;
 #if MOBILEGL_PIPE_PUSH
-        BackendSlotTable<StateObject, BackendObject, kKind> m_slotTable;
+        // P14 S6: the ONE bucket per {session, share group}, hidden behind the same member names
+        // the single table had, so every handle-keyed call site above and every caller below is
+        // unchanged. See SlotTables.h's SlotTableRegistry for the resolution chain.
+        SlotTables m_slotTables;
+        // The legacy arm's stand-in for BucketForCallingThread(). Never read after a Reset and
+        // never populated: the two entries that could write it are the handle-keyed ones, which
+        // return early on that arm.
+        SlotTable m_legacyEmptyTable;
 #endif
    };

@@ -1004,6 +1063,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // (the UBO binding walk) skip the probe while their pre-pass epoch read
            // matches, exactly like the per-VAO memo stamps.
            Uint64 drawCleanEpoch = 0;
+            // Frame serial whose completion retires the last glCopyBufferSubData the upload
+            // ring staged INTO this store (the flush ladder's ring tier: FlushPendingRangesNow
+            // on the pull arm, FlushPendingRangesFrom on the push arm); 0 = none. Until it
+            // retires, both keep every partial range off the range-invalidating map: on
+            // Adreno 830 that map, on a store with a queued copy, loses the copy (see
+            // FlushPendingRangesNow and InvalidateFlushAccessFor). Draw thread only, like the
+            // flush.
+            Uint64 ringCopyRetireSerial = 0;
            // Zero-copy coherent persistent map (EXT_buffer_storage): the GL store is
            // immutable, persistently+coherently mapped, and persistentPtr is what the app
            // (and the frontend PipeResource) write into directly. While set, draw-time
@@ -1042,6 +1109,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // the legacy arm did; this member exists for the drains that have no object, which
            // in P3a is the readback flush and the fp64 narrowing.
            const Uint8* hostBytes = nullptr;
+#endif
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // P11 B2 (T0): the client's AHardwareBuffer this store IS - imported with
+            // glBufferStorageExternalEXT, so `id` is immutable, persistentMapped and
+            // persistentPtr is this server's own coherent map of the client's pages. The
+            // reference is this twin's own, released only after a fence that follows the
+            // store's last use has signalled (RetireT0Import). Survives an ES context loss:
+            // the next ensure re-imports it. Null for every store T0 did not adopt.
+            void* externalAhb = nullptr;
+            // P8-C: A SHADER WROTE THIS STORE ON THIS SIDE since the R-11 staged copy (or, for
+            // T0, the coherent map) was last known to match it - set by the draw / dispatch that
+            // bound it writable (storage block, atomic counter, writable buffer image), cleared
+            // by SplitHostBytesForCpuRead's refresh and by a respecify. Only the server's own
+            // CPU readers consult it; the client keeps its own set (GpuWritePending.h).
+            Bool serverGpuWritten = false;
 #endif
        };

@@ -1058,8 +1140,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // HandleOf, no ForEachLive), and the handle overloads never look at it. Death is
        // announced by the family's own ResourceDestroy call, not by the shared death notice
        // (D-L), and the slot is freed by the CLIENT after that call returns.
+        //
+        // P14 S6: A BUCKET REGISTRY, like the six TwinRegistry tables. A buffer twin owns a
+        // driver buffer id, so the SAME {slot, gen} in two contexts that do not share a group
+        // must be two resources - and it was one process-wide table until this slice. The
+        // handle-keyed member-for-member forwarding in SlotTables.h is what leaves every
+        // `g_backendBufferResources.X(...)` call site below untouched.
        using BackendBufferResourceTable =
-            BackendSlotTable<MG_State::GLState::BufferObject, GLESBufferResource, MG_Pipe::MGPipeKind::Buffer>;
+            SlotTableRegistry<MG_State::GLState::BufferObject, GLESBufferResource, MG_Pipe::MGPipeKind::Buffer>;
        extern BackendBufferResourceTable g_backendBufferResources;

        // Resolved once per process and latched, exactly like EsprytSlotTablesEnabled() and
@@ -1104,6 +1192,34 @@ namespace MobileGL::MG_Backend::DirectGLES {
        void RequireStagedCoverage(GLESBufferResource& resource, const Uint8* hostBase, SizeT start,
                                   SizeT end, const char* site);

+        // P11 B2: THE CPU-READABLE BYTES OF A SPLIT SERVER'S BUFFER TWIN, for the handle-arm
+        // readers that have no frontend object (indirect commands, primitive-restart
+        // substitution, the multi-draw index rebase): the R-11 staged copy, or - for a T0 store,
+        // which has none - this server's own coherent map of the client's AHardwareBuffer, read
+        // exactly as the monolith arm reads the adopted map through MappedData().
+        inline const Uint8* SplitHostBytes(const GLESBufferResource& resource) {
+            if (resource.hostBytes != nullptr) return resource.hostBytes;
+            return resource.externalAhb != nullptr ? static_cast<const Uint8*>(resource.persistentPtr) : nullptr;
+        }
+
+        // P8-C: THE SAME BYTES, MADE CURRENT FIRST - for a reader that CONSUMES them (the restart
+        // rewrite, the multi-draw rebase, the *IndirectCount count, an indirect command's
+        // CPU-fed fields, the XFB scatter's pre-capture bytes). SplitHostBytes above is only
+        // "what uploads staged": a store a shader wrote on this side (serverGpuWritten) is
+        // refreshed from this server's own GL buffer - queued uploads flushed into it first,
+        // then the whole store read back into the staged copy, which also gives an orphaned
+        // store its first copy - and a T0 store lands its queued resident writes and finishes
+        // before its coherent map is read. A clean staged store costs one flag test.
+        // `res` names the record whose descriptor width the copy takes; `site` names the
+        // reader in the one-time diagnostic when the refresh cannot run.
+        const Uint8* SplitHostBytesForCpuRead(GLESBufferResource& resource, MG_Pipe::MGPipeHandle res,
+                                              const char* site);
+        // P8-C: bytes the SERVER itself put into a store's GL buffer and already holds on the CPU
+        // (an XFB capture it mapped for the client's writeback, a scatter it composed) go into
+        // the staged copy as well, so no refresh is owed for them. A T0 store has no copy.
+        void StageServerWrittenRange(GLESBufferResource& resource, MG_Pipe::MGPipeHandle res, const void* bytes,
+                                     SizeT offset, SizeT size);
+
        // M-3's WHOLE-STORE predicate, and the GUARD every caller of RequireStagedCoverage over
        // a whole-store read has to carry: true when the application DECLARED the store's
        // content (glBufferData(size, data)), so a coverage gap is a missing record; false when
@@ -1143,10 +1259,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // ("do not take tier 1"), which drops the range onto the staging ring and costs a copy
        // rather than a corruption.
        //
+        // `ringCopyInFlight`: the store still has an upload-ring glCopyBufferSubData queued that the
+        // GPU may not have run (GLESBufferResource::ringCopyRetireSerial not yet retired).
+        //
        // Returns the glMapBufferRange access bits, or 0 when tier 1 must not be taken.
        inline constexpr SizeT kEsprytInvalidateRangeMinBytes = 128u * 1024u;
        constexpr GLbitfield InvalidateFlushAccessFor(SizeT queuedStart, SizeT queuedEnd, SizeT mapStart,
-                                                      SizeT mapEnd, SizeT limit, SizeT storageSize) {
+                                                      SizeT mapEnd, SizeT limit, SizeT storageSize,
+                                                      Bool ringCopyInFlight) {
            if (mapEnd <= mapStart) return 0u;
            // THE WIDENING REFUSAL. Not >=, not "covers": exactly, in both directions. A map
            // narrower than the queued range leaves bytes unwritten inside a range it has just
@@ -1158,6 +1278,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // page-substitution fast path needs a page-coverable range to engage, and below it
            // the driver falls back to waiting out the WAR hazard on the CPU.
            if (!wholeBuffer && size < kEsprytInvalidateRangeMinBytes) return 0u;
+            // A partial range-invalidating map must not overtake a ring copy into this store the GPU
+            // has not run yet. Adreno 830 answers such a map on a store with a copy still queued by
+            // losing the copy (dev f973008c, the pull arm's FlushPendingRangesNow: create-instancing
+            // drew whole Embeddium terrain sections empty, SSIM 0.870). The range goes to the ring
+            // instead, GPU-ordered behind the copies. The whole-buffer orphan-map is exempt: it
+            // rewrites every byte from the shadow the queued copies were taken from.
+            if (!wholeBuffer && ringCopyInFlight) return 0u;
            return GL_MAP_WRITE_BIT |
                   (wholeBuffer ? GL_MAP_INVALIDATE_BUFFER_BIT : GL_MAP_INVALIDATE_RANGE_BIT);
        }
@@ -1243,6 +1370,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
        void UnregisterBufferBackendOps();
        // The ES context died: unregister ops, invalidate all outstanding GL ids
        // (they belonged to the dead context) and drop deferred deletes.
+        // P14: a server session ended while another session's native context is still alive (one
+        // process-wide share group): drops only the ending session's own staged bytes.
+        void OnServerSessionEndedWithOthersLive();
        void OnBackendContextDestroyed();

        // Get-or-create the backend resource and bring its storage up to date
@@ -1267,6 +1397,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // Deletes GL buffers whose owning frontend objects died (possibly on a
        // thread without a current ES context). Called from draw-time sync.
        void ProcessDeferredBufferReleases();
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // P11 B2: the T0 half of that drain - run right after it (retired AHardwareBuffers).
+        void ProcessDeferredT0Retires();
+#endif

        // glBindBuffer with a redundant-bind cache for GL_ARRAY_BUFFER.
        void BindBufferId(GLenum target, Uint id);
@@ -1810,6 +1944,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
            switch (target) {
            case TextureTarget::Texture1D:
            case TextureTarget::TextureRectangle:
+            case TextureTarget::External:
                return TextureTarget::Texture2D;
            case TextureTarget::Texture1DArray:
                return TextureTarget::Texture2DArray;
@@ -1963,15 +2098,33 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // allocation, or capture the old native contents, merge pending
            // upload regions and re-arm the server's whole-level replay set.
            // First allocations use staged bytes directly. No frontend dirty
-            // flag, allocator or pixel-store object participates.
+            // flag, allocator or pixel-store object participates. P9 W2: a level
+            // the driver does not hold (defined after the last sync) is never read
+            // back, and a readback the driver refuses falls back to the staged
+            // level when the store covers it (notes/p9/W2-REMINT.md).
            void RequireImageBindableStorageByHandle(MG_Pipe::MGPipeHandle res,
                                                     const MG_Pipe::MGPipeResourceRecord& record);
+#endif
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // SHARED IMAGES (MGPipeResourceRecord::SharedImageId). True when this call settled the
+            // storage - the ES name's level 0 IS the image (glEGLImageTargetTexture2DOES), or the
+            // image could not be bound and the name is left empty - and SyncMipmapsToBackend must
+            // not go on to allocate or upload. False when the record names no image: a twin that
+            // was bound to one has been told to re-mint, and the normal storage path takes over.
+            Bool SyncSharedImageStorage(const MG_Pipe::MGPipeResourceRecord& record);
+            // Pending sub-data written into the attached shared image (Managers.cpp).
+            void ReplaySharedImageUploads(const MG_Pipe::MGPipeResourceRecord& record);
+            // The attached image takes RGB texels only (an X format): uploads are converted.
+            Bool m_sharedImageUploadsRgb = false;
 #endif
            // Whether this texture's ES storage was minted in an image carrier rather than in the
            // frontend format's own layout - the readback has to ask, because for a NORMALIZED
            // carrier the storage is an integer texture holding codes and glGetTexImage still owes
            // the application floats.
            Bool RequiresImageBindableStorage() const { return m_imageBindableStorageRequired; }
+            // See m_driverMayHoldGpuWrites. Callers holding the STATE texture use
+            // NoteDriverSideTextureWrite, which also reaches the storage behind a view.
+            void NoteDriverSideWrite() { m_driverMayHoldGpuWrites = true; }
            void Bind(GLenum target, Uint unit = TempTextureUnit);
            Uint GetBackendTextureId() const;

@@ -1997,6 +2150,12 @@ namespace MobileGL::MG_Backend::DirectGLES {

            Bool IsDrawSyncClean(const MG_State::GLState::ITextureObject* t, Uint64 contextId,
                                 Uint64 samplingGeneration) const {
+#if MOBILEGL_BUILD_DISAGGREGATED
+                // A shared image's texels move with its producer's presents, which nothing here
+                // sees: every use goes through SyncMipmapsToBackend's shared-image arm, which waits
+                // for a write that landed since the last one (SharedImageImpl::AcquireForSampling).
+                if (m_sharedImage != nullptr) return false;
+#endif
                if (!m_isInitialized || m_syncedShapeContextId == 0 || m_syncedShapeContextId != contextId ||
                    m_syncedShapeGeneration != samplingGeneration) {
                    return false;
@@ -2088,6 +2247,23 @@ namespace MobileGL::MG_Backend::DirectGLES {
            Bool m_isInitialized = false;
            Bool m_imageBindableStorageRequired = false;
            Bool m_backendStorageImmutable = false;
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // The shared image the ES name's level 0 was last bound to (0 = none; image ids are
+            // never reused, so this is a lifetime key), and the EGLImage that binding went through,
+            // which holds the image for as long as the name samples it. Null with a nonzero id when
+            // the image could not be bound: the name is left empty until the record moves.
+            Uint64 m_sharedImageId = 0;
+            SharedImageImpl::EglImageRef m_sharedImage;
+#endif
+            // Whether the driver texture may hold texels the shadow never saw. Set the first time
+            // anything other than a texel upload can write it - an application framebuffer
+            // attachment, a writable image binding, a copy or blit into it, a driver-side mipmap
+            // generation - and never cleared. Until then the driver holds, outside a level's
+            // pending writes, exactly the shadow bytes it was last sent, so those writes may go up
+            // as their union box (one unpack-ring job, which is what Mali needs); from then on
+            // only the texels the client wrote may go up, since the shadow is stale everywhere
+            // else. See SyncMipmapsToBackend.
+            Bool m_driverMayHoldGpuWrites = false;
            // Latches the "this driver has no buffer textures" report to once per texture. The
            // report is emitted from the respecify path, which bails before recording the state
            // it was asked to apply - so without the latch the texture stays permanently dirty
@@ -2105,10 +2281,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // unconditionally bumps it), the content version on every CPU pixel mutation, and
            // the params version covers SetSamples/SetFixedSampleLocations, which bump neither
            // of the other two but feed the shape probe. The context id pins the generation to
-            // the context that produced it - generations restart at 0 with a new context, and a
-            // texture is owned by exactly one context (share groups are not implemented), so a
-            // mutation can never happen under a context this key does not name. 0 = never
-            // stamped (real context ids start at 1). Backend-side invalidation rides on
+            // the context that produced it - generations restart at 0 with a new context, so the
+            // id is part of the key. IT IS NOT AN OWNERSHIP CLAIM: a texture belongs to a SHARE
+            // GROUP (P14 S3/S6 - share groups ARE implemented, `ShareGroupState` holds the object
+            // tables and the twin tables are keyed by {session, share group}), so two contexts of
+            // one group both legitimately mutate this texture. What the id buys is only the
+            // cheaper half of the gate: a mutation made under a context sharing this texture
+            // stamps a different id and re-derives the shape, which is a spurious re-derivation
+            // (safe, and only on a shape mutation) rather than a missed one - the per-mutation
+            // versions below are what actually catch a change, and they are group-independent.
+            // 0 = never stamped (real context ids start at 1). Backend-side invalidation rides on
            // m_isInitialized: RequireImageBindableStorage and RecreateBackendTexture clear it.
            Uint64 m_syncedShapeContextId = 0;
            Uint64 m_syncedShapeGeneration = 0;
@@ -2257,6 +2439,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
        SharedPtr<BackendTextureObject>& SyncTextureObjectToBackend(
            const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
            Bool imageBindableStorageRequired = false);
+        // Every path that lets the DRIVER write a texture's texels calls this (see
+        // BackendTextureObject::m_driverMayHoldGpuWrites): an application framebuffer attachment,
+        // a writable image binding, a copy, blit or mipmap generation into it. A view writes its
+        // storage texture's image, so the storage texture is the one marked.
+        void NoteDriverSideTextureWrite(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject);
 #if MOBILEGL_PIPE_PUSH
        // ---- P5e SEAM (MG_Remote/CONTRACT-P5E.md §4.2, §5.2; declared by c0e, bodied by
        // id/tx2) --------------------------------------------------------------------------
@@ -2273,11 +2460,40 @@ namespace MobileGL::MG_Backend::DirectGLES {
        SharedPtr<BackendTextureObject>& SyncTextureToBackendByHandle(
            MG_Pipe::MGPipeHandle texture, Bool imageBindableStorageRequired = false);
        BackendTextureObject* ResolveTextureTwin(MG_Pipe::MGPipeHandle texture);
+        // NoteDriverSideTextureWrite's HANDLE ARM, for the same paths when the caller holds a
+        // Texture handle (a framebuffer surface's Res, an image unit's Res, VerbMipRes,
+        // VerbCopyTexDst, a copy-image endpoint): the twin the handle's record names is marked, or
+        // its storage owner's (Desc.ViewOf) for a view. Nothing on the apply thread reads a
+        // frontend object for it.
+        //
+        // P8-E: under a transport the same storage owner's staged store is marked driver-written
+        // too (StagedTextureStore::MarkDriverWritten), so a refused readback does not answer from
+        // bytes the driver has since overwritten - unless `storeFollowsTheWrite`, which only
+        // glCopyImageSubData passes: it tries FollowCopyImageInStagedStore first and marks the store
+        // itself when that fails.
+        void NoteDriverSideTextureWriteByHandle(MG_Pipe::MGPipeHandle texture, Bool storeFollowsTheWrite = false);
+#endif
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // P8-E (notes/p8/E.md): make glCopyImageSubData's texel-block move in the server's staged
+        // store too (StagedTextureStore::FollowCopy), so the destination level's bytes stay its
+        // content. Box in each image's own texel grid. False when the store cannot follow: a cube
+        // map or 1D array endpoint (copy-image puts faces / layers on z, the store does not), a view,
+        // two different internal formats, a level whose bytes did not all cross, a source the
+        // driver has written, or a monolith arm.
+        Bool FollowCopyImageInStagedStore(MG_Pipe::MGPipeHandle source, Uint32 sourceLevel, const IntVec3& sourceOrigin,
+                                          MG_Pipe::MGPipeHandle destination, Uint32 destinationLevel,
+                                          const IntVec3& destinationOrigin, const IntVec3& size);
 #endif
        // Brings every texture the next draw reads - the touched units' bindings and the draw
        // FBO's texture attachments - onto the backend, through the two borrowed-pair memos
        // documented at their definitions. Declared here so tests can drive those memos directly.
        void SyncNeccessaryTextures();
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // Every texture of this session bound to a shared image with sub-data still pending is
+        // synced now, so the uploads are in the images before a flush publishes them (an X
+        // server's PutImage into a window pixmap no later draw of its own samples).
+        void WritePendingSharedImageUploads();
+#endif
        extern Array<Array<BackendTextureObject*, (SizeT)TextureTarget::TextureTargetCount>,
                     MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
            g_boundTexturesCache;
@@ -3030,6 +3246,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
            // Same for gl_BaseVertex: only a program that reads it pays for the per-draw
            // uniform write, and only such a program needs the reset after one.
            Bool ReadsBaseVertex() const { return m_baseVertexUniformLocation >= 0; }
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // P8-C: same for the gl_BaseInstance uniform - which an indirect draw feeds from the
+            // command on the CPU only when the mg_IndirectParams view is absent.
+            Bool ReadsBaseInstance() const { return m_baseInstanceUniformLocation >= 0; }
+#endif
            // Which viewport indices the next draw's fragments may keep, one bit each. Written
            // once per replay pass; see ForEachViewportRoutingPass.
            void SetViewportPassMask(Uint32 indexMask) const;
@@ -3525,7 +3746,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // what makes HandleOf below resolve at all. Not one member that would dereference it is
        // instantiated - no Find(StateObject*), no ForEachLive - and the handle overloads never
        // look at it.
-        using BackendSamplerViewTable = BackendSlotTable<MG_State::GLState::ITextureObject,
+        // P14 S6: and a bucket registry for the same reason the buffer one above is - a sampler
+        // view's twin is a driver object of the texture's group.
+        using BackendSamplerViewTable = SlotTableRegistry<MG_State::GLState::ITextureObject,
                                                        BackendSamplerViewObject,
                                                        MG_Pipe::MGPipeKind::SamplerViewCso>;
        extern BackendSamplerViewTable g_backendSamplerViews;
@@ -191,15 +191,16 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
                if (question == IndexBufferQuestion::Presence) return view;

                if (question == IndexBufferQuestion::HostBytes) {
-                    // No Sync* pair here and none is missing: on this side the applier has
-                    // already consumed the persistent-map blocks and the shader writebacks for
-                    // this resource before the draw verb replayed, so the server's staged shadow
-                    // IS the synced copy. That is also why this arm cannot be expressed as "find
-                    // the object and run the monolith body".
+                    // The applier has already consumed the persistent-map blocks for this
+                    // resource before the draw verb replayed; the one Sync* this arm owes is the
+                    // monolith's SyncGpuWrites, and (P8-C) SplitHostBytesForCpuRead is it: a store
+                    // a shader wrote on this side is read back from this server's own GL buffer,
+                    // because no writeback ever reaches the server's staged copy. That is also why
+                    // this arm cannot be expressed as "find the object and run the monolith body".
                    auto* resource = BufferImpl::FindBufferResourceForHandle(res);
                    if (resource == nullptr) RefuseMissingIndexBufferRecord(entry, res, "backend resource");
                    view.Size = BufferImpl::ResourceWidthForHandle(res);
-                    view.HostBytes = resource->hostBytes;
+                    view.HostBytes = BufferImpl::SplitHostBytesForCpuRead(*resource, res, "multidraw_index_rebase");
                    if (view.HostBytes != nullptr && IndexBufferRecordHasDefinedContent(st, res)) {
                        BufferImpl::RequireStagedCoverage(*resource, view.HostBytes, 0, view.Size,
                                                          "multidraw_index_rebase");
@@ -11,6 +11,12 @@

 #include <MG_Pipe/MGPipeHandles.h>

+// P14 S6: the key this table is filed under includes the calling thread's {session, share
+// group}, and the probe that answers it lives on the backend's public surface beside S4's
+// native-tuple key (DirectGLES.h). Included here so this header is self-contained for the two
+// translation units that include it without copying Managers.h's include order.
+#include "DirectGLES.h"
+
 #if MOBILEGL_PIPE_PUSH
 #include <MG_Impl/Pipe/SlotAllocator.h>
 #endif
@@ -155,6 +161,250 @@ namespace MobileGL::MG_Backend::DirectGLES {
        return enabled;
    }

+    // =====================================================================================
+    // P14 S6 (docs/Disaggregated/design/11-state-ownership.md): THE TWIN TABLE IS PER
+    // {session, share group}, NOT PER PROCESS.
+    // =====================================================================================
+    //
+    // WHAT WAS WRONG, in one sentence: a twin is a DRIVER object - a GL buffer id, a program, a
+    // texture, the storage behind them - so it belongs to one native context's object namespace,
+    // and a table keyed by {slot, gen} ALONE resolves the same handle value to the same twin
+    // whichever context asked. The frontend gives two contexts that do NOT share a group two
+    // independent name spaces, so the same {slot, gen} in both is legal and means two different
+    // objects; one process-wide table made the second context adopt the first's twin (including
+    // its driver ids, its persistent map and its pooled store). At the same time two contexts of
+    // ONE group - which must see one set of objects - were served by whichever of them twinned
+    // first, i.e. by luck rather than by the group the two agree on.
+    //
+    // THE KEY IS the same pair the applier files its OBJECT RECORDS under: {session, share
+    // group}. Same probe, same `0 means this context started its own group` rule (see
+    // DirectGLES.h's TwinKey and SessionRuntime.cpp's ResolveThreadTwinKey), so a record and the
+    // twin it describes are always in the same bucket. Session is in the key because a twin
+    // names a driver id and no driver id crosses a connection - two connections in one process
+    // are two clients, and nothing may be shared with either from the other.
+    //
+    // THE HOLDER LIST STAYS PROCESS-WIDE AND THAT IS NOT A COMPROMISE. A notice is about an
+    // OBJECT, not about a bucket: OnFrontendObjectDestroyed and ReleaseTwinByHandle must reach
+    // the by-value copies a fixture or a context reset is holding, whichever bucket they sit in,
+    // and "which bucket holds it" is exactly the question that produced the two-holder leak the
+    // list exists to fix. Each SHARED holder of the list is ONE bucket; the walk therefore covers
+    // every bucket of every table.
+    //
+    // THE SAME TABLE OBJECT IS REUSED, NOT REPLACED, and that is what keeps ~200 call sites
+    // alive: `g_backendTextureObjects.Find(...)` still answers from the bucket the calling thread
+    // is in, because the object the global names multiplexes on the key at each lookup. A caller
+    // that must address a bucket OTHER than the calling thread's - the death walks, the
+    // session-teardown drop - takes Shared() or calls the static entries, which walk buckets;
+    // and a caller that wants the whole process's view (the teardown, the tests) takes Snapshot().
+    // FORWARD-DECLARED, because the registry below is DEFINED before the table it holds; the
+    // members that name the table are defined out of line at the bottom of this header.
+    template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
+    class BackendSlotTable;
+
+    template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
+    class SlotTableRegistry {
+    public:
+        using Table = BackendSlotTable<StateObject, BackendObject, kKind>;
+        using Key = TwinKey;
+        using StatePtr = SharedPtr<StateObject>;
+        using BackendPtr = SharedPtr<BackendObject>;
+        // The bound the two tables that ARE a bucket registry have always spelled by name; the
+        // six registries reach it through their own `SlotTable` alias.
+        static constexpr Uint32 kMaxHandleSlot = 1u << 20;
+
+        // One bucket. SHARED, so a caller may keep it across other lookups: the map owns the
+        // pointee and never moves it.
+        SharedPtr<Table> Shared(const Key& key) const {
+            // The common case on the hot path: the same key as last time. One compare, no probe.
+            if (m_memoTable != nullptr && m_memoKey == key) return m_memoTable;
+            const auto found = m_buckets.find(key);
+            if (found != m_buckets.end()) {
+                m_memoKey = key;
+                m_memoTable = found->second;
+                return found->second;
+            }
+            SharedPtr<Table> fresh = MakeShared<Table>();
+            fresh->SetOwnerSession(key.SessionKey);
+            m_buckets.emplace(key, fresh);
+            m_memoKey = key;
+            m_memoTable = fresh;
+            return fresh;
+        }
+
+        // The calling thread's bucket - the ONE entry every in-place call site below resolves
+        // through, and the process-wide answer on a thread that belongs to no session.
+        //
+        // CONST, AND IT ANSWERS A MUTABLE BUCKET: the index is `mutable` because a LOOKUP may
+        // have to create the bucket it is about to miss in (a handle with no twin yet is the
+        // ordinary case), and the bucket's own entries are the table's business rather than the
+        // registry's. The alternative - a const overload returning "no table" - would make every
+        // read-only caller handle a case that cannot arise.
+        Table& ForCallingThread() const {
+            Key key;
+            CurrentTwinKey(&key);
+            return *Shared(key);
+        }
+
+        // Every bucket, in no particular order.
+        Vector<SharedPtr<Table>> Snapshot() const {
+            Vector<SharedPtr<Table>> all;
+            all.reserve(m_buckets.size());
+            for (const auto& bucket : m_buckets) all.push_back(bucket.second);
+            return all;
+        }
+
+        SizeT BucketCount() const { return m_buckets.size(); }
+
+        // THE PER-BUCKET ONE-SHOT, because a bucket held by SharedPtr is never dropped by an
+        // assignment: a caller that means "this context/group is gone" has to be able to say so,
+        // and the key is how it says which one. The bucket LEAVES the index before its twins are
+        // released, so a twin destructor that re-entered Shared() cannot find a half-erased one.
+        Bool Drop(const Key& key) {
+            const auto found = m_buckets.find(key);
+            if (found == m_buckets.end()) return false;
+            SharedPtr<Table> dropped = found->second;
+            m_buckets.erase(found);
+            if (m_memoTable == dropped) {
+                m_memoKey = Key{};
+                m_memoTable = nullptr;
+            }
+            dropped->Reset();
+            return true;
+        }
+
+        // EVERY bucket of one session, for the teardown of a session that owned several groups.
+        // The session is the key's high half, so this is one pass and no per-group bookkeeping -
+        // and it is the difference between "this session's twins go" and "the process's do".
+        Uint32 DropSession(Uint64 sessionKey) {
+            Uint32 dropped = 0;
+            for (auto bucket = m_buckets.begin(); bucket != m_buckets.end();) {
+                if (bucket->first.SessionKey != sessionKey) {
+                    ++bucket;
+                    continue;
+                }
+                SharedPtr<Table> table = bucket->second;
+                if (m_memoTable == table) {
+                    m_memoKey = Key{};
+                    m_memoTable = nullptr;
+                }
+                bucket = m_buckets.erase(bucket);
+                table->Reset();
+                ++dropped;
+            }
+            return dropped;
+        }
+
+        // Replace this registry with an empty one, releasing every bucket's twins. This is what
+        // `registry = {}` meant when the registry WAS one table; a caller that must keep the
+        // tables alive past the call takes a Snapshot() first.
+        void Reset() {
+            const Vector<SharedPtr<Table>> all = Snapshot();
+            m_buckets.clear();
+            m_memoKey = Key{};
+            m_memoTable = nullptr;
+            for (const SharedPtr<Table>& table : all) table->Reset();
+        }
+
+        // ---- the calling thread's bucket, member for member ---------------------------------
+        //
+        // THIS IS WHY THE SHAPE COST THE CALL SITES NOTHING. Every handle-keyed entry the table
+        // has is repeated here and forwarded to the bucket the calling thread resolves to, so
+        // `g_backendTextureObjects.FindByHandle(h)` and `g_backendBufferResources.GetOrCreate(h)`
+        // - unchanged source - answer from the right context's object namespace. A caller that
+        // must name a bucket OTHER than its own (a death walk, a case) takes Shared()/Snapshot()
+        // and calls the table directly.
+        BackendPtr& GetOrCreate(MG_Pipe::MGPipeHandle handle) {
+            return ForCallingThread().GetOrCreate(handle);
+        }
+        BackendPtr* FindByHandle(MG_Pipe::MGPipeHandle handle) {
+            return ForCallingThread().FindByHandle(handle);
+        }
+        const BackendPtr* FindByHandle(MG_Pipe::MGPipeHandle handle) const {
+            return ForCallingThread().FindByHandle(handle);
+        }
+        // Resolve-or-create in a bucket the caller NAMES rather than the one it is in. The
+        // monolith-glue frontend-keyed entry (StateBackendObjectRegistry::GetOrCreate below)
+        // needs it: a state object handed over on the client side of a sync is described by a
+        // record in the CALLING thread's group, and a twin built for it must land in the same
+        // bucket as that record.
+        BackendPtr& GetOrCreateIn(const Key& key, MG_Pipe::MGPipeHandle handle) {
+            return Shared(key)->GetOrCreate(handle);
+        }
+        Uint32 LiveGenAt(Uint32 slot) const { return ForCallingThread().LiveGenAt(slot); }
+        MG_Pipe::MGPipeHandle HandleOf(const StateObject* stateObj) const {
+            return ForCallingThread().HandleOf(stateObj);
+        }
+        void NoteStateForHandle(MG_Pipe::MGPipeHandle handle, const StatePtr& stateObj) {
+            ForCallingThread().NoteStateForHandle(handle, stateObj);
+        }
+        StatePtr StateForHandle(MG_Pipe::MGPipeHandle handle) const {
+            return ForCallingThread().StateForHandle(handle);
+        }
+        template <typename Fn>
+        void ForEachLive(Fn&& fn) const {
+            // THE CALLING THREAD'S BUCKET, not every bucket, and that is the semantic answer
+            // rather than a shortcut: a caller walking twins is asking "which of the objects I
+            // can see need this", and what it can see is its own share group. Every bucket in
+            // one process would walk a neighbour session's objects.
+            ForCallingThread().ForEachLive(fn);
+        }
+        // Every bucket, in one pass. The caller receives the key with the handle so it can tell
+        // which context a twin belongs to - the only walks that need the whole process's view
+        // are a diagnostic and a teardown.
+        template <typename Fn>
+        void ForEachBucket(Fn&& fn) const {
+            for (const auto& bucket : m_buckets) {
+                bucket.second->ForEachLive(
+                    [&](MG_Pipe::MGPipeHandle handle, const BackendPtr& twin) { fn(bucket.first, handle, twin); });
+            }
+        }
+        // The process-wide answer for the two statics: a notice is about an OBJECT, so it must
+        // reach the twin in EVERY bucket of EVERY holder. They forward to the table's own
+        // statics, which walk the holder list - each SHARED holder of it is one bucket.
+        static Bool ReleaseTwinByHandle(MG_Pipe::MGPipeHandle handle) {
+            return Table::ReleaseTwinByHandle(handle);
+        }
+        static Bool OnFrontendObjectDestroyed(Uint64 lifetimeId) {
+            return Table::OnFrontendObjectDestroyed(lifetimeId);
+        }
+        // The instance spelling, and it HANDS THE TWIN OUT rather than answering a Bool: the two
+        // tables that ARE a bucket registry (buffers, sampler views) have always been called as
+        // `g_backendBufferResources.ReleaseByHandle(res)` and their callers still have to decide
+        // what happens to the driver id the twin owns. The bucket is the CALLING THREAD'S, which
+        // is the bucket that twin was created in.
+        BackendPtr ReleaseByHandle(MG_Pipe::MGPipeHandle handle) {
+            return ForCallingThread().ReleaseByHandle(handle);
+        }
+        static Uint32 HolderCount() { return Table::HolderCount(); }
+        Uint32 LiveCount() const { return ForCallingThread().LiveCount(); }
+        Uint32 CompositeLiveCount() const { return ForCallingThread().CompositeLiveCount(); }
+        SizeT OrdinaryCapacityForTest() const { return ForCallingThread().OrdinaryCapacityForTest(); }
+        SizeT CompositeCapacityForTest() const { return ForCallingThread().CompositeCapacityForTest(); }
+
+        // ---- the bucket index's own controls, for the two tables that ARE a registry --------
+        //
+        // The six TwinRegistry registries carry the same three under the names in Managers.h
+        // (DropBucket / DropBucketsOfSession / ResetBuckets); they are repeated here because
+        // SlotTableRegistry IS the public type of the buffer and sampler-view tables, so a caller
+        // of those has no registry object in between.
+        Bool DropBucket(const Key& key) { return Drop(key); }
+        // Drop every bucket of one session; answers how many groups that took with it.
+        Uint32 DropBucketsOfSession(Uint64 sessionKey) { return DropSession(sessionKey); }
+        // Empty every bucket (the process-shaped world's teardown, key {0, 0}).
+        void ResetBuckets() { Reset(); }
+
+    private:
+        // ONE BUCKET PER LIVE {session, share group}, held by SHARED pointer so a caller may
+        // snapshot one across other lookups and so `Drop` can take a bucket's twins away
+        // explicitly - an embedded table could do neither. `std::unordered_map` rather than the
+        // tree's own flat Vector: this map is probed once per twin lookup on the draw path.
+        mutable std::unordered_map<Key, SharedPtr<Table>, TwinKeyHash> m_buckets;
+        // Hot-path memo. A registry is per KIND and each kind is touched several times per draw;
+        // a thread only ever moves between buckets when its session binds another context.
+        mutable SharedPtr<Table> m_memoTable;
+        mutable Key m_memoKey;
+    };
+
    template <typename StateObject, typename BackendObject, MG_Pipe::MGPipeKind kKind>
    class BackendSlotTable {
    public:
@@ -209,6 +459,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // ENTRIES and the memo; the links are the table's own and are never copied.
        BackendSlotTable() { LinkHolder(); }
        BackendSlotTable(const BackendSlotTable& other):
+            m_ownerSession(other.m_ownerSession),
            m_slots(other.m_slots),
            m_band(other.m_band),
            m_nullTwin(other.m_nullTwin),
@@ -217,6 +468,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
            LinkHolder();
        }
        BackendSlotTable(BackendSlotTable&& other) noexcept:
+            m_ownerSession(other.m_ownerSession),
            m_slots(std::move(other.m_slots)),
            m_band(std::move(other.m_band)),
            m_nullTwin(std::move(other.m_nullTwin)),
@@ -252,6 +504,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
        }
        ~BackendSlotTable() { UnlinkHolder(); }

+        // The served session whose bucket this table is (SlotTableRegistry::Shared), 0 for a
+        // table that belongs to none (the process-wide registry on a sessionless thread, a
+        // fixture's copy of one).
+        void SetOwnerSession(Uint64 session) { m_ownerSession = session; }
+        Uint64 OwnerSession() const { return m_ownerSession; }
+
        // Resolve-or-create. The handle comes from the client allocator keyed on the frontend
        // object's lifetime id, so two calls for the same live object always land on the same
        // slot, and a successor object at the same heap address never does.
@@ -553,12 +811,24 @@ namespace MobileGL::MG_Backend::DirectGLES {
        // double Free would be refused by generation anyway. Idempotent against the kind's
        // own delete opcode, exactly as the notice arm is: a twin the delete already released
        // fails ReleaseTwinAt's generation check and the walk moves on.
+        //
+        // ONE SESSION'S HOLDERS, WHEN THE DEATH IS ONE SESSION'S. A handle is numbered by its own
+        // client's allocator, so with several clients served by one process the same {slot, gen}
+        // names a live, unrelated object in every other session: walking every holder released
+        // THOSE twins too - a client deleting a texture (or exiting) deleted the compositor's
+        // driver texture with the same handle, and its windows sampled freed memory. On a thread
+        // that belongs to a session only that session's buckets answer; a sessionless thread
+        // (the in-library backend, a unit fixture) keeps the process-wide walk.
        static Bool ReleaseTwinByHandle(MG_Pipe::MGPipeHandle handle) {
            if (MG_Pipe::MGPipeHandleIsNull(handle)) return false;
+            TwinKey key;
+            const Bool sessionScoped = CurrentTwinKey(&key) && key.SessionKey != 0;
            Bool released = false;
            for (BackendSlotTable* holder = s_firstHolder; holder != nullptr;) {
                BackendSlotTable* const next = holder->m_nextHolder;
-                released = holder->ReleaseTwinAt(handle) || released;
+                if (!sessionScoped || holder->m_ownerSession == key.SessionKey) {
+                    released = holder->ReleaseTwinAt(handle) || released;
+                }
                holder = next;
            }
            return released;
@@ -616,6 +886,24 @@ namespace MobileGL::MG_Backend::DirectGLES {
            }
        }

+        // Empty this table, releasing every twin it holds. THE SPACES ARE MOVED OUT FIRST and
+        // that is the whole of the care taken here: a twin's destructor is a driver call that
+        // could re-enter GetOrCreate on this table (SlotTables.h's ReleaseTwinAt says why), and
+        // destroying the entries in place would leave that re-entry resizing the vector the
+        // destructors are being walked out of. Here the table is already empty when the entries
+        // die, so a re-entrant insert grows a fresh vector.
+        //
+        // This is what `table = {}` used to mean, spelled so a caller cannot miss the ordering.
+        void Reset() {
+            Vector<Entry> slots = std::move(m_slots);
+            Vector<Entry> band = std::move(m_band);
+            BackendPtr nullTwin = std::move(m_nullTwin);
+            ForgetHandle();
+            slots.clear();
+            band.clear();
+            nullTwin.reset();
+        }
+
        Uint32 LiveCount() const {
            Uint32 count = 0;
            for (const Entry& entry : m_slots) {
@@ -745,6 +1033,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
        static inline BackendSlotTable* s_firstHolder = nullptr;
        BackendSlotTable* m_prevHolder = nullptr;
        BackendSlotTable* m_nextHolder = nullptr;
+        Uint64 m_ownerSession = 0;

        // Indexed by MGPipeHandle::Slot; [0] is the reserved slot and is never live.
        Vector<Entry> m_slots;
@@ -572,7 +572,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
            // The name is the marker: ShaderSourceProcessor only emits it when the source
-            // wrote gl_FragColor, and such a shader can have no other output.
+            // names gl_FragColor, and a shader that writes it can have no other output. One that
+            // names it only in an inactive preprocessor arm keeps the declaration unwritten, and
+            // the IO resolver moves that inactive output off location 0, so the match below
+            // (which wants location 0) does not take it.
            static const char* const kLoweredName = "mg_FragColor";
            if (shaderType != GL_FRAGMENT_SHADER || drawBufferCount <= 1) {
                return glslCode;
@@ -1,6 +1,7 @@
 // Included in DirectGLES.cpp after the native packed-word readback helper.
-// The server needs owned tight rows for a storage promotion, not client PACK
-// layout or a fallback into a frontend texture's potentially stale shadow.
+// A storage promotion needs the driver's own texels as owned tight rows, not
+// client PACK layout or a fallback into a frontend texture's potentially stale
+// shadow.
    class ScopedWireTextureReadLevel {
    public:
        ScopedWireTextureReadLevel(GLuint texture, GLenum target, GLint level)
@@ -280,6 +281,144 @@
        return true;
    }

+#if MOBILEGL_BUILD_DISAGGREGATED
+    // P9 W2-a. The store's Defined-ness follows the client's respecifies; the driver's follows the
+    // twin's last sync. A re-mint asks this before reading a level back, because a level defined
+    // after that sync (null data, or null data plus a partial upload) is not on the driver yet and
+    // its FBO attachment is incomplete - which used to end the session at
+    // Fatal{ResourceUnavailable, "image-promotion-readback"} on legal GL.
+    Bool NativeTextureLevelHasExtent(GLuint texture, TextureTarget target, TextureUploadTarget uploadTarget,
+                                     GLint level, const IntVec3& logicalExtent) {
+        if (!texture || level < 0 || !g_GLESFuncs.glGetTexLevelParameteriv) return true;
+        const IntVec3 extent = TextureImpl::GetBackendUploadSize(target, logicalExtent);
+        const GLenum nativeTarget = TextureImpl::ConvertTextureTargetToBackendGLEnum(target);
+        GLenum binding = GL_NONE;
+        switch (nativeTarget) {
+        case GL_TEXTURE_2D: binding = GL_TEXTURE_BINDING_2D; break;
+        case GL_TEXTURE_3D: binding = GL_TEXTURE_BINDING_3D; break;
+        case GL_TEXTURE_2D_ARRAY: binding = GL_TEXTURE_BINDING_2D_ARRAY; break;
+        case GL_TEXTURE_CUBE_MAP: binding = GL_TEXTURE_BINDING_CUBE_MAP; break;
+        case GL_TEXTURE_CUBE_MAP_ARRAY: binding = GL_TEXTURE_BINDING_CUBE_MAP_ARRAY; break;
+        default: return true;
+        }
+        // A cube's levels are per face; every other target answers for the whole level.
+        const GLenum queryTarget = nativeTarget == GL_TEXTURE_CUBE_MAP
+            ? TextureImpl::ConvertTextureUploadTargetToBackendGLEnum(uploadTarget) : nativeTarget;
+        DrainESErrors();
+        GLint previous = 0;
+        g_GLESFuncs.glGetIntegerv(binding, &previous);
+        g_GLESFuncs.glBindTexture(nativeTarget, texture);
+        GLint width = 0, height = 0, depth = 0;
+        g_GLESFuncs.glGetTexLevelParameteriv(queryTarget, level, GL_TEXTURE_WIDTH, &width);
+        g_GLESFuncs.glGetTexLevelParameteriv(queryTarget, level, GL_TEXTURE_HEIGHT, &height);
+        g_GLESFuncs.glGetTexLevelParameteriv(queryTarget, level, GL_TEXTURE_DEPTH, &depth);
+        const Bool answered = g_GLESFuncs.glGetError() == GL_NO_ERROR;
+        g_GLESFuncs.glBindTexture(nativeTarget, static_cast<GLuint>(previous));
+        if (!answered) return true;
+        const Bool layered = nativeTarget == GL_TEXTURE_3D || nativeTarget == GL_TEXTURE_2D_ARRAY ||
+                             nativeTarget == GL_TEXTURE_CUBE_MAP_ARRAY;
+        return width == extent.x() && height == extent.y() && (!layered || depth == extent.z());
+    }
+
+    // P8-E (notes/p8/E.md). MGITEST_ESPRYT_REFUSE_TEXTURE_READBACK_EXTENT=<w>x<h>, READ BY THE SERVER:
+    // a colour level of exactly that extent is treated as one the driver refused to attach - what an
+    // Adreno does for RGB16 and the 16-bit SNORM formats, and what no host driver can be made to do.
+    // Scoped by the extent so the tcp lane's shared supervisor can carry it for the whole lane: only
+    // the scenario that builds that extent (CopyImageStoreReadbackScenario) ever meets it.
+    static Bool WireTextureReadbackRefusedForTesting(const IntVec3& extent) {
+        static const IntVec3 s_refused = [] {
+            const char* value = std::getenv("MGITEST_ESPRYT_REFUSE_TEXTURE_READBACK_EXTENT");
+            if (value == nullptr) return IntVec3{0, 0, 0};
+            char* end = nullptr;
+            const long width = std::strtol(value, &end, 10);
+            if (end == value || *end != 'x') return IntVec3{0, 0, 0};
+            const long height = std::strtol(end + 1, nullptr, 10);
+            if (width <= 0 || height <= 0) return IntVec3{0, 0, 0};
+            return IntVec3{static_cast<Int>(width), static_cast<Int>(height), 1};
+        }();
+        return s_refused.x() > 0 && extent == s_refused;
+    }
+
+    // P8-E: THE STORE ROUTE, for a level the driver refused to read back. The server's staged store
+    // is the only CPU copy of the level on this side, and it answers only while its bytes ARE the
+    // level: every byte crossed (IsCovered), at the extent asked for, and the texture was never
+    // written by the driver except by copies the store followed (!IsDriverWritten - see
+    // StagedTextureStore's P8-E block and FollowCopyImageInStagedStore). The conversion is
+    // GetTexImageViaShadowConversion's, into tight rows: a packed format read with its own type
+    // takes the stored words verbatim, everything else decodes to wide RGBA and converts row by row.
+    // Anything else DECLINES BY NAME: the reply is an error, and the client's
+    // Fatal{ReplyError, "ReadPixels"} ends the session instead of handing back bytes the driver has
+    // since overwritten.
+    static Bool ReadWireLevelFromStagedStore(MG_Pipe::MGPipeHandle res, TextureInternalFormat logicalFormat,
+                                             TextureUploadTarget uploadTarget, GLint level, const IntVec3& extent,
+                                             GLenum format, GLenum type, Vector<Uint8>& bytes) {
+        auto& store = MG_Remote::Server::ServerStagedTexture();
+        const Uint64 key = MG_Remote::Server::StagedTextureStore::KeyForHandle(res);
+        const auto target = static_cast<Uint16>(uploadTarget);
+        const auto mip = static_cast<Uint16>(level);
+        ReadbackChannelMapping mapping{};
+        const SizeT pixelBytes = GetReadbackChannelMapping(format, mapping) ? GetReadbackDstPixelSize(mapping, type) : 0;
+        const SizeT texels = static_cast<SizeT>(std::max(extent.x(), 0)) * static_cast<SizeT>(std::max(extent.y(), 0)) *
+                             static_cast<SizeT>(std::max(extent.z(), 0));
+        const char* declined = nullptr;
+        if (!store.CopiesIntoServerStorage()) {
+            declined = "this arm keeps no staged store";
+        } else if (store.IsDriverWritten(key)) {
+            declined = "the driver has written the texture since its bytes crossed";
+        } else if (!store.IsCovered(key, target, mip)) {
+            declined = "not every byte of the level crossed";
+        } else if (store.LevelExtentOrUndefined(key, target, mip) != extent) {
+            declined = "the staged level is not the extent asked for";
+        } else if (pixelBytes == 0 || texels == 0) {
+            declined = "the format/type pair has no readback mapping";
+        }
+        Vector<Uint8> out;
+        if (declined == nullptr) {
+            const Uint8* staged = store.RequireLevelBytes(key, target, mip, "texture-readback-from-store");
+            if (MG_Util::PixelStoreProcessor::IsRawPackedPixelTransfer(
+                    logicalFormat, MG_Util::ConvertGLEnumToTextureInputFormat(format),
+                    MG_Util::ConvertGLEnumToTexturePixelDataType(type))) {
+                if (store.LevelByteSize(key, target, mip) != texels * pixelBytes) {
+                    declined = "the staged words are not the client's word size";
+                } else {
+                    out.assign(staged, staged + texels * pixelBytes);
+                }
+            } else {
+                Vector<Uint8> wide;
+                Bool isInteger = false;
+                Bool isSigned = false;
+                if (!MG_Util::PixelStoreProcessor::DecodeShadowDataToWideRGBA(logicalFormat, staged, texels, wide,
+                                                                             isInteger, isSigned)) {
+                    declined = "the staged format has no wide decoding";
+                } else if (mapping.isInteger != isInteger) {
+                    declined = "an integer read of a normalized level, or the reverse";
+                } else {
+                    const GLenum wideType = isInteger ? (isSigned ? GL_INT : GL_UNSIGNED_INT) : GL_FLOAT;
+                    const SizeT rowTexels = static_cast<SizeT>(extent.x());
+                    const SizeT rows = texels / rowTexels;
+                    const SizeT wideRowBytes = rowTexels * 4 * GetReadbackComponentSize(wideType);
+                    out.resize(texels * pixelBytes);
+                    for (SizeT row = 0; row < rows; ++row) {
+                        ReadbackImpl::ConvertWideReadbackRow(wide.data() + row * wideRowBytes,
+                                                             out.data() + row * rowTexels * pixelBytes, rowTexels,
+                                                             wideType, mapping, type);
+                    }
+                }
+            }
+        }
+        if (declined != nullptr) {
+            MGLOG_E_ONCE("MGPipe: texture {%u,%u} level %d: the driver refused the readback and the staged store "
+                         "cannot answer it (%s) - DECLINED (texture-readback-store-declined)",
+                         res.Slot, res.Gen, level, declined);
+            return false;
+        }
+        MGLOG_W_ONCE("MGPipe: texture {%u,%u} level %d: the driver refused the readback (internal format %u); "
+                     "answered from the server's staged store (texture-readback-from-store)",
+                     res.Slot, res.Gen, level, static_cast<unsigned>(logicalFormat));
+        bytes = Move(out);
+        return true;
+    }
+
    Bool ReadTextureImageWire(const MG_Pipe::MGPReadbackInfo& info, Vector<Uint8>& bytes) {
        const auto* record = PipeTextureRecordForHandle(info.Res);
        if (!record) return false;
@@ -293,17 +432,32 @@
        const Bool wholeCube = info.Target == GL_TEXTURE_CUBE_MAP;
        const IntVec3 extent{static_cast<Int>(info.Box.W), static_cast<Int>(info.Box.H),
                            wholeCube ? 1 : static_cast<Int>(info.Box.D)};
+        const auto logicalFormat = static_cast<TextureInternalFormat>(record->Desc.InternalFormat);
+        // The R32UI word leg never goes through a framebuffer, so the knob leaves it alone (it is
+        // ReadTextureLevelTight's first choice for exactly these reads).
+        const Bool packedWordRead =
+            !twin->RequiresImageBindableStorage() &&
+            MG_Util::PixelStoreProcessor::IsRawPackedPixelTransfer(
+                logicalFormat, MG_Util::ConvertGLEnumToTextureInputFormat(info.Format),
+                MG_Util::ConvertGLEnumToTexturePixelDataType(info.Type)) &&
+            MG_Util::GetInputBytesPerPixel(MG_Util::ConvertGLEnumToTextureInputFormat(info.Format),
+                                           MG_Util::ConvertGLEnumToTexturePixelDataType(info.Type)) == sizeof(Uint32);
+        const Bool refusedForTesting = !packedWordRead && WireTextureReadbackRefusedForTesting(extent);
        Vector<Uint8> result;
        for (Int face = 0; face < (wholeCube ? 6 : 1); ++face) {
            const auto uploadTarget = wholeCube
                ? static_cast<TextureUploadTarget>(static_cast<Int>(TextureUploadTarget::CubeMapPositiveX) + face)
                : MG_Util::ConvertGLEnumToTextureUploadTarget(info.Target);
            Vector<Uint8> slice;
-            if (!ReadTextureLevelTight(twin->GetBackendTextureId(), target, uploadTarget,
-                    static_cast<TextureInternalFormat>(record->Desc.InternalFormat), info.Level, extent,
-                    twin->RequiresImageBindableStorage(), info.Format, info.Type, slice)) return false;
+            if (refusedForTesting ||
+                !ReadTextureLevelTight(twin->GetBackendTextureId(), target, uploadTarget, logicalFormat, info.Level,
+                    extent, twin->RequiresImageBindableStorage(), info.Format, info.Type, slice)) {
+                if (!ReadWireLevelFromStagedStore(info.Res, logicalFormat, uploadTarget, info.Level, extent,
+                                                  info.Format, info.Type, slice)) return false;
+            }
            result.insert(result.end(), slice.begin(), slice.end());
        }
        bytes = Move(result);
        return true;
    }
+#endif // MOBILEGL_BUILD_DISAGGREGATED
@@ -332,6 +332,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {

 #if MOBILEGL_BUILD_DISAGGREGATED
    BackendObject_DirectVulkan::~BackendObject_DirectVulkan() {
+        // Everything below reaches this session's renderer slot, whichever thread destroys it.
+        MagmaSession* const ownSession = m_magmaSession.get();
+        MagmaSession* const previousSession = BoundMagmaSession();
+        if (ownSession != nullptr) BindMagmaSessionToThisThread(ownSession);
        // P12 review fix: A SERVER SESSION THAT ENDS WITHOUT A CLEAN TEARDOWN LEAVES NO RENDERER ON
        // THE WINDOW. The renderer - swapchain and VkSurfaceKHR on the window - is a process global
        // that only ReleaseEGLResources / a released surface drops, and a client that went away
@@ -345,6 +349,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            pVulkanRenderer.reset();
            ClearProgramResourceCaches();
        }
+        // And the swap interval it asked for (DirectVulkan.h).
+        if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
+            ForgetRequestedSwapInterval();
+        }
+        // The session's state goes with it, and no thread may keep answering with it.
+        m_magmaSession.reset();
+        if (ownSession != nullptr) {
+            BindMagmaSessionToThisThread(previousSession == ownSession ? nullptr : previousSession);
+        }
    }
 #else
    BackendObject_DirectVulkan::~BackendObject_DirectVulkan() = default;
@@ -353,6 +366,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    BackendObject_DirectVulkan::BackendObject_DirectVulkan() : m_rendererInfo{GetRendererIdentity()} {}

    Bool BackendObject_DirectVulkan::InitWindowSurface() {
+        // A suspended surface activates its placeholder - never the window, which is gone.
+        const auto suspended = m_suspendedWindowSurfaces.find(m_activatingSurface);
+        if (suspended != m_suspendedWindowSurfaces.end()) {
+            if (UsesSurfaceTargets() && pVulkanRenderer) {
+                return ActivateSuspendedPlaceholder(m_activatingSurface, suspended->second);
+            }
+            return InitPbufferSurface(static_cast<EGLint>(suspended->second.Width),
+                                      static_cast<EGLint>(suspended->second.Height));
+        }
        if (!m_windowHandle.Handle) {
            MGLOG_E("Cannot initialize DirectVulkan window surface: native window handle is null");
            return false;
@@ -360,12 +382,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        auto nativeWindow = reinterpret_cast<NativeWindowType>(m_windowHandle.Handle);

+        VulkanRendererConfig config;
+        config.SwapInterval = GetRequestedSwapInterval();
+        if (UsesSurfaceTargets() && pVulkanRenderer) {
+            return pVulkanRenderer->ActivateSurfaceTarget(SurfaceTargetKey(m_activatingSurface), nativeWindow, config);
+        }
        // Any renderer instance this assignment replaces is destroyed here;
        // fence/timer-query handles stamped with the old generation go stale.
        BumpRendererGeneration();
-        pVulkanRenderer = MakeUnique<MG_Backend::DirectVulkan::VulkanRenderer>(nativeWindow);
+        pVulkanRenderer = MakeUnique<MG_Backend::DirectVulkan::VulkanRenderer>(nativeWindow, config);
        MOBILEGL_ASSERT(pVulkanRenderer != nullptr, "InitWindowSurface: VulkanRenderer creation failed");
        pVulkanRenderer->Initialize();
+        pVulkanRenderer->SetActiveSurfaceTargetKey(SurfaceTargetKey(m_activatingSurface));
        return true;
    }

@@ -373,15 +401,35 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VulkanRendererConfig config;
        config.SurfaceWidth = static_cast<Uint32>(std::max<EGLint>(width, 1));
        config.SurfaceHeight = static_cast<Uint32>(std::max<EGLint>(height, 1));
+        if (UsesSurfaceTargets() && pVulkanRenderer) {
+            return pVulkanRenderer->ActivateSurfaceTarget(SurfaceTargetKey(m_activatingSurface), NativeWindowType{},
+                                                          config);
+        }
        // Any renderer instance this assignment replaces is destroyed here;
        // fence/timer-query handles stamped with the old generation go stale.
        BumpRendererGeneration();
        pVulkanRenderer = MakeUnique<MG_Backend::DirectVulkan::VulkanRenderer>(NativeWindowType{}, config);
        MOBILEGL_ASSERT(pVulkanRenderer != nullptr, "InitPbufferSurface: VulkanRenderer creation failed");
        pVulkanRenderer->Initialize();
+        pVulkanRenderer->SetActiveSurfaceTargetKey(SurfaceTargetKey(m_activatingSurface));
        return true;
    }

+    void BackendObject_DirectVulkan::BindSessionStateToThisThread() {
+        if (m_magmaSession == nullptr) m_magmaSession = MakeUnique<MagmaSession>();
+        BindMagmaSessionToThisThread(m_magmaSession.get());
+        MGLOG_I("DirectVulkan: session state %llu bound to this apply thread",
+                static_cast<unsigned long long>(m_magmaSession->Id()));
+    }
+
+    Bool BackendObject_DirectVulkan::LatchIfGpuFaulted() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return pVulkanRenderer != nullptr && pVulkanRenderer->LatchIfGpuHung("apply");
+#else
+        return false;
+#endif
+    }
+
    void BackendObject_DirectVulkan::Initialize() {
 #if MOBILEGL_BUILD_DISAGGREGATED
        VkBufferManager::RegisterWireResourceOps();
@@ -457,6 +505,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (!BackendObject::ResizeEGLWindowSurface(surface, width, height)) {
            return false;
        }
+        const EGLSurfaceState* state = GetRegisteredEGLSurface(surface);
+        if (UsesSurfaceTargets() && pVulkanRenderer && state != nullptr && state->Kind == SurfaceKind::Pbuffer) {
+            // A pbuffer-backed window (a Wayland window presented through wl_shm) whose target is an
+            // image reader of FIXED size: a swapchain rebuild cannot follow the resize, so the target
+            // is rebuilt at the new size - as DirectGLES replaces its native pbuffer. A parked one is
+            // rebuilt when it is next activated, from the size registered above.
+            const Bool active = m_eglSurfaceInitialized && m_eglSurface == surface;
+            pVulkanRenderer->DestroySurfaceTarget(SurfaceTargetKey(surface));
+            if (active) {
+                VulkanRendererConfig config;
+                config.SurfaceWidth = static_cast<Uint32>(state->Width);
+                config.SurfaceHeight = static_cast<Uint32>(state->Height);
+                (void)pVulkanRenderer->ActivateSurfaceTarget(SurfaceTargetKey(surface), NativeWindowType{}, config);
+            }
+            return true;
+        }
        if (pVulkanRenderer && m_eglSurface == surface) {
            pVulkanRenderer->RequestSwapchainResize(width, height);
        }
@@ -474,7 +538,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {

    Bool BackendObject_DirectVulkan::MakeEGLCurrent(EGLDisplay dpy, EGLSurface draw, EGLSurface read, EGLContext ctx) {
        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
-        return BackendObject::MakeEGLCurrent(dpy, draw, read, ctx);
+        const Bool ok = BackendObject::MakeEGLCurrent(dpy, draw, read, ctx);
+        if (ok && UsesSurfaceTargets() && ctx != EGL_NO_CONTEXT && draw != EGL_NO_SURFACE) {
+            m_contextBindings[ctx] = EGLCurrentState{.Display = dpy, .DrawSurface = draw, .ReadSurface = read,
+                                                     .Context = ctx};
+        }
+        return ok;
+    }
+
+    void BackendObject_DirectVulkan::OnClientContextBound(Uint64 token) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (!UsesSurfaceTargets() || token == 0) return;
+        // Each of a multi-window client's contexts draws to the surface IT was made current with
+        // (Qt renders every window on its own thread and context); the ring interleaves them with
+        // a bind_context between, and no make-current. Follow the context back to its surface.
+        const auto found = m_contextBindings.find(reinterpret_cast<EGLContext>(token));
+        if (found == m_contextBindings.end()) return;
+        const EGLCurrentState binding = found->second;
+        if (!GetRegisteredEGLSurface(binding.DrawSurface)) {
+            m_contextBindings.erase(found);
+            return;
+        }
+        const auto current = m_eglCurrentThreads.find(std::this_thread::get_id());
+        if (current != m_eglCurrentThreads.end() && current->second.Context == binding.Context &&
+            current->second.DrawSurface == binding.DrawSurface && m_eglSurface == binding.DrawSurface) {
+            return;
+        }
+        (void)BackendObject::MakeEGLCurrent(binding.Display, binding.DrawSurface, binding.ReadSurface, binding.Context);
    }

    Bool BackendObject_DirectVulkan::SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw) {
@@ -486,6 +576,56 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return BackendObject::SwapEGLBuffers(dpy, draw);
    }

+    Bool BackendObject_DirectVulkan::BlitDefaultFramebufferToSharedImage(const SharedImageView& image,
+                                                                         const MG_Util::Damage::Region& region) {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // Same lock as SwapEGLBuffers: the copy reads the current surface target's image.
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (!pVulkanRenderer) {
+            MGLOG_E("DirectVulkan: shared image %llu: the renderer is not initialized",
+                    static_cast<unsigned long long>(image.Id));
+            return false;
+        }
+        return pVulkanRenderer->BlitDefaultFramebufferToSharedImage(image, region);
+#else
+        (void)image, (void)region;
+        return false;
+#endif
+    }
+
+    Bool BackendObject_DirectVulkan::PublishSharedImageAccesses() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        return pVulkanRenderer ? pVulkanRenderer->PublishSharedImageAccesses() : true;
+#else
+        return true;
+#endif
+    }
+
+    Bool BackendObject_DirectVulkan::ExportNativeFence(int* fence) {
+        *fence = -1;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        return pVulkanRenderer ? pVulkanRenderer->ExportNativeFence(fence) : true;
+#else
+        return true;
+#endif
+    }
+
+    // A swapchain image keeps its content through a present whatever region is drawn next: the
+    // damage region the client may declare is not needed here.
+    Int32 BackendObject_DirectVulkan::QueryCurrentBufferAge(Bool damageRegionFollows) {
+        (void)damageRegionFollows;
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        return pVulkanRenderer ? pVulkanRenderer->CurrentDrawBufferAge() : 0;
+    }
+
+    void BackendObject_DirectVulkan::SetEGLSwapInterval(Int interval) {
+        // Same lock as SwapEGLBuffers: the renderer reads the request in Present.
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        BackendObject::SetEGLSwapInterval(interval);
+    }
+
    void BackendObject_DirectVulkan::ReleaseEGLSurface(EGLSurface surface) {
        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
        BackendObject::ReleaseEGLSurface(surface);
@@ -500,11 +640,131 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // The reflection cache is file-scope, not renderer-owned; without this the
        // deleted programs' reflection strings survive full context teardown.
        ClearProgramResourceCaches();
+        m_suspendedWindowSurfaces.clear();
        BackendObject::ReleaseEGLResources();
    }

+    Bool BackendObject_DirectVulkan::ActivateSuspendedPlaceholder(EGLSurface surface,
+                                                                  const SuspendedWindowSurface& extent) {
+        // A pbuffer target at the window's extent: every default-framebuffer path (render passes,
+        // read-back, the shared-image blit, the acquire and the present) keeps a live swapchain to
+        // work on, and the extent it publishes is the one the client already has.
+        VulkanRendererConfig config;
+        config.SurfaceWidth = std::max<Uint32>(extent.Width, 1);
+        config.SurfaceHeight = std::max<Uint32>(extent.Height, 1);
+        return pVulkanRenderer->ActivateSurfaceTarget(SurfaceTargetKey(surface), NativeWindowType{}, config);
+    }
+
+    Bool BackendObject_DirectVulkan::SuspendServerWindow(void* window) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        // The monolith's one renderer IS its window; only surface targets can be swapped under it.
+        if (window == nullptr || !UsesSurfaceTargets()) return false;
+        Vector<EGLSurface> onWindow;
+        for (const auto& [client, state] : m_eglSurfaces) {
+            if (state.Kind != SurfaceKind::Window || state.Window.Handle != window) continue;
+            if (m_suspendedWindowSurfaces.find(client) != m_suspendedWindowSurfaces.end()) continue;
+            onWindow.push_back(client);
+        }
+        for (const EGLSurface client : onWindow) {
+            const auto state = m_eglSurfaces.find(client);
+            SuspendedWindowSurface extent;
+            extent.Width = static_cast<Uint32>(std::max<EGLint>(state->second.Width, 1));
+            extent.Height = static_cast<Uint32>(std::max<EGLint>(state->second.Height, 1));
+            if (pVulkanRenderer) {
+                const Uint64 key = SurfaceTargetKey(client);
+                const VkExtent2D built = pVulkanRenderer->SurfaceTargetExtent(key);
+                if (built.width != 0 && built.height != 0) {
+                    extent.Width = built.width;
+                    extent.Height = built.height;
+                }
+                const Bool active = pVulkanRenderer->IsActiveSurfaceTarget(key);
+                // Flushes, waits the device idle, purges the render passes on the target and drops its
+                // swapchain and VkSurfaceKHR - nothing references the window after this.
+                pVulkanRenderer->DestroySurfaceTarget(key);
+                m_suspendedWindowSurfaces[client] = extent;
+                // A placeholder that cannot be built is a suspend that failed: the window is already let
+                // go of, and the caller ends the session by name instead (its EGL resources go).
+                if (active && !ActivateSuspendedPlaceholder(client, extent)) return false;
+            } else {
+                m_suspendedWindowSurfaces[client] = extent;
+            }
+        }
+        MGLOG_I("DirectVulkan: %zu window surface(s) on server window %p now draw into placeholder pbuffer targets; "
+                "the renderer, the device and every object stay",
+                onWindow.size(), window);
+        return true;
+    }
+
+    Bool BackendObject_DirectVulkan::ResumeServerWindow(void* window, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (window == nullptr || !UsesSurfaceTargets()) return false;
+        UnorderedMap<EGLSurface, SuspendedWindowSurface> suspended;
+        suspended.swap(m_suspendedWindowSurfaces);
+        for (const auto& [client, extent] : suspended) {
+            (void)extent;
+            const auto state = m_eglSurfaces.find(client);
+            if (state == m_eglSurfaces.end()) continue; // released while suspended
+            state->second.Window.Handle = window;
+            state->second.Window.Width = width;
+            state->second.Window.Height = height;
+            state->second.Width = static_cast<EGLint>(std::max<Uint32>(width, 1));
+            state->second.Height = static_cast<EGLint>(std::max<Uint32>(height, 1));
+            if (m_eglSurface == client) m_windowHandle = state->second.Window;
+            if (!pVulkanRenderer) continue;
+            const Uint64 key = SurfaceTargetKey(client);
+            const Bool active = pVulkanRenderer->IsActiveSurfaceTarget(key);
+            pVulkanRenderer->DestroySurfaceTarget(key); // the placeholder
+            // A parked surface is built on the new window when it is next activated (InitWindowSurface).
+            if (!active) continue;
+            // The creation's own path: a new VkSurfaceKHR and swapchain with the surface's current
+            // pre-transform, a new target serial (render passes rebuilt), and the extent published.
+            VulkanRendererConfig config;
+            config.SurfaceWidth = std::max<Uint32>(width, 1);
+            config.SurfaceHeight = std::max<Uint32>(height, 1);
+            config.SwapInterval = GetRequestedSwapInterval();
+            if (!pVulkanRenderer->ActivateSurfaceTarget(key, reinterpret_cast<NativeWindowType>(window), config)) {
+                return false;
+            }
+        }
+        MGLOG_I("DirectVulkan: %zu suspended window surface(s) rebuilt on server window %p (%ux%u)", suspended.size(),
+                window, width, height);
+        return true;
+    }
+
+    Bool BackendObject_DirectVulkan::ServerWindowResized(void* window, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (window == nullptr || !UsesSurfaceTargets()) return false;
+        Bool followed = false;
+        for (auto& [client, state] : m_eglSurfaces) {
+            if (state.Kind != SurfaceKind::Window || state.Window.Handle != window) continue;
+            if (m_suspendedWindowSurfaces.find(client) != m_suspendedWindowSurfaces.end()) continue;
+            state.Window.Width = width;
+            state.Window.Height = height;
+            state.Width = static_cast<EGLint>(std::max<Uint32>(width, 1));
+            state.Height = static_cast<EGLint>(std::max<Uint32>(height, 1));
+            if (m_eglSurface == client) m_windowHandle = state.Window;
+            // Magma's extent is its swapchain's, published when it is built: the active target follows
+            // the window now (or at the half-recorded frame's present); a parked one at its next present.
+            if (pVulkanRenderer && pVulkanRenderer->IsActiveSurfaceTarget(SurfaceTargetKey(client))) {
+                followed = pVulkanRenderer->FollowActiveWindowResize() || followed;
+            }
+        }
+        return followed;
+    }
+
+    Bool BackendObject_DirectVulkan::UsesSurfaceTargets() const {
+        return MG_Config::Transport != MG_Config::TransportMode::Monolith;
+    }
+
+    void BackendObject_DirectVulkan::OnEGLSurfaceForgotten(EGLSurface surface) {
+        m_suspendedWindowSurfaces.erase(surface);
+        if (UsesSurfaceTargets() && pVulkanRenderer) pVulkanRenderer->DestroySurfaceTarget(SurfaceTargetKey(surface));
+    }
+
    void BackendObject_DirectVulkan::OnEGLSurfaceReleased(EGLSurface surface) {
-        (void)surface;
+        // The renderer outlives its surfaces here: only this surface's target goes (in
+        // OnEGLSurfaceForgotten, which follows), and every fence stays valid.
+        if (UsesSurfaceTargets() && pVulkanRenderer) return;
        // Outstanding fence/timer-query handles now refer to a dead renderer;
        // treat them as signaled/available with zero results from here on.
        BumpRendererGeneration();
@@ -553,6 +813,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, V_OpenGL41, V_OpenGL42, V_OpenGL43,
            V_OpenGL44, V_OpenGL45, V_OpenGL46,
            E_GL_ARB_draw_buffers_blend,
+            // The GLSL 1.00/1.10-era ARB trio plus NPOT textures: all core since GL 2.0 and
+            // trivially true for every context this backend claims (4.6). KWin 6's
+            // EglContext::checkSupported() gates desktop-GL compositing on exactly these four
+            // strings, so withholding them fails its renderer bring-up. Same block as DirectGLES.
+            E_GL_ARB_shader_objects, E_GL_ARB_vertex_shader, E_GL_ARB_fragment_shader,
+            E_GL_ARB_texture_non_power_of_two,
            E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_draw_indirect,
            E_GL_ARB_multi_draw_indirect,
@@ -727,6 +993,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        static Bool funcsTableInitialized = false;
        if (!funcsTableInitialized) {
            funcsTable.Present = Present;
+            funcsTable.SetSwapInterval = SetSwapInterval;
            funcsTable.GL.DrawArrays = DrawArrays;
            funcsTable.GL.DrawElements = DrawElements;
            funcsTable.GL.DrawElementsBaseVertex = DrawElementsBaseVertex;
@@ -9,6 +9,7 @@
 #pragma once
 #include <Includes.h>
 #include "../BackendObject.h"
+#include "MagmaSession.h"
 #include <MG_Util/BackendLoaders/Vulkan/Loader.h>

 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -26,6 +27,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        ~BackendObject_DirectVulkan() override;

        void Initialize() override;
+        void BindSessionStateToThisThread() override;
+        // The GPU hang watch's verdict on this session's device (VulkanRenderer::LatchIfGpuHung).
+        Bool LatchIfGpuFaulted() override;
+        void OnClientContextBound(Uint64 token) override;
        Bool InitWindowSurface() override;
        Bool InitCapabilities() override;
        Bool InitializeEGLDisplay(EGLDisplay dpy, EGLint* major, EGLint* minor) override;
@@ -34,8 +39,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        Bool CreateEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height) override;
        Bool MakeEGLCurrent(EGLDisplay dpy, EGLSurface draw, EGLSurface read, EGLContext ctx) override;
        Bool SwapEGLBuffers(EGLDisplay dpy, EGLSurface draw) override;
+        void SetEGLSwapInterval(Int interval) override;
+        Bool BlitDefaultFramebufferToSharedImage(const SharedImageView& image,
+                                                 const MG_Util::Damage::Region& region) override;
+        // The active surface target's swapchain image ages (VulkanRenderer::CurrentDrawBufferAge).
+        Int32 QueryCurrentBufferAge(Bool damageRegionFollows) override;
+        // A glFlush that ends a shared-image producer's frame (VulkanRenderer::PublishSharedImageAccesses).
+        Bool PublishSharedImageAccesses() override;
+        // An EGL native fence of the session's work (VulkanRenderer::ExportNativeFence).
+        Bool ExportNativeFence(int* fence) override;
        void ReleaseEGLSurface(EGLSurface surface) override;
        void ReleaseEGLResources() override;
+        // The server window going and coming back: the window surface's target (VkSurfaceKHR,
+        // swapchain) is destroyed and a pbuffer target of the same extent stands in for it, and back.
+        Bool SuspendServerWindow(void* window) override;
+        Bool ResumeServerWindow(void* window, Uint32 width, Uint32 height) override;
+        Bool ServerWindowResized(void* window, Uint32 width, Uint32 height) override;

        const RendererInfo& GetRendererInfo() const override;
        String GetBackendAPIVersionString() const override;
@@ -47,6 +66,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    private:
        Bool InitPbufferSurface(EGLint width, EGLint height) override;
        void OnEGLSurfaceReleased(EGLSurface surface) override;
+        void OnEGLSurfaceForgotten(EGLSurface surface) override;
+        // A served session keeps one renderer across all its client's surfaces and switches the
+        // renderer's surface target instead (VulkanRenderer::ActivateSurfaceTarget).
+        Bool UsesSurfaceTargets() const;
+        static Uint64 SurfaceTargetKey(EGLSurface surface) { return reinterpret_cast<Uint64>(surface); }
        void UpdateAdvertisedExtensions();
        void UpdateDynamicBackendParameters();

@@ -54,6 +78,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        DynamicBackendParameters m_dynamicParameters;
        MG_External::VulkanCapabilities m_vulkanCaps;
        RendererInfo m_rendererInfo;
+        // A served session's own copy of Magma's per-process state (MagmaSession.h), bound to its
+        // apply thread. Null in the monolith and every other single-session shape.
+        UniquePtr<MagmaSession> m_magmaSession;
+        // The binding each client context was last made current with, for OnClientContextBound.
+        UnorderedMap<EGLContext, EGLCurrentState> m_contextBindings;
+        // Window surfaces whose window is gone (SuspendServerWindow), with the extent their
+        // placeholder pbuffer target is built at - the window's, so the default framebuffer the
+        // client knows keeps its size.
+        struct SuspendedWindowSurface {
+            Uint32 Width = 1;
+            Uint32 Height = 1;
+        };
+        UnorderedMap<EGLSurface, SuspendedWindowSurface> m_suspendedWindowSurfaces;
+        // The placeholder pbuffer target of a suspended surface, made active. False: it could not be built.
+        Bool ActivateSuspendedPlaceholder(EGLSurface surface, const SuspendedWindowSurface& extent);
    };

    // Single-source-of-truth helpers shared with the driver POST
@@ -0,0 +1,29 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/DescriptorPoolCensus.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+
+#include <cstdint>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+    // A count of Magma's per-frame descriptor pools and the sets they hold (UniformManager), for
+    // tests. Dependency-free on purpose: the integration scenarios include it next to the system GL
+    // headers. They link the static library on desktop; the shipping shared library is built with
+    // hidden visibility, so on Android nothing outside the library can call GetDescriptorPoolCensus.
+    struct DescriptorPoolCensus {
+        bool available = false;         // false when no Magma renderer is live (e.g. the DirectGLES backend)
+        uint32_t frameSlots = 0;
+        uint32_t pools = 0;             // descriptor pools across every frame slot
+        uint32_t maxPoolsInOneSlot = 0;
+        uint64_t capacitySets = 0;      // sum of maxSets over those pools
+        uint64_t cachedSets = 0;        // sets allocated from them and held in the per-layout caches
+        uint64_t slotTrims = 0;         // whole-slot trims since the renderer came up
+    };
+
+    DescriptorPoolCensus GetDescriptorPoolCensus();
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -35,8 +35,9 @@
 #include <spirv_reflect.h>

 namespace MobileGL::MG_Backend::DirectVulkan {
-    // Leak-at-exit storage; see GlobalObjects.cpp.
-    UniquePtr<VulkanRenderer>& pVulkanRenderer = *new UniquePtr<VulkanRenderer>();
+    // Per Magma session (MagmaSession.h); the process-wide session is leak-at-exit storage, see
+    // GlobalObjects.cpp.
+    RendererSlot pVulkanRenderer;

    namespace {
 #if MOBILEGL_BUILD_DISAGGREGATED
@@ -59,7 +60,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // current one (a new renderer restarts its frame-serial counter and
        // reuses pool indices). Atomic because handles may be polled from a
        // thread other than the EGL thread that recreates the renderer.
-        std::atomic<Uint64> g_rendererGeneration{1};
+        //
+        // One per Magma session, drawn from one process-wide source: a handle stamped by one
+        // session's renderer must read as stale against every other session's, never equal.
+        std::atomic<Uint64> g_rendererGenerationSource{1};
+        struct RendererGeneration {
+            std::atomic<Uint64> value{g_rendererGenerationSource.fetch_add(1, std::memory_order_acq_rel)};
+        };
+        SessionLocal<RendererGeneration> g_rendererGeneration;
    } // namespace

 #if MOBILEGL_BUILD_DISAGGREGATED && MOBILEGL_PIPE_PUSH
@@ -123,11 +131,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #endif // MOBILEGL_BUILD_DISAGGREGATED && MOBILEGL_PIPE_PUSH

    Uint64 GetRendererGeneration() {
-        return g_rendererGeneration.load(std::memory_order_acquire);
+        return g_rendererGeneration->value.load(std::memory_order_acquire);
    }

    void BumpRendererGeneration() {
-        g_rendererGeneration.fetch_add(1, std::memory_order_acq_rel);
+        g_rendererGeneration->value.store(g_rendererGenerationSource.fetch_add(1, std::memory_order_acq_rel),
+                                          std::memory_order_release);
    }

    namespace {
@@ -370,16 +379,36 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #endif // !MOBILEGL_BUILD_DISAGGREGATED

 #if MOBILEGL_BUILD_DISAGGREGATED
-        // The verb's handles identify server stores. Readback orders GPU-produced
-        // command/count bytes before CPU tier expansion; never inspect a client binding.
+        // The verb's handles identify server stores; never inspect a client binding.
+        //
+        // P8-D: THE GPU READS THE WORDS (VulkanRenderer::DrawWireIndirectNative, WireDraw.inc), as
+        // on the monolith arm. The CPU expansion below is kept for the one shape the monolith arm
+        // also expands - a COUNT form the device cannot issue natively (no VK_KHR_draw_indirect_count,
+        // or maxdrawcount > 1 without multiDrawIndirect: VulkanRenderer.cpp's
+        // MultiDrawElementsIndirectCount) - where ReadWireBuffer orders GPU-produced command/count
+        // bytes before the expansion.
        void DrawWireIndirect(GLenum mode, GLenum type, const void* indirect, GLsizei drawcount,
                              GLsizei stride, Bool indexed, Bool counted = false, GLintptr countOffset = 0) {
-            if (drawcount <= 0) return;
+            if (drawcount <= 0 || countOffset < 0) return;
+            if (pVulkanRenderer->DrawWireIndirectNative(mode, type, reinterpret_cast<Uint64>(indirect), drawcount,
+                    stride, indexed, counted, static_cast<Uint64>(countOffset)))
+                return;
            auto& buffers = pVulkanRenderer->GetWireBufferManager();
            const auto& state = MG_Pipe::MGPipeApplier();
+            // P8-D: this call reads its words on the CPU; the waits those reads take are its own.
+            struct ExpansionCount {
+                VkBufferManager& buffers;
+                const Uint64 waitsBefore = buffers.GetWireHostWaitCount();
+                ~ExpansionCount() {
+                    if (!MG_Util::PipeStats::Enabled()) return;
+                    MG_Util::PipeStats::AddCalls(MG_Util::PipeStats::CallClass::WireIndirectCpuExpansions, 1);
+                    MG_Util::PipeStats::AddCalls(MG_Util::PipeStats::CallClass::WireHostWaitsIndirect,
+                                                 buffers.GetWireHostWaitCount() - waitsBefore);
+                }
+            } expansionCount{buffers};
            if (counted) {
                Uint32 count = 0;
-                if (countOffset < 0 || !buffers.ReadWireBuffer(state.VerbIndirectParameterBuffer,
+                if (!buffers.ReadWireBuffer(state.VerbIndirectParameterBuffer,
                        static_cast<Uint64>(countOffset), sizeof(count), &count)) return;
                drawcount = static_cast<GLsizei>(std::min<Uint32>(count, static_cast<Uint32>(drawcount)));
                if (!drawcount) return;
@@ -917,12 +946,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {

    void DispatchComputeIndirect(GLintptr indirect) {
 #if MOBILEGL_BUILD_DISAGGREGATED
+        // P8-SV: THE GPU READS THE GROUP COUNTS on the wire arm too - VulkanRenderer's
+        // DispatchComputeIndirect takes its transport branch into DispatchWireComputeIndirect
+        // (WireDraw.inc), which issues vkCmdDispatchIndirect from the verb's store. The CPU read
+        // that was here (ReadWireBuffer, a whole-GPU wait once a shader had written the store) is gone.
        if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
-            Uint32 groups[3]{};
-            if (indirect >= 0 && pVulkanRenderer->GetWireBufferManager().ReadWireBuffer(
-                    MG_Pipe::MGPipeApplier().VerbDispatchIndirectBuffer,
-                    static_cast<Uint64>(indirect), sizeof(groups), groups))
-                pVulkanRenderer->DispatchCompute(groups[0], groups[1], groups[2]);
+            pVulkanRenderer->DispatchComputeIndirect(indirect);
            return;
        }
 #endif
@@ -1254,7 +1283,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        payload.mode = mode;

        // TODO: allocate draw cmd buf elsewhere
-        static Vector<DrawCmdParam> params;
+        static thread_local Vector<DrawCmdParam> params;
        params.clear();
        params.resize(drawcount);

@@ -1327,7 +1356,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        const Uint32 indexSizeShift = static_cast<Uint32>(std::countr_zero(indexSize));

        // TODO: allocate draw cmd buf elsewhere
-        static Vector<DrawIndexedCmdParam> params;
+        static thread_local Vector<DrawIndexedCmdParam> params;
        params.clear();
        params.resize(drawcount);

@@ -1731,7 +1760,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    void Present() {
-        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::Present called with null VulkanRenderer");
+        // A served session's renderer is gone once its client released its resources; a present
+        // that still reaches it has nothing to show and is dropped rather than dereferenced.
+        if (!pVulkanRenderer) {
+            MGLOG_E_ONCE("DirectVulkan: a present reached a session whose renderer is already released; dropped");
+            return;
+        }
        pVulkanRenderer->Present();
        // THE frame boundary for the MGPipe counters, at the backend entry point rather
        // than inside VulkanRenderer::Present: that function has an early return for the
@@ -1741,4 +1775,53 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            MG_Util::PipeStats::OnPresent();
        }
    }
+
+    namespace {
+        // Per session: each client asks for its own interval.
+        SessionLocal<Optional<Int>> g_requestedSwapInterval;
+    } // namespace
+
+    void SetSwapInterval(Int interval) {
+        *g_requestedSwapInterval = interval;
+        if (pVulkanRenderer) {
+            pVulkanRenderer->SetSwapInterval(interval);
+        }
+    }
+
+    Optional<Int> GetRequestedSwapInterval() {
+        return *g_requestedSwapInterval;
+    }
+
+#if MOBILEGL_BUILD_DISAGGREGATED
+    void ForgetRequestedSwapInterval() {
+        g_requestedSwapInterval->reset();
+    }
+#endif
+
+    DescriptorPoolCensus GetDescriptorPoolCensus() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // P11 M2: UNDER A TRANSPORT THE POOLS ARE THE SERVER RENDERER'S, and its apply thread
+        // grows, rewinds and trims them while the test thread would be reading - a census taken
+        // after a swap the client ran ahead of, or mid-burst, walks a set cache BeginFrame may be
+        // clearing. Inproc, the server is a thread of this process: the read is posted to it as
+        // the test probe, which runs between drain batches, so it sees every record published
+        // before it and none half-applied - without waiting on the GPU, which is the point of the
+        // trim case (it gives a burst back with other slots' work in flight). Spawn and tcp keep
+        // their pools in another process; the client has no renderer and answers unavailable.
+        if (MG_Config::Transport != MG_Config::TransportMode::Monolith &&
+            !MG_Remote::Server::ServerLoop::OnApplyThread()) {
+            if (MG_Config::Transport != MG_Config::TransportMode::InProcess) return DescriptorPoolCensus{};
+            DescriptorPoolCensus census{};
+            const MobileGLResult result = MG_Remote::Server::ServerLoopInstance().RunProbeOnApplyThreadForTesting(
+                +[](void* user) -> MobileGLResult {
+                    *static_cast<DescriptorPoolCensus*>(user) =
+                        pVulkanRenderer ? pVulkanRenderer->GetDescriptorPoolCensus() : DescriptorPoolCensus{};
+                    return MOBILEGL_OK;
+                },
+                &census);
+            return result == MOBILEGL_OK ? census : DescriptorPoolCensus{};
+        }
+#endif
+        return pVulkanRenderer ? pVulkanRenderer->GetDescriptorPoolCensus() : DescriptorPoolCensus{};
+    }
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -9,10 +9,32 @@
 #pragma once
 #include <Includes.h>
 #include <MG_Backend/BackendObject.h>
+#include "MagmaSession.h"
 #include "Renderer/VulkanRenderer.h"

 namespace MobileGL::MG_Backend::DirectVulkan {
-    extern UniquePtr<VulkanRenderer>& pVulkanRenderer;
+    // The renderer of the calling thread's Magma session (MagmaSession.h): a served session's own on
+    // its apply thread, the process-wide one everywhere else. Spelled like the UniquePtr it used to
+    // be so the call sites read the same.
+    class RendererSlot {
+    public:
+        UniquePtr<VulkanRenderer>& Slot() const { return m_slot.Get(); }
+        VulkanRenderer* get() const { return Slot().get(); }
+        VulkanRenderer* operator->() const { return get(); }
+        VulkanRenderer& operator*() const { return *get(); }
+        explicit operator bool() const { return get() != nullptr; }
+        RendererSlot& operator=(UniquePtr<VulkanRenderer>&& renderer) {
+            Slot() = std::move(renderer);
+            return *this;
+        }
+        void reset() { Slot().reset(); }
+        friend Bool operator==(const RendererSlot& slot, std::nullptr_t) { return slot.get() == nullptr; }
+        friend Bool operator!=(const RendererSlot& slot, std::nullptr_t) { return slot.get() != nullptr; }
+
+    private:
+        SessionLocal<UniquePtr<VulkanRenderer>> m_slot;
+    };
+    extern RendererSlot pVulkanRenderer;

    // Generation of the live VulkanRenderer instance, mirroring DirectGLES's
    // g_syncContextGeneration. BackendObject_DirectVulkan bumps it wherever
@@ -147,4 +169,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    // only exist as vkCmdWriteTimestamp results); the frontend falls back.
    Int64 GetGpuTimestampNs();
    void Present();
+    // eglSwapInterval. Remembered across renderers: one created for a later window surface
+    // starts from the app's last request (GetRequestedSwapInterval) instead of the default.
+    void SetSwapInterval(Int interval);
+    Optional<Int> GetRequestedSwapInterval();
+#if MOBILEGL_BUILD_DISAGGREGATED
+    // The swap interval an ended server session asked for is not the next one's, the same rule as
+    // DirectGLES::ForgetRequestedSwapInterval: under a transport ~BackendObject_DirectVulkan forgets
+    // it, so the next session starts at the default until it asks.
+    void ForgetRequestedSwapInterval();
+#endif
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,87 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/MagmaSession.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "MagmaSession.h"
+
+#include <atomic>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+    namespace {
+        struct SessionLocalEntry {
+            Detail::SessionLocalConstruct construct;
+            Detail::SessionLocalDestroy destroy;
+        };
+
+        // Construct-on-first-use: SessionLocals are namespace-scope objects in several translation
+        // units and register during static initialization, in no particular order.
+        Vector<SessionLocalEntry>& Registry() {
+            static auto* registry = new Vector<SessionLocalEntry>();
+            return *registry;
+        }
+
+        std::atomic<Uint64> g_nextSessionId{1};
+
+        thread_local MagmaSession* t_boundSession = nullptr;
+
+        // Leak-at-exit, like the renderer it holds (GlobalObjects.cpp): nothing may tear the
+        // monolith's renderer down from a static destructor after the driver is gone.
+        MagmaSession& ProcessSession() {
+            static auto* session = new MagmaSession();
+            return *session;
+        }
+    } // namespace
+
+    Uint32 Detail::RegisterSessionLocal(SessionLocalConstruct construct, SessionLocalDestroy destroy) {
+        auto& registry = Registry();
+        registry.push_back({construct, destroy});
+        return static_cast<Uint32>(registry.size() - 1);
+    }
+
+    MagmaSession& Detail::CurrentMagmaSession() {
+        MagmaSession* bound = t_boundSession;
+        return bound != nullptr ? *bound : ProcessSession();
+    }
+
+    void* Detail::SessionLocalSlot(MagmaSession& session, Uint32 index) {
+        if (index >= session.m_slots.size()) session.m_slots.resize(Registry().size(), nullptr);
+        void*& slot = session.m_slots[index];
+        if (slot == nullptr) slot = Registry()[index].construct();
+        return slot;
+    }
+
+    MagmaSession::MagmaSession() : m_id(g_nextSessionId.fetch_add(1, std::memory_order_relaxed)) {
+        // Every slot is built up front: the process-wide session is reached from whichever thread
+        // the app renders on, and a lazily built slot would be a first-touch race between two.
+        const auto& registry = Registry();
+        m_slots.resize(registry.size(), nullptr);
+        for (SizeT i = 0; i < registry.size(); ++i) m_slots[i] = registry[i].construct();
+    }
+
+    MagmaSession::~MagmaSession() {
+        MagmaSession* const previous = t_boundSession;
+        t_boundSession = this;
+        const auto& registry = Registry();
+        for (SizeT i = m_slots.size(); i-- > 0;) {
+            if (m_slots[i] != nullptr) registry[i].destroy(m_slots[i]);
+            m_slots[i] = nullptr;
+        }
+        // Never back to this session: it is gone.
+        t_boundSession = previous == this ? nullptr : previous;
+    }
+
+    MagmaSessionScope::MagmaSessionScope(MagmaSession* session) : m_previous(t_boundSession) {
+        t_boundSession = session;
+    }
+
+    MagmaSessionScope::~MagmaSessionScope() { t_boundSession = m_previous; }
+
+    void BindMagmaSessionToThisThread(MagmaSession* session) { t_boundSession = session; }
+
+    MagmaSession* BoundMagmaSession() { return t_boundSession; }
+
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,100 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/MagmaSession.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// MAGMA'S PER-SESSION STATE. Magma was written for one renderer per process: the renderer slot,
+// the active buffer manager, the dynamic-state shadow, the render-pass manager's statics and the
+// renderer generation are all file-scope. A multi-session server (several clients served from one
+// process, each on its own apply thread with its own BackendObject_DirectVulkan) needs one copy of
+// each per session, or one session's work lands in another's tables.
+//
+// SessionLocal<T> is that copy. It resolves through the CALLING THREAD's bound MagmaSession, and to
+// the process-wide default session on every thread nothing bound - which is the monolith, the client
+// role and every unit case, i.e. exactly the old behaviour, on whatever thread the app renders from.
+// A served session's backend owns its MagmaSession and binds it on its apply thread (the only thread
+// that runs that session's backend), so the two never mix.
+
+#pragma once
+#include <Includes.h>
+
+#include <cstddef>
+#include <new>
+#include <utility>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+
+    class MagmaSession;
+
+    namespace Detail {
+        using SessionLocalConstruct = void* (*)();
+        using SessionLocalDestroy = void (*)(void*);
+        // Registers one SessionLocal<T> and returns its slot. Called from static initialization.
+        Uint32 RegisterSessionLocal(SessionLocalConstruct construct, SessionLocalDestroy destroy);
+        // The calling thread's session, or the process-wide default.
+        MagmaSession& CurrentMagmaSession();
+        void* SessionLocalSlot(MagmaSession& session, Uint32 index);
+    } // namespace Detail
+
+    // One served session's copy of every SessionLocal. Slots are built lazily on first use and
+    // destroyed in reverse registration order with the session still bound, so a destructor that
+    // reaches another SessionLocal finds this session's copy and not the binding thread's.
+    class MagmaSession {
+    public:
+        MagmaSession();
+        ~MagmaSession();
+        MagmaSession(const MagmaSession&) = delete;
+        MagmaSession& operator=(const MagmaSession&) = delete;
+
+        // A monotonic id, unique for the process lifetime. Never an address: a later session can
+        // be allocated where an ended one lived.
+        Uint64 Id() const { return m_id; }
+
+    private:
+        friend void* Detail::SessionLocalSlot(MagmaSession& session, Uint32 index);
+        Vector<void*> m_slots;
+        Uint64 m_id = 0;
+    };
+
+    // Makes `session` the calling thread's Magma session for the scope's lifetime, restoring the
+    // previous binding (usually none) afterwards.
+    class MagmaSessionScope {
+    public:
+        explicit MagmaSessionScope(MagmaSession* session);
+        ~MagmaSessionScope();
+        MagmaSessionScope(const MagmaSessionScope&) = delete;
+        MagmaSessionScope& operator=(const MagmaSessionScope&) = delete;
+
+    private:
+        MagmaSession* m_previous = nullptr;
+    };
+
+    // Binds `session` to the calling thread until the next call (nullptr unbinds). For the apply
+    // thread, which binds its backend once and runs it for the rest of its life.
+    void BindMagmaSessionToThisThread(MagmaSession* session);
+    // The calling thread's bound session, or nullptr (the process-wide default answers).
+    MagmaSession* BoundMagmaSession();
+
+    template <class T>
+    class SessionLocal {
+    public:
+        SessionLocal() : m_index(Detail::RegisterSessionLocal(&Construct, &Destroy)) {}
+        SessionLocal(const SessionLocal&) = delete;
+        SessionLocal& operator=(const SessionLocal&) = delete;
+
+        T& Get() const {
+            return *static_cast<T*>(Detail::SessionLocalSlot(Detail::CurrentMagmaSession(), m_index));
+        }
+        T& operator*() const { return Get(); }
+        T* operator->() const { return &Get(); }
+
+    private:
+        static void* Construct() { return new T(); }
+        static void Destroy(void* object) { delete static_cast<T*>(object); }
+        const Uint32 m_index;
+    };
+
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -66,7 +66,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                FreeRetiredCommandBuffers(frame);
            }
            vkFreeCommandBuffers(device, commandPool, frameCount * 2, commandBuffers.data());
+            if (!m_spareCommandBuffers.empty())
+                vkFreeCommandBuffers(device, commandPool, static_cast<Uint32>(m_spareCommandBuffers.size()),
+                                     m_spareCommandBuffers.data());
        }
+        m_spareCommandBuffers.clear();
        m_frames.clear();
        currentFrameIndex = 0;
        m_device = VK_NULL_HANDLE;
@@ -300,6 +304,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return packet;
    }

+    VkResult FrameContext::AcquireNextImageInCurrentSlot(VkDevice device, VkSwapchainKHR swapchain,
+                                                         Uint32& outImageIndex) {
+        auto& frame = GetCurrent();
+        const VkResult result = vkAcquireNextImageKHR(device, swapchain, UINT64_MAX, frame.imageAvailableSemaphore,
+                                                      VK_NULL_HANDLE, &outImageIndex);
+        if (result == VK_SUCCESS || result == VK_SUBOPTIMAL_KHR) frame.imageAvailableSemaphoreConsumed = false;
+        return result;
+    }
+
    VkResult FrameContext::WaitAndAcquireNextImage(VkDevice device, VkSwapchainKHR swapchain, Uint32& outImageIndex,
                                                   Uint64 timeout, VkFence acquireFence) {
        auto& frame = GetCurrent();
@@ -353,21 +366,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MOBILEGL_ASSERT(!frame.isPreCommandRecording,
                        "RetireCurrentCommandBuffer called while the pre-pass stream is still recording");

-        VkCommandBufferAllocateInfo allocInfo{};
-        allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
-        allocInfo.commandPool = m_commandPool;
-        allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
-        allocInfo.commandBufferCount = 1;
        VkCommandBuffer replacement = VK_NULL_HANDLE;
-        VkResult result = vkAllocateCommandBuffers(m_device, &allocInfo, &replacement);
+        VkResult result = TakeCommandBuffer(replacement);
        if (result != VK_SUCCESS) {
            return result;
        }
        if (retirePreCommandBuffer) {
            VkCommandBuffer preReplacement = VK_NULL_HANDLE;
-            result = vkAllocateCommandBuffers(m_device, &allocInfo, &preReplacement);
+            result = TakeCommandBuffer(preReplacement);
            if (result != VK_SUCCESS) {
-                vkFreeCommandBuffers(m_device, m_commandPool, 1, &replacement);
+                RecycleCommandBuffer(replacement);
                return result;
            }
            frame.retiredCommandBuffers.push_back({frame.preCommandBuffer, frame.lastSubmitIndex});
@@ -402,8 +410,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            SizeT completedCount = 0;
            while (completedCount < frame.retiredCommandBuffers.size() &&
                   frame.retiredCommandBuffers[completedCount].submitIndex <= completedSubmitIndex) {
-                vkFreeCommandBuffers(m_device, m_commandPool, 1,
-                                     &frame.retiredCommandBuffers[completedCount].commandBuffer);
+                RecycleCommandBuffer(frame.retiredCommandBuffers[completedCount].commandBuffer);
                ++completedCount;
            }
            if (completedCount > 0) {
@@ -413,6 +420,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }
    }

+    VkResult FrameContext::TakeCommandBuffer(VkCommandBuffer& out) {
+        if (!m_spareCommandBuffers.empty()) {
+            out = m_spareCommandBuffers.back();
+            m_spareCommandBuffers.pop_back();
+            return VK_SUCCESS;
+        }
+        VkCommandBufferAllocateInfo allocInfo{};
+        allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
+        allocInfo.commandPool = m_commandPool;
+        allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+        allocInfo.commandBufferCount = 1;
+        return vkAllocateCommandBuffers(m_device, &allocInfo, &out);
+    }
+
+    void FrameContext::RecycleCommandBuffer(VkCommandBuffer commandBuffer) {
+        // A handful covers the buffers in flight between two completions; beyond that, free.
+        constexpr SizeT kSpareCommandBuffers = 8;
+        if (commandBuffer == VK_NULL_HANDLE) return;
+        if (m_spareCommandBuffers.size() < kSpareCommandBuffers) {
+            m_spareCommandBuffers.push_back(commandBuffer);
+            return;
+        }
+        vkFreeCommandBuffers(m_device, m_commandPool, 1, &commandBuffer);
+    }
+
    void FrameContext::FreeAllRetiredCommandBuffers() {
        for (auto& frame : m_frames) {
            FreeRetiredCommandBuffers(frame);
@@ -101,6 +101,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 VkImageLayout presentLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR);
        SubmitInfoPacket GetSubmitInfo(Bool shouldSubmitCommandBuffer, Uint32 swapchainImageIndex) const;
        PresentInfoPacket GetPresentInfo(VkSwapchainKHR swapchain, Uint32 imageIndex) const;
+        // Acquires into the current slot's semaphore WITHOUT the slot-fence wait: for a surface
+        // brought up mid-frame, whose slot fence this frame has already waited and reset. The
+        // slot's semaphore must be unsignaled with no pending signal (consumed).
+        VkResult AcquireNextImageInCurrentSlot(VkDevice device, VkSwapchainKHR swapchain, Uint32& outImageIndex);
+        // The per-swapchain-image render-finished semaphores belong to one swapchain; a renderer
+        // that parks a surface parks them with it.
+        void ExchangeSwapchainSemaphores(Vector<VkSemaphore>& other) { std::swap(m_swapchainImageRenderFinishedSemaphores, other); }
        VkResult WaitAndAcquireNextImage(VkDevice device, VkSwapchainKHR swapchain, Uint32& outImageIndex,
                                         Uint64 timeout = UINT64_MAX, VkFence acquireFence = VK_NULL_HANDLE);

@@ -143,5 +150,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // Stored at Initialize for retired-command-buffer management.
        VkDevice m_device = VK_NULL_HANDLE;
        VkCommandPool m_commandPool = VK_NULL_HANDLE;
+        // Completed command buffers kept for the next retirement instead of being freed and
+        // reallocated: the pool resets a buffer implicitly at its next vkBeginCommandBuffer.
+        Vector<VkCommandBuffer> m_spareCommandBuffers;
+        VkResult TakeCommandBuffer(VkCommandBuffer& out);
+        void RecycleCommandBuffer(VkCommandBuffer commandBuffer);
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,202 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/Renderer/GpuProgressMarkers.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "GpuProgressMarkers.h"
+
+#include <MG_Util/Debug/Log.h>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+
+    namespace Watch = MG_Backend::GpuHangWatch;
+
+    void GpuProgressMarkers::Init(VkDevice device, Uint32 queueFamilyIndex) {
+        if (m_device != VK_NULL_HANDLE || device == VK_NULL_HANDLE) return;
+        if (Watch::BudgetMs() == 0) {
+            MGLOG_I("Magma: GPU hang watch off (MOBILEGL_GPU_HANG_BUDGET_MS=0)");
+            return;
+        }
+        VkCommandPoolCreateInfo poolInfo{VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO};
+        poolInfo.queueFamilyIndex = queueFamilyIndex;
+        if (vkCreateCommandPool(device, &poolInfo, nullptr, &m_pool) != VK_SUCCESS) {
+            MGLOG_W("Magma: GPU hang watch off - no command pool for its markers");
+            m_pool = VK_NULL_HANDLE;
+            return;
+        }
+        m_device = device;
+        {
+            const std::lock_guard<std::mutex> lock(m_mutex);
+            if (!AddSlot()) {
+                MGLOG_W("Magma: GPU hang watch off - the device cannot make its progress markers");
+                vkDestroyCommandPool(m_device, m_pool, nullptr);
+                m_pool = VK_NULL_HANDLE;
+                m_device = VK_NULL_HANDLE;
+                return;
+            }
+        }
+        m_watch = Watch::Register(this, &Sample, &OnHung);
+        MGLOG_I("Magma: GPU hang watch on - a submission the GPU started and has not finished after %u ms ends "
+                "its session",
+                Watch::BudgetMs());
+    }
+
+    void GpuProgressMarkers::Shutdown() {
+        // The watch first, outside m_mutex: Unregister waits out a sample in progress, which takes it.
+        Watch::Unregister(m_watch);
+        m_watch = 0;
+        if (m_device == VK_NULL_HANDLE) return;
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        for (Slot& slot : m_slots) {
+            if (slot.Started != VK_NULL_HANDLE) vkDestroyEvent(m_device, slot.Started, nullptr);
+            if (slot.Finished != VK_NULL_HANDLE) vkDestroyEvent(m_device, slot.Finished, nullptr);
+        }
+        m_slots.clear();
+        if (m_pool != VK_NULL_HANDLE) vkDestroyCommandPool(m_device, m_pool, nullptr); // frees the buffers
+        m_pool = VK_NULL_HANDLE;
+        m_device = VK_NULL_HANDLE;
+    }
+
+    Bool GpuProgressMarkers::AddSlot() {
+        if (m_slots.size() >= kMaxSlots) return false;
+        Slot slot;
+        const VkEventCreateInfo eventInfo{VK_STRUCTURE_TYPE_EVENT_CREATE_INFO};
+        VkCommandBuffer buffers[2] = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        VkCommandBufferAllocateInfo allocInfo{VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO};
+        allocInfo.commandPool = m_pool;
+        allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+        allocInfo.commandBufferCount = 2;
+        const auto fail = [&] {
+            if (slot.Started != VK_NULL_HANDLE) vkDestroyEvent(m_device, slot.Started, nullptr);
+            if (slot.Finished != VK_NULL_HANDLE) vkDestroyEvent(m_device, slot.Finished, nullptr);
+            if (buffers[0] != VK_NULL_HANDLE) vkFreeCommandBuffers(m_device, m_pool, 2, buffers);
+            return false;
+        };
+        if (vkCreateEvent(m_device, &eventInfo, nullptr, &slot.Started) != VK_SUCCESS ||
+            vkCreateEvent(m_device, &eventInfo, nullptr, &slot.Finished) != VK_SUCCESS ||
+            vkAllocateCommandBuffers(m_device, &allocInfo, buffers) != VK_SUCCESS)
+            return fail();
+        // Prerecorded once and resubmitted for every submission the slot brackets; SIMULTANEOUS_USE
+        // because the slot is reused once its `finished` event is set, which can be a moment before
+        // the submission that set it formally leaves the pending state.
+        VkCommandBufferBeginInfo beginInfo{VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO};
+        beginInfo.flags = VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT;
+        if (vkBeginCommandBuffer(buffers[0], &beginInfo) != VK_SUCCESS) return fail();
+        vkCmdSetEvent(buffers[0], slot.Started, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+        if (vkEndCommandBuffer(buffers[0]) != VK_SUCCESS) return fail();
+        if (vkBeginCommandBuffer(buffers[1], &beginInfo) != VK_SUCCESS) return fail();
+        vkCmdSetEvent(buffers[1], slot.Finished, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
+        if (vkEndCommandBuffer(buffers[1]) != VK_SUCCESS) return fail();
+        slot.Begin = buffers[0];
+        slot.End = buffers[1];
+        m_slots.push_back(slot);
+        return true;
+    }
+
+    GpuProgressMarkers::Bracket GpuProgressMarkers::Acquire() {
+        Bracket bracket;
+        if (m_device == VK_NULL_HANDLE) return bracket;
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        if (m_lost) return bracket;
+        Slot* free = nullptr;
+        for (Slot& slot : m_slots) {
+            if (slot.Claimed) continue;
+            // Retire what finished since the watch last looked: frames come faster than its polls.
+            if (slot.InFlight && vkGetEventStatus(m_device, slot.Finished) == VK_EVENT_SET) slot.InFlight = false;
+            if (!slot.InFlight && free == nullptr) free = &slot;
+        }
+        if (free == nullptr) {
+            if (!AddSlot()) return bracket;
+            free = &m_slots.back();
+        }
+        if (free->Dirty) {
+            if (vkResetEvent(m_device, free->Started) != VK_SUCCESS ||
+                vkResetEvent(m_device, free->Finished) != VK_SUCCESS)
+                return bracket;
+            free->Dirty = false;
+        }
+        free->Claimed = true;
+        bracket.Begin = free->Begin;
+        bracket.End = free->End;
+        bracket.Slot = static_cast<Uint32>(free - m_slots.data());
+        return bracket;
+    }
+
+    void GpuProgressMarkers::Submitted(const Bracket& bracket) {
+        if (!bracket.Valid()) return;
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        if (bracket.Slot >= m_slots.size()) return;
+        Slot& slot = m_slots[bracket.Slot];
+        slot.Claimed = false;
+        slot.InFlight = true;
+        slot.Dirty = true;
+        slot.Id = m_nextId++;
+    }
+
+    void GpuProgressMarkers::Abandon(const Bracket& bracket) {
+        if (!bracket.Valid()) return;
+        const std::lock_guard<std::mutex> lock(m_mutex);
+        if (bracket.Slot >= m_slots.size()) return;
+        // A failed vkQueueSubmit may still have consumed nothing or something: reset before reuse.
+        m_slots[bracket.Slot].Claimed = false;
+        m_slots[bracket.Slot].Dirty = true;
+    }
+
+    Bool GpuProgressMarkers::Sample(void* self, Vector<MG_Backend::GpuHangWatch::SubmissionProgress>& out) {
+        auto* markers = static_cast<GpuProgressMarkers*>(self);
+        const std::lock_guard<std::mutex> lock(markers->m_mutex);
+        if (markers->m_lost || markers->m_device == VK_NULL_HANDLE) return false;
+        for (Slot& slot : markers->m_slots) {
+            if (!slot.InFlight) continue;
+            const VkResult finished = vkGetEventStatus(markers->m_device, slot.Finished);
+            const VkResult started = vkGetEventStatus(markers->m_device, slot.Started);
+            if ((finished != VK_EVENT_SET && finished != VK_EVENT_RESET) ||
+                (started != VK_EVENT_SET && started != VK_EVENT_RESET)) {
+                // VK_ERROR_DEVICE_LOST: the driver already reset the device and says so itself.
+                markers->m_lost = true;
+                return false;
+            }
+            if (finished == VK_EVENT_SET) {
+                slot.InFlight = false;
+                continue;
+            }
+            out.push_back({slot.Id, started == VK_EVENT_SET, false});
+        }
+        return true;
+    }
+
+    void GpuProgressMarkers::OnHung(void* self, Uint64 submissionId, Uint64 runningMs) {
+        auto* markers = static_cast<GpuProgressMarkers*>(self);
+        markers->m_hungRunningMs.store(runningMs, std::memory_order_release);
+        markers->m_hung.store(true, std::memory_order_release);
+        MGLOG_E("Magma: submission %llu of this session's VkDevice has run %llu ms on the GPU and has not finished - "
+                "it is what stops the GPU for every process; the session ends at its next record and submits "
+                "nothing more (budget %u ms, MOBILEGL_GPU_HANG_BUDGET_MS)",
+                static_cast<unsigned long long>(submissionId), static_cast<unsigned long long>(runningMs),
+                MG_Backend::GpuHangWatch::BudgetMs());
+    }
+
+    VkResult GpuProgressMarkers::SubmitBracketed(VkQueue queue, VkSubmitInfo& info, VkFence fence,
+                                                 const Bracket& bracket) {
+        constexpr Uint32 kMaxOwn = 6;
+        if (!bracket.Valid() || info.commandBufferCount == 0 || info.commandBufferCount > kMaxOwn)
+            return vkQueueSubmit(queue, 1, &info, fence);
+        VkCommandBuffer buffers[kMaxOwn + 2];
+        Uint32 count = 0;
+        buffers[count++] = bracket.Begin;
+        for (Uint32 i = 0; i < info.commandBufferCount; ++i) buffers[count++] = info.pCommandBuffers[i];
+        buffers[count++] = bracket.End;
+        const VkCommandBuffer* own = info.pCommandBuffers;
+        const Uint32 ownCount = info.commandBufferCount;
+        info.pCommandBuffers = buffers;
+        info.commandBufferCount = count;
+        const VkResult result = vkQueueSubmit(queue, 1, &info, fence);
+        info.pCommandBuffers = own;
+        info.commandBufferCount = ownCount;
+        return result;
+    }
+
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,95 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/Renderer/GpuProgressMarkers.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+
+#include "../VkIncludes.h"
+#include <Includes.h>
+#include <MG_Backend/GpuHangWatch.h>
+
+#include <atomic>
+#include <mutex>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+
+    // WHICH OF THIS DEVICE'S SUBMISSIONS THE GPU HAS STARTED, AND WHICH IT HAS FINISHED
+    // (GpuHangWatch.h says why that is the whole attribution of a GPU hang).
+    //
+    // A frame's submission is bracketed by two prerecorded one-command buffers: the first sets a
+    // `started` event at TOP_OF_PIPE, the last sets a `finished` event once everything before it in
+    // the queue has completed. The host reads both with vkGetEventStatus from the watch thread - no
+    // fence is shared with the renderer's own waits and resets. On Adreno 750 a submission whose
+    // fragment shader never ends reads started=SET from its first poll until the kernel resets the
+    // GPU, and a submission queued behind it reads started=RESET for the whole stall.
+    //
+    // The renderer's thread brackets submissions (Acquire / Submitted / Abandon) and asks Hung(); the
+    // watch thread samples. Everything Vulkan the two share - the events - is touched under m_mutex.
+    class GpuProgressMarkers {
+    public:
+        struct Bracket {
+            VkCommandBuffer Begin = VK_NULL_HANDLE;
+            VkCommandBuffer End = VK_NULL_HANDLE;
+            Uint32 Slot = ~0u;
+            Bool Valid() const { return Slot != ~0u; }
+        };
+
+        GpuProgressMarkers() = default;
+        ~GpuProgressMarkers() { Shutdown(); }
+        GpuProgressMarkers(const GpuProgressMarkers&) = delete;
+        GpuProgressMarkers& operator=(const GpuProgressMarkers&) = delete;
+
+        // Off (no brackets, never hung) when the budget is 0 or the device cannot make events.
+        void Init(VkDevice device, Uint32 queueFamilyIndex);
+        // Before the device goes; its queue must be idle. Stops the watch first.
+        void Shutdown();
+
+        // The two buffers to put first and last in one submission (an invalid bracket when off or
+        // when every slot is in flight: the submission then goes unwatched).
+        Bracket Acquire();
+        // The submission carrying `bracket` was accepted by vkQueueSubmit / was not submitted.
+        void Submitted(const Bracket& bracket);
+        void Abandon(const Bracket& bracket);
+
+        // The watch named one of this device's submissions as hanging the GPU.
+        Bool Hung() const { return m_hung.load(std::memory_order_acquire); }
+        Uint64 HungRunningMs() const { return m_hungRunningMs.load(std::memory_order_acquire); }
+        Bool Active() const { return m_watch != 0; }
+
+        // vkQueueSubmit of `info` with `bracket` wrapped around its command buffers (which it puts
+        // back before returning). At most 6 buffers of the caller's.
+        static VkResult SubmitBracketed(VkQueue queue, VkSubmitInfo& info, VkFence fence, const Bracket& bracket);
+
+    private:
+        struct Slot {
+            VkEvent Started = VK_NULL_HANDLE;
+            VkEvent Finished = VK_NULL_HANDLE;
+            VkCommandBuffer Begin = VK_NULL_HANDLE;
+            VkCommandBuffer End = VK_NULL_HANDLE;
+            Uint64 Id = 0;
+            Bool InFlight = false; // submitted and not yet seen finished
+            Bool Claimed = false;  // handed out by Acquire, not yet submitted or abandoned
+            Bool Dirty = false;    // its events were set by a past submission: reset before reuse
+        };
+        static constexpr Uint32 kMaxSlots = 16;
+
+        Bool AddSlot();
+        static Bool Sample(void* self, Vector<MG_Backend::GpuHangWatch::SubmissionProgress>& out);
+        static void OnHung(void* self, Uint64 submissionId, Uint64 runningMs);
+
+        std::mutex m_mutex;
+        VkDevice m_device = VK_NULL_HANDLE;
+        VkCommandPool m_pool = VK_NULL_HANDLE;
+        Vector<Slot> m_slots;
+        Uint64 m_nextId = 1;
+        Uint64 m_watch = 0;
+        Bool m_lost = false;
+        std::atomic<Bool> m_hung{false};
+        std::atomic<Uint64> m_hungRunningMs{0};
+    };
+
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,37 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/Renderer/OffscreenSurfaceRoute.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// How an offscreen (pbuffer) surface target gets its VkSurfaceKHR on Android, decided from what the
+// INSTANCE has, not from what the loader could offer. A renderer created for a window enabled
+// VK_KHR_android_surface and nothing headless, so its later offscreen targets (a second client
+// surface, or the placeholder a suspended server window draws into) must take the AImageReader
+// route; asking that instance for vkCreateHeadlessSurfaceEXT answers null. Pure, so a host case pins
+// it without a driver.
+
+#pragma once
+#include <Includes.h>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+
+    enum class OffscreenSurfaceRoute : Uint8 {
+        Headless,           // vkCreateHeadlessSurfaceEXT
+        AndroidImageReader, // vkCreateAndroidSurfaceKHR on an AImageReader's window nobody displays
+        Unavailable,        // neither: the target cannot be built (the caller refuses, by name)
+    };
+
+    // `headlessEnabled`: the instance was created with VK_EXT_headless_surface.
+    // `headlessEntryPoint`: vkGetInstanceProcAddr answered vkCreateHeadlessSurfaceEXT.
+    // `androidSurfaceEnabled`: the instance was created with VK_KHR_android_surface.
+    constexpr OffscreenSurfaceRoute ChooseOffscreenSurfaceRoute(Bool headlessEnabled, Bool headlessEntryPoint,
+                                                                Bool androidSurfaceEnabled) {
+        if (headlessEnabled && headlessEntryPoint) return OffscreenSurfaceRoute::Headless;
+        if (androidSurfaceEnabled) return OffscreenSurfaceRoute::AndroidImageReader;
+        return OffscreenSurfaceRoute::Unavailable;
+    }
+
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -391,6 +391,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {

    VkPipeline PipelineFactory::CreatePipeline(const PipelineCreatePayload& payload) const {
        MOBILEGL_ASSERT(payload.stages != nullptr && !payload.stages->empty(), "PipelineFactory: stages are empty");
+        // The assert compiles out of shipping builds, and a program declined for an invalid
+        // module (VkProgramObject::invalidSpirvDeclined) has no stages at all: a caller that
+        // forgot to refuse it must get a skipped draw, not a stageCount=0 pipeline.
+        if (payload.stages == nullptr || payload.stages->empty()) {
+            MGLOG_E_ONCE("PipelineFactory::CreatePipeline: refusing a pipeline with no shader stages "
+                         "(programHash=0x%llx); its draws are skipped",
+                         static_cast<unsigned long long>(payload.programHash));
+            return VK_NULL_HANDLE;
+        }
        MOBILEGL_ASSERT(payload.vertexInputState != nullptr, "PipelineFactory: vertexInputState is null");
        MOBILEGL_ASSERT(payload.pipelineLayout != VK_NULL_HANDLE, "PipelineFactory: pipelineLayout is null");
        MOBILEGL_ASSERT(payload.renderPass != VK_NULL_HANDLE, "PipelineFactory: renderPass is null");
@@ -199,7 +199,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        UnorderedMap<HashType, PipelineCacheEntry> m_cache;
        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
        Uint64 m_frameCounter = 0;
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline thread_local XXH64_state_t* m_hashState = XXH64_createState();
        static inline Bool s_suppressBlendedDepthWrite = false;
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -12,7 +12,9 @@
 #include "MG_Util/ShaderTranspiler/ShaderCompiler.h"
 #include "MG_Util/ShaderTranspiler/SpvcSession.h"
 #include "MG_Util/ShaderTranspiler/Types.h"
+#include <Config.h>
 #include <algorithm>
+#include <atomic>
 #include <bit>
 #include <cmath>
 #include <cstring>
@@ -1313,6 +1315,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                    }
                });
                if (!ok) return false;
+                // An access chain into the builtin (gl_FragCoord.y is one) carries an Input
+                // pointer type. Based on the Private copy it has to carry a Private one, and so
+                // does every chain built on top of it - otherwise the module hands the driver a
+                // pointer whose storage class disagrees with its base, which is invalid SPIR-V
+                // and reads an undefined value. Plan every retype before touching anything, so a
+                // shape this pass does not model declines the whole rewrite.
+                Vector<Pair<Instruction*, Uint32>> retypes;
+                Vector<Instruction*> chains;
+                for (Instruction* user : users) {
+                    if (IsAccessChain(user->opcode())) chains.push_back(user);
+                }
+                while (!chains.empty()) {
+                    Instruction* chain = chains.back();
+                    chains.pop_back();
+                    const Uint32 privateType = PrivatePointerTypeFor(chain->type_id());
+                    if (privateType == 0) return false;
+                    retypes.emplace_back(chain, privateType);
+                    context()->get_def_use_mgr()->ForEachUser(chain, [&](Instruction* next) {
+                        if (IsAccessChain(next->opcode()) && next->NumInOperands() > 0 &&
+                            next->GetSingleWordInOperand(0) == chain->result_id()) {
+                            chains.push_back(next);
+                        }
+                    });
+                }
                for (Instruction* user : users) {
                    for (Uint32 i = 0; i < user->NumInOperands(); ++i) {
                        auto& operand = user->GetInOperand(i);
@@ -1323,9 +1349,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                    }
                    context()->AnalyzeUses(user);
                }
+                for (auto& [chain, privateType] : retypes) {
+                    chain->SetResultType(privateType);
+                    context()->AnalyzeUses(chain);
+                }
                return true;
            }

+            static Bool IsAccessChain(spv::Op opcode) {
+                return opcode == spv::Op::OpAccessChain || opcode == spv::Op::OpInBoundsAccessChain ||
+                       opcode == spv::Op::OpPtrAccessChain || opcode == spv::Op::OpInBoundsPtrAccessChain;
+            }
+
+            // The Private-storage twin of an Input pointer type, or 0 when `pointerTypeId` is not
+            // an Input pointer (then the chain is not the shape this pass rewrites).
+            Uint32 PrivatePointerTypeFor(Uint32 pointerTypeId) {
+                using namespace spvtools::opt;
+                Instruction* pointerType = context()->get_def_use_mgr()->GetDef(pointerTypeId);
+                if (!pointerType || pointerType->opcode() != spv::Op::OpTypePointer) return 0;
+                if (static_cast<spv::StorageClass>(pointerType->GetSingleWordInOperand(0)) !=
+                    spv::StorageClass::Input) {
+                    return 0;
+                }
+                return context()->get_type_mgr()->FindPointerToType(pointerType->GetSingleWordInOperand(1),
+                                                                     spv::StorageClass::Private);
+            }
+
            Uint32 m_framebufferHeight = 0;
        };

@@ -1861,35 +1910,38 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return spvtools::Optimizer::PassToken(MakeUnique<GlToVulkanPositionFixPass>(transformFlags));
        }

-        Bool TransformSpirvForFragCoordYFlip(const Vector<Uint>& input, Vector<Uint>& output,
-                                             Uint32 framebufferHeight) {
-            if (input.empty()) {
-                output.clear();
-                return true;
-            }
-            if (framebufferHeight == 0) {
-                output = input;
-                return true;
-            }
+    } // namespace

-            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
-            spvtools::OptimizerOptions options;
-            options.set_run_validator(false); // see TransformSpirvForExplicitLod0Sampling
-            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
-                                            const char* message) {
-                MGLOG_E("Vulkan: fragcoord y-flip pass: %s", message != nullptr ? message : "");
-            });
-            optimizer.RegisterPass(
-                spvtools::Optimizer::PassToken(MakeUnique<GlFragCoordYFlipPass>(framebufferHeight)));
-
-            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
-            if (!success) {
-                MGLOG_E("Vulkan: failed to run the gl_FragCoord y-flip pass; keeping the original module");
-                output = input;
-            }
-            return success;
+    Bool TransformSpirvForFragCoordYFlip(const Vector<Uint>& input, Vector<Uint>& output,
+                                         Uint32 framebufferHeight) {
+        if (input.empty()) {
+            output.clear();
+            return true;
+        }
+        if (framebufferHeight == 0) {
+            output = input;
+            return true;
        }

+        spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+        spvtools::OptimizerOptions options;
+        options.set_run_validator(false); // see TransformSpirvForExplicitLod0Sampling
+        optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                        const char* message) {
+            MGLOG_E("Vulkan: fragcoord y-flip pass: %s", message != nullptr ? message : "");
+        });
+        optimizer.RegisterPass(
+            spvtools::Optimizer::PassToken(MakeUnique<GlFragCoordYFlipPass>(framebufferHeight)));
+
+        const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+        if (!success) {
+            MGLOG_E("Vulkan: failed to run the gl_FragCoord y-flip pass; keeping the original module");
+            output = input;
+        }
+        return success;
+    }
+
+    namespace {
        Bool TransformSpirvForXfbCapture(const Vector<Uint>& input, Vector<Uint>& output,
                                         const MagmaProgramSource& program) {
            if (input.empty()) {
@@ -2223,8 +2275,84 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                return false;
            }
        }
+
+        // ScreenModuleForDriver's memory. Keyed by content, never by program: the same bytes get
+        // the same answer whichever program, variant or session carries them.
+        struct DriverModuleVerdictCache {
+            struct Entry {
+                Uint64 wordCount = 0;
+                Bool valid = true;
+                String reason;
+            };
+            // Generous for any one application's distinct stage modules; past it the whole table
+            // is dropped rather than tracked per entry - a re-validation is the only cost.
+            static constexpr SizeT kMaxEntries = 8192;
+            std::mutex mutex;
+            UnorderedMap<Uint64, Entry> entries;
+            std::atomic<Uint64> validatorRuns{0};
+        };
+
+        DriverModuleVerdictCache& GetDriverModuleVerdictCache() {
+            // Leaked on purpose: screened from session threads that may outlive static destruction.
+            static auto* cache = new DriverModuleVerdictCache();
+            return *cache;
+        }
    } // namespace

+    ProgramFactory::DriverModuleVerdict ProgramFactory::ScreenModuleForDriver(const Vector<Uint>& spirv,
+                                                                              Bool allowInvalid, String* outReason) {
+        if (outReason != nullptr) outReason->clear();
+        if (spirv.empty()) return DriverModuleVerdict::Valid;
+
+        auto& cache = GetDriverModuleVerdictCache();
+        const Uint64 key = XXH64(spirv.data(), spirv.size() * sizeof(Uint), 0);
+        Bool valid = true;
+        String reason;
+        Bool known = false;
+        {
+            const std::lock_guard<std::mutex> lock(cache.mutex);
+            const auto it = cache.entries.find(key);
+            if (it != cache.entries.end() && it->second.wordCount == spirv.size()) {
+                known = true;
+                valid = it->second.valid;
+                reason = it->second.reason;
+            }
+        }
+        if (!known) {
+            // Outside the lock: two threads screening the same new module both validate, and
+            // agree; nothing waits on another session's spirv-val.
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrepareSpirvValidation();
+            cache.validatorRuns.fetch_add(1, std::memory_order_relaxed);
+            spv_context context = spvContextCreate(GetSpirvTargetEnv(spirv));
+            spv_validator_options options = spvValidatorOptionsCreate();
+            spvValidatorOptionsSetFriendlyNames(options, true);
+            spv_const_binary_t binary = {spirv.data(), spirv.size()};
+            spv_diagnostic diagnostic = nullptr;
+            const spv_result_t result = spvValidateWithOptions(context, options, &binary, &diagnostic);
+            valid = result == SPV_SUCCESS;
+            if (!valid) {
+                reason = diagnostic != nullptr && diagnostic->error != nullptr ? diagnostic->error : "<no diagnostic>";
+                // The process-wide counter the validating test lanes assert on.
+                MG_Util::ShaderTranspiler::ShaderCompiler::NoteSpirvValidationFailure();
+            }
+            spvDiagnosticDestroy(diagnostic);
+            spvValidatorOptionsDestroy(options);
+            spvContextDestroy(context);
+
+            const std::lock_guard<std::mutex> lock(cache.mutex);
+            if (cache.entries.size() >= DriverModuleVerdictCache::kMaxEntries) cache.entries.clear();
+            cache.entries[key] = DriverModuleVerdictCache::Entry{spirv.size(), valid, reason};
+        }
+
+        if (valid) return DriverModuleVerdict::Valid;
+        if (outReason != nullptr) *outReason = Move(reason);
+        return allowInvalid ? DriverModuleVerdict::AdmittedInvalid : DriverModuleVerdict::DeclinedInvalid;
+    }
+
+    Uint64 ProgramFactory::DriverModuleValidatorRuns() {
+        return GetDriverModuleVerdictCache().validatorRuns.load(std::memory_order_relaxed);
+    }
+
    // A shader that assigns gl_FragDepth (SPIR-V DepthReplacing) supplies depth itself
    // instead of taking the pipeline's interpolated Z, so a driver that varies the vertex
    // position math between pipelines cannot desynchronize it; the blended depth-write
@@ -2401,6 +2529,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return SamplerNumericDomain::UnsignedInteger;
        case GL_SAMPLER_1D:
        case GL_SAMPLER_2D:
+        case GL_SAMPLER_EXTERNAL_OES:
        case GL_SAMPLER_3D:
        case GL_SAMPLER_CUBE:
        case GL_SAMPLER_2D_RECT:
@@ -2591,6 +2720,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        case GL_UNSIGNED_INT_IMAGE_CUBE_MAP_ARRAY:
            return TextureTarget::TextureCubeMapArray;
        case GL_SAMPLER_2D:
+        case GL_SAMPLER_EXTERNAL_OES: // a 2D texture below the frontend
        case GL_SAMPLER_2D_SHADOW:
        case GL_INT_SAMPLER_2D:
        case GL_UNSIGNED_INT_SAMPLER_2D:
@@ -3764,27 +3894,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        const Bool remapOk = RemapDescriptorBindingsForVulkan(moduleSpirvs, m_maxBindings, moduleSpirvs);
        MOBILEGL_ASSERT(remapOk, "ProgramFactory::GetOrCreateProgram: descriptor binding remap failed");

+        // Last look at the exact bytes the driver receives, in EVERY build rather than only in
+        // DEBUG or with MOBILEGL_ENABLE_SPIRV_VALIDATION armed, and for every stage BEFORE any
+        // module is created. By here the descriptor bindings have been remapped and the layout
+        // about to be reflected describes the remapped module, so there is no module left that
+        // is both valid and consistent with it to fall back to (the recovery that exists lives
+        // one step earlier, at the clip/XFB fixups). What is left is to keep the bytes away from
+        // the driver: a driver is not a validating entry point, and an invalid module it accepts
+        // can fault the GPU into device loss (a frontend translation once put a sampler inside
+        // the default uniform block, and Adreno page-faulted reading its "descriptor" out of
+        // uniform-buffer bytes). One invalid stage declines the whole program: no shader module
+        // of it is created, and every draw or dispatch of it is refused by name
+        // (VkProgramObject::invalidSpirvDeclined). MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV hands the
+        // bytes over anyway, for diagnosing a driver that copes.
        for (SizeT i = 0; i < moduleCount; ++i) {
+            String reason;
+            const DriverModuleVerdict verdict = ScreenModuleForDriver(
+                moduleSpirvs[i], MG_Config::Features.MagmaAllowInvalidSpirv, &reason);
+            if (verdict == DriverModuleVerdict::DeclinedInvalid) {
+                MGLOG_E_ONCE("ProgramFactory: declining program %u - its stage %d module fails SPIR-V validation (%s); "
+                             "no shader module is created and its draws are refused. "
+                             "MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV=1 hands such modules to the driver anyway.",
+                             program.GetExternalIndex(), static_cast<Int>(stages[i]), reason.c_str());
+                entry.invalidSpirvDeclined = true;
+                break;
+            }
+            if (verdict == DriverModuleVerdict::AdmittedInvalid) {
+                MGLOG_E_ONCE("ProgramFactory: handing vkCreateShaderModule an INVALID module for program %u stage %d "
+                             "because MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV is set (%s)",
+                             program.GetExternalIndex(), static_cast<Int>(stages[i]), reason.c_str());
+            }
+        }
+
+        for (SizeT i = 0; i < moduleCount && !entry.invalidSpirvDeclined; ++i) {
            auto& moduleSpv = moduleSpirvs[i];
            if (moduleSpv.empty()) continue;

-            // Last look at the exact bytes the driver receives, in EVERY build rather than only
-            // in DEBUG or with MOBILEGL_ENABLE_SPIRV_VALIDATION armed. This one only reports:
-            // by here the descriptor bindings have been remapped and the layout about to be
-            // reflected describes the remapped module, so there is no module left that is both
-            // valid and consistent with it to fall back to. The recovery lives one step earlier,
-            // at the clip/XFB fixups (see the revert there) - which is where a transform can
-            // introduce a reference to a delisted interface variable, the failure this whole
-            // guard exists for. Anything that reaches this line names itself in the log of a
-            // shipping build instead of dying anonymously inside the driver.
-            SpirvValidationFailure finalFailure{};
-            if (!ValidateTransformedSpirv(moduleSpv, stages[i], program.GetExternalIndex(), &finalFailure)) {
-                MGLOG_E_ONCE("ProgramFactory: handing vkCreateShaderModule an INVALID module for program %u stage %d - "
-                             "a backend transform after the clip/XFB fixups broke it (%s)",
-                             program.GetExternalIndex(), static_cast<Int>(stages[i]),
-                             finalFailure.message.c_str());
-            }
-
            // Does the stage the driver will treat as the last pre-rasterization one actually
            // carry Xfb? Asked of the FINAL bytes, so it answers for whatever the whole transform
            // chain produced - a rewound clip/XFB backstop, a capture pass that resolved no
@@ -9,6 +9,7 @@
 #pragma once

 #include "../VkIncludes.h"
+#include "../MagmaSession.h"
 #include "PipelineFactory.h"
 #include "MagmaProgramSource.h"
 #include "MG_State/GLState/ProgramState/ShaderObject.h"
@@ -228,6 +229,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            // the cache entry, so every later draw of the same program variant reads the same
            // answer instead of re-deciding it.
            Bool pointSizeCapabilityUnsupported = false;
+            // One of the program's final modules fails SPIR-V validation, so none of its modules
+            // was created and every draw or dispatch of it is refused by name (the ProgramInvalidSpirv /
+            // ComputeProgramInvalidSpirv wire declines) - see ProgramFactory::ScreenModuleForDriver.
+            // Sticky like the flag above: it lives on the cache entry, so the one log line is honest.
+            Bool invalidSpirvDeclined = false;
            Bool needsPassthroughTessControl = false;
            // ...and the pass-through this renderer can synthesize carries gl_Position and
            // nothing else, so it is only correct when the evaluation stage's inputs are
@@ -251,7 +257,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            // entry pointer re-stamps use through a const reference (StampProgramUse).
            mutable Uint64 lastUsedFrame = 0;

-            static inline VkDevice s_device = VK_NULL_HANDLE;
+            static inline SessionLocal<VkDevice> s_device;  // per Magma session (MagmaSession.h)

            VkProgramObject() = default;
            VkProgramObject(const VkProgramObject&) = delete;
@@ -288,6 +294,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
                hasStorageImages = other.hasStorageImages;
                declinedDescriptors = other.declinedDescriptors;
+                // The three refusal flags travel like every other field: a robin-hood insert swaps
+                // entries, and a flag left behind would refuse (or admit) the wrong program.
+                invalidSpirvDeclined = other.invalidSpirvDeclined;
+                xfbCaptureDeclined = other.xfbCaptureDeclined;
+                pointSizeCapabilityUnsupported = other.pointSizeCapabilityUnsupported;
                globalUboBinding = other.globalUboBinding;
                activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                vertexInputTypes = other.vertexInputTypes;
@@ -309,6 +320,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                other.pipelineLayout = VK_NULL_HANDLE;
                other.hasStorageImages = false;
                other.declinedDescriptors = false;
+                other.invalidSpirvDeclined = false;
+                other.xfbCaptureDeclined = false;
+                other.pointSizeCapabilityUnsupported = false;
                other.globalUboBinding = -1;
                other.activeVertexInputLocationMask = 0;
                other.activeFragmentOutputLocationMask = 0;
@@ -352,6 +366,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
                hasStorageImages = other.hasStorageImages;
                declinedDescriptors = other.declinedDescriptors;
+                // The three refusal flags travel like every other field: a robin-hood insert swaps
+                // entries, and a flag left behind would refuse (or admit) the wrong program.
+                invalidSpirvDeclined = other.invalidSpirvDeclined;
+                xfbCaptureDeclined = other.xfbCaptureDeclined;
+                pointSizeCapabilityUnsupported = other.pointSizeCapabilityUnsupported;
                globalUboBinding = other.globalUboBinding;
                activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                vertexInputTypes = other.vertexInputTypes;
@@ -373,6 +392,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                other.pipelineLayout = VK_NULL_HANDLE;
                other.hasStorageImages = false;
                other.declinedDescriptors = false;
+                other.invalidSpirvDeclined = false;
+                other.xfbCaptureDeclined = false;
+                other.pointSizeCapabilityUnsupported = false;
                other.globalUboBinding = -1;
                other.activeVertexInputLocationMask = 0;
                other.activeFragmentOutputLocationMask = 0;
@@ -395,18 +417,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        private:
            void Destroy() {
-                if (s_device != VK_NULL_HANDLE) {
+                if (s_device.Get() != VK_NULL_HANDLE) {
                    if (pipelineLayout != VK_NULL_HANDLE) {
-                        vkDestroyPipelineLayout(s_device, pipelineLayout, nullptr);
+                        vkDestroyPipelineLayout(s_device.Get(), pipelineLayout, nullptr);
                        pipelineLayout = VK_NULL_HANDLE;
                    }
                    if (descriptorSetLayout != VK_NULL_HANDLE) {
-                        vkDestroyDescriptorSetLayout(s_device, descriptorSetLayout, nullptr);
+                        vkDestroyDescriptorSetLayout(s_device.Get(), descriptorSetLayout, nullptr);
                        descriptorSetLayout = VK_NULL_HANDLE;
                    }
                    for (auto module : modules) {
                        if (module != VK_NULL_HANDLE) {
-                            vkDestroyShaderModule(s_device, module, nullptr);
+                            vkDestroyShaderModule(s_device.Get(), module, nullptr);
                        }
                    }
                }
@@ -463,7 +485,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
              m_enableSpirvValidation(enableSpirvValidation),
              m_updateAfterBindLimits(updateAfterBindLimits),
              m_subgroupPolicy(subgroupPolicy) {
-            VkProgramObject::s_device = device;
+            VkProgramObject::s_device.Get() = device;
        }
        // Destroys the pass-through tessellation control modules. Runs while the device is
        // still alive for the same reason ~VkProgramObject's does: this factory outlives
@@ -503,6 +525,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // VkRenderPassManager::OnPresent's sweep.
        void OnFrameBoundary();

+        // What the last look at a module's final bytes - the exact words vkCreateShaderModule
+        // would receive - decided.
+        enum class DriverModuleVerdict : Uint8 {
+            Valid,           // validates: handed to the driver
+            DeclinedInvalid, // fails validation: never handed over, the program's work is declined
+            AdmittedInvalid, // fails validation, handed over anyway (MOBILEGL_MAGMA_ALLOW_INVALID_SPIRV)
+        };
+        // Validates `spirv` for the Vulkan environment its version word names and turns the
+        // result into a verdict. A driver is not a validating entry point: Adreno page-faulted
+        // into device loss on a module that carried a sampler inside a uniform block, Mali
+        // SIGSEGVs inside pipeline creation. So an invalid module is declined unless
+        // `allowInvalid` (the escape hatch) says to hand it over regardless.
+        //
+        // The validator's answer is remembered per module content (64-bit hash plus word count),
+        // process-wide: a program variant, an evicted-then-rebuilt entry and two programs that
+        // share a stage all re-screen the same bytes, and only the first pays for spirv-val.
+        // `outReason` receives the validator's message for an invalid module.
+        static DriverModuleVerdict ScreenModuleForDriver(const Vector<Uint>& spirv, Bool allowInvalid,
+                                                         String* outReason = nullptr);
+        // How many times ScreenModuleForDriver actually ran the validator (cache misses), for
+        // the tests that pin the cache.
+        static Uint64 DriverModuleValidatorRuns();
+
        static VkShaderStageFlagBits ToVkStage(ShaderStage stage);
        static VkFormat ConvertSpirvImageFormatToVkFormat(SpvImageFormat format);
        static SamplerNumericDomain UniformTypeToSamplerNumericDomain(GLenum glType);
@@ -655,6 +700,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // and adds only a handful of values, so it does not move the cap in practice.
        static constexpr SizeT kMaxPassthroughTessControlStages = 64;
        UnorderedMap<Uint64, VkPipelineShaderStageCreateInfo> m_passthroughTessControlStages;
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline thread_local XXH64_state_t* m_hashState = XXH64_createState();
    };
+
+    // The FragCoordYFlip variant's module transform: every read of gl_FragCoord goes through a
+    // Private copy whose y is (framebufferHeight - y). Fails open (output = input, false). Exposed
+    // so a unit case can validate the bytes it hands the driver.
+    Bool TransformSpirvForFragCoordYFlip(const Vector<Uint>& input, Vector<Uint>& output, Uint32 framebufferHeight);
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -0,0 +1,27 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/Renderer/RenderPassGuard.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include "../VkIncludes.h"
+#include <Includes.h>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+    // Ends the render pass that is active on `commandBuffer`, if there is one. Called right before
+    // recording a command a render pass may not contain (a barrier, a copy, a blit, a clear of an
+    // image) into a command buffer that may hold an open pass: the wire draw path keeps its pass
+    // open across consecutive draws to the same attachments, so the resource work that falls
+    // between two draws must close it first. A pass on another command buffer is left alone.
+    void EndActiveRenderPassOn(VkCommandBuffer commandBuffer);
+
+    // A counter that moves whenever images may have been written since it was last read: a render
+    // pass ended (its attachments), resource work recorded outside a pass (copies, clears, blits,
+    // dispatches all close the pass first), or a draw with storage images. A sampled image needs
+    // the memory barrier that makes earlier writes visible only when this moved since the last one.
+    Uint64 WireImageWriteEpoch();
+    void BumpWireImageWriteEpoch();
+} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -8,6 +8,8 @@

 #include "SwapchainObject.h"

+#include <span>
+
 #include "MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h"
 #include "MG_State/GLState/TextureState/TextureObject2D.h"

@@ -140,12 +142,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    VkPresentModeKHR SwapchainObject::ChooseSwapchainPresentMode(
-        const Vector<VkPresentModeKHR>& availablePresentModes) {
-        for (auto desiredPresentMode : s_desiredPresentModes) {
-            for (const auto& presentMode : availablePresentModes) {
-                if (presentMode == desiredPresentMode) {
-                    return presentMode;
-                }
+        const Vector<VkPresentModeKHR>& availablePresentModes, Optional<Int> swapInterval) {
+        static constexpr VkPresentModeKHR s_unthrottledPresentModes[]{
+            VK_PRESENT_MODE_IMMEDIATE_KHR, VK_PRESENT_MODE_MAILBOX_KHR, VK_PRESENT_MODE_FIFO_KHR};
+        static constexpr VkPresentModeKHR s_vsyncPresentModes[]{VK_PRESENT_MODE_FIFO_KHR};
+        static constexpr VkPresentModeKHR s_adaptiveVsyncPresentModes[]{VK_PRESENT_MODE_FIFO_RELAXED_KHR,
+                                                                         VK_PRESENT_MODE_FIFO_KHR};
+        std::span<const VkPresentModeKHR> ranked = s_desiredPresentModes;
+        if (swapInterval) {
+            if (*swapInterval == 0) {
+                ranked = s_unthrottledPresentModes;
+            } else if (*swapInterval > 0) {
+                ranked = s_vsyncPresentModes;
+            } else {
+                ranked = s_adaptiveVsyncPresentModes;
+            }
+        }
+        for (auto desiredPresentMode : ranked) {
+            if (std::find(availablePresentModes.begin(), availablePresentModes.end(), desiredPresentMode) !=
+                availablePresentModes.end()) {
+                return desiredPresentMode;
            }
        }

@@ -153,9 +169,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return availablePresentModes[0];
    }

+    const char* SwapchainObject::GetPresentModeName(VkPresentModeKHR presentMode) {
+        return string_VkPresentModeKHR(presentMode);
+    }
+
    void SwapchainObject::Create(VkDevice device, VkPhysicalDevice physicalDevice, VkSurfaceKHR surface,
                                 Uint32 graphicsQueueFamily, Uint32 presentQueueFamily, Uint32 minImageCountHint,
-                                 VkExtent2D desiredExtent) {
+                                 VkExtent2D desiredExtent, Optional<Int> swapInterval) {
        const auto swapchainCapabilities = GetSwapchainCapabilities(physicalDevice, surface);
        MOBILEGL_ASSERT(swapchainCapabilities.IsComplete(),
                        "SwapchainObject::Create failed: incomplete swapchain capabilities");
@@ -174,8 +194,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            MGLOG_D("    %s", string_VkPresentModeKHR(pm));
        }

-        const auto presentMode = ChooseSwapchainPresentMode(swapchainCapabilities.presentModes);
+        const auto presentMode = ChooseSwapchainPresentMode(swapchainCapabilities.presentModes, swapInterval);
        MGLOG_I("Picked present mode: %s", string_VkPresentModeKHR(presentMode));
+        m_presentMode = presentMode;
+        m_supportedPresentModes = swapchainCapabilities.presentModes;

        const auto& swapchainCaps = swapchainCapabilities.capabilities;
        Uint32 targetImageCount = std::max<Uint32>(minImageCountHint, swapchainCaps.minImageCount);
@@ -270,6 +292,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // Fresh swapchain images hold garbage until a render pass stores into them.
        m_imageContentDefined.assign(imageCount, false);
        m_depthStencilContentDefined.assign(imageCount, false);
+        // Nothing presented from this chain yet: every buffer age is unknown.
+        m_ages.Reset(imageCount);

        CreateImageViews(device);
        CreateDepthStencilResources(device, physicalDevice);
@@ -277,6 +301,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MGLOG_I("Swapchain created, extent = %dx%d, swapchain imageCount = %d", m_extent.width, m_extent.height,
                imageCount);

+        PublishDefaultFramebufferInfo();
+    }
+
+    void SwapchainObject::PublishDefaultFramebufferInfo() const {
+        const VkExtent2D defaultFramebufferExtent = m_surfaceExtent;
        // Properly initialize Default FBO here
        auto& defaultFBOInfo = MG_Impl::GLImpl::FramebufferImpl::pDefaultFramebufferInfo;
        const Int extentWidth = static_cast<Int>(defaultFramebufferExtent.width);
@@ -466,6 +495,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        m_imageLayouts.clear();
        m_imageContentDefined.clear();
        m_depthStencilContentDefined.clear();
+        m_ages.Reset(0);
        m_preTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
    }

@@ -9,6 +9,7 @@
 #pragma once

 #include <Includes.h>
+#include <MG_Util/Damage/Damage.h>
 #include "../VkIncludes.h"

 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -26,13 +27,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        static SwapchainCapabilities GetSwapchainCapabilities(VkPhysicalDevice physicalDevice, VkSurfaceKHR surface);
        static VkSurfaceFormatKHR ChooseSwapchainSurfaceFormat(const Vector<VkSurfaceFormatKHR>& availableFormats);
-        static VkPresentModeKHR ChooseSwapchainPresentMode(const Vector<VkPresentModeKHR>& availablePresentModes);
+        // swapInterval is the app's eglSwapInterval, empty until it asks (s_desiredPresentModes
+        // ranks the modes then). 0 wants IMMEDIATE, else MAILBOX, else FIFO; >= 1 wants FIFO,
+        // which cannot skip vblanks, so 2 and up pace like 1; a negative interval (adaptive
+        // vsync) wants FIFO_RELAXED, else FIFO.
+        static VkPresentModeKHR ChooseSwapchainPresentMode(const Vector<VkPresentModeKHR>& availablePresentModes,
+                                                           Optional<Int> swapInterval = Nullopt);
+        static const char* GetPresentModeName(VkPresentModeKHR presentMode);

        void Create(VkDevice device, VkPhysicalDevice physicalDevice, VkSurfaceKHR surface, Uint32 graphicsQueueFamily,
-                    Uint32 presentQueueFamily, Uint32 minImageCountHint, VkExtent2D desiredExtent);
+                    Uint32 presentQueueFamily, Uint32 minImageCountHint, VkExtent2D desiredExtent,
+                    Optional<Int> swapInterval);
        void Shutdown(VkDevice device);
+        // Tells the frontend (or, under a transport, the client) the default framebuffer's size
+        // and formats. Create does this; a renderer switching between surfaces repeats it.
+        void PublishDefaultFramebufferInfo() const;

        VkSwapchainKHR GetHandle() const { return m_swapchain; }
+        VkPresentModeKHR GetPresentMode() const { return m_presentMode; }
+        // The surface's present modes as queried when this swapchain was created.
+        const Vector<VkPresentModeKHR>& GetSupportedPresentModes() const { return m_supportedPresentModes; }
        const VkSurfaceFormatKHR& GetSurfaceFormat() const { return m_surfaceFormat; }
        VkExtent2D GetExtent() const { return m_extent; }
        // Surface-space extent (before the pre-rotation quarter-turn swap) this swapchain was
@@ -67,6 +81,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        void SetDepthStencilContentDefined(Uint32 index, Bool defined);
        void SetAllDepthStencilContentUndefined();

+        // EGL_EXT_buffer_age over this chain's images. Ages start unknown with the chain (Create).
+        // Until a client asks for an age (NoteBufferAgeAsked) a presented image's content is
+        // treated as undefined, as before; once one has, a presented image keeps it - its age is
+        // counted and its content stays defined, so the next frame loads it.
+        void NoteBufferAgeAsked() { m_bufferAgeAsked = true; }
+        Bool BufferAgeAsked() const { return m_bufferAgeAsked; }
+        Int32 BufferAgeOf(Uint32 index) const { return m_ages.AgeOf(index); }
+        // Image `index` was presented.
+        void OnImagePresented(Uint32 index) { m_ages.OnPresented(index, m_bufferAgeAsked); }
+        // A swapped frame reached no image: every age is unknown.
+        void OnFrameDropped() { m_ages.OnFrameDropped(); }
+
    private:
        void CreateImageViews(VkDevice device);
        void CreateDepthStencilResources(VkDevice device, VkPhysicalDevice physicalDevice);
@@ -79,6 +105,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        };

        VkSwapchainKHR m_swapchain = VK_NULL_HANDLE;
+        VkPresentModeKHR m_presentMode = VK_PRESENT_MODE_FIFO_KHR;
+        Vector<VkPresentModeKHR> m_supportedPresentModes;
        VkSurfaceFormatKHR m_surfaceFormat{};
        VkExtent2D m_extent{};
        VkExtent2D m_surfaceExtent{};
@@ -94,5 +122,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        Vector<VkImageLayout> m_depthStencilImageLayouts;
        Vector<Bool> m_imageContentDefined;
        Vector<Bool> m_depthStencilContentDefined;
+        MG_Util::Damage::SwapchainAgeTracker m_ages;
+        // Kept across a rebuild: it is the application's habit, not the chain's.
+        Bool m_bufferAgeAsked = false;
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -6,6 +6,7 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header

+#include "RenderPassGuard.h"
 #include "UniformManager.h"

 #include "MG_Backend/DirectVulkan/DirectVulkanResourceState.h"
@@ -251,18 +252,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            WireDescriptorFatal("image-descriptor-transition");
        // A second descriptor/draw can keep GENERAL after a shader write. A layout
        // equality fast return is not a memory dependency for that prior write.
-        VkMemoryBarrier memory{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
-        memory.srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT;
-        memory.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
-        vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-                             0, 1, &memory, 0, nullptr, 0, nullptr);
+        // Only when something may have written an image since the last such barrier
+        // (WireImageWriteEpoch): with nothing written in between, consecutive draws keep their
+        // render pass open instead of closing it for a barrier that orders nothing.
+        if (WireImageWriteEpoch() != m_wireImageBarrierEpoch) {
+            VkMemoryBarrier memory{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
+            memory.srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT;
+            memory.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+            EndActiveRenderPassOn(commandBuffer);
+            vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                                 0, 1, &memory, 0, nullptr, 0, nullptr);
+            m_wireImageBarrierEpoch = WireImageWriteEpoch();
+        }
        VkImageViewCreateInfo info{VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO};
        info.image = resource->image;
        info.viewType = type;
        info.format = format;
        info.subresourceRange = {aspect, level, levels, layer, layers};
        if (!storage) {
-            const Bool alphaIsOne = ResolveTextureFormatInfo(
+            // An X-format shared image carries undefined alpha bytes, exactly like an expanded RGB.
+            const Bool alphaIsOne = resource->sharedImageAlphaOne || ResolveTextureFormatInfo(
                static_cast<TextureInternalFormat>(record.Desc.InternalFormat)).expandRgbToRgba;
            const auto swizzle = [&](Uint32 channel) {
                const auto value = record.Params.Swizzle[channel];
@@ -293,11 +302,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        } else {
            auto samplerHandle = state.BoundSamplerStates[unit];
            if (MG_Pipe::MGPipeHandleIsNull(samplerHandle)) samplerHandle = record.Params.BuiltinSampler;
-            if (MG_Pipe::MGPipeHandleIsNull(samplerHandle) || samplerHandle.Slot >= state.SamplerCsos.size())
-                WireDescriptorFatal("sampler-record");
-            const auto& sampler = state.SamplerCsos[samplerHandle.Slot];
-            if (!sampler.Live || sampler.Gen != samplerHandle.Gen) WireDescriptorFatal("sampler-record-generation");
-            out.sampler = m_samplerManager->GetOrCreateSamplerFromParameters(sampler.Params,
+            // A texture's built-in sampler CSO can be missing or gone while the texture still names
+            // it: sampler CSOs are content-addressed and recycled at capacity, and the texture's
+            // record is only re-pointed by its next set_texture_params (a texture whose params never
+            // crossed names none at all). Sample with GL's default parameters rather than end the
+            // session - the DirectGLES arm declines the push in both cases.
+            using SamplerCsoRecord = std::decay_t<decltype(state.SamplerCsos[0])>;
+            static const decltype(SamplerCsoRecord::Params) kDefaultSamplerParams{};
+            const Bool known = !MG_Pipe::MGPipeHandleIsNull(samplerHandle) &&
+                               samplerHandle.Slot < state.SamplerCsos.size();
+            const SamplerCsoRecord* sampler = known ? &state.SamplerCsos[samplerHandle.Slot] : nullptr;
+            const Bool stale = sampler == nullptr || !sampler->Live || sampler->Gen != samplerHandle.Gen;
+            if (stale) {
+                MGLOG_E_ONCE("Magma wire: sampler CSO {%u, %u} named by texture unit %u is missing or no longer "
+                             "live; sampling with default parameters", samplerHandle.Slot, samplerHandle.Gen, unit);
+            }
+            out.sampler = m_samplerManager->GetOrCreateSamplerFromParameters(stale ? kDefaultSamplerParams : sampler->Params,
                static_cast<TextureInternalFormat>(record.Desc.InternalFormat),
                domain == SamplerNumericDomain::SignedInteger || domain == SamplerNumericDomain::UnsignedInteger,
                levels);
@@ -591,6 +611,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        m_samplerResolveMemoHighWater = 0;
        m_setsPerFrame = setsPerFrame;
        m_peakDescriptorSetsObserved = 0;
+        m_trimQuietEpochs = MG_Config::Features.MagmaDescriptorTrimFrames;
+        // The renderer records its first frame into slot 0 before any BeginFrame runs.
+        m_epochSlot = 0;
+        m_epochOpen = true;
+        m_descriptorSlotTrims = 0;
        m_textureManager = textureManager;
        m_samplerManager = samplerManager;

@@ -600,6 +625,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            frame.activeDescriptorPoolIndex = 0;
            frame.allocatedSetsThisFrame = 0;
            frame.peakAllocatedSetsThisFrame = 0;
+            frame.capacitySets = m_setsPerFrame;
+            frame.quietEpochs = 0;
            frame.descriptorPools.clear();

            VkDescriptorPool initialPool = VK_NULL_HANDLE;
@@ -654,6 +681,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            frame.activeDescriptorPoolIndex = 0;
            frame.allocatedSetsThisFrame = 0;
            frame.peakAllocatedSetsThisFrame = 0;
+            frame.capacitySets = 0;
+            frame.quietEpochs = 0;
        }
        m_frames.clear();
 #if MOBILEGL_BUILD_DISAGGREGATED
@@ -662,6 +691,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        m_wirePrivateStoragePlaceholders = 0;
        m_wireInvalidStorageImagesBindNull = false;
 #endif
+        m_epochSlot = 0;
+        m_epochOpen = false;

        m_bufferManager = nullptr;
        m_programFactory = nullptr;
@@ -701,6 +732,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                "UniformDescriptorBinder: new descriptor set peak observed=%u (base setsPerFrame=%u, frame=%u, pools=%zu)",
                m_peakDescriptorSetsObserved, m_setsPerFrame, frameIndex, frame.descriptorPools.size());
        }
+        // The epoch that just ended is the one of the slot begun last - this slot again on the
+        // drain path, the previous frame's slot on the Present path. Its peak is still unreset.
+        if (m_epochOpen && m_epochSlot < m_frames.size()) {
+            NoteEndedEpochForTrim(m_frames[m_epochSlot].peakAllocatedSetsThisFrame);
+        }
+        m_epochSlot = frameIndex;
+        m_epochOpen = true;
+        // No pending command buffer references this slot's sets here (see TrimFrameDescriptorPools
+        // for why that holds at both of BeginFrame's call sites), so a quiet grown slot can be
+        // handed back whole.
+        if (m_trimQuietEpochs != 0 && frame.descriptorPools.size() > 1 &&
+            frame.capacitySets >= kDescriptorTrimMinCapacity && frame.quietEpochs >= m_trimQuietEpochs) {
+            TrimFrameDescriptorPools(frame, frameIndex);
+        }
        frame.activeDescriptorPoolIndex = 0;
        frame.allocatedSetsThisFrame = 0;
        frame.peakAllocatedSetsThisFrame = 0;
@@ -709,11 +754,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }
        // The frame's descriptor sets are recycled above, so last frame's reuse targets
        // are gone: start the per-draw descriptor-reuse cache fresh this frame.
-        for (auto& entry : m_descriptorReuseMemo) {
-            entry.valid = false;
-        }
-        m_fastRebindMemo.valid = false;
-        m_lastBindValid = false;
+        InvalidateDescriptorSetMemos();
        // Re-fingerprint the bound sampler set fresh this frame so any GL object address
        // reuse cannot outlive a single frame (see SamplerResolveMemo). Only the entries a
        // resolve has actually written can be valid, so the high-water mark bounds the
@@ -747,10 +788,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            entry.second.cursor = 0;
        }
        frame.allocatedSetsThisFrame = 0;
-        frame.peakAllocatedSetsThisFrame = 0;
-        for (auto& entry : m_descriptorReuseMemo) entry.valid = false;
-        m_fastRebindMemo.valid = false;
-        m_lastBindValid = false;
+        // The PEAK is not reset: a rewind ends a window of the epoch, not the epoch. The trim scores
+        // an epoch by the most sets the slot held at once in it (NoteEndedEpochForTrim at the next
+        // BeginFrame), and on this arm that is the largest window - resetting here would score a
+        // burst frame by its last window alone and trim a slot that re-grows every frame.
+        InvalidateDescriptorSetMemos();
        const Uint32 touchedBindings =
            std::min<Uint32>(m_samplerResolveMemoHighWater, static_cast<Uint32>(m_samplerResolveMemo.size()));
        for (Uint32 binding = 0; binding < touchedBindings; ++binding) {
@@ -761,6 +803,68 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }
 #endif

+    void UniformManager::InvalidateDescriptorSetMemos() {
+        // Every VkDescriptorSet handle held outside the per-layout caches - in this backend they all
+        // live in this class. A trim can hand a destroyed set's handle value to a fresh allocation,
+        // so the bind-dedup shadow must drop too, not only the two reuse memos.
+        for (auto& entry : m_descriptorReuseMemo) {
+            entry.valid = false;
+        }
+        m_fastRebindMemo.valid = false;
+        m_lastBindValid = false;
+    }
+
+    void UniformManager::NoteEndedEpochForTrim(Uint32 epochPeakSets) {
+        for (auto& slot : m_frames) {
+            if (static_cast<Uint64>(epochPeakSets) * kDescriptorTrimHeadroom > slot.capacitySets) {
+                slot.quietEpochs = 0;
+            } else if (slot.quietEpochs < std::numeric_limits<Uint32>::max()) {
+                ++slot.quietEpochs;
+            }
+        }
+    }
+
+    void UniformManager::TrimFrameDescriptorPools(FrameResources& frame, Uint32 frameIndex) {
+        const SizeT poolsBefore = frame.descriptorPools.size();
+        const Uint64 capacityBefore = frame.capacitySets;
+        // Holders first: the per-layout cache and every memo that may name one of these sets.
+        frame.descriptorSetCacheByLayout.clear();
+        InvalidateDescriptorSetMemos();
+        for (SizeT i = 1; i < frame.descriptorPools.size(); ++i) {
+            if (frame.descriptorPools[i].handle != VK_NULL_HANDLE) {
+                vkDestroyDescriptorPool(m_device, frame.descriptorPools[i].handle, nullptr);
+            }
+        }
+        frame.descriptorPools.resize(1);
+        auto& base = frame.descriptorPools[0];
+        // Frees every set allocated from it; the spec defines no failure for vkResetDescriptorPool.
+        vkResetDescriptorPool(m_device, base.handle, 0);
+        base.allocatedSets = 0;
+        frame.activeDescriptorPoolIndex = 0;
+        frame.capacitySets = base.maxSets;
+        frame.quietEpochs = 0;
+        ++m_descriptorSlotTrims;
+        MGLOG_D("UniformDescriptorBinder: frame %u trimmed its descriptor pools (%zu pools / %llu sets -> 1 / %u)",
+                frameIndex, poolsBefore, static_cast<unsigned long long>(capacityBefore), base.maxSets);
+    }
+
+    DescriptorPoolCensus UniformManager::GetDescriptorPoolCensus() const {
+        DescriptorPoolCensus census;
+        census.available = true;
+        census.frameSlots = static_cast<uint32_t>(m_frames.size());
+        for (const auto& frame : m_frames) {
+            const auto poolCount = static_cast<uint32_t>(frame.descriptorPools.size());
+            census.pools += poolCount;
+            census.maxPoolsInOneSlot = std::max(census.maxPoolsInOneSlot, poolCount);
+            census.capacitySets += frame.capacitySets;
+            for (const auto& cacheEntryPair : frame.descriptorSetCacheByLayout) {
+                census.cachedSets += cacheEntryPair.second.sets.size();
+            }
+        }
+        census.slotTrims = m_descriptorSlotTrims;
+        return census;
+    }
+
    void UniformManager::OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout) {
        SizeT purgedSets = 0;
        for (auto& frame : m_frames) {
@@ -2700,9 +2804,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        auto& bindingPoint = MGB_CTX->GetBufferBindingPoint(BufferTarget::Uniform, frontendBinding);
        const auto& bufferObject = bindingPoint.GetBoundObject();
-        MOBILEGL_ASSERT(bufferObject != nullptr,
-                        "ResolveUniformBufferPayload: no UBO bound at frontend binding %u for block '%s'",
-                        frontendBinding, program.GetUniformBlockName(static_cast<Uint32>(blockIndex)).c_str());
+        static thread_local Vector<Uint8> paddedUbo;
+        if (bufferObject == nullptr) {
+            // A block the program declares at a point with nothing bound - never bound, or its
+            // buffer was deleted while bound, which unbinds it (UnboundUniformBlockScenario,
+            // DeletedBoundBufferScenario). GL leaves what the shader READS undefined and raises no
+            // error; this used to dereference the null object. Answer it with a zeroed block of the
+            // reflected size, as the wire arm does (ResolveWireUniformBufferPayload).
+            const VkDeviceSize blockSize =
+                static_cast<VkDeviceSize>(program.GetUBOSizeAt(static_cast<Uint32>(blockIndex)));
+            paddedUbo.assign(static_cast<SizeT>(blockSize > 0 ? blockSize : 16), 0);
+            out.payload = paddedUbo.data();
+            out.payloadSize = static_cast<VkDeviceSize>(paddedUbo.size());
+            return true;
+        }
        bufferObject->SyncPersistentMappedRange();

        MOBILEGL_ASSERT(bufferObject->MappedData() != nullptr && bufferObject->GetSize() != 0,
@@ -2735,7 +2850,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        outSize = blockSize;
        outData = bufferObject->MappedData() + static_cast<SizeT>(rangeStart);
        if (available < blockSize) {
-            static thread_local Vector<Uint8> paddedUbo;
            paddedUbo.assign(static_cast<SizeT>(blockSize), 0);
            Memcpy(paddedUbo.data(), outData, static_cast<SizeT>(available));
            outData = paddedUbo.data();
@@ -2831,32 +2945,49 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return true;
    }

+    Uint32 UniformManager::NextDescriptorPoolMaxSets(const Vector<DescriptorPoolBucket>& pools, Bool updateAfterBind,
+                                                     Uint32 initialMaxSets) {
+        const Uint32 minSets = std::max<Uint32>(1, initialMaxSets);
+        const Uint32 maxSets = std::max(minSets, kMaxDescriptorPoolSets);
+        Uint32 largest = 0;
+        for (const auto& pool : pools) {
+            if (pool.updateAfterBind == updateAfterBind) {
+                largest = std::max(largest, pool.maxSets);
+            }
+        }
+        if (largest == 0) {
+            return minSets;
+        }
+        const Uint64 doubled = static_cast<Uint64>(largest) * 2;
+        return static_cast<Uint32>(std::clamp<Uint64>(doubled, minSets, maxSets));
+    }
+
    Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind) {
        if (frame.descriptorPools.empty()) {
            return false;
        }

-        const auto matchingBucket = std::find_if(
-            frame.descriptorPools.begin(), frame.descriptorPools.end(),
-            [updateAfterBind](const DescriptorPoolBucket& candidate) { return candidate.updateAfterBind == updateAfterBind; });
-        const Uint32 currentMaxSets = matchingBucket != frame.descriptorPools.end()
-                                          ? std::max<Uint32>(1, matchingBucket->maxSets)
-                                          : m_setsPerFrame;
-        const Uint32 grownMaxSets = currentMaxSets <= (std::numeric_limits<Uint32>::max() / 2) ? (currentMaxSets * 2)
-                                                                                                 : currentMaxSets;
-
+        Uint32 grownMaxSets = NextDescriptorPoolMaxSets(frame.descriptorPools, updateAfterBind, m_setsPerFrame);
        VkDescriptorPool grownPool = VK_NULL_HANDLE;
-        if (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
-            MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u -> %u sets)",
-                    currentMaxSets, grownMaxSets);
-            return false;
+        // A large pool is a large driver allocation; if the device refuses one, settle for less
+        // rather than failing the draw - halving down to the base size, where the old policy lived.
+        while (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
+            if (grownMaxSets <= m_setsPerFrame) {
+                MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u sets)",
+                             grownMaxSets);
+                return false;
+            }
+            grownMaxSets = std::max(m_setsPerFrame, grownMaxSets / 2);
        }

        frame.descriptorPools.push_back({grownPool, grownMaxSets, 0, updateAfterBind});
        frame.activeDescriptorPoolIndex = static_cast<Uint32>(frame.descriptorPools.size() - 1);
-        MGLOG_D(
-            "UniformDescriptorBinder: frame %u descriptor pool exhausted, grew pool (%u -> %u sets), poolCount=%zu",
-            frameIndex, currentMaxSets, grownMaxSets, frame.descriptorPools.size());
+        frame.capacitySets += grownMaxSets;
+        // Growing is the opposite of quiet: a trim must wait a full quiet run from here.
+        frame.quietEpochs = 0;
+        MGLOG_D("UniformDescriptorBinder: frame %u descriptor pool exhausted, grew a %u-set pool (updateAfterBind=%d), "
+                "poolCount=%zu",
+                frameIndex, grownMaxSets, updateAfterBind ? 1 : 0, frame.descriptorPools.size());
        return true;
    }

@@ -3072,7 +3203,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                    // preparing the texture would decline a storage-less one and lose the pass.
                    if (storage && WireShaderImageNamesNoTexel(*m_textureManager, state.BoundShaderImages[unit]))
                        continue;
-                    if (!m_textureManager->SyncTextureResourceByHandle(handle, false, storage)) return false;
+                    const auto* resource = m_textureManager->SyncTextureResourceByHandle(handle, false, storage);
+                    if (resource == nullptr) return false;
+                    // Another session writes it: the renderer acquires it before this use.
+                    if (resource->sharedImageId != 0 || resource->yuvImageId != 0)
+                        m_textureManager->NoteSharedImageUse(*resource);
                }
            }
        }
@@ -14,6 +14,7 @@
 #include "VkSamplerManager.h"
 #include "VkTextureManager.h"
 #include "WirePlaceholderKey.h"
+#include "../DescriptorPoolCensus.h"
 #include "../VkIncludes.h"
 #include <Includes.h>

@@ -84,9 +85,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // FREE_DESCRIPTOR_SET_BIT) and the pool accounting is credited, so program
        // churn recycles pool capacity instead of abandoning it. GPU-safe: the layout
        // only dies after >1024 idle frame boundaries, so no in-flight command buffer
-        // references its sets. This is the only eviction path for the per-layout
-        // caches - a live layout's entry must never be purged (its sets would be
-        // unreachable pool slots), so there is deliberately no age-based sweep here.
+        // references its sets. This is the only PER-LAYOUT eviction path for the
+        // caches - a live layout's entry must never be purged on its own (its sets would be
+        // unreachable pool slots), so there is deliberately no age-based sweep here. The
+        // other path, TrimFrameDescriptorPools, drops a slot's whole cache together with
+        // the pools behind it.
        void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
        // One record per visited CombinedImageSampler DESCRIPTOR (post fallback substitution,
        // in binding order, and within a binding in array-element order): the resolved texture
@@ -158,13 +161,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        static Bool ProgramSamplesOnlySingleLevelTextures(const MagmaProgramSource& program,
                                                          const ProgramFactory::VkProgramObject& programObj);

-    private:
+        // ---- per-frame descriptor pools ----------------------------------------------------
+        // Each frame slot owns a list of pools. The first is the base pool (non-update-after-bind,
+        // setsPerFrame sets) and is never destroyed while the manager lives; every later one was
+        // added by GrowFrameDescriptorPool when an allocation found no room. A pool serves one
+        // updateAfterBind flavour, because a set can only come from a pool created with the flag
+        // its layout was.
        struct DescriptorPoolBucket {
            VkDescriptorPool handle = VK_NULL_HANDLE;
            Uint32 maxSets = 0;
            Uint32 allocatedSets = 0;
            Bool updateAfterBind = false;
        };
+        // One pool's maxSets never exceeds this. It bounds one pool's driver memory: CreateDescriptorPool
+        // declares 8 descriptors of each of 6 types per set, so a 4096-set pool asks for 196,608
+        // descriptors - about 12 MiB at 64 bytes each on a driver that backs every declared one.
+        static constexpr Uint32 kMaxDescriptorPoolSets = 4096;
+        // The growth policy, pure so the host tests can drive it: the size of the next pool a slot
+        // adds for `updateAfterBind` layouts, given the slot's existing pools. Twice the LARGEST pool
+        // of that flavour, capped at kMaxDescriptorPoolSets; `initialMaxSets` when the slot has no
+        // pool of that flavour yet. So a frame that needs N sets adds O(log N) pools while N stays
+        // under the cap, and one more pool per kMaxDescriptorPoolSets sets beyond it. Taking the
+        // largest (not the first) pool is the point: the first same-flavour pool is the oldest,
+        // smallest one, and doubling IT made every grown pool the same size.
+        static Uint32 NextDescriptorPoolMaxSets(const Vector<DescriptorPoolBucket>& pools, Bool updateAfterBind,
+                                                Uint32 initialMaxSets);
+        // Trim policy (see TrimFrameDescriptorPools). A slot is "quiet" in an epoch whose peak set
+        // count times kDescriptorTrimHeadroom still fits its capacity; a slot that has been quiet for
+        // MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES consecutive epochs, and holds at least
+        // kDescriptorTrimMinCapacity sets, is trimmed. The headroom is what keeps the trim from
+        // thrashing: geometric growth leaves a re-grown slot at least a third full (the pool a growth
+        // adds is at most twice the largest before it, so at most twice everything the slot held, and
+        // the epoch that grew it needed more than that), so the epoch that grew a slot is never quiet
+        // for it. The floor keeps a slot sized for ordinary
+        // frames from being trimmed at all - the trim exists to give back a burst.
+        static constexpr Uint32 kDescriptorTrimHeadroom = 4;
+        static constexpr Uint32 kDescriptorTrimMinCapacity = 1024;
+        // For tests (DescriptorPoolCensus.h): pools, capacity, cached sets and trims over every slot.
+        DescriptorPoolCensus GetDescriptorPoolCensus() const;
+
+    private:

        // A cached descriptor set together with the pool it was allocated from, so a
        // layout-destroyed purge can vkFreeDescriptorSets it back and credit the
@@ -239,6 +275,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        void DestroyWirePlaceholderImage(WirePlaceholderImage& image) const;
 #endif

+        // descriptorSetCacheByLayout is the per-frame set cache: BeginFrame rewinds every cursor, so
+        // a slot re-writes the sets it allocated in earlier frames instead of allocating new ones, and
+        // the cache only ever grows - except through TrimFrameDescriptorPools, which gives a slot's
+        // whole burst back at once (and OnDescriptorSetLayoutDestroyed, one dead layout at a time).
        struct FrameResources {
            Vector<DescriptorPoolBucket> descriptorPools;
            UnorderedMap<VkDescriptorSetLayout, DescriptorSetCacheEntry> descriptorSetCacheByLayout;
@@ -252,6 +292,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            Uint32 activeDescriptorPoolIndex = 0;
            Uint32 allocatedSetsThisFrame = 0;
            Uint32 peakAllocatedSetsThisFrame = 0;
+            // Sum of maxSets over descriptorPools, kept in step by Initialize, Grow and Trim.
+            Uint64 capacitySets = 0;
+            // Consecutive epochs (see NoteEndedEpochForTrim) that were quiet for THIS slot's capacity.
+            Uint32 quietEpochs = 0;
        };

        static Bool ResolveSamplerTexture(const MagmaProgramSource& program,
@@ -395,6 +439,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                      const Vector<Uint32>& dynamicOffsets);
        Bool CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const;
        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind);
+        // An epoch is the stretch between two BeginFrame calls: one frame, or the part of one between
+        // two drains (VulkanRenderer::TryDrainFrameTransients). Called by BeginFrame with the epoch
+        // that just ended, it updates every slot's quiet count - a burst in any slot counts against
+        // all of them, so bursts that land in the other slots still hold a slot's trim off.
+        void NoteEndedEpochForTrim(Uint32 epochPeakSets);
+        // Gives a slot's whole descriptor burst back: clears its per-layout cache, destroys every
+        // grown pool and resets the base pool, so the slot re-grows geometrically from setsPerFrame.
+        // Whole-slot rather than per-set: vkFreeDescriptorSets fragments a pool and can never empty
+        // a grown one. Called only from BeginFrame, which runs only when no pending command buffer
+        // references the slot's sets:
+        //   - Present path (VulkanRenderer::Present, after WaitAndAcquireNextImage): the slot's
+        //     imageInFlightFence has been waited. That fence rides the slot's LAST submission - the
+        //     Present submit - and every mid-frame submission of the slot (FlushPendingCommands,
+        //     SplitOversizedRecording, readbacks) went to the same queue earlier, so the fence's
+        //     signal covers them (vkQueueSubmit's fence scope includes all earlier submissions).
+        //   - Drain path (VulkanRenderer::TryDrainFrameTransients): every submission has been
+        //     observed complete and no recording is open.
+        // Sets are per slot: a slot's sets are only ever bound into recordings made while it is the
+        // current slot, so trimming one slot never touches another slot's in-flight work.
+        void TrimFrameDescriptorPools(FrameResources& frame, Uint32 frameIndex);
+        // Drops every handle this class holds to a descriptor set outside the per-layout caches.
+        void InvalidateDescriptorSetMemos();
        VkResult AllocateDescriptorSetsFromActivePool(
            Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet);
        VkResult AcquireDescriptorSet(Uint32 frameIndex,
@@ -412,6 +478,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        Uint32 m_maxBindings = 0;
        Uint32 m_setsPerFrame = 0;
        Uint32 m_peakDescriptorSetsObserved = 0;
+        // MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES, latched at Initialize (0 = never trim).
+        Uint32 m_trimQuietEpochs = 0;
+        // The slot whose epoch is open (the last one BeginFrame was called for); valid once
+        // m_epochOpen is set. Its peak is the ended epoch's peak at the next BeginFrame.
+        Uint32 m_epochSlot = 0;
+        Bool m_epochOpen = false;
+        Uint64 m_descriptorSlotTrims = 0;
        VkTextureManager* m_textureManager = nullptr;
        VkSamplerManager* m_samplerManager = nullptr;
        mutable SharedPtr<MG_State::GLState::ITextureObject> m_fallbackTexture2D;
@@ -587,5 +660,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                m_samplerResolveMemoHighWater = binding + 1;
            }
        }
+        // WireImageWriteEpoch() when ResolveWireImageDescriptor last made image writes visible.
+        mutable Uint64 m_wireImageBarrierEpoch = 0;
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -19,6 +19,7 @@
 #include "VertexInputStateBuilder.h"
 #include "MG_State/GLState/VertexArrayState/VertexArrayObject.h"
 #include <Includes.h>
+#include <atomic>
 #include "../VkIncludes.h"

 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -217,11 +218,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // memo table is owned by this factory and dies with it, so a per-instance counter is
        // enough there and the epoch shrinks back to what it looks like it should be.
 #if MOBILEGL_PIPE_LEGACY_MEMOS
-        static inline Uint64 s_evictionEpochSource = 0;
+        static inline std::atomic<Uint64> s_evictionEpochSource{0};
        Uint64 m_evictionEpoch = ++s_evictionEpochSource;
 #else
        Uint64 m_evictionEpoch = 1;
 #endif
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline thread_local XXH64_state_t* m_hashState = XXH64_createState();
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -6,6 +6,7 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header

+#include "RenderPassGuard.h"
 #include "VkBufferManager.h"
 #include "../DirectVulkan.h"
 #include "VulkanRenderer.h"
@@ -15,7 +16,14 @@
 #include "MG_Pipe/MGPipeCallbacks.h"
 #include "MG_Pipe/PipeApply.h"
 #include "MG_Remote/Server/StagedShadow.h"
+#if MOBILEGL_BUILD_DISAGGREGATED
+#include "MG_Remote/Transport/AdoptT0.h"
+#endif
+#if MOBILEGL_BUILD_DISAGGREGATED && defined(__ANDROID__)
+#include <android/hardware_buffer.h>
+#endif
 #include <Config.h>
+#include <chrono>
 #include <cstdlib>
 #endif

@@ -106,30 +114,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        using MG_State::GLState::BufferObject;

        // The manager owned by the active VulkanRenderer; immediate ops route here.
-        VkBufferManager* g_activeBufferManager = nullptr;
+        SessionLocal<VkBufferManager*> g_activeBufferManager;  // per Magma session (MagmaSession.h)

        void Ops_Respecify(BufferObject& bufferObject) {
-            if (g_activeBufferManager) {
-                g_activeBufferManager->OnRespecify(bufferObject);
+            if (g_activeBufferManager.Get()) {
+                g_activeBufferManager.Get()->OnRespecify(bufferObject);
            }
        }

        void Ops_SubData(BufferObject& bufferObject, SizeT offset, SizeT size) {
-            if (g_activeBufferManager) {
-                g_activeBufferManager->OnSubData(bufferObject, offset, size);
+            if (g_activeBufferManager.Get()) {
+                g_activeBufferManager.Get()->OnSubData(bufferObject, offset, size);
            }
        }

        void Ops_ResidentSubData(BufferObject& bufferObject, SizeT offset, DataPtr data) {
-            if (g_activeBufferManager) {
-                g_activeBufferManager->OnResidentSubData(bufferObject, offset, data);
+            if (g_activeBufferManager.Get()) {
+                g_activeBufferManager.Get()->OnResidentSubData(bufferObject, offset, data);
            }
        }

        void Ops_FlushMappedRange(BufferObject& bufferObject, Range1D range,
                                  Flags<BufferMappingAccessBit> appAccess) {
-            if (g_activeBufferManager) {
-                g_activeBufferManager->OnFlushMappedRange(bufferObject, range, appAccess);
+            if (g_activeBufferManager.Get()) {
+                g_activeBufferManager.Get()->OnFlushMappedRange(bufferObject, range, appAccess);
            }
        }

@@ -147,15 +155,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        void* Ops_AcquirePersistentMap(BufferObject& bufferObject) {
-            if (g_activeBufferManager) {
-                return g_activeBufferManager->AcquirePersistentMap(bufferObject);
+            if (g_activeBufferManager.Get()) {
+                return g_activeBufferManager.Get()->AcquirePersistentMap(bufferObject);
            }
            return nullptr;
        }

        void Ops_OnDestroy(SharedPtr<BackendBufferResource>&& resource) {
-            if (g_activeBufferManager) {
-                g_activeBufferManager->OnResourceDestroyed(std::move(resource));
+            if (g_activeBufferManager.Get()) {
+                g_activeBufferManager.Get()->OnResourceDestroyed(std::move(resource));
            }
            // No active manager: the device/allocator is gone or going away and
            // Shutdown() already destroyed the storage; dropping the handle here
@@ -174,38 +182,54 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        };

 #if MOBILEGL_BUILD_DISAGGREGATED
-        VkBufferManager& WireManager() {
-            if (g_activeBufferManager == nullptr) {
-                MGLOG_F("Magma: Fatal{ResourceUnavailable, \"wire-buffer-manager\"}");
-                std::abort();
+        // The session's manager, or null once its renderer is gone - a client's records can still
+        // arrive after its eglTerminate released it (or, surfaceless, before its first surface built
+        // one). Such a record is declined with one line rather than taking the whole server down.
+        VkBufferManager* WireManager(const char* op) {
+            VkBufferManager* manager = g_activeBufferManager.Get();
+            if (manager == nullptr) {
+                MGLOG_E_ONCE("Magma: a wire buffer %s reached a session with no renderer; declined", op);
            }
-            return *g_activeBufferManager;
+            return manager;
        }

        const MG_Pipe::MGPipeResourceOps g_vulkanWireResourceOps = {
-            .Create = [](auto res, const auto& desc) { WireManager().CreateWireBuffer(res, desc); },
+            .Create = [](auto res, const auto& desc) { if (auto* m = WireManager("create")) m->CreateWireBuffer(res, desc); },
            .Respecify = [](auto res, const auto& desc, const void* bytes) {
-                WireManager().RespecifyWireBuffer(res, desc, bytes);
+                if (auto* m = WireManager("respecify")) m->RespecifyWireBuffer(res, desc, bytes);
            },
            .SubData = [](auto res, const auto& record, const void* bytes) {
-                WireManager().WriteWireBuffer(res, MG_Pipe::MGPipeSubDataBufferOffset(record),
-                                               MG_Pipe::MGPipeSubDataBufferSize(record), bytes);
+                if (auto* m = WireManager("write"))
+                    m->WriteWireBuffer(res, MG_Pipe::MGPipeSubDataBufferOffset(record),
+                                       MG_Pipe::MGPipeSubDataBufferSize(record), bytes);
            },
            .SubDataResident = [](auto res, const auto& record, const void* bytes) {
-                WireManager().WriteWireBuffer(res, MG_Pipe::MGPipeSubDataBufferOffset(record),
-                                               MG_Pipe::MGPipeSubDataBufferSize(record), bytes);
+                if (auto* m = WireManager("write"))
+                    m->WriteWireBuffer(res, MG_Pipe::MGPipeSubDataBufferOffset(record),
+                                       MG_Pipe::MGPipeSubDataBufferSize(record), bytes);
            },
            .FlushRange = [](auto res, const auto& record, const void* bytes) {
-                WireManager().FlushWireBuffer(res, record.Offset, record.Size, bytes);
+                if (auto* m = WireManager("flush")) m->FlushWireBuffer(res, record.Offset, record.Size, bytes);
            },
            .Readback = [](auto res, const auto& record) {
-                WireManager().ReadbackWireBuffer(res, record.Offset, record.Size);
+                if (auto* m = WireManager("readback")) m->ReadbackWireBuffer(res, record.Offset, record.Size);
            },
-            .Destroy = [](auto res) { WireManager().DestroyWireBuffer(res); },
+            .Destroy = [](auto res) { if (auto* m = WireManager("destroy")) m->DestroyWireBuffer(res); },
            // Split mapping remains T2: the client owns its map and pushes exact
            // modified ranges. A server pointer is never donated across the wire.
            .MapPersistent = [](MG_Pipe::MGPipeHandle, Uint64, const void*) -> void* { return nullptr; },
            .UnmapPersistent = [](MG_Pipe::MGPipeHandle) {},
+#if defined(__ANDROID__)
+            // P11 B2 (T0): the one door a server pointer is still never donated through stays shut;
+            // what T0 adds is the other direction - the CLIENT's pages imported as the store.
+            .ImportExternal = [](MG_Pipe::MGPipeHandle res, void* ahb, Uint64 size) -> Bool {
+                auto* m = WireManager("import");
+                return m != nullptr && m->ImportWireBuffer(res, ahb, size);
+            },
+            .SelfTestExternal = [](char* why, Uint64 whyBytes) -> Bool {
+                return VkBufferManager::SelfTestWireImport(why, whyBytes);
+            },
+#endif
        };
 #endif
    } // namespace
@@ -217,6 +241,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }
    }

+    namespace {
+        // Diagnostic memory for DescribeWireBufferHistory: the last handles created and destroyed.
+        struct WireBufferHistory {
+            static constexpr SizeT kDepth = 256;
+            MG_Pipe::MGPipeHandle created[kDepth]{};
+            MG_Pipe::MGPipeHandle destroyed[kDepth]{};
+            SizeT createdCount = 0;
+            SizeT destroyedCount = 0;
+        };
+        SessionLocal<WireBufferHistory> g_wireBufferHistory;
+    } // namespace
+
    VkBufferManager::WireBufferResource* VkBufferManager::FindWireBuffer(MG_Pipe::MGPipeHandle res) {
        const auto found = m_wireBuffers.find(WireBufferKey(res));
        return found == m_wireBuffers.end() ? nullptr : &found->second;
@@ -229,6 +265,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            std::abort();
        }
        m_wireBuffers.try_emplace(WireBufferKey(res));
+        {
+            auto& history = *g_wireBufferHistory;
+            history.created[history.createdCount++ % 256] = res;
+        }
        RespecifyWireBuffer(res, desc, nullptr);
    }

@@ -245,11 +285,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // to wait for a submission (DeferredWireRelease), and three lines from now the record
        // no longer carries it.
        const Uint64 orphanedUseSerial = resource->lastUseSerial;
-        DeferWireRelease(std::move(resource->buffer), orphanedUseSerial);
+        // P11 B2: a T0 store is ALWAYS released through the serial-gated list (never "at once"),
+        // so the client's AHardwareBuffer outlives every submission that can still name it however
+        // the client's own release (at this respecify) raced the frames in flight.
+        DeferWireRelease(std::move(resource->buffer),
+                         resource->imported ? T0ReleaseSerial(orphanedUseSerial) : orphanedUseSerial);
+        resource->imported = false;
        resource->size = desc.Width;
        resource->lastUseSerial = 0;
        resource->lastUseSubmitIndex = 0;  // M2 r2: the new store carries no submission yet (B3's stamp is per store)
        resource->gpuWritesPending = false;
+        resource->indirectReadBarrierPending = false;
        resource->stagedCoverage.clear();
        ++m_sliceEpochCounter;
        if (resource->size == 0) return;
@@ -292,9 +338,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkMemoryBarrier barrier{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        barrier.srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT;
        barrier.dstAccessMask = VK_ACCESS_HOST_READ_BIT | VK_ACCESS_HOST_WRITE_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_HOST_BIT,
                             0, 1, &barrier, 0, nullptr, 0, nullptr);
-        if (!pVulkanRenderer->WaitForSubmitIndex(pVulkanRenderer->GetSyncPointSubmitIndex(), UINT64_MAX, true)) {
+        // P8-D: counted here, past the idle early-out, so `whw` is waits TAKEN, not calls.
+        ++m_wireHostWaits;
+        const Bool timed = MG_Util::PipeStats::Enabled();
+        const auto waitStart = timed ? std::chrono::steady_clock::now() : std::chrono::steady_clock::time_point{};
+        const Bool waited =
+            pVulkanRenderer->WaitForSubmitIndex(pVulkanRenderer->GetSyncPointSubmitIndex(), UINT64_MAX, true);
+        if (timed) {
+            MG_Util::PipeStats::AddCalls(MG_Util::PipeStats::CallClass::WireHostWaits, 1);
+            MG_Util::PipeStats::AddCalls(MG_Util::PipeStats::CallClass::WireHostWaitMicros,
+                static_cast<Uint64>(std::chrono::duration_cast<std::chrono::microseconds>(
+                    std::chrono::steady_clock::now() - waitStart).count()));
+        }
+        if (!waited) {
            return false;
        }
        resource.lastUseSerial = 0;
@@ -368,7 +427,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                uploaded = StagedWireRangeCopy(*resource, bytes, static_cast<SizeT>(offset),
                                                static_cast<SizeT>(size));
            }
-            if (!uploaded && !WaitForWireBufferHostAccess(*resource)) WireBufferSyncFatal("host-write");
+            if (!uploaded && !WaitForWireBufferHostAccess(*resource)) {
+                if (LatchedOnDeviceLoss("buffer-host-write")) return;
+                WireBufferSyncFatal("host-write");
+            }
        }
        if (!uploaded) {
            uploaded = resource->buffer.Upload(bytes, size, offset);
@@ -377,6 +439,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            }
        }
        if (!uploaded) {
+            if (LatchedOnDeviceLoss("buffer-upload")) return;
            MGLOG_F("Magma: Fatal{ResourceUnavailable, \"buffer-upload\"}");
            std::abort();
        }
@@ -470,13 +533,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkMemoryBarrier before{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        before.srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT | VK_ACCESS_HOST_WRITE_BIT;
        before.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT | VK_PIPELINE_STAGE_HOST_BIT,
                             VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1, &before, 0, nullptr, 0, nullptr);
        const VkBufferCopy copy{offset, dst.offset, size};
+        EndActiveRenderPassOn(commands);
        vkCmdCopyBuffer(commands, resource->buffer.GetHandle(), dst.buffer, 1, &copy);
        VkMemoryBarrier after{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        after.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
        after.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                             0, 1, &after, 0, nullptr, 0, nullptr);
        resource->lastUseSerial = m_frameSerial;
@@ -514,22 +580,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkMemoryBarrier before{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        before.srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT | VK_ACCESS_HOST_WRITE_BIT;
        before.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT | VK_PIPELINE_STAGE_HOST_BIT,
                             VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1, &before, 0, nullptr, 0, nullptr);
        const VkBufferCopy widen{alignedOffset, staging.offset, stagedSize};
+        EndActiveRenderPassOn(commands);
        vkCmdCopyBuffer(commands, resource->buffer.GetHandle(), staging.buffer, 1, &widen);
        // The shift reads what the widening copy just wrote, so it needs its own edge even
        // though both halves are transfers on one command buffer.
        VkMemoryBarrier staged{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        staged.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
        staged.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
                             0, 1, &staged, 0, nullptr, 0, nullptr);
        const VkBufferCopy shift{staging.offset + headPad, dst.offset + dstSkip, size};
+        EndActiveRenderPassOn(commands);
        vkCmdCopyBuffer(commands, staging.buffer, dst.buffer, 1, &shift);
        VkMemoryBarrier after{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        after.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
        after.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+        EndActiveRenderPassOn(commands);
        vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                             0, 1, &after, 0, nullptr, 0, nullptr);
        resource->lastUseSerial = m_frameSerial;
@@ -537,11 +608,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    void VkBufferManager::ReadbackWireBuffer(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size) {
-        Vector<Uint8> bytes(static_cast<SizeT>(size));
-        if (!ReadWireBuffer(res, offset, size, bytes.data())) {
+        // P11 B2: A T0 STORE'S READBACK IS "DONE, NO BYTES". Its pages are the client's
+        // AHardwareBuffer, so what the client needs before it reads them is this server's GPU
+        // having finished every write to them - the resident copies WriteWireBuffer recorded and
+        // any shader write - and nothing crossing SEG_EVENT. The wait is the store's own host-access
+        // wait (a barrier to host, then the sync point's submission); the reply then says "done".
+        // One death for both shapes: the store could not be made readable.
+        auto* resource = FindWireBuffer(res);
+        const Bool t0 = resource != nullptr && resource->imported;
+        Vector<Uint8> bytes(t0 ? 0 : static_cast<SizeT>(size));
+        if (t0 ? !WaitForWireBufferHostAccess(*resource) : !ReadWireBuffer(res, offset, size, bytes.data())) {
+            // A lost device: the session is latched and closing; its client reads a lost context.
+            if (LatchedOnDeviceLoss("buffer-readback")) return;
            MGLOG_F("Magma: Fatal{ResourceUnavailable, \"buffer-readback\"} {slot=%u, gen=%u}", res.Slot, res.Gen);
            std::abort();
        }
+        if (t0) {
+            ++m_sliceEpochCounter;
+            return;
+        }
        if (MG_Pipe::gMGPipeCallbacks.OnBufferWriteback == nullptr) {
            MGLOG_F("Magma: Fatal{ResourceUnavailable, \"buffer-writeback-callback\"}");
            std::abort();
@@ -561,6 +646,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        resource->lastUseSerial = m_frameSerial;
        resource->stagedCoverage.clear();
        resource->gpuWritesPending = true;
+        resource->indirectReadBarrierPending = true;
        ++m_sliceEpochCounter;
        // The existing client dirty-state protocol accepts one whole-store
        // notification. The store remains on the GPU and exact later SubData
@@ -573,10 +659,393 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MG_Pipe::gMGPipeCallbacks.OnGpuWritten(res, 1, &range);
    }

+    Bool VkBufferManager::TakeWireIndirectReadBarrier(MG_Pipe::MGPipeHandle res) {
+        auto* resource = FindWireBuffer(res);
+        if (!resource || !resource->indirectReadBarrierPending) return false;
+        resource->indirectReadBarrierPending = false;
+        return true;
+    }
+
+    // ---- P11 B2: T0 - THE CLIENT'S AHardwareBuffer AS A WIRE STORE -----------------------------
+
+    // The serial a T0 store is released under: its last use, or - when no GPU command has named it
+    // since a host-access wait - the current frame's, so the release still waits for the sync
+    // point rather than taking DeferWireRelease's "never named, destroy now" arm. Conservative by
+    // at most one submission.
+    Uint64 VkBufferManager::T0ReleaseSerial(Uint64 lastUseSerial) const {
+        if (lastUseSerial != 0) return lastUseSerial;
+        return m_frameSerial != 0 ? m_frameSerial : 1;
+    }
+
+#if defined(__ANDROID__)
+    namespace {
+        Int32 T0PickMemoryType(const VkPhysicalDeviceMemoryProperties& props, Uint32 bits,
+                               VkMemoryPropertyFlags want) {
+            for (Uint32 i = 0; i < props.memoryTypeCount; ++i) {
+                if ((bits & (1u << i)) != 0 && (props.memoryTypes[i].propertyFlags & want) == want)
+                    return static_cast<Int32>(i);
+            }
+            return -1;
+        }
+
+        struct T0ImportedBuffer {
+            VkBuffer buffer = VK_NULL_HANDLE;
+            VkDeviceMemory memory = VK_NULL_HANDLE;
+            void* mapped = nullptr;
+            Bool coherent = true;
+        };
+
+        void T0DestroyImported(VkDevice device, T0ImportedBuffer& imported) {
+            if (imported.mapped != nullptr) vkUnmapMemory(device, imported.memory);
+            if (imported.buffer != VK_NULL_HANDLE) vkDestroyBuffer(device, imported.buffer, nullptr);
+            if (imported.memory != VK_NULL_HANDLE) vkFreeMemory(device, imported.memory, nullptr);
+            imported = T0ImportedBuffer{};
+        }
+
+        // A VkBuffer on a dedicated allocation IMPORTED from `ahb`, in a host-visible memory type
+        // (coherent preferred), mapped whole. The server's own host map is what WriteWireBuffer's
+        // idle path and every CPU reader of a wire store use, so a type without HOST_VISIBLE is a
+        // refusal, not a degraded import. Nothing is left behind on failure.
+        Bool T0ImportAhb(const VulkanRenderer::WireAhbImport& ctx, void* ahb, Uint64 size, VkBufferUsageFlags usage,
+                         T0ImportedBuffer& out, String& why) {
+            out = T0ImportedBuffer{};
+            const auto getProperties =
+                reinterpret_cast<PFN_vkGetAndroidHardwareBufferPropertiesANDROID>(ctx.getAhbProperties);
+            VkAndroidHardwareBufferPropertiesANDROID properties{
+                VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_PROPERTIES_ANDROID};
+            VkResult result = getProperties(ctx.device, static_cast<const AHardwareBuffer*>(ahb), &properties);
+            if (result != VK_SUCCESS) {
+                why = "vkGetAndroidHardwareBufferPropertiesANDROID -> " + std::to_string(static_cast<int>(result));
+                return false;
+            }
+            VkExternalMemoryBufferCreateInfo external{VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO};
+            external.handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID;
+            VkBufferCreateInfo bufferInfo{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
+            bufferInfo.pNext = &external;
+            bufferInfo.size = size;
+            bufferInfo.usage = usage;
+            bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+            result = vkCreateBuffer(ctx.device, &bufferInfo, nullptr, &out.buffer);
+            if (result != VK_SUCCESS) {
+                out.buffer = VK_NULL_HANDLE;
+                why = "vkCreateBuffer(external AHB) -> " + std::to_string(static_cast<int>(result));
+                return false;
+            }
+            Int32 type = T0PickMemoryType(ctx.memory, properties.memoryTypeBits,
+                                          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
+            out.coherent = type >= 0;
+            if (type < 0)
+                type = T0PickMemoryType(ctx.memory, properties.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
+            if (type < 0) {
+                why = "no host-visible memory type can import it (memoryTypeBits 0x" +
+                      std::to_string(properties.memoryTypeBits) + ")";
+                T0DestroyImported(ctx.device, out);
+                return false;
+            }
+            VkMemoryDedicatedAllocateInfo dedicated{VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO};
+            dedicated.buffer = out.buffer;
+            VkImportAndroidHardwareBufferInfoANDROID import{VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID};
+            import.buffer = static_cast<AHardwareBuffer*>(ahb);
+            import.pNext = &dedicated;
+            VkMemoryAllocateInfo allocateInfo{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
+            allocateInfo.pNext = &import;
+            allocateInfo.allocationSize = properties.allocationSize;
+            allocateInfo.memoryTypeIndex = static_cast<Uint32>(type);
+            result = vkAllocateMemory(ctx.device, &allocateInfo, nullptr, &out.memory);
+            if (result != VK_SUCCESS) {
+                out.memory = VK_NULL_HANDLE;
+                why = "vkAllocateMemory(import AHB) -> " + std::to_string(static_cast<int>(result));
+                T0DestroyImported(ctx.device, out);
+                return false;
+            }
+            result = vkBindBufferMemory(ctx.device, out.buffer, out.memory, 0);
+            if (result == VK_SUCCESS) result = vkMapMemory(ctx.device, out.memory, 0, VK_WHOLE_SIZE, 0, &out.mapped);
+            if (result != VK_SUCCESS || out.mapped == nullptr) {
+                out.mapped = nullptr;
+                why = "vkBindBufferMemory/vkMapMemory -> " + std::to_string(static_cast<int>(result));
+                T0DestroyImported(ctx.device, out);
+                return false;
+            }
+            return true;
+        }
+    } // namespace
+#endif
+
+    Bool VkBufferManager::ImportWireBuffer(MG_Pipe::MGPipeHandle res, void* ahb, Uint64 size) {
+#if defined(__ANDROID__)
+        auto* resource = FindWireBuffer(res);
+        if (resource == nullptr || ahb == nullptr || size == 0 || size != resource->size) return false;
+        VulkanRenderer::WireAhbImport ctx;
+        if (pVulkanRenderer == nullptr || !pVulkanRenderer->GetWireAhbImport(ctx)) {
+            MGLOG_W_ONCE("Magma T0: this device did not take VK_ANDROID_external_memory_android_hardware_buffer");
+            return false;
+        }
+        VkBufferUsageFlags usage = kPersistentBackedUsage | VK_BUFFER_USAGE_TRANSFER_DST_BIT;
+        if (m_initInfo.transformFeedbackUsageEnabled) usage |= VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
+        T0ImportedBuffer imported;
+        String why;
+        if (!T0ImportAhb(ctx, ahb, size, usage, imported, why)) {
+            MGLOG_W("Magma T0: store {slot=%u, gen=%u} (%llu bytes) not imported - %s", res.Slot, res.Gen,
+                    static_cast<unsigned long long>(size), why.c_str());
+            return false;
+        }
+        // The T2 store it replaces goes the way a respecify's does (and any draw still naming it
+        // keeps it until its submission completes).
+        DeferWireRelease(std::move(resource->buffer), resource->lastUseSerial);
+        MG_Remote::Transport::AdoptT0::AcquireImported(ahb);
+        resource->buffer.AdoptExternal(ctx.device, imported.buffer, imported.memory, size, imported.mapped,
+                                       imported.coherent, ahb);
+        resource->imported = true;
+        resource->lastUseSerial = 0;
+        resource->lastUseSubmitIndex = 0;
+        resource->gpuWritesPending = false;
+        // The whole store holds bytes the client supplied (its AHB, seeded from its shadow).
+        resource->stagedCoverage.clear();
+        MG_Remote::Server::StagedShadowStore::CoverageAdd(resource->stagedCoverage, 0, static_cast<SizeT>(size));
+        ++m_sliceEpochCounter;
+        ++m_wireStoreCount;
+        NoteWireStorePeaks();
+        PublishWireReclaimGauges();
+        MGLOG_D("Magma T0: imported {slot=%u, gen=%u} (%llu bytes, %s memory)", res.Slot, res.Gen,
+                static_cast<unsigned long long>(size), imported.coherent ? "coherent" : "non-coherent");
+        return true;
+#else
+        (void)res;
+        (void)ahb;
+        (void)size;
+        return false;
+#endif
+    }
+
+    // THE POST (B2; decided by probe, never by driver name). The AHB API promises CPU coherence only
+    // across an unlock/lock, and T0 holds the client's lock for the store's life, so this exercises
+    // that pattern on THIS device before kCapAdoptT0 may be published:
+    //   hold - a 64 KiB BLOB allocated and locked by the test, never unlocked while it runs;
+    //   read - this server's own host map of the import reads what the CPU wrote through the lock;
+    //   gpu  - one submission copies that pattern out of the import (the GPU read) and fills another
+    //          region of it (the GPU write), then a transfer->host barrier and a fence wait;
+    //   back - the fill is read through the HELD pointer, the copy from a host-visible staging buffer;
+    //   steady - the same again with bytes the CPU wrote AFTER the import and its first GPU use (the
+    //          block at the end says why the first round alone is not enough).
+    Bool VkBufferManager::SelfTestWireImport(char* whyOut, Uint64 whyBytes) {
+        const auto answer = [&](Bool ok, const String& text) {
+            std::snprintf(whyOut, static_cast<SizeT>(whyBytes), "%s", text.c_str());
+            return ok;
+        };
+#if defined(__ANDROID__)
+        VulkanRenderer::WireAhbImport ctx;
+        if (pVulkanRenderer == nullptr || !pVulkanRenderer->GetWireAhbImport(ctx))
+            return answer(false, "the device did not take VK_ANDROID_external_memory_android_hardware_buffer");
+        constexpr Uint64 kBytes = 64 * 1024;
+        constexpr Uint32 kWords = 1024;          // one 4 KiB region
+        constexpr Uint64 kFillAt = 4 * 4096;     // region 4
+        constexpr Uint32 kFillWord = 0x6C0FFEE0u;
+        Vector<Uint32> pattern(kBytes / 4);
+        for (SizeT i = 0; i < pattern.size(); ++i) pattern[i] = 0x5A000000u ^ static_cast<Uint32>(i * 2654435761u);
+        MG_Remote::Transport::AdoptT0::HeldStore held;
+        String why;
+        if (!MG_Remote::Transport::AdoptT0::AllocateHeld(kBytes, pattern.data(), held, why))
+            return answer(false, "hold: " + why);
+        T0ImportedBuffer imported, staging;
+        VkCommandPool pool = VK_NULL_HANDLE;
+        VkFence fence = VK_NULL_HANDLE;
+        String verdict;
+        Bool ok = false;
+        do {
+            if (!T0ImportAhb(ctx, held.ahb, kBytes,
+                             VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+                                 VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
+                             imported, why)) {
+                verdict = "import: " + why;
+                break;
+            }
+            if (!imported.coherent) {
+                VkMappedMemoryRange range{VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE};
+                range.memory = imported.memory;
+                range.size = VK_WHOLE_SIZE;
+                vkInvalidateMappedMemoryRanges(ctx.device, 1, &range);
+            }
+            if (std::memcmp(imported.mapped, pattern.data(), kWords * 4) != 0) {
+                verdict = "read: the server's host map of the import does not hold the CPU's pattern";
+                break;
+            }
+            // A host-visible staging buffer for the GPU read's destination.
+            VkBufferCreateInfo stagingInfo{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
+            stagingInfo.size = kWords * 4;
+            stagingInfo.usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT;
+            stagingInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+            if (vkCreateBuffer(ctx.device, &stagingInfo, nullptr, &staging.buffer) != VK_SUCCESS) {
+                staging.buffer = VK_NULL_HANDLE;
+                verdict = "staging buffer";
+                break;
+            }
+            VkMemoryRequirements requirements{};
+            vkGetBufferMemoryRequirements(ctx.device, staging.buffer, &requirements);
+            const Int32 stagingType = T0PickMemoryType(ctx.memory, requirements.memoryTypeBits,
+                VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
+            VkMemoryAllocateInfo stagingAlloc{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
+            stagingAlloc.allocationSize = requirements.size;
+            stagingAlloc.memoryTypeIndex = static_cast<Uint32>(stagingType < 0 ? 0 : stagingType);
+            if (stagingType < 0 || vkAllocateMemory(ctx.device, &stagingAlloc, nullptr, &staging.memory) != VK_SUCCESS ||
+                vkBindBufferMemory(ctx.device, staging.buffer, staging.memory, 0) != VK_SUCCESS ||
+                vkMapMemory(ctx.device, staging.memory, 0, VK_WHOLE_SIZE, 0, &staging.mapped) != VK_SUCCESS) {
+                verdict = "staging memory";
+                break;
+            }
+            std::memset(staging.mapped, 0, kWords * 4);
+            VkCommandPoolCreateInfo poolInfo{VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO};
+            poolInfo.queueFamilyIndex = ctx.queueFamily;
+            poolInfo.flags = VK_COMMAND_POOL_CREATE_TRANSIENT_BIT;
+            VkCommandBuffer commands = VK_NULL_HANDLE;
+            VkCommandBufferAllocateInfo commandInfo{VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO};
+            commandInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+            commandInfo.commandBufferCount = 1;
+            VkFenceCreateInfo fenceInfo{VK_STRUCTURE_TYPE_FENCE_CREATE_INFO};
+            if (vkCreateCommandPool(ctx.device, &poolInfo, nullptr, &pool) != VK_SUCCESS) {
+                pool = VK_NULL_HANDLE;
+                verdict = "command pool";
+                break;
+            }
+            commandInfo.commandPool = pool;
+            if (vkAllocateCommandBuffers(ctx.device, &commandInfo, &commands) != VK_SUCCESS ||
+                vkCreateFence(ctx.device, &fenceInfo, nullptr, &fence) != VK_SUCCESS) {
+                verdict = "command buffer / fence";
+                break;
+            }
+            VkCommandBufferBeginInfo begin{VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO};
+            begin.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
+            vkBeginCommandBuffer(commands, &begin);
+            VkMemoryBarrier hostWrites{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
+            hostWrites.srcAccessMask = VK_ACCESS_HOST_WRITE_BIT;
+            hostWrites.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
+            EndActiveRenderPassOn(commands);
+            vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_HOST_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1,
+                                 &hostWrites, 0, nullptr, 0, nullptr);
+            VkBufferCopy region{0, 0, kWords * 4};
+            EndActiveRenderPassOn(commands);
+            vkCmdCopyBuffer(commands, imported.buffer, staging.buffer, 1, &region);
+            EndActiveRenderPassOn(commands);
+            vkCmdFillBuffer(commands, imported.buffer, kFillAt, kWords * 4, kFillWord);
+            VkMemoryBarrier toHost{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
+            toHost.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+            toHost.dstAccessMask = VK_ACCESS_HOST_READ_BIT;
+            EndActiveRenderPassOn(commands);
+            vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0, 1, &toHost,
+                                 0, nullptr, 0, nullptr);
+            vkEndCommandBuffer(commands);
+            VkSubmitInfo submit{VK_STRUCTURE_TYPE_SUBMIT_INFO};
+            submit.commandBufferCount = 1;
+            submit.pCommandBuffers = &commands;
+            if (vkQueueSubmit(ctx.queue, 1, &submit, fence) != VK_SUCCESS ||
+                vkWaitForFences(ctx.device, 1, &fence, VK_TRUE, 2'000'000'000ull) != VK_SUCCESS) {
+                verdict = "gpu: the submission did not complete within 2 s";
+                break;
+            }
+            Uint32 gpuRead = 0, heldSaw = 0;
+            const auto* copied = static_cast<const Uint32*>(staging.mapped);
+            const auto* heldWords = static_cast<const Uint32*>(held.ptr);
+            for (Uint32 i = 0; i < kWords; ++i) {
+                if (copied[i] == pattern[i]) ++gpuRead;
+                if (heldWords[kFillAt / 4 + i] == kFillWord) ++heldSaw;
+            }
+            ok = gpuRead == kWords && heldSaw == kWords;
+            verdict = "64 KiB AHB, lock held throughout: host map read ok (" +
+                      String(imported.coherent ? "coherent" : "non-coherent") + " memory), GPU read " +
+                      std::to_string(gpuRead) + "/1024 words of the CPU's pattern, GPU write seen through the "
+                      "held lock " + std::to_string(heldSaw) + "/1024 words";
+            if (!ok) break;
+
+            // STEADY STATE. The round above reads bytes the CPU wrote BEFORE the import, and an import
+            // can clean the CPU's caches on its way in, so it passes on a device whose GPU never sees a
+            // write made through the held lock afterwards - which is all of T0's real traffic (Mali-G1
+            // on a Dimensity: the first round passed, then Flywheel's staging scatter read stale bytes
+            // and the create-indirect trace lost every instanced object). So the CPU now writes a fresh
+            // pattern into region 1 and reads region 4 back into its cache, and one more submission
+            // copies region 1 out and overwrites region 4.
+            constexpr Uint64 kSteadyAt = 1 * 4096; // region 1
+            constexpr Uint32 kSteadyFillWord = 0x3D5EA11Eu;
+            auto* heldWrite = static_cast<volatile Uint32*>(held.ptr);
+            Uint32 cached = 0;
+            for (Uint32 i = 0; i < kWords; ++i) {
+                heldWrite[kSteadyAt / 4 + i] = 0xA5000000u ^ (i * 40503u);
+                cached ^= heldWrite[kFillAt / 4 + i];
+            }
+            (void)cached;
+            std::memset(staging.mapped, 0, kWords * 4);
+            if (vkResetFences(ctx.device, 1, &fence) != VK_SUCCESS ||
+                vkResetCommandPool(ctx.device, pool, 0) != VK_SUCCESS ||
+                vkBeginCommandBuffer(commands, &begin) != VK_SUCCESS) {
+                ok = false;
+                verdict = "steady: the second submission could not be recorded";
+                break;
+            }
+            EndActiveRenderPassOn(commands);
+            vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_HOST_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1,
+                                 &hostWrites, 0, nullptr, 0, nullptr);
+            VkBufferCopy steadyRegion{kSteadyAt, 0, kWords * 4};
+            EndActiveRenderPassOn(commands);
+            vkCmdCopyBuffer(commands, imported.buffer, staging.buffer, 1, &steadyRegion);
+            EndActiveRenderPassOn(commands);
+            vkCmdFillBuffer(commands, imported.buffer, kFillAt, kWords * 4, kSteadyFillWord);
+            EndActiveRenderPassOn(commands);
+            vkCmdPipelineBarrier(commands, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0, 1, &toHost,
+                                 0, nullptr, 0, nullptr);
+            vkEndCommandBuffer(commands);
+            if (vkQueueSubmit(ctx.queue, 1, &submit, fence) != VK_SUCCESS ||
+                vkWaitForFences(ctx.device, 1, &fence, VK_TRUE, 2'000'000'000ull) != VK_SUCCESS) {
+                ok = false;
+                verdict = "steady: the second submission did not complete within 2 s";
+                break;
+            }
+            Uint32 steadyRead = 0, steadySaw = 0;
+            for (Uint32 i = 0; i < kWords; ++i) {
+                if (copied[i] == (0xA5000000u ^ (i * 40503u))) ++steadyRead;
+                if (heldWords[kFillAt / 4 + i] == kSteadyFillWord) ++steadySaw;
+            }
+            ok = steadyRead == kWords && steadySaw == kWords;
+            verdict += "; with the import in use, GPU read " + std::to_string(steadyRead) +
+                       "/1024 words the CPU wrote after it, GPU write seen through the held lock " +
+                       std::to_string(steadySaw) + "/1024 words";
+        } while (false);
+        if (fence != VK_NULL_HANDLE) vkDestroyFence(ctx.device, fence, nullptr);
+        if (pool != VK_NULL_HANDLE) vkDestroyCommandPool(ctx.device, pool, nullptr);
+        T0DestroyImported(ctx.device, staging);
+        T0DestroyImported(ctx.device, imported);
+        MG_Remote::Transport::AdoptT0::ReleaseHeld(held);
+        return answer(ok, verdict);
+#else
+        return answer(false, "no AHardwareBuffer on this platform");
+#endif
+    }
+
+
+    void VkBufferManager::DescribeWireBufferHistory(MG_Pipe::MGPipeHandle res) const {
+        const auto& history = *g_wireBufferHistory;
+        Int createdAgo = -1, destroyedAgo = -1;
+        for (SizeT i = 0; i < std::min(history.createdCount, WireBufferHistory::kDepth); ++i) {
+            const auto& h = history.created[(history.createdCount - 1 - i) % WireBufferHistory::kDepth];
+            if (h.Slot == res.Slot && h.Gen == res.Gen) { createdAgo = static_cast<Int>(i); break; }
+        }
+        for (SizeT i = 0; i < std::min(history.destroyedCount, WireBufferHistory::kDepth); ++i) {
+            const auto& h = history.destroyed[(history.destroyedCount - 1 - i) % WireBufferHistory::kDepth];
+            if (h.Slot == res.Slot && h.Gen == res.Gen) { destroyedAgo = static_cast<Int>(i); break; }
+        }
+        MGLOG_E("Magma wire buffer {slot=%u, gen=%u}: live stores %zu; created %d creates ago, destroyed %d "
+                "destroys ago (-1 = not in the last %zu); %zu creates, %zu destroys in this session",
+                res.Slot, res.Gen, m_wireBuffers.size(), createdAgo, destroyedAgo, WireBufferHistory::kDepth,
+                history.createdCount, history.destroyedCount);
+    }
+
    void VkBufferManager::DestroyWireBuffer(MG_Pipe::MGPipeHandle res) {
+        {
+            auto& history = *g_wireBufferHistory;
+            history.destroyed[history.destroyedCount++ % WireBufferHistory::kDepth] = res;
+        }
        const auto found = m_wireBuffers.find(WireBufferKey(res));
        if (found == m_wireBuffers.end()) return;
-        DeferWireRelease(std::move(found->second.buffer), found->second.lastUseSerial);
+        DeferWireRelease(std::move(found->second.buffer),
+                         found->second.imported ? T0ReleaseSerial(found->second.lastUseSerial)
+                                                : found->second.lastUseSerial); // P11 B2, as Respecify
        m_wireBuffers.erase(found);
        ++m_sliceEpochCounter;
        PublishWireReclaimGauges();
@@ -629,6 +1098,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // budget's worth of orphans, not one per glBufferData.
        if (pVulkanRenderer == nullptr ||
            !pVulkanRenderer->WaitForSubmitIndex(pVulkanRenderer->GetSyncPointSubmitIndex(), UINT64_MAX, true)) {
+            if (LatchedOnDeviceLoss("buffer-deferred-watermark")) return;
            WireBufferSyncFatal("deferred-watermark");
        }
        ++m_wireDeferredSyncs;
@@ -649,6 +1119,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MG_Util::PipeStats::PublishGauge(Gauge::WireDeferredSyncs, m_wireDeferredSyncs);
    }

+    Bool VkBufferManager::LatchedOnDeviceLoss(const char* site) {
+        return pVulkanRenderer != nullptr && pVulkanRenderer->LatchWireDeviceLoss(site);
+    }
+
    void VkBufferManager::WireBufferSyncFatal(const char* site) {
        MGLOG_F("Magma: Fatal{ResourceUnavailable, \"buffer-write-sync\"} {site=%s}", site);
        std::abort();
@@ -722,7 +1196,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (!InitializeTransientArenas()) {
            return false;
        }
-        g_activeBufferManager = this;
+        g_activeBufferManager.Get() = this;
        MG_State::GLState::SetBufferBackendOps(&g_vulkanBufferBackendOps);
 #if MOBILEGL_BUILD_DISAGGREGATED
        RegisterWireResourceOps();
@@ -731,10 +1205,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    void VkBufferManager::Shutdown() {
-        if (g_activeBufferManager == this) {
-            g_activeBufferManager = nullptr;
+        if (g_activeBufferManager.Get() == this) {
+            g_activeBufferManager.Get() = nullptr;
 #if MOBILEGL_BUILD_DISAGGREGATED
-            if (MG_Pipe::MGPipeGetResourceOps() == &g_vulkanWireResourceOps) {
+            // Under a transport the table is the PROCESS's and serves every session: it dispatches
+            // to the calling session's own manager (WireManager), so one session's renderer going
+            // away must not take it from the others - their buffer creates would be dropped from
+            // then on, and every buffer born after that would be a handle the server never saw.
+            if (MG_Config::Transport == MG_Config::TransportMode::Monolith &&
+                MG_Pipe::MGPipeGetResourceOps() == &g_vulkanWireResourceOps) {
                MG_Pipe::MGPipeSetResourceOps(nullptr);
            }
 #endif
@@ -1027,6 +1506,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkMemoryBarrier beforeBarrier{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        beforeBarrier.srcAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
        beforeBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1,
                             &beforeBarrier, 0, nullptr, 0, nullptr);

@@ -1034,11 +1514,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        region.srcOffset = staging.offset;
        region.dstOffset = static_cast<VkDeviceSize>(offset);
        region.size = static_cast<VkDeviceSize>(size);
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdCopyBuffer(commandBuffer, staging.buffer, resource.buffer.GetHandle(), 1, &region);

        VkMemoryBarrier afterBarrier{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        afterBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
        afterBarrier.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, 1,
                             &afterBarrier, 0, nullptr, 0, nullptr);

@@ -1073,6 +1555,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkMemoryBarrier beforeBarrier{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        beforeBarrier.srcAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
        beforeBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1,
                             &beforeBarrier, 0, nullptr, 0, nullptr);

@@ -1080,11 +1563,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        region.srcOffset = staging.offset;
        region.dstOffset = static_cast<VkDeviceSize>(offset);
        region.size = static_cast<VkDeviceSize>(size);
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdCopyBuffer(commandBuffer, staging.buffer, resource.buffer.GetHandle(), 1, &region);

        VkMemoryBarrier afterBarrier{VK_STRUCTURE_TYPE_MEMORY_BARRIER};
        afterBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
        afterBarrier.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, 1,
                             &afterBarrier, 0, nullptr, 0, nullptr);

@@ -110,6 +110,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // frontend BufferObject. Acquires expose the full GPU store, without a CPU
        // pointer: CPU consumers must use ReadWireBuffer for ordered, current bytes.
        Bool AcquireWireSlice(BufferKind kind, MG_Pipe::MGPipeHandle res, BufferSlice& outSlice);
+        // The handle names a store this session created (whatever its size, zero included).
+        Bool IsKnownWireBuffer(MG_Pipe::MGPipeHandle res) { return FindWireBuffer(res) != nullptr; }
+        // Names, for a Fatal about an unknown handle, whether this manager ever saw it.
+        void DescribeWireBufferHistory(MG_Pipe::MGPipeHandle res) const;
        Bool ReadWireBuffer(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size, void* dst);
        Bool CopyWireBufferRangeToSlice(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size,
                                        const BufferSlice& dst);
@@ -125,6 +129,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        Bool CopyWireBufferSubWordRangeToSlice(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size,
                                               Uint32 frameIndex, const BufferSlice& dst, Uint64 dstSkip);
        void MarkWireBufferGpuWritten(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size);
+        // P8-D: whether a shader write was marked on this store since the last indirect draw
+        // read it natively, clearing the mark. The caller records the INDIRECT_COMMAND_READ
+        // barrier before the draw when this answers true. Transfer writes need no mark: their own
+        // after-barriers already make them visible to MEMORY_READ at ALL_COMMANDS.
+        Bool TakeWireIndirectReadBarrier(MG_Pipe::MGPipeHandle res);
+        // P8-D: every WaitForWireBufferHostAccess that actually waited, over this manager's life -
+        // the attribution base a caller reads before and after its own ReadWireBuffer.
+        Uint64 GetWireHostWaitCount() const { return m_wireHostWaits; }

        // Resource-op entry points. All run on the server apply owner.
        void CreateWireBuffer(MG_Pipe::MGPipeHandle res, const MG_Pipe::MGPResourceDesc& desc);
@@ -134,6 +146,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        void FlushWireBuffer(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size, const void* bytes);
        void ReadbackWireBuffer(MG_Pipe::MGPipeHandle res, Uint64 offset, Uint64 size);
        void DestroyWireBuffer(MG_Pipe::MGPipeHandle res);
+        // P11 B2 (T0): make the client's AHardwareBuffer `ahb` (a BLOB of `size` bytes) buffer
+        // `res`'s store - a VkBuffer on memory imported through
+        // VK_ANDROID_external_memory_android_hardware_buffer, host-visible and mapped like every
+        // wire store, so every wire path (draw binds, WriteWireBuffer's ordered copy, CPU readers)
+        // applies unchanged. The T2 store it replaces goes through DeferWireRelease. False = not
+        // imported (the T2 store is untouched and the client is DECLINED).
+        Bool ImportWireBuffer(MG_Pipe::MGPipeHandle res, void* ahb, Uint64 size);
+        // The POST self-test of the sustained-lock pattern on this device (see the definition).
+        static Bool SelfTestWireImport(char* why, Uint64 whyBytes);
 #endif

        // Recreate all per-frame transient arenas
@@ -245,12 +266,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            // Only ranges actually submitted by resource_subdata are covered. No
            // shadow is retained: flush cannot replay stale bytes over GPU writes.
            Vector<Range1D> stagedCoverage;
+            // P11 B2 (T0): `buffer` is the client's imported AHardwareBuffer. Its readback posts
+            // no bytes (the client reads its own pages once the wait below it is done).
+            Bool imported = false;
+            // P8-D: a shader write was marked (MarkWireBufferGpuWritten) and no indirect draw has
+            // recorded its INDIRECT_COMMAND_READ barrier since (TakeWireIndirectReadBarrier).
+            Bool indirectReadBarrierPending = false;
        };
+        // P8-D: see GetWireHostWaitCount.
+        Uint64 m_wireHostWaits = 0;
        static Uint64 WireBufferKey(MG_Pipe::MGPipeHandle res) {
            return (static_cast<Uint64>(res.Gen) << 32) | res.Slot;
        }
        WireBufferResource* FindWireBuffer(MG_Pipe::MGPipeHandle res);
        Bool WaitForWireBufferHostAccess(WireBufferResource& resource);
+        // P11 B2: the serial a T0 store's release waits behind (never "destroy at once").
+        Uint64 T0ReleaseSerial(Uint64 lastUseSerial) const;

        // P7 wave 4 (M2), ID-P7-27: THE WIRE ARM'S ORPHANS NEED A RECLAIM THAT IS NOT A FRAME
        // BOUNDARY.
@@ -331,6 +362,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // for the census to count), shared by the host write that cannot wait for the GPU and
        // the watermark that cannot. `site` says which.
        [[noreturn]] static void WireBufferSyncFatal(const char* site);
+        // True when the wait or copy that just failed failed because the device is lost: the
+        // session has been latched (VulkanRenderer::LatchWireDeviceLoss) and the caller returns
+        // instead of taking its Fatal.
+        static Bool LatchedOnDeviceLoss(const char* site);
        // The wbuf[] gauges (PipeStats.h, Gauge::WireBuffers..WireDeferredSyncs). The peaks are
        // taken at the two points the numbers can rise - a park and a mint - and published when
        // the stats channel is on; MagmaWireReclaimScenario reads them off the server's line.
@@ -8,13 +8,45 @@

 #include "VkBufferObject.h"

+#if MOBILEGL_BUILD_DISAGGREGATED
+#include <MG_Remote/Transport/AdoptT0.h>
+#endif
+
 namespace MobileGL::MG_Backend::DirectVulkan {
+#if MOBILEGL_BUILD_DISAGGREGATED
+    void VkBufferObject::MoveExternalFrom(VkBufferObject& other) {
+        m_device = other.m_device;
+        m_externalMemory = other.m_externalMemory;
+        m_externalAhb = other.m_externalAhb;
+        m_externalCoherent = other.m_externalCoherent;
+        other.m_device = VK_NULL_HANDLE;
+        other.m_externalMemory = VK_NULL_HANDLE;
+        other.m_externalAhb = nullptr;
+        other.m_externalCoherent = true;
+    }
+
+    void VkBufferObject::AdoptExternal(VkDevice device, VkBuffer buffer, VkDeviceMemory memory, VkDeviceSize size,
+                                       void* mapped, Bool coherent, void* ahb) {
+        Destroy();
+        m_device = device;
+        m_buffer = buffer;
+        m_externalMemory = memory;
+        m_mappedData = mapped;
+        m_size = size;
+        m_externalCoherent = coherent;
+        m_externalAhb = ahb;
+    }
+#endif
+
    VkBufferObject::VkBufferObject(VkBufferObject&& other) noexcept {
        m_allocator = other.m_allocator;
        m_buffer = other.m_buffer;
        m_allocation = other.m_allocation;
        m_mappedData = other.m_mappedData;
        m_size = other.m_size;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        MoveExternalFrom(other);
+#endif

        other.m_allocator = nullptr;
        other.m_buffer = VK_NULL_HANDLE;
@@ -35,6 +67,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        m_allocation = other.m_allocation;
        m_mappedData = other.m_mappedData;
        m_size = other.m_size;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        MoveExternalFrom(other);
+#endif

        other.m_allocator = nullptr;
        other.m_buffer = VK_NULL_HANDLE;
@@ -89,6 +124,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    void VkBufferObject::Destroy() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        if (m_externalMemory != VK_NULL_HANDLE) {
+            // P11 B2 (T0): an imported AHardwareBuffer store. Reached through the wire arm's
+            // deferred release, i.e. once no submission can still name it.
+            if (m_mappedData != nullptr) vkUnmapMemory(m_device, m_externalMemory);
+            if (m_buffer != VK_NULL_HANDLE) vkDestroyBuffer(m_device, m_buffer, nullptr);
+            vkFreeMemory(m_device, m_externalMemory, nullptr);
+            MG_Remote::Transport::AdoptT0::ReleaseImported(m_externalAhb);
+            m_mappedData = nullptr;
+            m_buffer = VK_NULL_HANDLE;
+            m_externalMemory = VK_NULL_HANDLE;
+            m_externalAhb = nullptr;
+            m_device = VK_NULL_HANDLE;
+            m_externalCoherent = true;
+            m_size = 0;
+            return;
+        }
+#endif
        Unmap();
        if (m_allocator != nullptr && m_buffer != VK_NULL_HANDLE && m_allocation != nullptr) {
            vmaDestroyBuffer(m_allocator, m_buffer, m_allocation);
@@ -117,6 +170,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    void VkBufferObject::Unmap() {
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // An imported store stays mapped for its life (Destroy unmaps it).
+        if (m_externalMemory != VK_NULL_HANDLE) return;
+#endif
        if (!IsValid() || m_mappedData == nullptr) {
            m_mappedData = nullptr;
            return;
@@ -143,6 +200,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        Memcpy(static_cast<Uint8*>(mapped) + offset, data, static_cast<SizeT>(size));
+#if MOBILEGL_BUILD_DISAGGREGATED
+        if (m_externalMemory != VK_NULL_HANDLE) {
+            if (m_externalCoherent) return true;
+            VkMappedMemoryRange range{VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE};
+            range.memory = m_externalMemory;
+            range.offset = 0;
+            range.size = VK_WHOLE_SIZE;
+            return vkFlushMappedMemoryRanges(m_device, 1, &range) == VK_SUCCESS;
+        }
+#endif
        const VkResult flushResult = vmaFlushAllocation(m_allocator, m_allocation, offset, size);
        if (flushResult != VK_SUCCESS) {
            MGLOG_E_ONCE("VkBufferObject::Upload failed: vmaFlushAllocation returned %d", flushResult);
@@ -167,6 +234,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (resolvedSize == 0) {
            return true;
        }
+#if MOBILEGL_BUILD_DISAGGREGATED
+        if (m_externalMemory != VK_NULL_HANDLE) {
+            if (m_externalCoherent) return true;
+            VkMappedMemoryRange range{VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE};
+            range.memory = m_externalMemory;
+            range.offset = 0;
+            range.size = VK_WHOLE_SIZE;
+            return vkInvalidateMappedMemoryRanges(m_device, 1, &range) == VK_SUCCESS;
+        }
+#endif

        const VkResult result = vmaInvalidateAllocation(m_allocator, m_allocation, offset, resolvedSize);
        if (result != VK_SUCCESS) {
@@ -64,7 +64,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }
        void* GetMappedData() const { return m_mappedData; }
        Bool IsMapped() const { return m_mappedData != nullptr; }
-        Bool IsValid() const { return m_allocator != nullptr && m_buffer != VK_NULL_HANDLE && m_allocation != nullptr; }
+        Bool IsValid() const {
+#if MOBILEGL_BUILD_DISAGGREGATED
+            if (m_externalMemory != VK_NULL_HANDLE) return m_buffer != VK_NULL_HANDLE;
+#endif
+            return m_allocator != nullptr && m_buffer != VK_NULL_HANDLE && m_allocation != nullptr;
+        }
+
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // P11 B2 (T0): a store whose memory is NOT this allocator's - a client's AHardwareBuffer
+        // imported with VK_ANDROID_external_memory_android_hardware_buffer into `memory` (a
+        // dedicated allocation bound to `buffer`, mapped persistently at `mapped`). It behaves
+        // like any mapped store (GetSlice, Upload, Invalidate), and Destroy frees the buffer and
+        // the memory and drops `ahb` - which is why the store goes through the wire arm's
+        // serial-gated deferred release like every other: the AHardwareBuffer is let go only once
+        // no submission can still name it. `coherent` = the memory type is HOST_COHERENT.
+        void AdoptExternal(VkDevice device, VkBuffer buffer, VkDeviceMemory memory, VkDeviceSize size,
+                           void* mapped, Bool coherent, void* ahb);
+        Bool IsExternal() const { return m_externalMemory != VK_NULL_HANDLE; }
+#endif

    private:
        VmaAllocator m_allocator = nullptr;
@@ -72,5 +90,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VmaAllocation m_allocation = nullptr;
        void* m_mappedData = nullptr;
        VkDeviceSize m_size = 0;
+#if MOBILEGL_BUILD_DISAGGREGATED
+        VkDevice m_device = VK_NULL_HANDLE;
+        VkDeviceMemory m_externalMemory = VK_NULL_HANDLE;
+        void* m_externalAhb = nullptr;
+        Bool m_externalCoherent = true;
+        void MoveExternalFrom(VkBufferObject& other);
+#endif
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -7,6 +7,8 @@
 // End of Source File Header

 #include "VkRenderPassManager.h"
+#include "RenderPassGuard.h"
+#include <atomic>

 #include "MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h"
 #include "MG_State/GLState/TextureState/TextureObject2D.h"
@@ -235,11 +237,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VkClearManager& clearManager, VkTextureManager& textureManager, SwapchainObject& swapchainObject):
        m_device(device), m_physicalDevice(physicalDevice), m_allocator(allocator), m_config(config),
        m_clearManager(clearManager), m_textureManager(textureManager), m_swapchainObject(swapchainObject) {
-        RenderPassEntry::s_device = m_device;
-        s_clearManager = &m_clearManager;
-        s_textureManager = &m_textureManager;
-        s_swapchainObject = &m_swapchainObject;
-        s_renderPassManager = this;
+        RenderPassEntry::s_device.Get() = m_device;
+        s_clearManager.Get() = &m_clearManager;
+        s_textureManager.Get() = &m_textureManager;
+        s_swapchainObject.Get() = &m_swapchainObject;
+        s_renderPassManager.Get() = this;
    }

    VkRenderPassManager::~VkRenderPassManager() {}
@@ -257,8 +259,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        CollectDeferredRenderbufferReleases(/*destroyAll=*/true); // caller guarantees device idle
        m_pendingRenderbufferClears.clear();
        RenderPassEntry::s_textureResourcesScratch.clear();
-        s_activeRenderPass = {};
-        s_hasActiveRenderPass = false;
+        s_activeRenderPass.Get() = {};
+        s_hasActiveRenderPass.Get() = false;
        m_rpFastValid = false;
    }

@@ -607,6 +609,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        const Bool isDefaultFbo = fbo.IsDefaultFramebuffer();
        if (isDefaultFbo) {
            XXHASH_VERIFY(XXH64_update(m_hashState, &swapchainImageIndex, sizeof(swapchainImageIndex)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &m_defaultFramebufferTarget, sizeof(m_defaultFramebufferTarget)));
        }
        // sRGB attachments switch between their sRGB and UNORM-twin views with this
        // capability (ResolveSrgbAttachmentWriteFormat), changing the render pass formats.
@@ -1528,6 +1531,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return &insertedIt->second;
    }

+    void VkRenderPassManager::PurgeRenderPasses() {
+        Vector<VkRenderPass> destroyedRenderPasses;
+        destroyedRenderPasses.reserve(m_renderPasses.size());
+        for (auto& [_, entry] : m_renderPasses) destroyedRenderPasses.push_back(entry.renderPass);
+        m_renderPasses.clear();
+        m_rpFastValid = false;
+        if (!destroyedRenderPasses.empty() && m_evictionObserver != nullptr) {
+            m_evictionObserver->OnRenderPassesDestroyed(destroyedRenderPasses);
+        }
+    }
+
    void VkRenderPassManager::OnPresent() {
        ++m_frameCounter;

@@ -1554,9 +1568,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // immediately - and a single batched notification costs one pipeline-cache
        // scan instead of one per evicted pass.
        Vector<VkRenderPass> destroyedRenderPasses;
-        const Uint64 activeHash = s_hasActiveRenderPass ? s_activeRenderPass.hash : 0;
+        const Uint64 activeHash = s_hasActiveRenderPass.Get() ? s_activeRenderPass.Get().hash : 0;
        for (auto it = m_renderPasses.begin(); it != m_renderPasses.end();) {
-            const Bool isActive = s_hasActiveRenderPass && it->first == activeHash;
+            const Bool isActive = s_hasActiveRenderPass.Get() && it->first == activeHash;
            if (!isActive && m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
                if (m_rpFastValid && m_rpFastRenderPassHash == it->first) {
                    m_rpFastValid = false;
@@ -1601,8 +1615,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                // texture path below is immune because it re-reads the live payload).
                // Same defense as ClearAttachmentsOnActiveRenderPass: prefer the live
                // pending clear, fall back to the snapshot only when none is queued.
-                if (s_renderPassManager != nullptr &&
-                    s_renderPassManager->GetPendingRenderbufferClear(pending.renderbuffer, clearPayload)) {
+                if (s_renderPassManager.Get() != nullptr &&
+                    s_renderPassManager.Get()->GetPendingRenderbufferClear(pending.renderbuffer, clearPayload)) {
                    if ((clearPayload.mask & GL_COLOR_BUFFER_BIT) != 0 && pending.renderbuffer != nullptr &&
                        MG_Util::GetBaseInternalFormatComponentCount(pending.renderbuffer->GetInternalFormat()) ==
                            3) {
@@ -1614,7 +1628,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                }
            } else {
                if (pending.key.texture == nullptr ||
-                    !s_clearManager->GetPendingClear(pending.key, clearPayload, liveTexture)) {
+                    !s_clearManager.Get()->GetPendingClear(pending.key, clearPayload, liveTexture)) {
                    continue;
                }
            }
@@ -1636,48 +1650,58 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);
        // Pre-pass stream bookkeeping: this pass's attachment images are now
        // referenced by the open frame recording.
-        if (s_textureManager != nullptr) {
+        if (s_textureManager.Get() != nullptr) {
            for (const auto& tracked : renderPassEntry.trackedAttachmentLayouts) {
                if (tracked.target == TrackedAttachmentTarget::Texture) {
                    if (const auto texture = tracked.texture.lock()) {
-                        s_textureManager->StampTextureRecordingUse(texture.get());
+                        s_textureManager.Get()->StampTextureRecordingUse(texture.get());
                    }
                }
            }
        }
        for (const auto& pending: renderPassEntry.pendingClearAttachments) {
            if (pending.hasInlinePayload) {
-                if (s_renderPassManager != nullptr) {
-                    s_renderPassManager->PopPendingRenderbufferClear(pending.renderbuffer);
+                if (s_renderPassManager.Get() != nullptr) {
+                    s_renderPassManager.Get()->PopPendingRenderbufferClear(pending.renderbuffer);
                }
            } else {
-                s_clearManager->PopPendingClear(pending.key);
+                s_clearManager.Get()->PopPendingClear(pending.key);
            }
        }
-        s_activeRenderPass.hash = renderPassEntry.hash;
-        s_activeRenderPass.compatibilityHash = renderPassEntry.compatibilityHash;
-        s_activeRenderPass.trackedAttachmentLayouts = renderPassEntry.trackedAttachmentLayouts;
-        s_activeRenderPass.extent = renderPassEntry.extent;
-        s_hasActiveRenderPass = true;
+        s_activeRenderPass.Get().commandBuffer = commandBuffer;
+        s_activeRenderPass.Get().hash = renderPassEntry.hash;
+        s_activeRenderPass.Get().compatibilityHash = renderPassEntry.compatibilityHash;
+        s_activeRenderPass.Get().trackedAttachmentLayouts = renderPassEntry.trackedAttachmentLayouts;
+        s_activeRenderPass.Get().extent = renderPassEntry.extent;
+        s_hasActiveRenderPass.Get() = true;

        return true;
    }

    Bool VkRenderPassManager::EndRenderPass(VkCommandBuffer commandBuffer) {
+        BumpWireImageWriteEpoch();
        auto* activeRenderPass = GetActiveRenderPass();
+        Vector<TrackedAttachmentLayoutInfo> trackedAttachmentLayouts;
+        if (activeRenderPass != nullptr) {
+            trackedAttachmentLayouts = Move(activeRenderPass->trackedAttachmentLayouts);
+        }
        vkCmdEndRenderPass(commandBuffer);
+        // Layout reconciliation below may record barriers, whose guard closes an active
+        // pass. Detach the ended pass first so that guard cannot recursively end it again.
+        s_activeRenderPass.Get() = {};
+        s_hasActiveRenderPass.Get() = false;
        // The fast-path memo reuses the ACTIVE render pass; once the pass ends it must not carry
        // over (the next span may be a different FBO resolved before its render pass is begun).
-        if (s_renderPassManager != nullptr) {
-            s_renderPassManager->m_rpFastValid = false;
+        if (s_renderPassManager.Get() != nullptr) {
+            s_renderPassManager.Get()->m_rpFastValid = false;
        }
-        if (activeRenderPass != nullptr && !activeRenderPass->trackedAttachmentLayouts.empty()) {
-            for (const auto& trackedAttachment : activeRenderPass->trackedAttachmentLayouts) {
+        if (!trackedAttachmentLayouts.empty()) {
+            for (const auto& trackedAttachment : trackedAttachmentLayouts) {
                switch (trackedAttachment.target) {
                    case TrackedAttachmentTarget::Texture:
-                        MOBILEGL_ASSERT(s_textureManager != nullptr, "EndRenderPass: texture manager is null");
+                        MOBILEGL_ASSERT(s_textureManager.Get() != nullptr, "EndRenderPass: texture manager is null");
                        if (const auto texture = trackedAttachment.texture.lock()) {
-                            s_textureManager->UpdateTrackedImageLayoutAfterAttachmentWrite(
+                            s_textureManager.Get()->UpdateTrackedImageLayoutAfterAttachmentWrite(
                                commandBuffer,
                                texture.get(),
                                trackedAttachment.textureMipLevel,
@@ -1685,27 +1709,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                        }
                        break;
                    case TrackedAttachmentTarget::Renderbuffer:
-                        MOBILEGL_ASSERT(s_renderPassManager != nullptr, "EndRenderPass: render pass manager is null");
+                        MOBILEGL_ASSERT(s_renderPassManager.Get() != nullptr, "EndRenderPass: render pass manager is null");
                        if (const auto renderbuffer = trackedAttachment.renderbuffer.lock()) {
                            auto resourceIt =
-                                s_renderPassManager->m_renderbufferResources.find(renderbuffer.get());
-                            if (resourceIt != s_renderPassManager->m_renderbufferResources.end()) {
+                                s_renderPassManager.Get()->m_renderbufferResources.find(renderbuffer.get());
+                            if (resourceIt != s_renderPassManager.Get()->m_renderbufferResources.end()) {
                                resourceIt->second.layout = trackedAttachment.finalLayout;
                            }
                        }
                        break;
                    case TrackedAttachmentTarget::SwapchainColor:
-                        MOBILEGL_ASSERT(s_swapchainObject != nullptr, "EndRenderPass: swapchain object is null");
-                        s_swapchainObject->SetImageLayout(trackedAttachment.swapchainImageIndex, trackedAttachment.finalLayout);
+                        MOBILEGL_ASSERT(s_swapchainObject.Get() != nullptr, "EndRenderPass: swapchain object is null");
+                        s_swapchainObject.Get()->SetImageLayout(trackedAttachment.swapchainImageIndex, trackedAttachment.finalLayout);
                        // The pass stored into the attachment: its content is defined
                        // until the image is next presented.
-                        s_swapchainObject->SetImageContentDefined(trackedAttachment.swapchainImageIndex, true);
+                        s_swapchainObject.Get()->SetImageContentDefined(trackedAttachment.swapchainImageIndex, true);
                        break;
                    case TrackedAttachmentTarget::SwapchainDepthStencil:
-                        MOBILEGL_ASSERT(s_swapchainObject != nullptr, "EndRenderPass: swapchain object is null");
-                        s_swapchainObject->SetDepthStencilImageLayout(trackedAttachment.swapchainImageIndex,
+                        MOBILEGL_ASSERT(s_swapchainObject.Get() != nullptr, "EndRenderPass: swapchain object is null");
+                        s_swapchainObject.Get()->SetDepthStencilImageLayout(trackedAttachment.swapchainImageIndex,
                                                                      trackedAttachment.finalLayout);
-                        s_swapchainObject->SetDepthStencilContentDefined(trackedAttachment.swapchainImageIndex, true);
+                        s_swapchainObject.Get()->SetDepthStencilContentDefined(trackedAttachment.swapchainImageIndex, true);
                        break;
                    default:
                        MOBILEGL_ASSERT(false, "EndRenderPass: unsupported tracked attachment target=%d",
@@ -1714,12 +1738,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                }
            }
        }
-        s_activeRenderPass = {};
-        s_hasActiveRenderPass = false;
        return true;
    }

    ActiveRenderPassInfo* VkRenderPassManager::GetActiveRenderPass() {
-        return s_hasActiveRenderPass ? &s_activeRenderPass : nullptr;
+        return s_hasActiveRenderPass.Get() ? &s_activeRenderPass.Get() : nullptr;
    }
+    namespace {
+        std::atomic<Uint64> g_wireImageWriteEpoch{1};
+    }
+    Uint64 WireImageWriteEpoch() { return g_wireImageWriteEpoch.load(std::memory_order_relaxed); }
+    void BumpWireImageWriteEpoch() { g_wireImageWriteEpoch.fetch_add(1, std::memory_order_relaxed); }
+
+    void EndActiveRenderPassOn(VkCommandBuffer commandBuffer) {
+        BumpWireImageWriteEpoch();
+        const auto* active = VkRenderPassManager::GetActiveRenderPass();
+        if (active != nullptr && active->commandBuffer == commandBuffer && commandBuffer != VK_NULL_HANDLE)
+            VkRenderPassManager::EndRenderPass(commandBuffer);
+    }
+
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -66,8 +66,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VkExtent2D swapchainExtent);

    struct RenderPassEntry {
-        static inline VkDevice s_device;
-        static inline Vector<VkTextureManager::TextureResource*> s_textureResourcesScratch;
+        // Per Magma session (MagmaSession.h): each session's renderer is its own device.
+        static inline SessionLocal<VkDevice> s_device;
+        static inline thread_local Vector<VkTextureManager::TextureResource*> s_textureResourcesScratch;
        Uint64 hash = 0;
        VkRenderPass renderPass = VK_NULL_HANDLE;
        VkFramebuffer framebuffer = VK_NULL_HANDLE;
@@ -165,10 +166,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        ~RenderPassEntry() {
            if (renderPass != VK_NULL_HANDLE) {
-                vkDestroyRenderPass(s_device, renderPass, nullptr);
+                vkDestroyRenderPass(s_device.Get(), renderPass, nullptr);
            }
            if (framebuffer != VK_NULL_HANDLE) {
-                vkDestroyFramebuffer(s_device, framebuffer, nullptr);
+                vkDestroyFramebuffer(s_device.Get(), framebuffer, nullptr);
            }
        }

@@ -184,6 +185,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    struct ActiveRenderPassInfo {
        Uint64 hash = 0;
        Uint64 compatibilityHash = 0;
+        VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
        Vector<TrackedAttachmentLayoutInfo> trackedAttachmentLayouts;
        IntVec2 extent = {0, 0};

@@ -226,6 +228,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {

        Bool Initialize();
        void Shutdown();
+        // The surface target the default framebuffer currently names (VulkanRenderer's surface
+        // targets): a default-framebuffer pass of one surface is not another's.
+        void SetDefaultFramebufferTarget(Uint64 serial) {
+            m_defaultFramebufferTarget = serial;
+            m_rpFastValid = false;
+        }
+        // Destroys every cached pass and framebuffer (not the renderbuffers). The device must be
+        // idle and no pass may be open: a surface target's image views are going away.
+        void PurgeRenderPasses();

        HashType ComputeHash(
            const MG_State::GLState::FramebufferObject& fbo,
@@ -302,6 +313,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // is reused WITHOUT recomputing the expensive per-draw hash. Invalidated by FBO switch /
        // version change, swapchain rotation, any attachment image recreation (the two epochs),
        // or a pending clear. Portable to Vulkan 1.1 (no dynamic_rendering / imageless FB needed).
+        Uint64 m_defaultFramebufferTarget = 0;
        Bool m_rpFastValid = false;
        const MG_State::GLState::FramebufferObject* m_rpFastFbo = nullptr;
        // The FBO's never-reused lifetime id joins the raw pointer + Uint16 version:
@@ -411,12 +423,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        void DeferRenderbufferBackingRelease(RenderbufferResource& resource);
        void CollectDeferredRenderbufferReleases(Bool destroyAll);

-        static inline XXH64_state_t* m_hashState = XXH64_createState();
-        static inline ActiveRenderPassInfo s_activeRenderPass{};
-        static inline Bool s_hasActiveRenderPass = false;
-        static inline VkClearManager* s_clearManager = nullptr;
-        static inline VkTextureManager* s_textureManager = nullptr;
-        static inline SwapchainObject* s_swapchainObject = nullptr;
-        static inline VkRenderPassManager* s_renderPassManager = nullptr;
+        static inline thread_local XXH64_state_t* m_hashState = XXH64_createState();
+        // Per Magma session (MagmaSession.h): one session's open render pass and managers are not
+        // another's.
+        static inline SessionLocal<ActiveRenderPassInfo> s_activeRenderPass;
+        static inline SessionLocal<Bool> s_hasActiveRenderPass;
+        static inline SessionLocal<VkClearManager*> s_clearManager;
+        static inline SessionLocal<VkTextureManager*> s_textureManager;
+        static inline SessionLocal<SwapchainObject*> s_swapchainObject;
+        static inline SessionLocal<VkRenderPassManager*> s_renderPassManager;
    };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -138,6 +138,6 @@ private:
    UnorderedMap<Uint64, SamplerCacheEntry> m_samplers;
    // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
    Uint64 m_frameBoundaryCounter = 0;
-    static inline XXH64_state_t* m_hashState = XXH64_createState();
+    static inline thread_local XXH64_state_t* m_hashState = XXH64_createState();
 };
 } // namespace MobileGL::MG_Backend::DirectVulkan
@@ -6,6 +6,7 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header

+#include "RenderPassGuard.h"
 #include "VkTextureManager.h"

 #include "ProgramFactory.h"
@@ -22,10 +23,16 @@
 // (shape) and the server's staged-texture store (texels).
 #include <MG_Pipe/PipeApply.h>
 #include <MG_Remote/Server/StagedTextureStore.h>
+// Shared images: a texture whose level 0 is a server-allocated AHardwareBuffer.
+#include <MG_Remote/Server/SharedImageRegistry.h>
 #include "../DirectVulkan.h"
 // P7 wave 2 package B3: rule I's tally for the silent exits on this file's wire arm.
 #include "WireDeclineTally.h"
+#if defined(__ANDROID__)
+#include <android/hardware_buffer.h>
 #endif
+#endif
+#include <string>
 #include <algorithm>
 #include <cstdio>
 #include <cstdlib>
@@ -42,9 +49,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    // their feature is off (VUID-vkCmdPipelineBarrier-srcStageMask-04090/-04091), and ALL_GRAPHICS
    // would also serialize against non-shader stages. The default only matters before a device
    // exists, when nothing records barriers.
-    static VkPipelineStageFlags s_sampledReadStages =
-        VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-        VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+    // Per Magma session: each session's renderer is its own device.
+    struct SampledReadStages {
+        VkPipelineStageFlags value = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
+                                     VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+    };
+    static SessionLocal<SampledReadStages> s_sampledReadStagesSlot;
+
+    static VkImageUsageFlags ResolveTextureUsage(VkImageAspectFlags aspect,
+                                                 VkFormatFeatureFlags formatFeatures,
+                                                 Bool supportsStorageImage) {
+        VkImageUsageFlags usage = VK_IMAGE_USAGE_SAMPLED_BIT;
+        if (supportsStorageImage) usage |= VK_IMAGE_USAGE_STORAGE_BIT;
+        // A color aspect does not imply renderability (for example, RGB9_E5).
+        if ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) &&
+            (formatFeatures & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT)) {
+            usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+        }
+        if (aspect & (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
+            usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
+        }
+        return usage;
+    }

    static Uint32 ComputeFullMipLevelCount(const IntVec3& baseTexelSize) {
        Int maxDimension = std::max<Int>(baseTexelSize.x(),
@@ -203,7 +229,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        case VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL:
        case VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL:
        case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
-            outSrcStageMask = s_sampledReadStages;
+            outSrcStageMask = s_sampledReadStagesSlot->value;
            outSrcAccessMask = VK_ACCESS_SHADER_READ_BIT;
            return;
        case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL:
@@ -267,7 +293,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        case VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL:
        case VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL:
        case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
-            outDstStageMask = s_sampledReadStages;
+            outDstStageMask = s_sampledReadStagesSlot->value;
            outDstAccessMask = VK_ACCESS_SHADER_READ_BIT;
            return;
        case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL:
@@ -367,6 +393,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            copyRegions.push_back(copy);
        }

+        EndActiveRenderPassOn(commandBuffer);
        vkCmdCopyImage(commandBuffer,
                       oldResource.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                       newResource.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
@@ -646,7 +673,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        m_commandPool = initInfo.commandPool;
        m_graphicsQueue = initInfo.graphicsQueue;
        m_imageFormatListSupported = initInfo.imageFormatListSupported;
-        s_sampledReadStages = initInfo.sampledReadStageMask;
+        s_sampledReadStagesSlot->value = initInfo.sampledReadStageMask;
        m_currentFrameIndex = 0;
        m_deferredReleases.clear();
        m_deferredReleases.resize(initInfo.frameCount);
@@ -659,8 +686,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MOBILEGL_ASSERT(initInfo.frameCount > 0,
                        "VkTextureManager::Initialize failed: frameCount must be > 0");

-        TextureResource::s_device = m_device;
-        TextureResource::s_allocator = m_allocator;
+        TextureResource::s_device.Get() = m_device;
+        TextureResource::s_allocator.Get() = m_allocator;

        // Own pool for the recycled upload-batch command buffers. Parking a
        // dozen reset-but-alive command buffers in the renderer's shared pool
@@ -698,6 +725,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // Destroy images while the device/allocator still exist, including after context death.
        m_wireTextureResources.clear();
        m_wireRenderbufferResources.clear();
+        m_sharedImageUses.clear();
 #endif
        m_aliveObjects.clear();
        m_storageImageTextures.clear();
@@ -1333,6 +1361,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }
    }

+    VkTextureManager::TextureResource* VkTextureManager::FindTextureResource(
+        MG_State::GLState::ITextureObject& textureOrView) {
+        auto it = m_textureResources.find(MakeTextureIdentity(&StorageTextureOf(textureOrView)));
+        return it != m_textureResources.end() ? &it->second : nullptr;
+    }
+
+    Bool VkTextureManager::HasLiveImage(MG_State::GLState::ITextureObject& textureOrView) const {
+        auto it = m_textureResources.find(MakeTextureIdentity(&StorageTextureOf(textureOrView)));
+        return it != m_textureResources.end() && it->second.image != VK_NULL_HANDLE;
+    }
+
+    Bool VkTextureManager::HasPendingTexelUpload(MG_State::GLState::ITextureObject& textureOrView) const {
+        auto& storage = StorageTextureOf(textureOrView);
+        auto it = m_textureResources.find(MakeTextureIdentity(&storage));
+        if (it == m_textureResources.end() || it->second.image == VK_NULL_HANDLE) {
+            return false;
+        }
+        return it->second.syncedShapeVersion == storage.GetShapeVersion() &&
+               it->second.syncedContentVersion != storage.GetContentVersion();
+    }
+
+    Bool VkTextureManager::WasTextureWrittenThisRecording(MG_State::GLState::ITextureObject& textureOrView) const {
+        auto it = m_textureResources.find(MakeTextureIdentity(&StorageTextureOf(textureOrView)));
+        return it != m_textureResources.end() && it->second.lastRecordingWriteGeneration == m_recordingGeneration;
+    }
+
+    void VkTextureManager::StampTextureRecordingWrite(MG_State::GLState::ITextureObject* texture) {
+        if (texture == nullptr) {
+            return;
+        }
+        auto it = m_textureResources.find(MakeTextureIdentity(&StorageTextureOf(*texture)));
+        if (it != m_textureResources.end()) {
+            StampResourceRecordingWrite(it->second);
+        }
+    }
+
    void VkTextureManager::UpdateTrackedImageLayout(MG_State::GLState::ITextureObject* texture, VkImageLayout newLayout) {
        MOBILEGL_ASSERT(texture != nullptr, "UpdateTrackedImageLayout: texture is null");
        auto it = m_textureResources.find(MakeTextureIdentity(texture));
@@ -1361,7 +1425,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                        "UpdateTrackedImageLayoutAfterAttachmentWrite: textureId=%d mipLevel=%u out of range %u",
                        texture->GetExternalIndex(), writtenMipLevel, resource.mipLevels);
        // Pre-pass stream bookkeeping: the render pass that just ended wrote this image.
-        StampResourceRecordingUse(resource);
+        StampResourceRecordingWrite(resource);

        if (resource.layout != newLayout && resource.mipLevels > 1) {
            VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
@@ -1441,7 +1505,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        const Bool ok = TransitionImageLayout(commandBuffer, resource->image, resource->layout, targetLayout, srcStageMask,
-                                              s_sampledReadStages, srcAccessMask,
+                                              s_sampledReadStagesSlot->value, srcAccessMask,
                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
        MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
        // Pre-pass stream bookkeeping: a command referencing the image was recorded.
@@ -1475,8 +1539,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                              resource->aspect, 0, resource->mipLevels);
        MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                        texture.GetExternalIndex());
-        // Pre-pass stream bookkeeping: a command referencing the image was recorded.
-        StampResourceRecordingUse(*resource);
+        // Pre-pass stream bookkeeping: a command referencing the image was recorded, and an
+        // image unit may store to it.
+        StampResourceRecordingWrite(*resource);
        return ok;
    }

@@ -1611,6 +1676,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                           std::max(source->depth >> level, 1u)};
            copyRegions.push_back(copy);
        }
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdCopyImage(commandBuffer, source->image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, snapshot.image,
                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<Uint32>(copyRegions.size()), copyRegions.data());

@@ -1725,6 +1791,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // slices" but will read as "depth slice 0" once VK_KHR_maintenance9 is enabled. The
        // validation layer warns about that literal 1 by name.
        barrier.subresourceRange.layerCount = VK_REMAINING_ARRAY_LAYERS;
+        EndActiveRenderPassOn(commandBuffer);
        vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr, 1, &barrier);

        trackedLayout = newLayout;
@@ -2068,6 +2135,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    Bool VkTextureManager::SyncWireTextureShape(const MG_Pipe::MGPipeResourceRecord& record,
                                                TextureResource& resource, Bool requireStorage) {
        const MG_Pipe::MGPResourceDesc& desc = record.Desc;
+        // A YUV shared image is converted INTO the texture's own RGBA8 storage, which the rest of
+        // this function defines as for any texture; the image is only named here, and the
+        // renderer converts it at its first use in each frame (WireYuvImage.inc).
+        if (record.SharedImageId != resource.yuvImageId && !m_syncingYuvTextureStorage) {
+            // Respecified away from (or to another) YUV image: its source goes with it, once the
+            // GPU is done with it (a released resource's Reset is deferred).
+            if (resource.yuvSource != nullptr) {
+                TextureResource retired;
+                retired.yuvSource = std::move(resource.yuvSource);
+                DeferResourceRelease(Move(retired));
+            }
+            resource.yuvImageId = 0;
+            resource.yuvOwner.reset();
+            resource.yuvSource.reset();
+            resource.yuvConvertedFrame = 0;
+        }
+        if (record.SharedImageId != 0 && !m_syncingYuvTextureStorage) {
+            const auto yuv = MG_Remote::Server::SharedImages::Find(record.SharedImageId);
+            if (yuv != nullptr && MG_Remote::Server::SharedImages::FourccIsYuv(yuv->Fourcc)) {
+                m_syncingYuvTextureStorage = true;
+                const Bool defined = SyncWireTextureShape(record, resource, requireStorage);
+                m_syncingYuvTextureStorage = false;
+                if (!defined) return false;
+                if (resource.yuvImageId != record.SharedImageId) {
+                    resource.yuvImageId = record.SharedImageId;
+                    resource.yuvOwner = yuv;
+                    resource.yuvSource.reset();
+                    resource.yuvConvertedFrame = 0;
+                }
+                return true;
+            }
+        }
+        // Level 0 is a shared image: bind its buffer. An image that cannot be bound here leaves
+        // the texture the storage its NULL-data definition asked for, below.
+        if (record.SharedImageId != 0 && !m_syncingYuvTextureStorage) {
+            const SharedImageBind bound = SyncWireSharedImage(record, resource, requireStorage);
+            if (bound != SharedImageBind::Unresolved) return bound == SharedImageBind::Bound;
+        }
        if (desc.StorageKind != static_cast<Uint8>(TextureStorageType::Mipmap)) {
            MGLOG_W_ONCE("Magma wire texture {slot=%u, gen=%u}: buffer-backed textures are not migrated; declined",
                         desc.Resource.Slot, desc.Resource.Gen);
@@ -2167,13 +2272,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        VkImageUsageFlags desiredUsage =
-            VK_IMAGE_USAGE_SAMPLED_BIT | (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
-            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) &&
-             (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT)
-                 ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
-            (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT))
-                 ? VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
-                 : 0);
+            ResolveTextureUsage(aspect, formatProperties.optimalTilingFeatures, supportsStorageImage);
        // Multisample attachments are also transfer sources for vkCmdResolveImage.
        desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;

@@ -2236,7 +2335,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                          : (desc.Levels > 1 ? std::min(std::max<Uint32>(desc.Levels, fullMipLevels), fullMipLevels)
                                             : 1u);

-        const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
+        // A shared image never stands in for allocated storage: the attach is over (or never
+        // resolved), so the texture gets storage of its own, and none of the image's texels.
+        const Bool compatible = resource.image != VK_NULL_HANDLE && resource.sharedImageId == 0 &&
+                                resource.format == format &&
                                resource.extent.width == desc.Width && resource.extent.height == imageHeight &&
                                resource.depth == shapeInfo.depth && resource.arrayLayers == shapeInfo.arrayLayers &&
                                resource.viewType == shapeInfo.viewType &&
@@ -2251,7 +2353,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // A named-level respecify changes ONLY that level; a BindMask update changes no
        // storage at all. Preserve every compatible old level when growing/upgrading the
        // allocation. A whole-store redefinition permits (but does not require) old texels.
-        const Bool preserve = resource.image != VK_NULL_HANDLE && resource.format == format &&
+        const Bool preserve = resource.image != VK_NULL_HANDLE && resource.sharedImageId == 0 &&
+                              resource.format == format &&
                              resource.extent.width == desc.Width && resource.extent.height == imageHeight &&
                              resource.depth == shapeInfo.depth && resource.arrayLayers == shapeInfo.arrayLayers &&
                              resource.viewType == shapeInfo.viewType &&
@@ -2645,9 +2748,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                stencilCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
                stencilCopy.bufferOffset = stagingBase + item.offset + static_cast<VkDeviceSize>(texelCount) * 4;
                const VkBufferImageCopy copies[2] = {depthCopy, stencilCopy};
+                EndActiveRenderPassOn(commandBuffer);
                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, resource.image,
                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 2, copies);
            } else {
+                EndActiveRenderPassOn(commandBuffer);
                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, resource.image,
                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &copy);
            }
@@ -2656,7 +2761,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        const VkImageLayout finalLayout = ResolveSampledReadOnlyLayout(aspectMask);
        VkImageLayout uploadLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
        ok = TransitionImageLayout(commandBuffer, resource.image, uploadLayout, finalLayout,
-                                   VK_PIPELINE_STAGE_TRANSFER_BIT, s_sampledReadStages,
+                                   VK_PIPELINE_STAGE_TRANSFER_BIT, s_sampledReadStagesSlot->value,
                                   VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, aspectMask, 0,
                                   resource.mipLevels);
        MOBILEGL_ASSERT(ok, "UploadPendingWireLevels: transition to the sampled layout failed");
@@ -2729,6 +2834,201 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        return MG_Pipe::kMGPipeNullHandle;
    }

+    Bool VkTextureManager::ImportSharedImage(void* nativeBuffer, Uint32 width, Uint32 height,
+                                             ImportedSharedImage& out, String& why) {
+        out = ImportedSharedImage{};
+#if defined(__ANDROID__)
+        auto* ahb = static_cast<AHardwareBuffer*>(nativeBuffer);
+        if (ahb == nullptr) {
+            why = "the image has no AHardwareBuffer";
+            return false;
+        }
+        VulkanRenderer::WireAhbImport ctx;
+        if (pVulkanRenderer == nullptr || !pVulkanRenderer->GetWireAhbImport(ctx)) {
+            why = "the device did not take VK_ANDROID_external_memory_android_hardware_buffer";
+            return false;
+        }
+        AHardwareBuffer_Desc desc{};
+        AHardwareBuffer_describe(ahb, &desc);
+        if (desc.width != width || desc.height != height || desc.layers != 1) {
+            why = "the buffer is " + std::to_string(desc.width) + "x" + std::to_string(desc.height) + "x" +
+                  std::to_string(desc.layers) + ", not " + std::to_string(width) + "x" + std::to_string(height);
+            return false;
+        }
+        const auto getProperties =
+            reinterpret_cast<PFN_vkGetAndroidHardwareBufferPropertiesANDROID>(ctx.getAhbProperties);
+        VkAndroidHardwareBufferFormatPropertiesANDROID formatProperties{
+            VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_FORMAT_PROPERTIES_ANDROID};
+        VkAndroidHardwareBufferPropertiesANDROID properties{VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_PROPERTIES_ANDROID};
+        properties.pNext = &formatProperties;
+        VkResult result = getProperties(ctx.device, ahb, &properties);
+        if (result != VK_SUCCESS) {
+            why = "vkGetAndroidHardwareBufferPropertiesANDROID -> " + std::to_string(static_cast<int>(result));
+            return false;
+        }
+        // The RGBA8/RGBX8 buffers the registry allocates have a Vulkan format; an external-only
+        // format would need a YCbCr conversion and is not a colour buffer this path can write.
+        if (formatProperties.format == VK_FORMAT_UNDEFINED) {
+            why = "the buffer has no Vulkan format (external format only)";
+            return false;
+        }
+        // Usage is bounded by the buffer's own usage (the AHardwareBuffer usage equivalence);
+        // transfers need no buffer usage bit.
+        VkImageUsageFlags usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
+        if ((desc.usage & AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE) != 0 &&
+            (formatProperties.formatFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT) != 0)
+            usage |= VK_IMAGE_USAGE_SAMPLED_BIT;
+        if ((desc.usage & AHARDWAREBUFFER_USAGE_GPU_COLOR_OUTPUT) != 0 &&
+            (formatProperties.formatFeatures & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT) != 0)
+            usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+        VkExternalMemoryImageCreateInfo external{VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO};
+        external.handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID;
+        VkImageCreateInfo imageInfo{VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO};
+        imageInfo.pNext = &external;
+        imageInfo.imageType = VK_IMAGE_TYPE_2D;
+        imageInfo.format = formatProperties.format;
+        imageInfo.extent = {width, height, 1};
+        imageInfo.mipLevels = 1;
+        imageInfo.arrayLayers = 1;
+        imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
+        imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
+        imageInfo.usage = usage;
+        imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+        imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+        result = vkCreateImage(ctx.device, &imageInfo, nullptr, &out.image);
+        if (result != VK_SUCCESS) {
+            out.image = VK_NULL_HANDLE;
+            why = "vkCreateImage(external AHB) -> " + std::to_string(static_cast<int>(result));
+            return false;
+        }
+        Int32 type = -1;
+        for (Uint32 i = 0; i < ctx.memory.memoryTypeCount && type < 0; ++i) {
+            if ((properties.memoryTypeBits & (1u << i)) != 0 &&
+                (ctx.memory.memoryTypes[i].propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) != 0)
+                type = static_cast<Int32>(i);
+        }
+        for (Uint32 i = 0; i < ctx.memory.memoryTypeCount && type < 0; ++i) {
+            if ((properties.memoryTypeBits & (1u << i)) != 0) type = static_cast<Int32>(i);
+        }
+        if (type < 0) {
+            why = "no memory type can import it";
+            DestroyImportedSharedImage(ctx.device, out);
+            return false;
+        }
+        VkMemoryDedicatedAllocateInfo dedicated{VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO};
+        dedicated.image = out.image;
+        VkImportAndroidHardwareBufferInfoANDROID import{VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID};
+        import.pNext = &dedicated;
+        import.buffer = ahb;
+        VkMemoryAllocateInfo allocateInfo{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
+        allocateInfo.pNext = &import;
+        allocateInfo.allocationSize = properties.allocationSize;
+        allocateInfo.memoryTypeIndex = static_cast<Uint32>(type);
+        result = vkAllocateMemory(ctx.device, &allocateInfo, nullptr, &out.memory);
+        if (result != VK_SUCCESS) {
+            out.memory = VK_NULL_HANDLE;
+            why = "vkAllocateMemory(import AHB) -> " + std::to_string(static_cast<int>(result));
+            DestroyImportedSharedImage(ctx.device, out);
+            return false;
+        }
+        result = vkBindImageMemory(ctx.device, out.image, out.memory, 0);
+        if (result != VK_SUCCESS) {
+            why = "vkBindImageMemory -> " + std::to_string(static_cast<int>(result));
+            DestroyImportedSharedImage(ctx.device, out);
+            return false;
+        }
+        out.format = formatProperties.format;
+        out.usage = usage;
+        out.features = formatProperties.formatFeatures;
+        out.queueFamily = ctx.queueFamily;
+        return true;
+#else
+        (void)nativeBuffer, (void)width, (void)height;
+        why = "shared images are imported on Android only";
+        return false;
+#endif
+    }
+
+    void VkTextureManager::DestroyImportedSharedImage(VkDevice device, ImportedSharedImage& image) {
+        if (image.image != VK_NULL_HANDLE) vkDestroyImage(device, image.image, nullptr);
+        if (image.memory != VK_NULL_HANDLE) vkFreeMemory(device, image.memory, nullptr);
+        image = ImportedSharedImage{};
+    }
+
+    VkTextureManager::SharedImageBind VkTextureManager::SyncWireSharedImage(
+        const MG_Pipe::MGPipeResourceRecord& record, TextureResource& resource, Bool requireStorage) {
+        const MG_Pipe::MGPResourceDesc& desc = record.Desc;
+        const Uint64 id = record.SharedImageId;
+        if (requireStorage) {
+            // The buffer's usage has no storage equivalent; a reallocation would drop the image.
+            MGLOG_E_ONCE("Magma wire texture {slot=%u, gen=%u}: shared image %llu is bound as a shader image, "
+                         "which it cannot back; declined",
+                         desc.Resource.Slot, desc.Resource.Gen, static_cast<unsigned long long>(id));
+            return SharedImageBind::Failed;
+        }
+        // Already this image: the serial moved for something else (params, the bind mask).
+        if (resource.sharedImageId == id && resource.image != VK_NULL_HANDLE) return SharedImageBind::Bound;
+
+        const auto image = MG_Remote::Server::SharedImages::Find(id);
+        if (image == nullptr) {
+            MGLOG_E_ONCE("Magma wire texture {slot=%u, gen=%u}: shared image %llu is no longer live; the texture "
+                         "keeps its own (undefined) level 0",
+                         desc.Resource.Slot, desc.Resource.Gen, static_cast<unsigned long long>(id));
+            return SharedImageBind::Unresolved;
+        }
+        const Bool flat2D = desc.Target == static_cast<Uint8>(MG_Pipe::MGPipeResourceTarget::Tex2D) ||
+                            desc.Target == static_cast<Uint8>(MG_Pipe::MGPipeResourceTarget::TexRect);
+        if (!flat2D || desc.Width != image->Width || desc.Height != image->Height ||
+            !MG_Pipe::MGPipeHandleIsNull(desc.ViewOf)) {
+            MGLOG_E_ONCE("Magma wire texture {slot=%u, gen=%u}: pipe target %u at %ux%u cannot take shared image "
+                         "%llu (%ux%u); the texture keeps its own level 0",
+                         desc.Resource.Slot, desc.Resource.Gen, static_cast<Uint32>(desc.Target), desc.Width,
+                         desc.Height, static_cast<unsigned long long>(id), image->Width, image->Height);
+            return SharedImageBind::Unresolved;
+        }
+        ImportedSharedImage imported;
+        String why;
+        if (!ImportSharedImage(image->Native, image->Width, image->Height, imported, why)) {
+            MGLOG_E_ONCE("Magma wire texture {slot=%u, gen=%u}: shared image %llu not imported - %s; the texture "
+                         "keeps its own level 0",
+                         desc.Resource.Slot, desc.Resource.Gen, static_cast<unsigned long long>(id), why.c_str());
+            return SharedImageBind::Unresolved;
+        }
+
+        // NOT ACQUIRED HERE. The writer (a client session's present, on its own device) releases
+        // the image to VK_QUEUE_FAMILY_FOREIGN_EXT in GENERAL - also the layout every wire
+        // descriptor samples in - once per frame it writes. Each use acquires it again when the
+        // write generation moved, waiting for that write's fence on the GPU, and the frame boundary
+        // releases it back (VulkanRenderer::AcquireSharedImage, WireSharedImage.inc).
+
+        TextureResource replacement;
+        replacement.image = imported.image;
+        replacement.importedMemory = imported.memory;
+        replacement.layout = VK_IMAGE_LAYOUT_GENERAL;
+        replacement.extent = {image->Width, image->Height};
+        replacement.depth = 1;
+        replacement.arrayLayers = 1;
+        replacement.mipLevels = 1;
+        replacement.sampledBaseMipLevel = 0;
+        replacement.sampledLevelCount = 1;
+        replacement.format = imported.format;
+        replacement.aspect = VK_IMAGE_ASPECT_COLOR_BIT;
+        replacement.viewType = VK_IMAGE_VIEW_TYPE_2D;
+        replacement.sampleCount = VK_SAMPLE_COUNT_1_BIT;
+        replacement.imageCreateFlags = 0;
+        replacement.usageFlags = imported.usage;
+        replacement.sharedImageId = id;
+        replacement.sharedImageOwner = image;
+        replacement.sharedImageAlphaOne = MG_Remote::Server::SharedImages::FourccIgnoresAlpha(image->Fourcc);
+        ++m_textureImageEpoch; // a new attachment image invalidates cached render passes
+        DeferResourceRelease(Move(resource));
+        std::destroy_at(&resource);
+        std::construct_at(&resource, Move(replacement));
+        MGLOG_D("Magma wire texture {slot=%u, gen=%u}: level 0 is shared image %llu (%ux%u)", desc.Resource.Slot,
+                desc.Resource.Gen, static_cast<unsigned long long>(id), image->Width, image->Height);
+        return SharedImageBind::Bound;
+    }
+
    VkTextureManager::TextureResource* VkTextureManager::SyncTextureResourceByHandle(
        MG_Pipe::MGPipeHandle handle, Bool renderbuffer, Bool requireStorage) {
        MOBILEGL_ASSERT(m_device != VK_NULL_HANDLE, "SyncTextureResourceByHandle: m_device == VK_NULL_HANDLE");
@@ -2805,6 +3105,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                handle.Slot, handle.Gen);
            return nullptr;
        }
+        resource.sharedImageKey = resource.sharedImageId != 0 || resource.yuvImageId != 0 ? key : 0;
        resource.syncedWireSerial = record.Serial;
        return &resource;
    }
@@ -2926,6 +3227,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                if ((aspect & copyAspect) == 0) continue;
                copy.srcSubresource.aspectMask = copyAspect;
                copy.dstSubresource.aspectMask = copyAspect;
+                EndActiveRenderPassOn(commandBuffer);
                vkCmdCopyImage(commandBuffer, old.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, grown.image,
                               VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &copy);
            }
@@ -3221,12 +3523,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        VkImageUsageFlags desiredUsage =
-            VK_IMAGE_USAGE_SAMPLED_BIT |
-            (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
-            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
-            (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
-                 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
-                 0);
+            ResolveTextureUsage(aspect, formatProperties.optimalTilingFeatures, supportsStorageImage);
        if (!isMultisampleTexture) {
            desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
        }
@@ -3521,7 +3818,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            resource.sampledView == VK_NULL_HANDLE &&
            resource.perMipViews.empty() && resource.perMipSampledViews.empty() &&
            resource.attachmentViews.empty() && resource.alternateSampledViews.empty() &&
-            resource.storageImageViews.empty()) {
+            resource.storageImageViews.empty()
+#if MOBILEGL_BUILD_DISAGGREGATED
+            && resource.yuvSource == nullptr
+#endif
+        ) {
            return;
        }

@@ -4077,18 +4378,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            VkDeviceSize offset = 0;
            // Sub-region upload (a small sprite in a big atlas): only the dirty box
            // is staged and copied. texelSize keeps the LEVEL extent - the staging
-            // row copy needs it for the shadow's stride. Plain color formats only;
-            // the RGB-expand and depth(+stencil) conversion passes rewrite whole
-            // levels and stay full-size.
+            // row copy needs it for the stride. Color formats only (an RGB-expanded
+            // level is cut from its expanded bytes); the depth(+stencil) conversion
+            // passes rewrite whole levels and stay full-size.
            Bool subRegion = false;
            IntVec3 regionLo = {0, 0, 0};
            IntVec3 regionSize = {0, 0, 0};
            SizeT texelBytes = 0;
-            // Scatter refinement of the single dirty box: when the storage's rect
-            // list reports the writes' true footprint (~100 sprites whose union box
-            // spans the whole atlas), each rect is staged tightly and copied with
-            // its own VkBufferImageCopy in ONE vkCmdCopyBufferToImage. Empty means
-            // "stage the one box above". Only set while subRegion.
+            // The writes' exact footprint when it is more than one box (~100 sprites
+            // whose union box spans the whole atlas): each rect is staged tightly and
+            // copied with its own VkBufferImageCopy in ONE vkCmdCopyBufferToImage.
+            // Empty means "stage the one box above". Only set while subRegion.
            Vector<MG_State::GLState::MipmapDirtyRegion> rects;
        };

@@ -4100,6 +4400,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            targets.push_back(uploadTarget);
        }
        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(mipmapTexture.GetFormat());
+        Vector<MG_State::GLState::MipmapDirtyRegion> dirtyFootprint;

        VkDeviceSize stagingSize = 0;
        for (const TextureUploadTarget target : targets) {
@@ -4142,34 +4443,48 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                uploadItem.source = source;
                uploadItem.offset = stagingSize;
                uploadItem.uploadByteSize = byteSize;
-                if (!formatInfo.expandRgbToRgba &&
-                    GetAspectMaskForFormat(outResource.format) == VK_IMAGE_ASPECT_COLOR_BIT) {
+                // The RGB -> RGBA expansion rewrites the level texel for texel into a tight buffer
+                // of its own, so it runs first and the dirty region below is taken from whichever
+                // bytes are staged. Uploading an expanded level WHOLE put the shadow's stale texels
+                // over whatever the GPU had drawn into the rest of it.
+                if (formatInfo.expandRgbToRgba) {
+                    const Bool expanded = ExpandRgbSourceToRgba(source, byteSize, texelSize, formatInfo,
+                                                                uploadItem.expandedData);
+                    MOBILEGL_ASSERT(expanded,
+                                    "UploadDirtyMipLevels: failed to expand RGB textureId=%d target=%s level=%u to RGBA staging data",
+                                    mipmapTexture.GetExternalIndex(),
+                                    MG_Util::ConvertTextureUploadTargetToString(target).c_str(), level);
+                    uploadItem.uploadByteSize = uploadItem.expandedData.size();
+                }
+                const SizeT levelBytes = uploadItem.uploadByteSize;
+                if (GetAspectMaskForFormat(outResource.format) == VK_IMAGE_ASPECT_COLOR_BIT) {
                    const auto region = mipmapTexture.GetStorageDirtyRegion(target, level);
                    const SizeT texelCount = static_cast<SizeT>(texelSize.x()) *
                                             static_cast<SizeT>(texelSize.y()) *
                                             static_cast<SizeT>(std::max(texelSize.z(), 1));
-                    if (!region.Empty() && !region.CoversWholeLevel(texelSize) && texelCount > 0 &&
-                        byteSize % texelCount == 0) {
+                    // Exactly the texels the client wrote, never their union box or the storage's
+                    // rect list (which merges touching writes into their bounding box, folds when
+                    // full and is withheld past 3/4 of the box): the image may hold what the GPU
+                    // drew or cleared anywhere else, and the shadow is stale there. Each box is one
+                    // more VkBufferImageCopy in the same vkCmdCopyBufferToImage, which costs next
+                    // to nothing, so there is no cheaper shape worth being wrong for. A union box
+                    // that spans the level with gaps in it is not a whole-level write either.
+                    mipmapTexture.GetStorageDirtyFootprint(target, level, dirtyFootprint);
+                    if (!region.Empty() && (dirtyFootprint.size() >= 2 || !region.CoversWholeLevel(texelSize)) &&
+                        texelCount > 0 && levelBytes % texelCount == 0) {
                        uploadItem.subRegion = true;
                        uploadItem.regionLo = region.lo;
                        uploadItem.regionSize = {region.hi.x() - region.lo.x(), region.hi.y() - region.lo.y(),
                                                 region.hi.z() - region.lo.z()};
-                        uploadItem.texelBytes = byteSize / texelCount;
+                        uploadItem.texelBytes = levelBytes / texelCount;
                        uploadItem.uploadByteSize = static_cast<SizeT>(uploadItem.regionSize.x()) *
                                                    static_cast<SizeT>(uploadItem.regionSize.y()) *
                                                    static_cast<SizeT>(uploadItem.regionSize.z()) *
                                                    uploadItem.texelBytes;
-                        // Scatter refinement: the storage only hands out its rect list
-                        // when the rects' summed area is materially smaller than the
-                        // union box (0 otherwise), so taking it always stages fewer
-                        // bytes than the box - the very amplification this path exists
-                        // to avoid paying twice.
-                        MG_State::GLState::MipmapDirtyRegion
-                            dirtyRects[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
-                        const SizeT dirtyRectCount = mipmapTexture.GetStorageDirtyRects(
-                            target, level, dirtyRects, MG_State::GLState::MipmapStorage::kMaxDirtyRects);
-                        if (dirtyRectCount >= 2) {
-                            uploadItem.rects.assign(dirtyRects, dirtyRects + dirtyRectCount);
+                        // Several boxes: each is staged tightly and copied on its own. One box is
+                        // the union box above.
+                        if (dirtyFootprint.size() >= 2) {
+                            uploadItem.rects.assign(dirtyFootprint.begin(), dirtyFootprint.end());
                            SizeT rectTexels = 0;
                            for (const auto& rect : uploadItem.rects) {
                                rectTexels += rect.TexelCount();
@@ -4195,15 +4510,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                        }
                    }
                }
-                if (formatInfo.expandRgbToRgba) {
-                    const Bool expanded = ExpandRgbSourceToRgba(source, byteSize, texelSize, formatInfo,
-                                                                uploadItem.expandedData);
-                    MOBILEGL_ASSERT(expanded,
-                                    "UploadDirtyMipLevels: failed to expand RGB textureId=%d target=%s level=%u to RGBA staging data",
-                                    mipmapTexture.GetExternalIndex(),
-                                    MG_Util::ConvertTextureUploadTargetToString(target).c_str(), level);
-                    uploadItem.uploadByteSize = uploadItem.expandedData.size();
-                }
                uploadItems.push_back(Move(uploadItem));
                if (!uploadItems.back().expandedData.empty()) {
                    uploadItems.back().source = uploadItems.back().expandedData.data();
@@ -4438,6 +4744,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        const Bool depthSelectsArrayLayer = outResource.viewType == VK_IMAGE_VIEW_TYPE_1D_ARRAY ||
                                            outResource.viewType == VK_IMAGE_VIEW_TYPE_2D_ARRAY ||
                                            outResource.viewType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
+        Vector<VkBufferImageCopy> rectCopies;
        for (const auto& item : uploadItems) {
            if (!item.rects.empty()) {
                // Multi-rect item: one VkBufferImageCopy per rect, all submitted in a
@@ -4445,7 +4752,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                // construction, so no two copies write the same texels. Multi-rect
                // implies subRegion, which implies a plain color aspect - the combined
                // depth-stencil split below can never see one of these.
-                VkBufferImageCopy rectCopies[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
+                rectCopies.resize(item.rects.size());
                Uint32 rectCopyCount = 0;
                VkDeviceSize runningOffset = item.offset;
                for (const auto& rect : item.rects) {
@@ -4475,8 +4782,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                    rectCopies[rectCopyCount++] = rectCopy;
                    runningOffset += static_cast<VkDeviceSize>(rect.TexelCount() * item.texelBytes);
                }
+                EndActiveRenderPassOn(commandBuffer);
                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
-                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, rectCopyCount, rectCopies);
+                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, rectCopyCount, rectCopies.data());
                continue;
            }
            const Uint32 depthOrLayers = item.texelSize.z() > 0 ? static_cast<Uint32>(item.texelSize.z()) : 1u;
@@ -4516,10 +4824,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                stencilCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
                stencilCopy.bufferOffset = stagingBase + item.offset + static_cast<VkDeviceSize>(texelCount) * 4;
                const VkBufferImageCopy copies[2] = {depthCopy, stencilCopy};
+                EndActiveRenderPassOn(commandBuffer);
                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 2, copies);
                continue;
            }
+            EndActiveRenderPassOn(commandBuffer);
            vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                                   1, &copy);
        }
@@ -4530,7 +4840,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                   uploadLayout,
                                   finalLayout,
                                   VK_PIPELINE_STAGE_TRANSFER_BIT,
-                                   s_sampledReadStages,
+                                   s_sampledReadStagesSlot->value,
                                   VK_ACCESS_TRANSFER_WRITE_BIT,
                                   VK_ACCESS_SHADER_READ_BIT,
                                   aspectMask, 0, outResource.mipLevels);
@@ -9,6 +9,7 @@
 #pragma once

 #include "../VkIncludes.h"
+#include "../MagmaSession.h"
 #include <Includes.h>
 #include <MG_State/GLState/FramebufferState/FramebufferObject.h>
 #include <MG_State/GLState/TextureState/TextureObject.h>
@@ -332,6 +333,11 @@ public:
        // into the frame's PRE command buffer - which executes strictly before the
        // frame's commands - instead of splitting the active render pass.
        Uint64 lastRecordingGeneration = 0;
+        // The same, for the last recorded command that WRITES the image (a render pass it is an
+        // attachment of, a clear, a blit/copy into it, a mip generation, a storage-image binding).
+        // An upload into an image written this recording has to be ordered after that write
+        // (VulkanRenderer::OrderPendingUploadAfterRecording).
+        Uint64 lastRecordingWriteGeneration = 0;
        // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
        // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
        Uint64 syncedContentVersion = 0;
@@ -351,6 +357,30 @@ public:
        // on every mutation, so the first handle sync of a fresh image always re-checks.
 #if MOBILEGL_BUILD_DISAGGREGATED
        Uint64 syncedWireSerial = 0;
+        // SHARED IMAGES: level 0 is a server shared image (an imported AHardwareBuffer), not a
+        // VMA allocation. The resource owns the image and its dedicated imported memory, and
+        // holds the registry's reference so the buffer outlives every use (deferred releases
+        // included). `sharedImageAlphaOne`: an X-format image, whose alpha bytes are undefined.
+        VkDeviceMemory importedMemory = VK_NULL_HANDLE;
+        Uint64 sharedImageId = 0;
+        SharedPtr<const void> sharedImageOwner;
+        Bool sharedImageAlphaOne = false;
+        // The reading side's ownership of the image (VulkanRenderer::AcquireSharedImage): whether
+        // this device holds it now - acquired from the foreign family and not yet released back at
+        // the frame boundary - the writer's generation it last acquired, and the
+        // m_wireTextureResources key it lives under, by which the frame boundary finds it again.
+        Bool sharedImageHeld = false;
+        Uint64 sharedImageGeneration = 0;
+        Uint64 sharedImageKey = 0;
+        // A YUV shared image (NV12/P010) is never the texture's image: the texture keeps its own
+        // RGBA8 level 0, into which the renderer converts the image at its first use in each frame
+        // (WireYuvImage.inc) - `yuvSource` being that conversion's import of the buffer, made on
+        // first use, and `yuvConvertedFrame` the frame it last ran in. `yuvOwner` holds the
+        // registry's image for as long as the texture names it.
+        Uint64 yuvImageId = 0;
+        SharedPtr<const void> yuvOwner;
+        SharedPtr<void> yuvSource;
+        Uint64 yuvConvertedFrame = 0;
 #endif

        TextureResource() = default;
@@ -381,49 +411,78 @@ public:
            std::swap(this->storageUsageResolved, that.storageUsageResolved);
            std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
            std::swap(this->lastRecordingGeneration, that.lastRecordingGeneration);
+            std::swap(this->lastRecordingWriteGeneration, that.lastRecordingWriteGeneration);
            std::swap(this->syncedContentVersion, that.syncedContentVersion);
            std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
            std::swap(this->syncedShapeVersion, that.syncedShapeVersion);
 #if MOBILEGL_BUILD_DISAGGREGATED
            std::swap(this->syncedWireSerial, that.syncedWireSerial);
+            std::swap(this->importedMemory, that.importedMemory);
+            std::swap(this->sharedImageId, that.sharedImageId);
+            std::swap(this->sharedImageOwner, that.sharedImageOwner);
+            std::swap(this->sharedImageAlphaOne, that.sharedImageAlphaOne);
+            std::swap(this->sharedImageHeld, that.sharedImageHeld);
+            std::swap(this->sharedImageGeneration, that.sharedImageGeneration);
+            std::swap(this->sharedImageKey, that.sharedImageKey);
+            std::swap(this->yuvImageId, that.yuvImageId);
+            std::swap(this->yuvOwner, that.yuvOwner);
+            std::swap(this->yuvSource, that.yuvSource);
+            std::swap(this->yuvConvertedFrame, that.yuvConvertedFrame);
 #endif
        }

        void Reset() {
            if (fullView != VK_NULL_HANDLE) {
-                vkDestroyImageView(s_device, fullView, nullptr);
+                vkDestroyImageView(s_device.Get(), fullView, nullptr);
            }
            if (sampledView != VK_NULL_HANDLE) {
-                vkDestroyImageView(s_device, sampledView, nullptr);
+                vkDestroyImageView(s_device.Get(), sampledView, nullptr);
            }
            for (const auto attachmentView : perMipViews) {
                if (attachmentView != VK_NULL_HANDLE) {
-                    vkDestroyImageView(s_device, attachmentView, nullptr);
+                    vkDestroyImageView(s_device.Get(), attachmentView, nullptr);
                }
            }
            for (const auto sampledView : perMipSampledViews) {
                if (sampledView != VK_NULL_HANDLE) {
-                    vkDestroyImageView(s_device, sampledView, nullptr);
+                    vkDestroyImageView(s_device.Get(), sampledView, nullptr);
                }
            }
            for (const auto& [_, attachmentView] : attachmentViews) {
                if (attachmentView != VK_NULL_HANDLE) {
-                    vkDestroyImageView(s_device, attachmentView, nullptr);
+                    vkDestroyImageView(s_device.Get(), attachmentView, nullptr);
                }
            }
            for (const auto& [_, sampledView] : alternateSampledViews) {
                if (sampledView != VK_NULL_HANDLE) {
-                    vkDestroyImageView(s_device, sampledView, nullptr);
+                    vkDestroyImageView(s_device.Get(), sampledView, nullptr);
                }
            }
            for (const auto& [_, storageImageView] : storageImageViews) {
                if (storageImageView != VK_NULL_HANDLE) {
-                    vkDestroyImageView(s_device, storageImageView, nullptr);
+                    vkDestroyImageView(s_device.Get(), storageImageView, nullptr);
                }
            }
            if (image != VK_NULL_HANDLE && allocation != nullptr) {
-                vmaDestroyImage(s_allocator, image, allocation);
+                vmaDestroyImage(s_allocator.Get(), image, allocation);
            }
+#if MOBILEGL_BUILD_DISAGGREGATED
+            if (importedMemory != VK_NULL_HANDLE) {
+                if (image != VK_NULL_HANDLE) vkDestroyImage(s_device.Get(), image, nullptr);
+                vkFreeMemory(s_device.Get(), importedMemory, nullptr);
+            }
+            importedMemory = VK_NULL_HANDLE;
+            sharedImageId = 0;
+            sharedImageOwner.reset();
+            sharedImageAlphaOne = false;
+            sharedImageHeld = false;
+            sharedImageGeneration = 0;
+            sharedImageKey = 0;
+            yuvImageId = 0;
+            yuvOwner.reset();
+            yuvSource.reset();
+            yuvConvertedFrame = 0;
+#endif
            fullView = VK_NULL_HANDLE;
            sampledView = VK_NULL_HANDLE;
            perMipViews.clear();
@@ -460,8 +519,8 @@ public:
            Reset();
        }

-        static inline VkDevice s_device = VK_NULL_HANDLE;
-        static inline VmaAllocator s_allocator = VK_NULL_HANDLE;
+        static inline SessionLocal<VkDevice> s_device;  // per Magma session (MagmaSession.h)
+        static inline SessionLocal<VmaAllocator> s_allocator;
    };

    struct SampledTextureSnapshot {
@@ -585,6 +644,41 @@ public:
    // ranges select cube faces; array and 3D levels share one staged level shadow.
    void MarkWireTextureGpuWritten(MG_Pipe::MGPipeHandle handle, Uint32 mipLevel,
                                    Uint32 baseArrayLayer = 0, Uint32 layerCount = 1);
+
+    // SHARED IMAGES (docs/Disaggregated/notes/anland/plan-ahb-dmabuf.md): a server shared image's
+    // AHardwareBuffer bound as a single-level 2D colour image on THIS session's device, on dedicated
+    // memory imported from the buffer. Android only (false with `why` elsewhere, or when the device
+    // did not take the AHardwareBuffer extension). The image is born UNDEFINED; whoever uses it
+    // acquires it from the foreign queue family first, and a writer releases it back after.
+    struct ImportedSharedImage {
+        VkImage image = VK_NULL_HANDLE;
+        VkDeviceMemory memory = VK_NULL_HANDLE;
+        VkFormat format = VK_FORMAT_UNDEFINED;
+        VkImageUsageFlags usage = 0;
+        VkFormatFeatureFlags features = 0;
+        Uint32 queueFamily = 0;
+    };
+    static Bool ImportSharedImage(void* nativeBuffer, Uint32 width, Uint32 height, ImportedSharedImage& out,
+                                  String& why);
+    static void DestroyImportedSharedImage(VkDevice device, ImportedSharedImage& image);
+    // The reading side's per-use hook: PrepareWireTextureResources notes every shared-image
+    // texture a draw or dispatch samples, and the renderer takes the list right after
+    // (VulkanRenderer::AcquireNotedSharedImages) to acquire each one ahead of the sampling.
+    void NoteSharedImageUse(const TextureResource& resource) {
+        if (std::find(m_sharedImageUses.begin(), m_sharedImageUses.end(), resource.sharedImageKey) ==
+            m_sharedImageUses.end())
+            m_sharedImageUses.push_back(resource.sharedImageKey);
+    }
+    Bool HasNotedSharedImageUses() const { return !m_sharedImageUses.empty(); }
+    void TakeNotedSharedImageUses(Vector<Uint64>& out) {
+        out.clear();
+        std::swap(out, m_sharedImageUses);
+    }
+    // The live handle-keyed texture resource under `key`, or null.
+    TextureResource* FindWireTextureResource(Uint64 key) {
+        const auto it = m_wireTextureResources.find(key);
+        return it != m_wireTextureResources.end() && it->second.image != VK_NULL_HANDLE ? &it->second : nullptr;
+    }
 #endif
    VkImageView GetOrCreateViewAtMipLevel(MG_State::GLState::ITextureObject& texture, Uint32 mipLevel);
    VkImageView GetOrCreateAttachmentViewAtMipLevel(MG_State::GLState::ITextureObject& texture, Uint32 mipLevel,
@@ -623,6 +717,26 @@ public:
    Bool WasTouchedThisRecording(const TextureResource& resource) const {
        return resource.lastRecordingGeneration == m_recordingGeneration;
    }
+    // A write is a use too. See TextureResource::lastRecordingWriteGeneration.
+    void StampResourceRecordingWrite(TextureResource& resource) const {
+        resource.lastRecordingGeneration = m_recordingGeneration;
+        resource.lastRecordingWriteGeneration = m_recordingGeneration;
+    }
+    void StampTextureRecordingWrite(MG_State::GLState::ITextureObject* texture);
+    // Upload-ordering queries (VulkanRenderer::OrderPendingUploadAfterRecording). None of them
+    // syncs, so nothing pending is uploaded by asking. All answer for the STORAGE texture.
+    //
+    // Uploads are recorded into a batch submitted AHEAD of the frame's recording
+    // (FlushPendingUploads), so an upload into an image the open recording already WROTE lands
+    // before that write, which then overwrites the newer texels. (An image the recording only
+    // read keeps the old trade: its earlier reads see the newer texels.)
+    Bool HasLiveImage(MG_State::GLState::ITextureObject& textureOrView) const;
+    // The next sync would record client texels into the image that is live now: same shape (a
+    // shape change mints a new image instead) and content moved since the last sync.
+    Bool HasPendingTexelUpload(MG_State::GLState::ITextureObject& textureOrView) const;
+    Bool WasTextureWrittenThisRecording(MG_State::GLState::ITextureObject& textureOrView) const;
+    // The live resource, or null.
+    TextureResource* FindTextureResource(MG_State::GLState::ITextureObject& textureOrView);
    // Records that this texture is bound to a GL image unit, so its image must carry
    // VK_IMAGE_USAGE_STORAGE_BIT. Must be called before NeedsStorageImagePreparation, and
    // therefore before the render pass is committed: an image that has to be upgraded is
@@ -751,10 +865,18 @@ private:
                              Bool requireStorage);
    Bool UploadPendingWireLevels(MG_Pipe::MGPipeHandle handle, MG_Pipe::MGPipeResourceRecord& record,
                                 TextureResource& resource);
+    // SyncWireTextureShape's arm for a record whose level 0 IS a shared image (SharedImageId):
+    // binds the imported buffer instead of allocating. Unresolved = the image is gone or cannot be
+    // imported here; `resource` is then untouched and the caller allocates as for any texture.
+    enum class SharedImageBind : Uint8 { Bound, Failed, Unresolved };
+    SharedImageBind SyncWireSharedImage(const MG_Pipe::MGPipeResourceRecord& record, TextureResource& resource,
+                                        Bool requireStorage);
    // The handle-keyed twin of m_textureResources, keyed by StagedTextureStore::KeyForHandle.
    // Node-based for the same reason m_renderbufferResources is: callers hold TextureResource*
    // across map-touching calls.
    std::unordered_map<Uint64, TextureResource> m_wireTextureResources;
+    // Shared-image textures sampled since the renderer last took the list (NoteSharedImageUse).
+    Vector<Uint64> m_sharedImageUses;
    // Handles are per-kind; a renderbuffer and texture may have identical slot/gen.
    std::unordered_map<Uint64, TextureResource> m_wireRenderbufferResources;
 #endif
@@ -777,6 +899,8 @@ private:
    // Active only between BeginDrawSyncScope/EndDrawSyncScope; identities of
    // textures already fully synced in the current draw (small N -> flat scan).
    Bool m_drawSyncScopeActive = false;
+    // SyncWireTextureShape defining the RGBA8 storage a YUV shared image is converted into.
+    Bool m_syncingYuvTextureStorage = false;
    // Per-draw sync memo: the identity plus the resolved resource pointer. The pointer is stable
    // across rehash in the node-based m_textureResources and stays valid for the draw (a texture
    // synced this draw is alive and is not erased mid-draw), so a repeat sync of the same texture
--- a/Show More
+++ b/Show More