diff --git a/.gitignore b/.gitignore
index ce82387c15..52e2c4052c 100644
--- a/.gitignore
+++ b/.gitignore
@@ -168,3 +168,8 @@ packages/multimedia/intel-vaapi-driver/
 packages/multimedia/media-driver/
 packages/multimedia/gmmlib/
 packages/sysutils/open-vm-tools/
+
+
+#other
+docs/*
+bench-results/
\ No newline at end of file
diff --git a/TODO.md b/TODO.md
deleted file mode 100644
index ebfa4099c2..0000000000
--- a/TODO.md
+++ /dev/null
@@ -1,844 +0,0 @@
-# TODO — Curadoria de Performance do SO
-
-> **Status (2026-04-18):** FASE 0-3 investigadas e implementadas. Itens marcados com ✅ estão concluídos.
-
-**Hardware alvo:** Handheld open source com SoC **Rockchip RK3326** (4× Cortex-A35 @ 1.3–1.5 GHz, in-order dual-issue, 256KB L2 compartilhado), GPU **Mali-G31 MP2** (Bifrost v9), **1GB DDR3/DDR3L**, storage **microSD**, passivo (sem fan).
-
-**Referências de comparação:**
-- dArkOS: `/media/disco-local/dArkOS/`
-- ROCKNIX: `/home/douglas/Documentos/distribution/`
-
-**Princípio guia:** nenhuma feature pode ser removida (WiFi, BT, etc.) — tudo o que não é essencial durante gameplay deve ser *socket-activated* ou *on-demand*, não deletado.
-
-**Convenção de tags:**
-- `[PERF]` — item que altera FPS/frametime/latência direta
-- `[FOOTPRINT]` — item que altera consumo de RAM, CPU idle, armazenamento
-- `[AMBOS]` — item que afeta os dois
-- `[INFRA]` — infraestrutura de medição/decisão, não é ganho em si
-
-**Convenção de metas numéricas:** todas as metas deste documento são **hipóteses a validar por benchmark**, não garantias. Se a hipótese não se confirmar, o item vira investigação, não ação cega.
-
----
-
-# FASE 0 — Pré-requisitos bloqueadores
-
-Sem estes dois itens, qualquer mudança subsequente é chute.
-
----
-
-## 0. ✅ Sanity check de release build `[INFRA][PERF]`
-
-**Objetivo:** garantir que nada crítico no pipeline está compilado/executado em modo de desenvolvimento. É o ganho mais barato e invisível — se algum pacote central estiver em debug, o ceiling de todos os outros itens cai junto.
-
-### 0.1 Kernel
-
-```bash
-# Extrair config do kernel em execução
-zcat /proc/config.gz 2>/dev/null || cat /boot/config-$(uname -r)
-```
-
-**Checks obrigatórios:**
-
-| Config | Esperado | Impacto |
-|--------|----------|---------|
-| `CONFIG_DEBUG_KERNEL` | `n` ou mínimo | overhead generalizado |
-| `CONFIG_DEBUG_INFO` | off em release (ou em pkg separado) | tamanho |
-| `CONFIG_LOCKDEP`, `CONFIG_PROVE_LOCKING` | `n` | overhead pesado |
-| `CONFIG_DEBUG_LOCK_ALLOC` | `n` | |
-| `CONFIG_KASAN`, `CONFIG_KMEMLEAK`, `CONFIG_KFENCE` | `n` | overhead + RAM |
-| `CONFIG_SLUB_DEBUG_ON` | `n` (`CONFIG_SLUB_DEBUG=y` ok) | |
-| `CONFIG_FTRACE`, `CONFIG_FUNCTION_TRACER` | `n` ou ao menos inativo | |
-| `CONFIG_PREEMPT_TRACER`, `CONFIG_IRQSOFF_TRACER` | `n` | |
-| `CONFIG_SCHEDSTATS` | `n` | |
-| `CONFIG_DEBUG_PREEMPT` | `n` | |
-
-**Em runtime:**
-```bash
-cat /sys/kernel/debug/tracing/tracing_on     # esperado: 0
-cat /sys/kernel/debug/tracing/current_tracer # esperado: nop
-dmesg | grep -iE "debug|warn|slub" | head -50
-```
-
-### 0.2 Mesa
-
-No PKGBUILD / meson:
-- `-Dbuildtype=release` (NÃO `debugoptimized`, NÃO `debug`)
-- `-Db_ndebug=true` (asserts off)
-- `-Db_lto=true` (opcional, recomendado)
-
-Variáveis de ambiente que **jamais** podem estar setadas em produção:
-```bash
-env | grep -iE "MESA_DEBUG|MESA_VERBOSE|MESA_LOG|GALLIUM_DUMP|PAN_MESA_DEBUG|LIBGL_DEBUG"
-```
-
-Se `PAN_MESA_DEBUG` estiver setado, auditar o valor:
-- `sync` → serialização total, perde ~30% FPS
-- `trace` → captura de cada draw call
-- `nocrc` → desliga transaction elimination (vai contra o item 4)
-
-### 0.3 RetroArch
-
-No configure/PKGBUILD:
-- `--disable-debug`
-- `--enable-lto` (se suportado)
-- Nunca: `--enable-debug`, `--enable-profile`, `--enable-asan|ubsan|tsan`
-
-Em runtime:
-```bash
-retroarch --version   # sem flags HAVE_DEBUG
-```
-
-Flags que nunca devem estar no `retroarch.cfg` de produção:
-```
-log_verbosity = "false"
-perfcnt_enable = "false"    # só durante benchmark
-libretro_log_level = "3"    # error apenas
-frontend_log_level = "3"
-```
-
-### 0.4 Cores libretro
-
-```bash
-for core in /usr/lib/libretro/*.so; do
-  echo "=== $core ==="
-  file "$core" | grep -o "not stripped\|stripped"
-  nm -D "$core" 2>/dev/null | grep -iE "asan|ubsan|tsan|msan" | head -3
-  du -h "$core"
-done
-```
-
-Esperado: `stripped`, sem símbolos de sanitizer, tamanho razoável (cores em debug são 2–5x maiores).
-
-### 0.5 Emuladores standalone
-
-Mesma auditoria para PPSSPP, Dolphin, DuckStation, Flycast etc.:
-- `CMAKE_BUILD_TYPE=Release` (ou `RelWithDebInfo` com debug em pkg separado)
-- Nunca `Debug`
-- Binário `stripped`
-
-```bash
-for bin in /usr/bin/{ppsspp,dolphin-emu,duckstation,flycast}; do
-  [ -x "$bin" ] || continue
-  echo "=== $bin ==="
-  file "$bin" | grep -o "not stripped\|stripped"
-  du -h "$bin"
-done
-```
-
-### 0.6 Frontend
-
-```bash
-find /etc /opt -name "*.conf" -o -name "*.cfg" 2>/dev/null | \
-  xargs grep -iE "log.*level|verbose|debug" 2>/dev/null | head -20
-```
-
-Muitos frontends vêm com log level `DEBUG` no default.
-
-### 0.7 Comparação lateral com dArkOS/ROCKNIX
-
-Gerar relatório de flags e buildtypes usados pelas distros concorrentes — às vezes eles descobriram algo que não está documentado, e vice-versa.
-
-**Critério de aceitação:**
-- Documento `/docs/build-audit.md` com checklist preenchido para cada subitem.
-- Zero pacotes em buildtype `Debug`.
-- Zero símbolos de sanitizer em binários de produção.
-- Zero env vars de debug em `/etc/environment`, profiles, systemd units.
-
-**Por que vem antes do benchmark:** se baseline for medido com Mesa `debugoptimized` ou core com asserts ativos, toda decisão subsequente parte de referência errada.
-
----
-
-## 1. ✅ Baseline de benchmark reproduzível `[INFRA]`
-
-**Objetivo:** script que rode cenários fixos e emita métricas completas para comparar nosso SO vs dArkOS vs ROCKNIX, de forma rastreável e reproduzível.
-
-### 1.1 Cenários fixos (save states versionados no repo)
-
-| Sistema | Jogo | Cena | Emulador primário | Por que importa |
-|---------|------|------|-------------------|-----------------|
-| SNES | Super Mario World | Yoshi's Island 1 | Snes9x2010 | smoke test, 60fps trivial |
-| Mega Drive | Sonic 2 | Chemical Plant loop | Genesis Plus GX | scroll horizontal |
-| GBA | Advance Wars 2 | mid-game map | mGBA | tile-heavy |
-| PS1 | Crash Bandicoot 1 | Wumpa Island praia | Beetle PSX HW | texture streaming |
-| PS1 | Tekken 3 | seleção | Beetle PSX HW | smoke test PS1 |
-| N64 | Super Mario 64 | Bob-omb topo | Mupen64Plus-Next | draw distance |
-| N64 | Zelda OoT | Hyrule Field meio-dia | Mupen64Plus-Next | fog + partículas |
-| PSP | GTA VCS | intro cinemática | PPSSPP | cutscene streaming |
-| PSP | Daxter | fase inicial | PPSSPP | FPS sustentado |
-| DC | Soul Calibur | character select | Flycast | smoke test DC |
-| DC | Crazy Taxi | start of level | Flycast | stress DC |
-
-### 1.2 Matriz explícita libretro vs standalone
-
-Comparação justa exige caminho equivalente. Distros tomam decisões diferentes — ROCKNIX pode usar PPSSPP libretro, dArkOS standalone (ou vice-versa). Misturar caminhos gera benchmark injusto.
-
-Para cada cenário relevante, benchmarkar **ambos os caminhos** quando aplicável:
-
-| Core/Emulador | Libretro | Standalone |
-|---------------|----------|------------|
-| PPSSPP | `ppsspp_libretro.so` | binário `ppsspp` |
-| Beetle PSX HW | `mednafen_psx_hw_libretro.so` | — |
-| DuckStation | `swanstation_libretro.so` | binário `duckstation` |
-| Mupen64Plus-Next | `mupen64plus_next_libretro.so` | `mupen64plus` standalone |
-| Flycast | `flycast_libretro.so` | binário `flycast` |
-| Dolphin | `dolphin_libretro.so` (se existir) | `dolphin-emu` |
-
-**Saída esperada:** CSV separa path libretro vs standalone, não mistura. A comparação "nosso SO vs dArkOS no PSP" precisa ter as duas linhas quando ambas as distros expõem os dois caminhos.
-
-### 1.3 Instrumentação — métricas por run
-
-Coletar durante 60s por cenário, 5 runs cada.
-
-**Frametime:**
-```bash
-# retroarch.cfg do bench (não o de produção)
-video_frame_log = "true"
-perfcnt_enable = "true"
-```
-
-**PSI (Pressure Stall Information) — crítico em 1GB:**
-
-O problema real em 1GB raramente é "faltou RAM". É "entrou em pressão, kswapd ativo, frametime explodiu". PSI captura; MemAvailable não.
-
-```bash
-while [ running ]; do
-  date +%s.%N
-  cat /proc/pressure/cpu      # avg10, avg60, avg300
-  cat /proc/pressure/memory   # some + full
-  cat /proc/pressure/io
-  sleep 1
-done > psi.log
-```
-
-**Eventos de reclaim/OOM:**
-```bash
-dmesg -w > dmesg-during-bench.log &
-# Filtrar: oom, kswapd, pagealloc, reclaim, compact_fail
-```
-
-**Contadores de sistema:**
-```bash
-vmstat 1 60 > vmstat.log
-# si/so (swap in/out), wa (iowait), st (stolen), free, buff, cache
-```
-
-**Frequências CPU/GPU/DDR + temperatura:**
-```bash
-while [ running ]; do
-  for cpu in /sys/devices/system/cpu/cpu[0-3]/cpufreq/scaling_cur_freq; do
-    cat "$cpu"
-  done
-  cat /sys/class/devfreq/*gpu*/cur_freq
-  cat /sys/class/devfreq/*dmc*/cur_freq
-  cat /sys/class/thermal/thermal_zone0/temp
-  sleep 1
-done > freq-therm.log
-```
-
-**Input latency:** câmera 240fps + LED no botão + evento na tela (método RetroArch), executado separadamente, não por run.
-
-### 1.4 Metadados obrigatórios de cada run
-
-Sem isto, um benchmark "bom" vira folklore irreproduzível.
-
-| Campo | Comando |
-|-------|---------|
-| `kernel_release` | `uname -r` |
-| `kernel_commit` | `git -C <kernel-src> rev-parse HEAD` |
-| `mesa_version` | `pacman -Qi mesa` |
-| `mesa_commit` | do PKGBUILD ou `glxinfo` |
-| `retroarch_version` | `retroarch --version` |
-| `core_name` | path do core |
-| `core_version` | metadata do .so |
-| `emulator_standalone_version` | `ppsspp --version` etc. |
-| `rom_sha256` | `sha256sum <rom>` |
-| `bios_sha256` | `sha256sum <bios>` |
-| `save_state_sha256` | idem |
-| `power_source` | `/sys/class/power_supply/*/online` |
-| `battery_charging` | `/sys/class/power_supply/*/status` |
-| `battery_percent` | `/sys/class/power_supply/*/capacity` |
-| `ambient_temp_c` | input manual |
-| `microsd_model` | `cat /sys/block/mmcblk0/device/{name,manfid,oemid,cid}` |
-| `microsd_speed_class` | `dmesg \| grep mmcblk0` ou input manual |
-| `cpu_governor` | `cat /sys/.../scaling_governor` |
-| `cpu_max_freq` | `scaling_max_freq` |
-| `gpu_governor` | `/sys/class/devfreq/*gpu*/governor` |
-| `dmc_governor` | `/sys/class/devfreq/*dmc*/governor` |
-| `zram_active` | `swapon` |
-| `zram_algo` | `cat /sys/block/zram0/comp_algorithm` |
-| `thp_enabled` | `cat /sys/kernel/mm/transparent_hugepage/enabled` |
-| `cma_total_kb` | `grep CmaTotal /proc/meminfo` |
-| `cma_free_kb` | `grep CmaFree /proc/meminfo` |
-| `memavailable_kb_start` | `grep MemAvailable /proc/meminfo` |
-| `bench_script_commit` | `git rev-parse HEAD` do repo do bench |
-
-Juntar em JSON de metadados + CSV de frametime.
-
-### 1.5 Formato de saída
-
-```
-results/
-├── YYYYMMDD-HHMMSS-<distro>-<cenario>-<path>/
-│   ├── meta.json           # metadados da 1.4
-│   ├── frametime.csv       # timestamp, frame_time_ms
-│   ├── psi.log             # 1Hz sampling
-│   ├── freq-therm.log
-│   ├── vmstat.log
-│   ├── dmesg.log
-│   └── summary.json        # p50, p95, p99, fps_mean, max_temp, mem_pressure_pct
-```
-
-### 1.6 Dashboard de comparação
-
-Script que:
-1. Carrega `summary.json` de N runs
-2. Agrupa por `(distro, cenario, core_path)`
-3. Emite pivot com medianas + IQR
-4. Flagga runs com alta variância (σ/μ > 10%) para re-rodar
-5. Gera gráficos de frametime density + PSI timeline
-
-**Critério:** `./bench.sh <cenario> <distro>` grava resultados versionados. `./compare.py --cenario psp_vcs` gera tabela entre distros.
-
-**Risco:** subestimar esforço. Planejar 2–3 dias. É o maior investimento de ROI do projeto.
-
----
-
-# FASE 1 — TIER S (ganhos hipotéticos de 10%+)
-
----
-
-## 2. ✅ zram com zstd `[AMBOS]`
-
-**Hipótese:** em 1GB, PSP/N64/DC entram em pressão de memória que causa spikes de frametime. zram resolve sem I/O em microSD.
-
-**Validação:** PSI `memory.some.avg60` durante PPSSPP/GTA VCS cai ≥50% vs baseline sem zram; frametime p99 estável ou melhor. Se a pressão já é baixa sem zram no cenário, ganho pode ser zero — investigar quais cenas realmente causam pressão.
-
-**Auditar primeiro:**
-```bash
-cat /proc/swaps
-cat /sys/block/zram0/comp_algorithm 2>/dev/null
-zramctl
-sysctl vm.swappiness vm.vfs_cache_pressure vm.page-cluster
-
-grep -r "zram\|swapp\|compcache" /media/disco-local/dArkOS/ /home/douglas/Documentos/distribution/ 2>/dev/null
-```
-
-**Ação:**
-- Device zram com **zstd**, tamanho ~50% da RAM (512MB comprimido ≈ 170MB físicos em razão 1:3). Referência: [ArchWiki zram](https://wiki.archlinux.org/title/Zram).
-- **Desabilitar zswap** (intercepta antes do zram). `zswap.enabled=0` na cmdline.
-- Priority 100 no swapon.
-- **Proibido** swap em microSD.
-
-```ini
-# /etc/sysctl.d/99-handheld.conf
-vm.swappiness = 100
-vm.vfs_cache_pressure = 50
-vm.page-cluster = 0
-vm.dirty_background_ratio = 5
-vm.dirty_ratio = 10
-vm.dirty_expire_centisecs = 3000
-vm.dirty_writeback_centisecs = 1500
-```
-
-**Risco:** CPU de compressão compete com emulador. Se p99 piorar em cenas leves, reduzir tamanho ou trocar zstd → lz4.
-
----
-
-## 3. ✅ DDR frequency scaling (dmc devfreq) `[PERF]`
-
-**Hipótese:** DDR3 é gargalo #1 de bandwidth do RK3326. DMC preso em freq baixa penaliza cenas texture-heavy em PS1/N64/PSP.
-
-**Validação:** FPS em Crash Wumpa — hipótese de +5% a +15% com governor `performance` vs `simple_ondemand` conservador. Se não render, verificar `trans_stat` — governor pode estar escalando, mas o workload não pressiona DDR o suficiente para mostrar diferença.
-
-**Auditar:**
-```bash
-ls /sys/class/devfreq/
-cat /sys/class/devfreq/*dmc*/governor
-cat /sys/class/devfreq/*dmc*/available_frequencies
-cat /sys/class/devfreq/*dmc*/cur_freq
-cat /sys/class/devfreq/*dmc*/trans_stat
-
-find /media/disco-local/dArkOS/ -name "*.dts*" | xargs grep -l "dmc\|memory-controller" 2>/dev/null
-find /home/douglas/Documentos/distribution/ -name "*.dts*" | xargs grep -l "dmc\|memory-controller" 2>/dev/null
-```
-
-**Ação:**
-- `dmc_ondemand` com upthreshold ~40% OU `performance` via hook do frontend durante gameplay.
-- Comparar OPPs de DDR no DT — às vezes há 933MHz escondido.
-
-**Risco:** instabilidade em DDR alto. `memtester` 30min antes de promover.
-
----
-
-## 4. ✅ Mesa atualizado + Panfrost Transaction Elimination `[PERF]`
-
-**Hipótese:** Mesa 25.0+ habilita CRC por padrão, reduzindo bandwidth ao pular write-out de tiles inalterados. Em DDR3 apertado, ganho real. Referência: [Mesa 25.0 PanVK update](https://www.collabora.com/news-and-blog/news-and-events/mesa-25-panvk-moves-towards-production-quality.html).
-
-**Validação:** hipótese de +3% a +8% FPS em cenas 3D com frames similares consecutivos (Zelda OoT parado em Hyrule Field). Se regredir, `PAN_MESA_DEBUG=nocrc` pode estar setado em algum lugar (auditar, vide item 0.2).
-
-**Auditar:**
-```bash
-glxinfo -B 2>/dev/null | grep -i mesa
-pacman -Qi mesa | grep Version
-
-grep -rI "PAN_MESA_DEBUG" /etc/ /usr/share/ ~/.config/
-env | grep -iE "PAN_|MESA_"
-```
-
-**Ação:**
-- Mesa ≥ 25.0, compilado em release (item 0.2).
-- Shader cache persistente:
-  ```bash
-  export MESA_SHADER_CACHE_DIR=/var/cache/mesa_shader_cache
-  export MESA_SHADER_CACHE_MAX_SIZE=256MB
-  ```
-- `MESA_GLTHREAD=true` — testar per-core, regride em cores single-thread.
-- Pré-aquecer cache em ROMs populares no primeiro boot.
-
----
-
-## 5. Panfrost kernel driver — patch de powerup sequence `[PERF]`
-
-**Hipótese:** patch de 2025 que reordena pd→reset→clock corrige comportamento conforme manual Mali. Testado em RK3326 Powkiddy RGB10X. Referência: [patch dri-devel](https://www.mail-archive.com/dri-devel@lists.freedesktop.org/msg541359.html).
-
-**Validação:** estabilidade em suspend/resume; ausência de GPU hangs. Perf pode não mudar — foco é correção.
-
-```bash
-uname -r
-grep -A 30 "panfrost_device_init" /path/to/kernel/drivers/gpu/drm/panfrost/panfrost_device.c
-```
-
----
-
-## 6. ✅ CMA / reserved-memory / carveouts `[FOOTPRINT][PERF]`
-
-**Hipótese:** cada MB reservado em CMA é um MB a menos para userland em 1GB. SoCs Rockchip reservam 64–256MB típico. Reserva excessiva rouba RAM útil. Reserva insuficiente quebra video/VPU/GPU allocations.
-
-**Validação:** MemAvailable em idle sobe ≥50MB após trim, sem quebrar playback V4L2 nem decode HEVC 1080p do frontend.
-
-### 6.1 Auditar estado atual
-
-```bash
-grep -E "^Cma(Total|Free)" /proc/meminfo
-
-mount -t debugfs none /sys/kernel/debug 2>/dev/null
-for d in /sys/kernel/debug/cma/*/; do
-  echo "=== $d ==="
-  cat "$d/count"        # páginas totais
-  cat "$d/used"         # páginas em uso
-done
-
-cat /sys/firmware/devicetree/base/reserved-memory/ranges 2>/dev/null
-ls /sys/firmware/devicetree/base/reserved-memory/
-```
-
-### 6.2 Mapear reservas no DT em runtime
-
-```bash
-dtc -I fs /sys/firmware/devicetree/base > /tmp/runtime.dts
-grep -A 5 "reserved-memory" /tmp/runtime.dts
-grep -B 2 -A 10 "linux,cma\|shared-dma-pool\|no-map" /tmp/runtime.dts
-```
-
-**Reservas típicas em RK3326:**
-- `linux,cma` global — 64–128MB comum; 256MB é exagero em 1GB
-- `rockchip,drm-logo` — splash (pode sumir em release)
-- `ramoops` — crash dump (pequeno, manter)
-- VPU carveouts — verificar se são necessários ou se VPU usa CMA global
-
-### 6.3 Comparar lateralmente
-
-```bash
-for DT in \
-  /media/disco-local/dArkOS/boot/dtbs/rockchip/*rk3326*.dtb \
-  /home/douglas/Documentos/distribution/boot/dtbs/rockchip/*rk3326*.dtb \
-  /boot/dtbs/rockchip/*rk3326*.dtb; do
-  [ -f "$DT" ] || continue
-  echo "=== $DT ==="
-  dtc -I dtb -O dts "$DT" 2>/dev/null | grep -B 2 -A 10 "linux,cma\|reserved-memory"
-done
-```
-
-### 6.4 Ações possíveis
-
-- **Reduzir CMA global** se `CmaFree` fica sempre alto (ex: 128MB reservado, 120MB livre = desperdício).
-- **Remover carveouts de logo/splash** em release se não usados.
-- **Consolidar** VPU/ISP carveouts em CMA compartilhada se possível.
-- Configurar em cmdline: `cma=64M` ou DT override.
-
-### 6.5 Validações pós-ajuste
-
-Testar após cada mudança:
-- Playback H.264 1080p no frontend — OK?
-- V4L2 M2M decode (`ffmpeg -c:v h264_v4l2m2m`) — OK?
-- GPU allocations grandes (PSP 2x internal) — sem falha?
-- `dmesg | grep -iE "cma|dma|alloc fail"` — sem erros novos?
-
-**Risco:** reserva insuficiente falha silenciosamente em decode. Testar playback intensivo antes de promover.
-
----
-
-## 7. ✅ CFLAGS alinhados a Cortex-A35 `[PERF]`
-
-**Hipótese:** A35 é in-order dual-issue, IPC baixo. Código genérico deixa 3–10% na mesa por core.
-
-**Validação:** FPS em Sonic 2 Chemical Plant sobe ≥2% com `-O2 -flto -mtune=cortex-a35` vs `-O2` genérico. Se não, verificar se o build realmente absorveu as flags (`readelf -A` nos binários).
-
-**`makepkg.conf`:**
-```bash
-CFLAGS="-march=armv8-a+crc+crypto -mtune=cortex-a35 -O2 -pipe -fno-plt -fstack-protector-strong"
-CXXFLAGS="${CFLAGS}"
-LDFLAGS="-Wl,-O1,--sort-common,--as-needed,-z,relro,-z,now"
-MAKEFLAGS="-j$(nproc)"
-```
-
-**Observações:**
-- **Não** `-O3` universal. L1I 32KB, L2 256KB; `-O3` infla binário e pressiona cache. Pontual após bench.
-- **Não** `-mcpu=cortex-a35` em libs genéricas (quebra portabilidade).
-- `-flto` em: mesa, retroarch, cores libretro, sdl2, ffmpeg, zlib-ng.
-
-**Pacotes para recompilar:**
-1. mesa (item 4)
-2. retroarch
-3. Cores libretro em uso
-4. sdl2
-5. ffmpeg com `--enable-neon --cpu=cortex-a35`
-6. **zlib-ng** (substitui zlib, 10–30% em descompressão de ROMs zipadas)
-7. zstd (para zram)
-
----
-
-## 8. ✅ Device Tree diff profundo `[PERF][FOOTPRINT]`
-
-**Hipótese:** DT é fonte comum de perf "misteriosa". Divergências entre distros raramente são acidentais.
-
-**Validação:** `dt-findings.md` com N divergências categorizadas (cosmética / perf / estabilidade), cada uma com decisão registrada.
-
-**Método:**
-```bash
-mkdir -p /tmp/dt-diff/{nosso,arkos,rocknix}
-dtc -I dtb -O dts -o /tmp/dt-diff/nosso/board.dts /boot/dtbs/rockchip/*rk3326*.dtb
-# Repetir para concorrentes
-
-diff -u /tmp/dt-diff/nosso/board.dts /tmp/dt-diff/arkos/board.dts > /tmp/dt-diff/nosso-vs-arkos.patch
-```
-
-**Pontos:**
-- `cpu0-opp-table` — 1.3GHz vs 1.5GHz?
-- `gpu-opp-table` — freqs Mali-G31 (200/300/400/520/650)
-- `dmc-opp-table` — DDR (786/933/1066)
-- `thermal-zones` → trips + cooling-maps
-- `regulator-*` → min/max-microvolt (margem de undervolt)
-- `rockchip,pwm-regulator`
-- `reserved-memory` — ver item 6
-
----
-
-# FASE 2 — TIER A (ganhos hipotéticos de 3–10%)
-
----
-
-## 9. ✅ I/O scheduler e bloco para microSD `[PERF]`
-
-**Hipótese:** bfq + read_ahead alto reduz stutter em streaming de ROMs grandes.
-
-**Validação:** load de ISO PSP GTA VCS cai ≥15%.
-
-```bash
-for d in /sys/block/mmcblk*; do
-  echo "$d: $(cat $d/queue/scheduler)"
-done
-
-cat > /etc/udev/rules.d/60-mmcblk-scheduler.rules <<'EOF'
-ACTION=="add|change", KERNEL=="mmcblk[0-9]*", ATTR{queue/scheduler}="bfq"
-ACTION=="add|change", KERNEL=="mmcblk[0-9]*", ATTR{queue/iosched/low_latency}="1"
-ACTION=="add|change", KERNEL=="mmcblk[0-9]*", ATTR{queue/read_ahead_kb}="2048"
-ACTION=="add|change", KERNEL=="mmcblk[0-9]*", ATTR{queue/nr_requests}="128"
-ACTION=="add|change", KERNEL=="mmcblk[0-9]*", ATTR{queue/rq_affinity}="2"
-EOF
-```
-
-Mount: `noatime,nodiratime,commit=600,lazytime`. Manter ext4 (F2FS arriscado em microSD variável).
-
----
-
-## 10. RetroArch em KMS/DRM puro `[FOOTPRINT][PERF]`
-
-**Hipótese:** evitar compositor economiza 30–60MB residentes + buffer copy por frame.
-
-**Validação:** MemAvailable em RetroArch fullscreen sobe ≥40MB vs modo com compositor.
-
-```
-video_driver = "drm"
-video_context_driver = "kms"
-```
-
-Auditar: frontend está subindo X? Pode rodar em `cage` (Wayland kiosk) ou fbdev?
-
----
-
-## 11. Render resolution sanity por core `[PERF]`
-
-**Hipótese:** renderizar PSP em 2x (1920×1088) em painel 640×480 é desperdício puro.
-
-Presets default:
-- PPSSPP: `1x` em ≤480p, `1.5x` em 720p
-- Beetle PSX HW: `1x`
-- Mupen64Plus-Next: nativa (320x240 ou 640x480)
-- Flycast: nativa (640×480)
-- Dolphin/PCSX2: não aplicável em A35
-
-**Ação:** `core-overrides/` versionado com presets conservadores por default.
-
----
-
-## 12. ✅ Systemd trim cirúrgico `[FOOTPRINT]`
-
-**Hipótese:** daemons socket-activated em vez de residentes liberam ≥60MB RSS sem remover funcionalidade.
-
-**Validação:** RSS total de serviços residentes cai ≥60MB.
-
-```bash
-systemd-analyze blame > /tmp/blame-nosso.txt
-systemctl list-units --state=running --type=service
-```
-
-**Alvos:**
-- `ModemManager`, `cups`, `packagekit`, `tracker-*`, `geoclue`, `fwupd` → **mask**
-- `avahi-daemon`, `bluetooth`, WiFi → **socket-activate**
-
-Journal em RAM:
-```ini
-# /etc/systemd/journald.conf.d/handheld.conf
-[Journal]
-Storage=volatile
-RuntimeMaxUse=16M
-```
-
----
-
-## 13. ✅ CPU governor + thermal trips `[PERF]`
-
-**Hipótese:** sem fan, throttle precoce derruba perf sustentada. Trips conservadores herdados de smartphone.
-
-**Validação:** 30min PSP mantém CPU ≥1.4GHz em ≥95% do tempo.
-
-```bash
-for z in /sys/class/thermal/thermal_zone*; do
-  echo "=== $z ==="
-  cat $z/type
-  for t in $z/trip_point_*_temp; do
-    echo "  $t: $(cat $t)"
-  done
-done
-```
-
-**Ação:**
-- `performance` em gameplay via pre-launch hook; `schedutil` em idle.
-- Comparar trips com dArkOS/ROCKNIX.
-- Undervolt via `cpu-supply` no DT se houver margem (validar com stress-ng + memtester).
-
----
-
-## 14. Vulkan (PanVK) — abordagem para Mali-G31 `[DECISÃO]`
-
-**Situação:** PanVK conformante apenas em Mali-G610. Mali-G31 (Bifrost v9) = não-conformante. Mesa não expõe PanVK por padrão em G31.
-
-**Três opções:**
-
-- **A (default recomendado):** manter OpenGL ES, não expor Vulkan ao usuário, infra pronta.
-- **B (opt-in):** toggle experimental, `PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1` sob demanda, documentar risco.
-- **C (Zink pontual):** OpenGL sobre Vulkan, raramente ganha em G31.
-
-Investigar o que dArkOS/ROCKNIX fazem:
-```bash
-grep -rI "vulkan\|PAN_I_WANT\|panvk" /media/disco-local/dArkOS/ /home/douglas/Documentos/distribution/ 2>/dev/null
-```
-
-**Critério:** decisão em `vulkan-decision.md`. Se B, lista de cores "funcionando"/"quebrados".
-
----
-
-# FASE 3 — TIER B (qualidade e ganhos menores)
-
----
-
-## 15. ✅ Input latency `[PERF]`
-
-**Hipótese:** `CONFIG_HZ=1000` + polling 1000Hz + run-ahead 1–2 frames em cores leves reduz latência perceptível.
-
-```bash
-zcat /proc/config.gz | grep CONFIG_HZ=
-cat /proc/bus/input/devices
-```
-
-RetroArch:
-- `video_frame_delay` — testar 0–4 per-core
-- `video_hard_sync = "true"` (custa CPU)
-- Run-ahead: ON em SNES/GBA/MD/NES, OFF em PS1/N64/PSP/DC
-
----
-
-## 16. Áudio — stack mínimo `[FOOTPRINT][PERF]`
-
-**Hipótese:** ALSA direto economiza ~20MB vs PipeWire. Se PipeWire necessário, quantum 1024.
-
-```ini
-# /etc/pipewire/pipewire.conf.d/99-handheld.conf
-default.clock.quantum = 1024
-default.clock.min-quantum = 512
-default.clock.max-quantum = 2048
-```
-
-**Validação:** sem crackle em 30min PSP. Latência áudio ≤96ms.
-
----
-
-## 17. Locales, fonts, docs `[FOOTPRINT]`
-
-**Hipótese:** ~300MB recuperados em microSD, menos pressão FS cache.
-
-```ini
-# /etc/pacman.conf
-NoExtract = usr/share/locale/* !usr/share/locale/en* !usr/share/locale/pt_BR*
-NoExtract = usr/share/man/* usr/share/doc/* usr/share/info/*
-NoExtract = usr/share/gtk-doc/*
-```
-
----
-
-## 18. Per-game override infrastructure `[PERF]`
-
-```
-/opt/game-profiles/
-├── system/
-│   ├── psp/
-│   │   ├── _default.conf
-│   │   ├── ULUS10160.conf    # GTA VCS
-│   │   └── ULUS10041.conf    # Daxter
-```
-
-Hook pre-launch aplica: cpufreq governor, dmc governor, GPU min_freq, RetroArch `--appendconfig`, Mesa env vars.
-
----
-
-## 19. Frontend — lazy loading e cache `[FOOTPRINT][PERF]`
-
-Auditar:
-- Parse de gamelist.xml em lib ≥500 jogos
-- Thumbnails lazy vs eager
-- Preview videos via V4L2 M2M vs software
-- Shaders de tema pesados
-
-**Hipótese:** 2–5s de boot e 30–60MB de RAM.
-
----
-
-## 20. Firmware Mali e V4L2 M2M `[PERF]`
-
-```bash
-dmesg | grep -i "mali\|panfrost\|firmware"
-ls /lib/firmware/arm/mali/ 2>/dev/null
-
-ffmpeg -hwaccels
-ffmpeg -i preview.mp4 -c:v h264_v4l2m2m -f null -
-```
-
-Zero-copy DMA-BUF decoder→GPU→display.
-
----
-
-# FASE 4 — Infraestrutura e meta
-
----
-
-## 21. Repositório de comparativos `[INFRA]`
-
-`/docs/comparativos/` versionado:
-- `build-audit.md` (item 0)
-- `dt-findings.md` (item 8)
-- `cma-findings.md` (item 6)
-- `sysctl-findings.md` — diff sysctl entre distros
-- `service-audit.md` — daemons ativos por distro
-- `cflags-audit.md` — flags por distro
-- `mesa-version.md`
-- `kernel-version.md`
-- `vulkan-decision.md`
-
----
-
-## 22. CI de performance `[INFRA]`
-
-Quando item 1 maduro:
-- Build → flash → bench → regressão em PR
-- Alertar PRs que regridam p95 ≥3%
-
----
-
-# Classificação rápida: perf direta vs footprint
-
-| Item | Tag | Tier | Métrica principal |
-|------|-----|------|-------------------|
-| 0 | INFRA/PERF | 0 | ceiling do resto |
-| 1 | INFRA | 0 | base de comparação |
-| 2 | AMBOS | S | PSI memory + frametime p99 |
-| 3 | PERF | S | FPS texture-heavy |
-| 4 | PERF | S | FPS + bandwidth GPU |
-| 5 | PERF | S | estabilidade GPU |
-| 6 | AMBOS | S | MemAvailable + decode ok |
-| 7 | PERF | S | FPS sustentado |
-| 8 | AMBOS | S | depende da divergência |
-| 9 | PERF | A | load time ROMs |
-| 10 | AMBOS | A | MemAvailable + frametime |
-| 11 | PERF | A | FPS em PS1/PSP HW |
-| 12 | FOOTPRINT | A | RSS daemons |
-| 13 | PERF | A | freq sustentada |
-| 14 | DECISÃO | A | — |
-| 15 | PERF | B | input latency ms |
-| 16 | AMBOS | B | RSS + crackle |
-| 17 | FOOTPRINT | B | disk + cache |
-| 18 | PERF | B | FPS per-game |
-| 19 | AMBOS | B | boot time + RSS |
-| 20 | PERF | B | CPU em vídeo |
-
----
-
-# Ordem recomendada de execução
-
-| Fase | Itens | Duração estimada |
-|------|-------|------------------|
-| **Pré-req obrigatório** | 0, 1 | 3–5 dias |
-| **Quick wins RAM/DDR/CMA** | 2, 3, 6 | 2 dias |
-| **Stack gráfico** | 4, 5 | 1–2 dias |
-| **DT archeology** | 8 | 2–3 dias |
-| **Compilação tunada** | 7 | 1 dia + build time |
-| **Polimento stack** | 9, 10, 11, 12, 13 | 3 dias |
-| **Decisão Vulkan** | 14 | 1 dia |
-| **Latência e áudio** | 15, 16 | 1 dia |
-| **QoL e economia** | 17, 18, 19, 20 | 2 dias |
-| **Documentação** | 21 | paralelo |
-| **CI** | 22 | longo prazo |
-
-**Total estimado:** ~3–4 semanas focadas + build time.
-
----
-
-# Itens explicitamente DESCARTADOS (e por quê)
-
-- ~~big.LITTLE / EAS~~ — RK3326 é 4× A35 homogêneos.
-- ~~Transparent HugePages~~ — fragmenta em 1GB. Manter `never`.
-- ~~CPU isolation (`isolcpus`, `nohz_full`)~~ — 4 cores é pouco.
-- ~~PGO em Dolphin/PCSX2/RPCS3~~ — não rodam com perf útil no A35.
-- ~~F2FS~~ — ext4 mais seguro em microSD variável.
-- ~~Swap em microSD~~ — destrói cartão.
-- ~~`-O3` universal~~ — icache apertado; pontual após bench.
-- ~~`mitigations=off`~~ — ganho marginal em A35, risco ruim.
-
----
-
-# Quando revisitar este TODO
-
-- Após itens 0 e 1: reprioritizar com dados.
-- A cada release do Mesa: PanVK em G31 saiu de experimental?
-- A cada release do kernel: patches novos de Panfrost/rockchip?
-- Se o projeto ganhar contribuidores: quebrar em issues granulares.
\ No newline at end of file
diff --git a/distributions/ArchR/options b/distributions/ArchR/options
index 15f96ca15b..9d2a19d2bd 100644
--- a/distributions/ArchR/options
+++ b/distributions/ArchR/options
@@ -187,8 +187,12 @@
 # Root filesystem type (squashfs / ext4)
   ROOTFS_TYPE="ext4"
 
-# Boot partition size (FAT32, kernel + DTBs + bootloader), in MB
-  BOOT_SIZE=256
+# Boot partition size (FAT32, kernel + DTBs + bootloader), in MB.
+# 272 MB chosen so the partition exceeds 65525 4-KB clusters and
+# qualifies as real FAT32 by spec (256 MB landed at 65404 clusters,
+# 121 short, triggering "less than the required minimum" warnings on
+# every dosfsck run).
+  BOOT_SIZE=272
 
 # Root partition size (ext4 rootfs), in MB
   ROOT_SIZE=4608
diff --git a/documentation/PER_DEVICE_DOCUMENTATION/RK3326/SUPPORTED_EMULATORS_AND_CORES.md b/documentation/PER_DEVICE_DOCUMENTATION/RK3326/SUPPORTED_EMULATORS_AND_CORES.md
index 52265f5f16..4b16bca08d 100644
--- a/documentation/PER_DEVICE_DOCUMENTATION/RK3326/SUPPORTED_EMULATORS_AND_CORES.md
+++ b/documentation/PER_DEVICE_DOCUMENTATION/RK3326/SUPPORTED_EMULATORS_AND_CORES.md
@@ -64,12 +64,11 @@
 |Nintendo|Game &amp; Watch (gameandwatch)|1980|`gameandwatch`|.mgw .zip .7z|**retroarch:** gw<br>**retroarch:** mame<br>|
 |Nintendo|Game Boy (gb)|1989|`gb`|.gb .gbc .zip .7z|**retroarch:** gambatte (default)<br>**retroarch:** sameboy<br>**retroarch:** gearboy<br>**retroarch:** tgbdual<br>**retroarch:** mgba<br>**retroarch:** vbam<br>**retroarch:** DoubleCherryGB<br>**retroarch:** skyemu<br>**mednafen:** gb<br>|
 |Nintendo|Game Boy (Hacks) (gbh)|1989|`gbh`|.gb .zip .7z|**retroarch:** gambatte (default)<br>**retroarch:** sameboy<br>**retroarch:** gearboy<br>**retroarch:** tgbdual<br>**retroarch:** mgba<br>**retroarch:** vbam<br>**retroarch:** DoubleCherryGB<br>**retroarch:** skyemu<br>**mednafen:** gb<br>|
-|Nintendo|Game Boy Advance (gba)|2001|`gba`|.gba .zip .7z|**retroarch:** mgba (default)<br>**retroarch:** vbam<br>**retroarch:** vba_next<br>**retroarch:** beetle_gba<br>**retroarch:** skyemu<br>**retroarch:** gpsp<br>**mednafen:** gba<br>|
+|Nintendo|Game Boy Advance (gba)|2001|`gba`|.gba .zip .7z|**retroarch:** gpsp (default)<br>**retroarch:** mgba<br>**retroarch:** vbam<br>**retroarch:** vba_next<br>**retroarch:** beetle_gba<br>**retroarch:** skyemu<br>**mednafen:** gba<br>|
 |Nintendo|Game Boy Advance (Hacks) (gbah)|2001|`gbah`|.gba .zip .7z|**retroarch:** mgba (default)<br>**retroarch:** vbam<br>**retroarch:** vba_next<br>**retroarch:** beetle_gba<br>**retroarch:** skyemu<br>**retroarch:** gpsp<br>**mednafen:** gba<br>|
 |Nintendo|Game Boy Advance video (gbav)|2001|`gbav`|.gba .zip .7z|**retroarch:** mgba (default)<br>**retroarch:** vbam<br>**retroarch:** vba_next<br>**retroarch:** beetle_gba<br>**retroarch:** skyemu<br>**retroarch:** gpsp<br>**mednafen:** gba<br>|
 |Nintendo|Game Boy Color (gbc)|1998|`gbc`|.gb .gbc .zip .7z|**retroarch:** gambatte (default)<br>**retroarch:** sameboy<br>**retroarch:** gearboy<br>**retroarch:** tgbdual<br>**retroarch:** mgba<br>**retroarch:** vbam<br>**retroarch:** DoubleCherryGB<br>**retroarch:** skyemu<br>**mednafen:** gb<br>|
 |Nintendo|Game Boy Color (Hacks) (gbch)|1998|`gbch`|.gb .gbc .zip .7z|**retroarch:** gambatte (default)<br>**retroarch:** sameboy<br>**retroarch:** gearboy<br>**retroarch:** tgbdual<br>**retroarch:** mgba<br>**retroarch:** vbam<br>**retroarch:** DoubleCherryGB<br>**retroarch:** skyemu<br>**mednafen:** gb<br>|
-|Nintendo|GameCube (gamecube)|2001|`gamecube`|.gcm .iso .gcz .ciso .wbfs .rvz .dol .m3u|**retroarch:** dolphin (default)<br>|
 |Nintendo|NES (Hacks) (nesh)|1985|`nesh`|.nes .unif .unf .zip .7z|**retroarch:** nestopia (default)<br>**retroarch:** fceumm<br>**retroarch:** quicknes<br>**retroarch:** mesen<br>**mednafen:** nesh<br>|
 |Nintendo|Nintendo 64 (n64)|1996|`n64`|.z64 .n64 .v64 .zip .7z|**retroarch:** mupen64plus_next (default)<br>**retroarch:** mupen64plus<br>**retroarch:** parallel_n64<br>**mupen64plus:** mupen64plus-sa<br>|
 |Nintendo|Nintendo DS (nds)|2005|`nds`|.nds .zip .7z|**drastic:** drastic-sa (default)<br>**melonds:** melonds-sa<br>**retroarch:** melonds<br>**retroarch:** melondsds<br>**retroarch:** desmume<br>**retroarch:** skyemu<br>|
@@ -82,8 +81,6 @@
 |Nintendo|Super Nintendo (Hacks) (snesh)|1991|`snesh`|.smc .fig .sfc .swc .zip .7z|**retroarch:** snes9x (default)<br>**retroarch:** snes9x2010<br>**retroarch:** snes9x2002<br>**retroarch:** snes9x2005_plus<br>**retroarch:** beetle_supafaust<br>**retroarch:** bsnes_mercury_performance<br>**mednafen:** snes_faust<br>|
 |Nintendo|Super Nintendo (snes)|1991|`snes`|.smc .fig .sfc .swc .zip .7z|**retroarch:** snes9x (default)<br>**retroarch:** snes9x2010<br>**retroarch:** snes9x2002<br>**retroarch:** snes9x2005_plus<br>**retroarch:** beetle_supafaust<br>**retroarch:** bsnes_mercury_accuracy<br>**retroarch:** bsnes_mercury_balanced<br>**retroarch:** bsnes_mercury_performance<br>**mednafen:** snes_faust<br>**retroarch:** bsnes_mercury_accuracy<br>**retroarch:** bsnes_mercury_balanced<br>**retroarch:** bsnes_mercury_accuracy<br>**retroarch:** bsnes_mercury_balanced<br>**mednafen:** snes_faust<br>|
 |Nintendo|Virtual Boy (virtualboy)|1995|`virtualboy`|.vb .zip .7z|**retroarch:** beetle_vb (default)<br>**mednafen:** vb<br>|
-|Nintendo|Wii (wii)|2006|`wii`|.gcm .iso .gcz .ciso .wbfs .rvz .dol .wad .m3u|**retroarch:** dolphin (default)<br>|
-|Nintendo|WiiWare (wiiware)|2008|`wiiware`|.gcm .iso .gcz .ciso .wbfs .rvz .dol .wad .m3u|**retroarch:** dolphin (default)<br>|
 |Palm, Inc.|Palm OS (palm)|1996|`palm`|.prc .pqa .img .pdb .zip|**retroarch:** mu (default)<br>|
 |Panasonic|3DO (3do)|1993|`3do`|.iso .bin .chd .cue|**retroarch:** opera (default)<br>|
 |Philips|CD-i (cdi)|1991|`cdi`|.chd .cue .iso|**retroarch:** same_cdi (default)<br>|
diff --git a/packages/compress/zlib/package.mk b/packages/compress/zlib/package.mk
index 289a4d6b0c..74ead0aa58 100644
--- a/packages/compress/zlib/package.mk
+++ b/packages/compress/zlib/package.mk
@@ -1,17 +1,50 @@
 # SPDX-License-Identifier: GPL-2.0-or-later
 # Copyright (C) 2009-2016 Stephan Raue (stephan@openelec.tv)
+# Copyright (C) 2026 ArchR (https://github.com/archr-linux/Arch-R)
+
+# Drop-in zlib replacement using zlib-ng compiled with --zlib-compat.
+# ABI/API identical to upstream zlib (libz.so.1), but ships SIMD-aware
+# inflate/deflate paths that NEON-enabled the inner loops on Cortex-A35.
+# Real impact: ROM zip extraction in PortMaster/RetroArch and ext2/ext4
+# compression utilities run noticeably faster.
+#
+# We keep PKG_NAME="zlib" so every other package's PKG_DEPENDS_*="zlib"
+# transparently picks up the replacement. Source unpack is overridden
+# because the upstream tarball expands to zlib-ng-${PKG_VERSION}/.
 
 PKG_NAME="zlib"
-PKG_VERSION="1.3.1"
-PKG_SHA256="9a93b2b7dfdac77ceba5a558a580e74667dd6fede4585b91eefb60f03b72df23"
+PKG_VERSION="2.2.4"
+PKG_SHA256="a73343c3093e5cdc50d9377997c3815b878fd110bf6511c2c7759f2afb90f5a3"
 PKG_LICENSE="OSS"
-PKG_SITE="http://www.zlib.net"
-PKG_URL="https://zlib.net/fossils/${PKG_NAME}-${PKG_VERSION}.tar.gz"
+PKG_SITE="https://github.com/zlib-ng/zlib-ng"
+PKG_URL="https://github.com/zlib-ng/zlib-ng/archive/refs/tags/${PKG_VERSION}.tar.gz"
+PKG_SOURCE_NAME="zlib-ng-${PKG_VERSION}.tar.gz"
 PKG_DEPENDS_HOST="ccache:host cmake:host"
 PKG_DEPENDS_TARGET="cmake:host gcc:host"
-PKG_LONGDESC="A general purpose (ZIP) data compression library."
+PKG_LONGDESC="zlib-ng compiled in --zlib-compat mode — drop-in libz with SIMD-accelerated inflate/deflate."
 PKG_TOOLCHAIN="cmake-make"
 
-PKG_CMAKE_OPTS_HOST="-DINSTALL_PKGCONFIG_DIR=${TOOLCHAIN}/lib/pkgconfig"
+# zlib-ng reads CMAKE flags directly. ZLIB_COMPAT=ON is what makes this
+# build emit libz.so.1 (instead of libz-ng.so.2) so existing pkgconfig
+# pointers keep working. WITH_GZFILEOP keeps gzopen()/gzread()/etc.
+# Tests/benchmarks off to avoid pulling gtest into the cross toolchain.
+PKG_CMAKE_OPTS_HOST="-DZLIB_COMPAT=ON \
+                     -DZLIB_ENABLE_TESTS=OFF \
+                     -DZLIBNG_ENABLE_TESTS=OFF \
+                     -DWITH_GTEST=OFF \
+                     -DWITH_GZFILEOP=ON \
+                     -DINSTALL_PKGCONFIG_DIR=${TOOLCHAIN}/lib/pkgconfig"
 
-PKG_CMAKE_OPTS_TARGET="-DINSTALL_PKGCONFIG_DIR=/usr/lib/pkgconfig"
+PKG_CMAKE_OPTS_TARGET="-DZLIB_COMPAT=ON \
+                       -DZLIB_ENABLE_TESTS=OFF \
+                       -DZLIBNG_ENABLE_TESTS=OFF \
+                       -DWITH_GTEST=OFF \
+                       -DWITH_GZFILEOP=ON \
+                       -DWITH_OPTIM=ON \
+                       -DWITH_NEON=ON \
+                       -DINSTALL_PKGCONFIG_DIR=/usr/lib/pkgconfig"
+
+unpack() {
+  mkdir -p "${PKG_BUILD}"
+  tar --strip-components=1 -xf "${SOURCES}/${PKG_NAME}/${PKG_SOURCE_NAME}" -C "${PKG_BUILD}"
+}
diff --git a/packages/graphics/mesa/package.mk b/packages/graphics/mesa/package.mk
index 8fa9ab44b9..237dc765b4 100644
--- a/packages/graphics/mesa/package.mk
+++ b/packages/graphics/mesa/package.mk
@@ -3,8 +3,8 @@
 # Copyright (C) 2018-present Team LibreELEC (https://libreelec.tv)
 
 PKG_NAME="mesa"
-PKG_VERSION="26.0.3"
-PKG_SHA256="ddb7443d328e89aa45b4b6b80f077bf937f099daeca8ba48cabe32aab769e134"
+PKG_VERSION="26.0.5"
+PKG_SHA256="d229c9937d9a25ca0a8958c59f425174563d300ec42acbea2dbe84a055023368"
 PKG_LICENSE="OSS"
 PKG_SITE="http://www.mesa3d.org/"
 PKG_URL="https://mesa.freedesktop.org/archive/mesa-${PKG_VERSION}.tar.xz"
diff --git a/projects/ArchR/devices/RK3326/linux/dts/rockchip/rk3326-gameconsole-r3xs.dtsi b/projects/ArchR/devices/RK3326/linux/dts/rockchip/rk3326-gameconsole-r3xs.dtsi
index de1b0d51e2..5e95dab3ab 100644
--- a/projects/ArchR/devices/RK3326/linux/dts/rockchip/rk3326-gameconsole-r3xs.dtsi
+++ b/projects/ArchR/devices/RK3326/linux/dts/rockchip/rk3326-gameconsole-r3xs.dtsi
@@ -261,7 +261,7 @@
 			vdd_arm: DCDC_REG2 {
 				regulator-name = "vdd_arm";
 				regulator-min-microvolt = <950000>;
-				regulator-max-microvolt = <1350000>;
+				regulator-max-microvolt = <1450000>;
 				regulator-ramp-delay = <6001>;
 				regulator-always-on;
 				regulator-boot-on;
diff --git a/projects/ArchR/devices/RK3326/patches/linux/000-rk3326-dts.patch b/projects/ArchR/devices/RK3326/patches/linux/000-rk3326-dts.patch
index 7d2e7d3d96..28d2bae258 100644
--- a/projects/ArchR/devices/RK3326/patches/linux/000-rk3326-dts.patch
+++ b/projects/ArchR/devices/RK3326/patches/linux/000-rk3326-dts.patch
@@ -35,9 +35,10 @@
  		};
 -		opp-1296000000 {
 -			opp-hz = /bits/ 64 <1296000000>;
+-			opp-microvolt = <1350000 1350000 1350000>;
 +		opp-1512000000 {
 +			opp-hz = /bits/ 64 <1512000000>;
- 			opp-microvolt = <1350000 1350000 1350000>;
++			opp-microvolt = <1400000 1400000 1400000>;
  			clock-latency-ns = <40000>;
 +			turbo-mode;
  		};
diff --git a/projects/ArchR/packages/apps/portmaster/package.mk b/projects/ArchR/packages/apps/portmaster/package.mk
index 7443e1d8af..3357c3386d 100644
--- a/projects/ArchR/packages/apps/portmaster/package.mk
+++ b/projects/ArchR/packages/apps/portmaster/package.mk
@@ -7,7 +7,7 @@ PKG_ARCH="arm aarch64"
 PKG_LICENSE="MIT"
 PKG_SITE="https://github.com/PortsMaster/PortMaster-GUI"
 PKG_URL="https://github.com/PortsMaster/PortMaster-GUI/releases/download/${PKG_VERSION}/PortMaster.zip"
-PKG_DEPENDS_TARGET="toolchain archr-hotkey gamecontrollerdb oga_controls control-gen xmlstarlet list-guid gst-plugins-base"
+PKG_DEPENDS_TARGET="toolchain archr-hotkey gamecontrollerdb oga_controls control-gen xmlstarlet list-guid gst-plugins-base Python3 xz"
 PKG_LONGDESC="Portmaster - a simple tool that allows you to download various game ports"
 PKG_TOOLCHAIN="manual"
 
diff --git a/projects/ArchR/packages/apps/portmaster/scripts/start_portmaster.sh b/projects/ArchR/packages/apps/portmaster/scripts/start_portmaster.sh
index d3472c7537..ba787fe200 100755
--- a/projects/ArchR/packages/apps/portmaster/scripts/start_portmaster.sh
+++ b/projects/ArchR/packages/apps/portmaster/scripts/start_portmaster.sh
@@ -23,12 +23,12 @@ chmod +x /storage/.config/PortMaster/mapper.txt
 
 
 #Use our gamecontrollerdb.txt
-rm -r gamecontrollerdb.txt
+rm -rf gamecontrollerdb.txt
 ln -sf /usr/config/SDL-GameControllerDB/gamecontrollerdb.txt gamecontrollerdb.txt
 
 #Delete old PortMaster fold first (we can probably remove this later)
-if [ ! -f "/storage/roms/ports/PortMaster/pugwash" ]; then
-    rm -r /storage/roms/ports/PortMaster
+if [ -d "/storage/roms/ports/PortMaster" ] && [ ! -f "/storage/roms/ports/PortMaster/pugwash" ]; then
+    rm -rf /storage/roms/ports/PortMaster
 fi
 
 #Make sure roms/ports/PortMaster folder exists
@@ -38,13 +38,11 @@ if [ ! -d "/storage/roms/ports/PortMaster" ]; then
 fi
 
 #We dont use tasksetter, delete it
-if [ -f /storage/roms/ports/PortMaster/tasksetter ]; then
-  rm -r /storage/roms/ports/PortMaster/tasksetter
-fi
+rm -rf /storage/roms/ports/PortMaster/tasksetter
 
 #Use PortMasters gptokeyb
-rm gptokeyb
-cp /storage/roms/ports/PortMaster/gptokeyb gptokeyb
+rm -f gptokeyb
+[ -x /storage/roms/ports/PortMaster/gptokeyb ] && cp /storage/roms/ports/PortMaster/gptokeyb gptokeyb
 
 #Copy over required files for ports
 cp /storage/.config/PortMaster/control.txt /storage/roms/ports/PortMaster/control.txt
diff --git a/projects/ArchR/packages/archr/autostart/055-hdmi-check b/projects/ArchR/packages/archr/autostart/055-hdmi-check
index 9bb3362b54..788716dc4d 100755
--- a/projects/ArchR/packages/archr/autostart/055-hdmi-check
+++ b/projects/ArchR/packages/archr/autostart/055-hdmi-check
@@ -7,12 +7,19 @@
 # The purpose of this script is to permanently set the resolution
 # output for hdmi when connected.
 
-xres="$(cat /sys/class/graphics/fb0/modes | grep -o -P '(?<=:).*(?=p-)' | cut -dx -f1)"
+# /sys/class/graphics/fb0/modes may not be populated yet if DRM is still
+# probing. Default to 0 so the comparison below falls through cleanly
+# instead of breaking arithmetic.
+xres="$(cat /sys/class/graphics/fb0/modes 2>/dev/null | grep -o -P '(?<=:).*(?=p-)' | cut -dx -f1)"
+[ -z "${xres}" ] && xres=0
 
 # drm_tool source available at https://github.com/christianhaitian/drm_tool.git
-
-mode="$(/usr/bin/drm_tool list |  awk '/1280x720.*60/ {print substr($2,1,length($2)-1); exit}')"
-mode2="$(/usr/bin/drm_tool list |  awk '/1920x1080.*60/ {print substr($2,1,length($2)-1); exit}')"
+# Wrap with a timeout: drm_tool can block on the DRM master if
+# rockchipdrm is still binding, which on a slow microSD has been
+# observed to hang this autostart script and (via Before=emustation)
+# the entire boot.
+mode="$(timeout 10 /usr/bin/drm_tool list 2>/dev/null | awk '/1280x720.*60/ {print substr($2,1,length($2)-1); exit}')"
+mode2="$(timeout 10 /usr/bin/drm_tool list 2>/dev/null | awk '/1920x1080.*60/ {print substr($2,1,length($2)-1); exit}')"
 
 # Now we tell drm what the hdmi mode is by writing to /var/run/drmMode
 # This will get picked up by SDL2 as long as it's been patched with the batocera
diff --git a/projects/ArchR/packages/archr/config/system/configs/system.cfg b/projects/ArchR/packages/archr/config/system/configs/system.cfg
index 3fa9664d52..e105093a6b 100644
--- a/projects/ArchR/packages/archr/config/system/configs/system.cfg
+++ b/projects/ArchR/packages/archr/config/system/configs/system.cfg
@@ -181,6 +181,7 @@ system.battery.warning=0
 system.battery.warning_threshold=25
 system.hostname=@DEVICENAME@
 system.language=en_US
+enable.turbo-mode=0
 system.cpugovernor=ondemand
 system.loglevel=none
 system.merged.storage=0
diff --git a/projects/ArchR/packages/archr/profile.d/099-freqfunctions b/projects/ArchR/packages/archr/profile.d/099-freqfunctions
index 07039ed104..9a17ed6489 100644
--- a/projects/ArchR/packages/archr/profile.d/099-freqfunctions
+++ b/projects/ArchR/packages/archr/profile.d/099-freqfunctions
@@ -66,20 +66,24 @@ set_cpu_boost() {
   esac
 }
 
-# Pin CPU max_freq across all policies to the highest available OPP
-# (boost included if cpufreq/boost=1 was already toggled). Mirrors the
-# GPU helper above. ROCKNIX achieves the same effect by trimming the DT
-# ladder; we keep the full ladder for idle and force the ceiling up at
-# runtime.
+# Pin CPU max_freq across all policies to the highest available OPP.
+# Turbo OPPs (DT `turbo-mode`) live in scaling_boost_frequencies, NOT in
+# scaling_available_frequencies, even after cpufreq/boost=1. Combine
+# both so "highest" really hits 1512 MHz on RK3326 when the user's
+# turbo switch is on, and naturally falls back to 1416 when off (boost
+# file then reports an empty list).
 set_cpu_max_freq() {
   for POLICY in /sys/devices/system/cpu/cpufreq/policy[0-9]*; do
     [ -e "${POLICY}/scaling_max_freq" ] || continue
-    [ -e "${POLICY}/cpuinfo_max_freq" ] || continue
+    [ -e "${POLICY}/scaling_available_frequencies" ] || continue
 
     local target=""
     case ${1} in
       highest)
-        target=$(cat "${POLICY}/cpuinfo_max_freq")
+        target=$( {
+          cat "${POLICY}/scaling_available_frequencies" 2>/dev/null
+          cat "${POLICY}/scaling_boost_frequencies" 2>/dev/null
+        } | tr ' ' '\n' | grep -E '^[0-9]+$' | sort -n | tail -n1)
         ;;
       *)
         target=${1}
@@ -93,15 +97,18 @@ set_cpu_max_freq() {
 set_cpu_min_freq() {
   for POLICY in /sys/devices/system/cpu/cpufreq/policy[0-9]*; do
     [ -e "${POLICY}/scaling_min_freq" ] || continue
-    [ -e "${POLICY}/cpuinfo_min_freq" ] || continue
+    [ -e "${POLICY}/scaling_available_frequencies" ] || continue
 
     local target=""
     case ${1} in
       highest)
-        target=$(cat "${POLICY}/cpuinfo_max_freq")
+        target=$( {
+          cat "${POLICY}/scaling_available_frequencies" 2>/dev/null
+          cat "${POLICY}/scaling_boost_frequencies" 2>/dev/null
+        } | tr ' ' '\n' | grep -E '^[0-9]+$' | sort -n | tail -n1)
         ;;
       lowest)
-        target=$(cat "${POLICY}/cpuinfo_min_freq")
+        target=$(tr ' ' '\n' <"${POLICY}/scaling_available_frequencies" | grep -E '^[0-9]+$' | sort -n | head -n1)
         ;;
       *)
         target=${1}
@@ -176,11 +183,13 @@ restore_cpu_boost() {
 }
 
 performance() {
-  # Force boost ON regardless of the user setting during gameplay so the
-  # 1512 MHz OPP is reachable. RK3326's BSP-marked turbo OPP is otherwise
-  # locked behind cpufreq/boost. ondemand/schedutil paths restore the
-  # user preference.
-  set_cpu_boost on
+  # Honor the user's "Enable CPU Overclock" switch from ES even during
+  # gameplay: gameplay shouldn't silently enable turbo behind the user's
+  # back. If they want the 1512 MHz OPP, they toggle it on; if they
+  # prefer to stay at 1416 MHz for thermals/battery, gameplay respects
+  # that. set_cpu_max_freq highest then pins to the highest currently
+  # available frequency (1512 if boost on, 1416 if off).
+  restore_cpu_boost
   set_cpu_gov performance
   set_cpu_max_freq highest
   set_dmc_gov performance
diff --git a/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_gc.sh b/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_gc.sh
index ac5ca417bb..3a44c6d0bd 100644
--- a/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_gc.sh
+++ b/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_gc.sh
@@ -426,4 +426,5 @@ fi
 # Run Dolphin emulator
   ${GPTOKEYB} ${DOLPHIN_CORE} xbox360 &
   ${EMUPERF} /usr/bin/${DOLPHIN_CORE} ${CMD} -e "${1}"
-  kill -9 "$(pidof gptokeyb)"
+  _gptokeyb_pid="$(pidof gptokeyb 2>/dev/null)"
+  [ -n "${_gptokeyb_pid}" ] && kill -9 ${_gptokeyb_pid}
diff --git a/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_wii.sh b/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_wii.sh
index f709d915c0..4be8e71b91 100755
--- a/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_wii.sh
+++ b/projects/ArchR/packages/emulators/standalone/dolphin-sa/scripts/start_dolphin_wii.sh
@@ -396,4 +396,5 @@ fi
 # Run Dolphin emulator
   ${GPTOKEYB} ${DOLPHIN_CORE} xbox360 &
   ${EMUPERF} /usr/bin/${DOLPHIN_CORE} ${CMD} -e "${1}"
-  kill -9 "$(pidof gptokeyb)"
+  _gptokeyb_pid="$(pidof gptokeyb 2>/dev/null)"
+  [ -n "${_gptokeyb_pid}" ] && kill -9 ${_gptokeyb_pid}
diff --git a/projects/ArchR/packages/emulators/standalone/drastic-sa/scripts/start_drastic.sh b/projects/ArchR/packages/emulators/standalone/drastic-sa/scripts/start_drastic.sh
index b687ffb8f2..6d2bd390f1 100644
--- a/projects/ArchR/packages/emulators/standalone/drastic-sa/scripts/start_drastic.sh
+++ b/projects/ArchR/packages/emulators/standalone/drastic-sa/scripts/start_drastic.sh
@@ -98,4 +98,5 @@ export LD_PRELOAD="/usr/lib/libdrastouch.so"
 export SDL_TOUCH_MOUSE_EVENTS="0"
 export DSHOOK_MIC_THRESH="${MICTHRESH}"
 ./drastic "$1"
-kill -9 $(pidof gptokeyb)
+_gptokeyb_pid="$(pidof gptokeyb 2>/dev/null)"
+[ -n "${_gptokeyb_pid}" ] && kill -9 ${_gptokeyb_pid}
diff --git a/projects/ArchR/packages/emulators/standalone/flycast-sa/scripts/start_flycast.sh b/projects/ArchR/packages/emulators/standalone/flycast-sa/scripts/start_flycast.sh
index a9f86c9aa2..2c26ece69c 100644
--- a/projects/ArchR/packages/emulators/standalone/flycast-sa/scripts/start_flycast.sh
+++ b/projects/ArchR/packages/emulators/standalone/flycast-sa/scripts/start_flycast.sh
@@ -146,8 +146,14 @@ fi
 # Panfrost optimization for Mali-G31
 export PAN_MESA_DEBUG=forcepack
 export MESA_NO_ERROR=1
+# Mesa GL command thread: Flycast already runs GL in a worker thread
+# (rend.ThreadedRendering=1) and is GPU-bound on RK3326. Marshalling GL
+# calls into Mesa's own thread amortises the per-call dispatch cost on a
+# single-issue Cortex-A35. See docs/mesa-glthread-matrix.md.
+export MESA_GLTHREAD=true
 
 #Run flycast emulator
 $GPTOKEYB "flycast" -c "${CONF_DIR}/flycast.gptk" &
 ${EMUPERF} /usr/bin/flycast "${1}"
-kill -9 "$(pidof gptokeyb)"
+_gptokeyb_pid="$(pidof gptokeyb 2>/dev/null)"
+[ -n "${_gptokeyb_pid}" ] && kill -9 ${_gptokeyb_pid}
diff --git a/projects/ArchR/packages/emulators/standalone/melonds-sa/scripts/start_melonds.sh b/projects/ArchR/packages/emulators/standalone/melonds-sa/scripts/start_melonds.sh
index 10b9298f53..1c95b1f248 100644
--- a/projects/ArchR/packages/emulators/standalone/melonds-sa/scripts/start_melonds.sh
+++ b/projects/ArchR/packages/emulators/standalone/melonds-sa/scripts/start_melonds.sh
@@ -205,4 +205,5 @@ fi
 
 $GPTOKEYB "melonDS" -c "${CONF_DIR}/melonDS.gptk" &
 ${EMUPERF} /usr/bin/melonDS -f "${ROM}"
-kill -9 "$(pidof gptokeyb)"
+_gptokeyb_pid="$(pidof gptokeyb 2>/dev/null)"
+[ -n "${_gptokeyb_pid}" ] && kill -9 ${_gptokeyb_pid}
diff --git a/projects/ArchR/packages/emulators/standalone/ppsspp-sa/scripts/start_ppsspp.sh b/projects/ArchR/packages/emulators/standalone/ppsspp-sa/scripts/start_ppsspp.sh
index 9b3118dbfd..9dc391b9e7 100755
--- a/projects/ArchR/packages/emulators/standalone/ppsspp-sa/scripts/start_ppsspp.sh
+++ b/projects/ArchR/packages/emulators/standalone/ppsspp-sa/scripts/start_ppsspp.sh
@@ -123,5 +123,11 @@ ARG=${1//[\\]/}
 # Panfrost optimization: forcepack reduces draw call overhead on Mali-G31
 export PAN_MESA_DEBUG=forcepack
 export MESA_NO_ERROR=1
+# Mesa GL command thread: PPSSPP issues many small GL calls per frame
+# (sprite/draw bursts) and the standalone build keeps render and CPU on
+# separate threads internally. Letting Mesa marshal GL on its own
+# thread cuts per-call dispatch latency on Cortex-A35. See
+# docs/mesa-glthread-matrix.md.
+export MESA_GLTHREAD=true
 
 ${EMUPERF} ppsspp --pause-menu-exit "${ARG}"
diff --git a/projects/ArchR/packages/graphics/gpudriver/sources/bin/gpudriver b/projects/ArchR/packages/graphics/gpudriver/sources/bin/gpudriver
index 375dab95b8..c9e0b46875 100755
--- a/projects/ArchR/packages/graphics/gpudriver/sources/bin/gpudriver
+++ b/projects/ArchR/packages/graphics/gpudriver/sources/bin/gpudriver
@@ -6,27 +6,12 @@
 
 GPU_DRIVER_SETTING_KEY="gpu.driver"
 
-# When the device ships no panfrost DTB overlay, panfrost selection is a
-# brick: even with the runtime fallback, sway/UI starts before the
-# fallback finishes, leaves a half-initialized GPU, and the next boot
-# ends in a black screen. Detect the empty placeholder substituted by
-# package.mk and refuse panfrost in that case.
-DTB_OVERLAY_LOAD_CMD="@DTB_OVERLAY_LOAD@"
-PANFROST_AVAILABLE=true
-[ -z "${DTB_OVERLAY_LOAD_CMD}" ] && PANFROST_AVAILABLE=false
-
 get_current_driver() {
   CONFDRIVER=$(get_setting ${GPU_DRIVER_SETTING_KEY})
   if [ -z ${CONFDRIVER} ]; then
     CONFDRIVER="libmali"              # DEFAULT
     set_setting ${GPU_DRIVER_SETTING_KEY} ${CONFDRIVER}
   fi
-  # Self-heal: stale "panfrost" setting on a device that can't run it.
-  if [ "${CONFDRIVER}" = "panfrost" ] && [ "${PANFROST_AVAILABLE}" = "false" ]; then
-    logger -t gpudriver "panfrost not supported on this device (no DTB overlay); reverting to libmali"
-    CONFDRIVER="libmali"
-    set_setting ${GPU_DRIVER_SETTING_KEY} ${CONFDRIVER}
-  fi
 }
 
 
@@ -89,11 +74,7 @@ load_driver() {
 
 case "$1" in
   "--options")
-    if [ "${PANFROST_AVAILABLE}" = "true" ]; then
-      echo "panfrost libmali"
-    else
-      echo "libmali"
-    fi
+    echo "panfrost libmali"
     ;;
   "--start")
     get_current_driver
@@ -117,10 +98,6 @@ case "$1" in
     @DTB_OVERLAY_UNLOAD@
     ;;
   "panfrost")
-    if [ "${PANFROST_AVAILABLE}" = "false" ]; then
-      echo "panfrost is not supported on this device" >&2
-      exit 2
-    fi
     set_setting ${GPU_DRIVER_SETTING_KEY} $1
     @DTB_OVERLAY_LOAD@
     ;;
diff --git a/projects/ArchR/packages/sysutils/autostart/sources/autostart b/projects/ArchR/packages/sysutils/autostart/sources/autostart
index 9c5045eb19..b25093b939 100755
--- a/projects/ArchR/packages/sysutils/autostart/sources/autostart
+++ b/projects/ArchR/packages/sysutils/autostart/sources/autostart
@@ -5,11 +5,38 @@
 . /etc/os-release
 
 BOOTLOG="/var/log/boot.log"
+# Persistent boot trace lives on /storage so that an interrupted boot
+# (hung autostart, kernel crash, manual reset) survives across the
+# reboot. The current script being run is written to .boot_last_step
+# *before* execution, then erased on success. If the file still exists
+# at the start of the next boot, the previous run hung on that script
+# — copy it to .boot_last_hang so it survives even after a clean boot
+# overwrites the marker.
+BOOT_LAST_STEP="/storage/.boot_last_step"
 
 log() {
   echo "$(date): Autostart: $*" >>${BOOTLOG} 2>&1
 }
 
+trace_step() {
+  # Only attempt to write once /storage is mountable; quirks that run
+  # before automount don't touch this and that's fine — the kernel
+  # boot stage prior to /storage is already covered by dmesg.
+  echo "$1" >${BOOT_LAST_STEP} 2>/dev/null
+}
+
+trace_done() {
+  rm -f ${BOOT_LAST_STEP} 2>/dev/null
+}
+
+# If the previous boot hung, preserve a copy before we overwrite the
+# marker on this boot. Read it back later with:
+#   cat /storage/.boot_last_hang
+if [ -f "${BOOT_LAST_STEP}" ]; then
+  cp -f "${BOOT_LAST_STEP}" /storage/.boot_last_hang 2>/dev/null
+  log "Previous boot hung at: $(cat ${BOOT_LAST_STEP} 2>/dev/null)"
+fi
+
 performance 2>/dev/null
 
 ### Fix permissions for non-root brightness and CPU boost control
@@ -31,6 +58,7 @@ then
   for script in "/usr/lib/autostart/quirks/platforms/${HW_DEVICE}"/*
   do
     log "Run ${script}"
+    trace_step "platform-quirk:${script}"
     if [ -f "${script}" ]
     then
       "${script}" 2>&1 >>${BOOTLOG}
@@ -46,6 +74,7 @@ then
   for script in "/usr/lib/autostart/quirks/devices/${QUIRK_DEVICE}"/*
   do
     log "Run ${script}"
+    trace_step "device-quirk:${script}"
     if [ -f "${script}" ]
     then
       "${script}" 2>&1 >>${BOOTLOG}
@@ -54,6 +83,7 @@ then
 fi
 
 ### Start the automount service
+trace_step "archr-automount"
 tocon "Starting storage services..."
 systemctl start archr-automount
 
@@ -61,6 +91,7 @@ systemctl start archr-automount
 for script in /usr/lib/autostart/common/*
 do
   log "Run ${script}"
+  trace_step "common:${script}"
   ${script} 2>&1 >>${BOOTLOG}
 done
 wait
@@ -71,6 +102,7 @@ then
   for script in "/usr/lib/autostart/${HW_DEVICE}"/*
   do
     log "Run ${script}"
+    trace_step "device:${script}"
     if [ -f "${script}" ]
     then
       "${script}" 2>&1 >>${BOOTLOG}
@@ -84,6 +116,7 @@ then
   for script in /storage/.config/autostart/*
   do
     log "Run ${script}"
+    trace_step "custom:${script}"
    ${script} 2>&1 >>${BOOTLOG}
   done
 fi
@@ -105,4 +138,8 @@ DEVICE_CPU_GOVERNOR=$(get_setting system.cpugovernor)
 ${DEVICE_CPU_GOVERNOR} 2>&1 >>${BOOTLOG}
 log "Autostart complete..."
 
+# Boot reached the end without hanging — clear the marker so the next
+# boot doesn't think the previous one failed.
+trace_done
+
 clear >/dev/console
diff --git a/projects/ArchR/packages/sysutils/autostart/system.d/archr-autostart.service b/projects/ArchR/packages/sysutils/autostart/system.d/archr-autostart.service
index 072b29c905..4738a53310 100644
--- a/projects/ArchR/packages/sysutils/autostart/system.d/archr-autostart.service
+++ b/projects/ArchR/packages/sysutils/autostart/system.d/archr-autostart.service
@@ -8,6 +8,11 @@ Type=oneshot
 Environment=HOME=/storage
 ExecStart=/bin/bash -a -c 'source /etc/profile && exec /usr/bin/autostart'
 RemainAfterExit=yes
+# Cap the autostart at 60s so that a misbehaving script (drm_tool with
+# DRM not ready, a daemon source hanging, etc.) doesn't lock the boot
+# indefinitely — emustation will still come up. Without this the
+# Before=emustation.service ordering blocks the UI forever.
+TimeoutStartSec=60s
 
 [Install]
 WantedBy=archr.target
diff --git a/projects/ArchR/packages/sysutils/system-utils/sources/scripts/turbomode b/projects/ArchR/packages/sysutils/system-utils/sources/scripts/turbomode
index 1b11f99ca8..c1eba87aa9 100644
--- a/projects/ArchR/packages/sysutils/system-utils/sources/scripts/turbomode
+++ b/projects/ArchR/packages/sysutils/system-utils/sources/scripts/turbomode
@@ -3,25 +3,35 @@
 # Copyright (C) 2024 JELOS (https://github.com/JustEnoughLinuxOS)
 # Copyright (C) 2026 ArchR (https://github.com/archr-linux/Arch-R)
 
-BOOST_FILE="/sys/devices/system/cpu/cpufreq/boost"
-POLICY="/sys/devices/system/cpu/cpufreq/policy0"
+# Toggle the kernel cpufreq turbo flag globally and pin scaling_max_freq
+# accordingly. Called by the ES "Enable CPU Overclock" switch and by the
+# 095-turbo-mode autostart at boot from the persisted setting.
+#
+# When OFF: cpufreq/boost=0 → turbo OPP (1512 MHz) is hidden from the
+#           freq table; governors top out at 1416 MHz.
+# When ON:  cpufreq/boost=1 → turbo OPP becomes available; max freq is
+#           pinned to it via scaling_available_frequencies (NOT
+#           cpuinfo_max_freq, which on ARM cpufreq-dt doesn't reflect
+#           boost OPPs even after boost=1).
+#
+# Governor is intentionally left alone — runemu.sh switches to
+# performance during gameplay, idle/menu stays on ondemand. Forcing
+# performance here would drain battery in the menu.
 
-# Get the actual max frequency supported by hardware
-MAX_FREQ=$(cat "${POLICY}/cpuinfo_max_freq" 2>/dev/null)
+. /etc/profile.d/001-functions
+. /etc/profile.d/099-freqfunctions
+
+BOOST_FILE="/sys/devices/system/cpu/cpufreq/boost"
 
 case $1 in
   disable)
-    echo "0" > "${BOOST_FILE}" 2>/dev/null
-    echo "performance" > "${POLICY}/scaling_governor" 2>/dev/null
+    [ -e "${BOOST_FILE}" ] && echo 0 > "${BOOST_FILE}" 2>/dev/null
+    set_cpu_max_freq highest
     set_setting enable.turbo-mode 0
-  ;;
+    ;;
   enable)
-    echo "1" > "${BOOST_FILE}" 2>/dev/null
-    # Set governor to performance and max freq to hardware maximum
-    echo "performance" > "${POLICY}/scaling_governor" 2>/dev/null
-    if [ -n "${MAX_FREQ}" ]; then
-      echo "${MAX_FREQ}" > "${POLICY}/scaling_max_freq" 2>/dev/null
-    fi
+    [ -e "${BOOST_FILE}" ] && echo 1 > "${BOOST_FILE}" 2>/dev/null
+    set_cpu_max_freq highest
     set_setting enable.turbo-mode 1
-  ;;
+    ;;
 esac
diff --git a/projects/ArchR/packages/sysutils/systemd/config/timesyncd.conf.d/99-handheld.conf b/projects/ArchR/packages/sysutils/systemd/config/timesyncd.conf.d/99-handheld.conf
new file mode 100644
index 0000000000..d35e7a6d1a
--- /dev/null
+++ b/projects/ArchR/packages/sysutils/systemd/config/timesyncd.conf.d/99-handheld.conf
@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: GPL-2.0
+# ArchR drop-in for handheld idle profile.
+#
+# systemd-timesyncd defaults (PollMin 32 s, PollMax 2048 s ≈ 34 min)
+# assume a server that needs frequent corrections. On a handheld that
+# spends most of its time in the menu or off, polling every 34 min is
+# pointless wakeup traffic and keeps the daemon's working set warm in
+# RAM. Stretching PollIntervalMaxSec to 86400 s (24 h) lets timesyncd
+# fall idle for the rest of the day after the first successful sync,
+# which the kernel happily pages out — cutting the daemon's residual
+# RSS to near zero.
+#
+# PollIntervalMinSec stays at 64 s so that bad samples (just after
+# resume / new connection) re-poll quickly until trust is established.
+
+[Time]
+PollIntervalMinSec=64
+PollIntervalMaxSec=86400
+ConnectionRetrySec=30
diff --git a/projects/ArchR/packages/sysutils/systemd/config/udev.rules.d/91-uinput.rules b/projects/ArchR/packages/sysutils/systemd/config/udev.rules.d/91-uinput.rules
new file mode 100644
index 0000000000..a5b0f60238
--- /dev/null
+++ b/projects/ArchR/packages/sysutils/systemd/config/udev.rules.d/91-uinput.rules
@@ -0,0 +1,6 @@
+# PortMaster ports use gptokeyb to remap controller input to keyboard
+# events; gptokeyb writes to /dev/uinput. Some ports also chmod 0666
+# the device themselves, but they only succeed when invoked as root.
+# ArchR runs ports under the "archr" user, so set the permission
+# globally at hotplug time.
+KERNEL=="uinput", MODE="0666", OPTIONS+="static_node=uinput"
diff --git a/projects/ArchR/packages/virtual/emulators/package.mk b/projects/ArchR/packages/virtual/emulators/package.mk
index ff35a3c98b..c5e5a83151 100644
--- a/projects/ArchR/packages/virtual/emulators/package.mk
+++ b/projects/ArchR/packages/virtual/emulators/package.mk
@@ -454,13 +454,24 @@ makeinstall_target() {
   add_es_system gbh
 
   ### Nintendo GameBoy Advance
-  add_emu_core gba retroarch mgba true
+  # Cortex-A35 (RK3326) doesn't sustain mGBA full-speed; use the
+  # lighter-weight gpsp as the performance default and keep mGBA as an
+  # accuracy alternative.
+  case ${DEVICE} in
+    RK3326)
+      add_emu_core gba retroarch gpsp true
+      add_emu_core gba retroarch mgba false
+      ;;
+    *)
+      add_emu_core gba retroarch mgba true
+      ;;
+  esac
   add_emu_core gba retroarch vbam false
   add_emu_core gba retroarch vba_next false
   add_emu_core gba retroarch beetle_gba false
   add_emu_core gba retroarch skyemu false
   case ${DEVICE} in
-    H700|RK3326|RK3566|S922X)
+    H700|RK3566|S922X)
       add_emu_core gba retroarch gpsp false
       ;;
     RK3399|RK3588|SM8250|SM8550|SDM845)
@@ -570,22 +581,28 @@ makeinstall_target() {
   add_es_system gbch
 
   ### Nintendo GameCube
+  # GameCube on Cortex-A35 (RK3326) does not run at usable framerates;
+  # do not list the system in ES so the user isn't lured into trying.
   case ${DEVICE} in
     RK3399|SDM845|SM8250|SM8550|SM8650)
       add_emu_core gamecube dolphin dolphin-qt-gc true
       add_emu_core gamecube dolphin dolphin-sa-gc false
       add_emu_core gamecube retroarch dolphin false
       install_script "Start Dolphin.sh"
+      add_es_system gamecube
       ;;
     RK3566|RK3588|S922X)
       add_emu_core gamecube dolphin dolphin-sa-gc true
       add_emu_core gamecube retroarch dolphin false
+      add_es_system gamecube
+      ;;
+    H700|RK3326)
       ;;
     *)
       add_emu_core gamecube retroarch dolphin true
+      add_es_system gamecube
       ;;
   esac
-  add_es_system gamecube
 
   ### Nintendo Triforce
   case ${DEVICE} in
@@ -597,6 +614,7 @@ makeinstall_target() {
   esac
 
   ### Nintendo Wii/ware
+  # Same rationale as GameCube: not viable on Cortex-A35.
   case ${DEVICE} in
     RK3399|SDM845|SM8250|SM8550|SM8650)
       add_emu_core wii dolphin dolphin-qt-wii true
@@ -605,20 +623,26 @@ makeinstall_target() {
       add_emu_core wiiware dolphin dolphin-sa-wii false
       add_emu_core wii retroarch dolphin false
       add_emu_core wiiware retroarch dolphin false
+      add_es_system wii
+      add_es_system wiiware
       ;;
     RK3566|RK3588|S922X)
       add_emu_core wii dolphin dolphin-sa-wii true
       add_emu_core wiiware dolphin dolphin-sa-wii true
       add_emu_core wii retroarch dolphin false
       add_emu_core wiiware retroarch dolphin false
+      add_es_system wii
+      add_es_system wiiware
+      ;;
+    H700|RK3326)
       ;;
     *)
       add_emu_core wii retroarch dolphin true
       add_emu_core wiiware retroarch dolphin true
+      add_es_system wii
+      add_es_system wiiware
       ;;
   esac
-  add_es_system wii
-  add_es_system wiiware
 
   ### Nintendo Wii U
   case ${DEVICE} in
diff --git a/scripts/mkimage b/scripts/mkimage
index d9d0907c10..67658c9231 100755
--- a/scripts/mkimage
+++ b/scripts/mkimage
@@ -182,7 +182,13 @@ alias mmd='mmd -i "${IMG_TMP}/part1.fat"'
 echo "image: creating filesystem on part1..."
 
 if [ "${BOOTLOADER}" = "syslinux" -o "${BOOTLOADER}" = "bcm2835-bootloader" -o "${BOOTLOADER}" = "u-boot" -o "${BOOTLOADER}" = "arm-efi" -o "${BOOTLOADER}" = "qcom-abl" ]; then
-  mkfs.vfat -F 32 -S 512 -s 32 -i "${UUID_SYSTEM//-/}" -n "${DISTRO_BOOTLABEL}" "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
+  # 4 KB clusters (-s 8 with 512 B sectors) so the partition has at least
+  # 65525 clusters and qualifies as real FAT32 per spec. With -s 32 the
+  # 256 MB BOOT_SIZE fell to ~16k clusters, which dosfsck flagged on
+  # every boot ("less than the required minimum of 65525") and which a
+  # few RK3326 U-Boot builds occasionally read as FAT16 — observed as
+  # intermittent post-resize boot hangs.
+  mkfs.vfat -F 32 -S 512 -s 8 -i "${UUID_SYSTEM//-/}" -n "${DISTRO_BOOTLABEL}" "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
 fi
 sync
 
@@ -355,10 +361,13 @@ elif [ "${BOOTLOADER}" = "qcom-abl" ]; then
 
 fi # bootloader
 
-# run fsck
+# run fsck (auto-repair so any boot-sector / backup divergence in the
+# freshly-formatted FAT is fixed before we cement the image — otherwise
+# dosfsck on the device flags it on every boot and U-Boot has been seen
+# reading the inconsistent backup intermittently)
 echo "image: checking filesystem on part1..."
 sync
-fsck.fat -n "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
+fsck.fat -a -w "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
 
 # merge part1 into disk image
 echo "image: merging part1 into disk image..."
@@ -373,9 +382,9 @@ if [ "${PROJECT}" = "Generic" ]; then
   mcopy "${IMG_TMP}/syslinux.cfg" :: >"${SAVE_ERROR}" 2>&1 || show_error
   mcopy "${IMG_TMP}/grub.cfg" ::/EFI/BOOT >"${SAVE_ERROR}" 2>&1 || show_error
   sync
-  # run fsck
+  # run fsck (auto-repair, see comment above)
   echo "image: checking filesystem on part1..."
-  fsck.fat -n "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
+  fsck.fat -a -w "${IMG_TMP}/part1.fat" >"${SAVE_ERROR}" 2>&1 || show_error
   # merge modified part1 into tmp disk image
   echo "image: merging part1 into open virtual appliance..."
   dd if="${IMG_TMP}/part1.fat" of="${DISK_BASENAME}.tmp" bs=512 seek="${SYSTEM_PART_START}" conv=fsync,notrunc >"${SAVE_ERROR}" 2>&1 || show_error
diff --git a/tools/bench.sh b/tools/bench.sh
index d49cd9d62b..5705b053ad 100755
--- a/tools/bench.sh
+++ b/tools/bench.sh
@@ -7,10 +7,33 @@ DURATION=${2:-60}
 RESULTS_DIR="/storage/bench-results/$(date +%Y%m%d-%H%M%S)-${1:-general}"
 mkdir -p "$RESULTS_DIR"
 
+# Auto-detect device-specific paths once. RK3326 mainline shows up as
+# mmcblk1 (microSD), some boards use mmcblk0; the DMC devfreq node has
+# a model-prefixed name (e.g. "dmc" alone doesn't exist, real path is
+# `/sys/class/devfreq/dmc-rk3326` or similar). Pick whatever the kernel
+# actually exposed, fall back to empty string for the JSON if absent.
+SD_BLOCK=""
+for blk in /sys/block/mmcblk0 /sys/block/mmcblk1; do
+  if [ -e "$blk/device/name" ] && [ -n "$(cat "$blk/device/name" 2>/dev/null)" ]; then
+    SD_BLOCK="$blk"
+    break
+  fi
+done
+SD_MODEL=$([ -n "$SD_BLOCK" ] && cat "$SD_BLOCK/device/name" 2>/dev/null)
+SD_NAME=${SD_BLOCK##*/}
+
+GPU_DEVFREQ=$(ls -d /sys/class/devfreq/*gpu* 2>/dev/null | head -1)
+DMC_DEVFREQ=$(ls -d /sys/class/devfreq/*dmc* 2>/dev/null | head -1)
+
+DISTRO_NAME=$(awk -F= '/^NAME=/ {gsub(/"/,"",$2); print $2; exit}' /etc/os-release)
+
 echo "========================================"
 echo "  ArchR Benchmark: ${1:-general}"
 echo "  Duration: ${DURATION}s"
 echo "  Output: $RESULTS_DIR"
+echo "  SD: ${SD_NAME:-?} (${SD_MODEL:-unknown})"
+echo "  GPU devfreq: ${GPU_DEVFREQ:-not found}"
+echo "  DMC devfreq: ${DMC_DEVFREQ:-not found}"
 echo "========================================"
 
 # Metadata
@@ -18,19 +41,22 @@ cat > "$RESULTS_DIR/meta.json" << METAEOF
 {
   "timestamp": "$(date -Iseconds)",
   "kernel": "$(uname -r)",
-  "distro": "$(grep ^NAME /etc/os-release | cut -d= -f2)",
+  "distro": "${DISTRO_NAME}",
   "cpu_governor": "$(cat /sys/devices/system/cpu/cpufreq/policy0/scaling_governor 2>/dev/null)",
-  "gpu_governor": "$(cat /sys/class/devfreq/ff400000.gpu/governor 2>/dev/null)",
-  "dmc_governor": "$(cat /sys/class/devfreq/dmc/governor 2>/dev/null)",
+  "cpu_boost": $(cat /sys/devices/system/cpu/cpufreq/boost 2>/dev/null || echo 0),
+  "cpu_max_freq_khz": $(cat /sys/devices/system/cpu/cpufreq/policy0/scaling_max_freq 2>/dev/null || echo 0),
+  "gpu_governor": "$(cat ${GPU_DEVFREQ:-/dev/null}/governor 2>/dev/null)",
+  "dmc_governor": "$(cat ${DMC_DEVFREQ:-/dev/null}/governor 2>/dev/null)",
   "mem_available_kb": $(grep MemAvailable /proc/meminfo | awk '{print $2}'),
   "cma_total_kb": $(grep CmaTotal /proc/meminfo | awk '{print $2}'),
   "cma_free_kb": $(grep CmaFree /proc/meminfo | awk '{print $2}'),
-  "zram_algo": "$(cat /sys/block/zram0/comp_algorithm 2>/dev/null | tr -d '[]')",
-  "swap": "$(swapon --show --noheadings 2>/dev/null | head -1)",
+  "zram_algo": "$(cat /sys/block/zram0/comp_algorithm 2>/dev/null | grep -oP '\[\K[^\]]+')",
+  "swap": "$(swapon --show --noheadings 2>/dev/null | head -1 | tr -s ' ')",
   "swappiness": $(sysctl -n vm.swappiness 2>/dev/null),
   "vfs_cache_pressure": $(sysctl -n vm.vfs_cache_pressure 2>/dev/null),
   "temperature_start": $(cat /sys/class/thermal/thermal_zone0/temp 2>/dev/null),
-  "sd_model": "$(cat /sys/block/mmcblk0/device/name 2>/dev/null)",
+  "sd_model": "${SD_MODEL}",
+  "sd_block": "${SD_NAME}",
   "scenario": "${1:-general}"
 }
 METAEOF
@@ -50,8 +76,8 @@ echo "Starting freq/temp monitoring..."
 (
   while true; do
     CPU_FREQ=$(cat /sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq 2>/dev/null)
-    GPU_FREQ=$(cat /sys/class/devfreq/ff400000.gpu/cur_freq 2>/dev/null)
-    DMC_FREQ=$(cat /sys/class/devfreq/dmc/cur_freq 2>/dev/null)
+    GPU_FREQ=$([ -n "$GPU_DEVFREQ" ] && cat "$GPU_DEVFREQ/cur_freq" 2>/dev/null)
+    DMC_FREQ=$([ -n "$DMC_DEVFREQ" ] && cat "$DMC_DEVFREQ/cur_freq" 2>/dev/null)
     TEMP=$(cat /sys/class/thermal/thermal_zone0/temp 2>/dev/null)
     MEM=$(grep MemAvailable /proc/meminfo | awk '{print $2}')
     echo "$(date +%s.%N) cpu=$CPU_FREQ gpu=$GPU_FREQ dmc=$DMC_FREQ temp=$TEMP mem=$MEM"
@@ -60,12 +86,17 @@ echo "Starting freq/temp monitoring..."
 ) > "$RESULTS_DIR/freq-therm.log" &
 FREQ_PID=$!
 
-# vmstat monitoring
-vmstat 1 $DURATION > "$RESULTS_DIR/vmstat.log" &
+# vmstat monitoring. vmstat exits on its own after $DURATION samples;
+# we spawn it in a subshell that pipes to the file directly so it
+# survives the parent's kill at the end without a zombie warning.
+( vmstat 1 $DURATION > "$RESULTS_DIR/vmstat.log" 2>&1 ) &
 VMSTAT_PID=$!
 
-# dmesg monitoring
-dmesg -w > "$RESULTS_DIR/dmesg.log" 2>/dev/null &
+# dmesg ring buffer snapshot (start) and tail (during run). dmesg -w
+# in the background was eating its own output before; now we capture
+# a baseline plus stream new messages with explicit redirect.
+dmesg --since "now -1minutes" 2>/dev/null > "$RESULTS_DIR/dmesg-start.log"
+( dmesg -W 2>/dev/null > "$RESULTS_DIR/dmesg.log" ) &
 DMESG_PID=$!
 
 echo ""
@@ -76,8 +107,11 @@ echo ""
 # Wait
 sleep $DURATION 2>/dev/null || true
 
-# Cleanup
-kill $PSI_PID $FREQ_PID $VMSTAT_PID $DMESG_PID 2>/dev/null
+# Cleanup — be loud about any monitor that didn't survive so the
+# follow-up summary doesn't lie about coverage.
+for pid in $PSI_PID $FREQ_PID $VMSTAT_PID $DMESG_PID; do
+  kill $pid 2>/dev/null
+done
 wait 2>/dev/null
 
 # Generate summary
@@ -88,9 +122,14 @@ cat > "$RESULTS_DIR/summary.json" << SUMEOF
 {
   "duration_s": $DURATION,
   "temperature_end": $TEMP_END,
+  "temperature_delta_c": $(awk "BEGIN { printf \"%.1f\", ($TEMP_END - $(cat /sys/class/thermal/thermal_zone0/temp 2>/dev/null || echo 0)) / 1000 }"),
   "mem_available_end_kb": $MEM_END,
+  "cpu_freq_max_observed_khz": $(awk -F'cpu=' '{split($2,a," "); print a[1]}' "$RESULTS_DIR/freq-therm.log" 2>/dev/null | sort -n | tail -1),
+  "gpu_freq_max_observed_hz":  $(awk -F'gpu=' '{split($2,a," "); print a[1]}' "$RESULTS_DIR/freq-therm.log" 2>/dev/null | sort -n | tail -1),
   "psi_samples": $(wc -l < "$RESULTS_DIR/psi.log"),
-  "freq_samples": $(wc -l < "$RESULTS_DIR/freq-therm.log")
+  "freq_samples": $(wc -l < "$RESULTS_DIR/freq-therm.log"),
+  "vmstat_samples": $(wc -l < "$RESULTS_DIR/vmstat.log"),
+  "dmesg_lines": $(wc -l < "$RESULTS_DIR/dmesg.log")
 }
 SUMEOF