diff --git a/.gitignore b/.gitignore
index 46b1d3a..8547ef5 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,5 +1,6 @@
 # binarios y estado local, no van al repo
 deathwatch.exe
+deathwatch-linux
 *.exe
 totals.json
 client-id.txt
diff --git a/CLAUDE.md b/CLAUDE.md
index da034f8..5c23955 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -2,19 +2,26 @@
 
 Lee en **solo lectura** la memoria del proceso de Elden Ring y expone el
 contador de muertes como overlay web para OBS. Soporta co-op: cada jugador
-corre el programa en su PC y uno hace de hub.
+corre el programa en su PC y uno hace de hub. Corre nativo en Windows y
+en Linux contra el juego bajo Proton (mismos offsets y firmas: es el
+mismo binario de Windows).
 
 ## Comandos
 
 ```bash
-GOOS=windows GOARCH=amd64 go build -o deathwatch.exe .   # el binario (hoy, unico soportado)
-go test ./...                                            # tests (corren en cualquier SO)
-gofmt -l *.go                                            # formato
-GOOS=windows GOARCH=amd64 go vet .                       # vet: SIEMPRE con GOOS=windows
+GOOS=windows GOARCH=amd64 go build -o deathwatch.exe .        # binario Windows
+GOOS=linux   GOARCH=amd64 go build -o deathwatch-linux .      # binario Linux (Proton)
+go test ./...                                                  # tests (corren en cualquier SO)
+gofmt -l *.go                                                  # formato
+GOOS=windows GOARCH=amd64 go vet .                             # vet en los dos GOOS
+GOOS=linux   GOARCH=amd64 go vet .
 ```
 
-`go vet` sin `GOOS=windows` falla con `syscall.Handle undefined`. No es un
-error del código: es que `main.go` e `i18n.go` solo compilan para Windows.
+Dos plataformas reales ahora: correr `go vet` (y `go build`) con los dos
+`GOOS` es la única forma de agarrar una rotura especifica de una
+plataforma antes de que la vea alguien que corre la otra. `process.go`
+es portable; `process_windows.go` y `process_linux.go` son cada uno
+solo para su SO (ver "Multiplataforma" más abajo).
 
 ## Idioma del código
 
@@ -35,8 +42,11 @@ idiomas del overlay, ni al revés.
 
 | Archivo | Qué hace | Plataforma |
 |---|---|---|
-| `main.go` | Lectura de memoria, escaneo de firmas, loop de polling, HTTP | solo Windows |
-| `i18n.go` | Carga de `locales/*.json`, idioma del sistema | solo Windows |
+| `main.go` | HTTP, startup, arma todo | portable |
+| `i18n.go` | Carga de `locales/*.json`, elige idioma | portable |
+| `process.go` | Escaneo de firmas, resolución de punteros, loop de polling | portable |
+| `process_windows.go` | Primitivas de SO: abrir proceso, leer memoria, version.dll, idioma | solo Windows |
+| `process_linux.go` | Lo mismo que `process_windows.go`, vía `/proc/<pid>/{maps,mem}` | solo Linux |
 | `counter.go` | **Contabilidad**: a qué personaje va cada muerte | portable |
 | `names.go` | `looksLikeName`: filtra basura binaria leída como nombre | portable |
 | `totals.go` | Persistencia por personaje (`totals.json`) | portable |
@@ -53,6 +63,71 @@ los tres bugs de conteo, y separarlo es lo que permite testearlo sin una PC
 con el juego abierto. Si agregás reglas de conteo, van ahí, con test.
 `names.go`, `tlscert.go`, `pin.go` e `invite.go` siguen el mismo principio.
 
+## Multiplataforma
+
+`process.go` no sabe nada de Windows ni de Linux: escanea firmas, resuelve
+punteros y corre el loop de polling contra siete funciones que cruzan la
+frontera con el SO, cada una implementada una vez por plataforma
+(`process_windows.go` / `process_linux.go`):
+
+```go
+type procHandle uintptr // opaco: en Windows es el HANDLE real; en Linux, el pid
+
+func findProcessID(name string) (uint32, error)
+func openProcess(pid uint32) (procHandle, error)
+func closeProcessHandle(h procHandle)
+func readMemory(h procHandle, addr uintptr, size int) ([]byte, bool)
+func findModuleBase(pid uint32, name string) (uintptr, uint32, string, error)
+func productVersion(path string) (major, minor uint16, label string, ok bool)
+func systemLang() string
+```
+
+En Linux, contra el juego corriendo bajo Proton (mismo binario de
+Windows, mismas firmas y offsets):
+
+- **Encontrar el proceso por sus mapeos, no por el nombre**
+  (`findProcessID`/`scanMaps`): Proton levanta varios procesos: se
+  recorre `/proc/*/maps` y se toma el pid que tenga mapeado un archivo
+  terminado en `eldenring.exe`. Los mismos mapeos dan la base y el
+  tamaño del módulo para `findModuleBase`: puede venir partido en varios
+  tramos (`.text`/`.rdata`/`.data`), así que se toma el span completo
+  (mínimo inicio, máximo final) — alcanza, porque el escáner ya lee por
+  chunks y saltea los que no puede leer.
+- **Leer memoria vía `/proc/<pid>/mem`** (`ReadAt`, sin dependencias
+  externas), no `process_vm_readv(2)` crudo: mismo resultado, sin tener
+  que hacer un syscall a mano con structs `iovec` sin `golang.org/x/sys`.
+- **El obstáculo real es `ptrace_scope`.** Sin la capacidad, abrir
+  `/proc/<pid>/mem` da `EPERM`. `openProcess` lo prueba una vez al
+  arrancar y, si falla, el error (que sale por el mismo camino que ya
+  existía: `st.setDisconnected(err.Error())` en el loop de polling) trae
+  el comando exacto con la ruta real del binario:
+  `sudo setcap cap_sys_ptrace+ep <ruta>`. **Nunca** sugiere
+  `sysctl kernel.yama.ptrace_scope=0` ni correr como root — eso baja la
+  defensa de todo el sistema, no solo la de este programa.
+- **`productVersion` devuelve `ok=false` siempre.** No hay equivalente a
+  `version.dll` en Linux, pero nunca hizo falta: la versión era solo una
+  corazonada para elegir qué offset de `PlayerIns` probar primero
+  (`playerInsCandidates`); el que vale se confirma leyendo memoria en
+  `isPlayerLoaded` igual, con o sin la corazonada.
+- **`systemLang`** sale de `$LC_ALL` / `$LC_MESSAGES` / `$LANG` en vez de
+  `GetUserDefaultLocaleName`, devolviendo el mismo formato que ya
+  devuelve la versión de Windows (el código corto: `"es"`, no
+  `"es-AR"` ni `"es_AR.UTF-8"`), para que `resolveLang` (`i18n.go`) no
+  tenga que distinguir de dónde vino.
+
+**Modo sólo-hub en Linux sale gratis, sin código extra.** `main()` llama
+`go pollLoop()` sin importar el modo. Si no hay ningún `eldenring.exe`
+local (el caso de una PC con el OBS en Linux mientras se juega en otra),
+`findProcessID` simplemente no encuentra nada y `pollLoop` reintenta cada
+3s sin nunca llegar a `openProcess` — `setcap`/`ptrace_scope` no entran
+en juego para nada en ese caso.
+
+**Pendiente, no parte de este cambio:** testear el escáner/poller con un
+lector de memoria falso. El split ya lo habilita, pero escribir esos
+tests (siguiendo el patrón de variable de paquete intercambiable que ya
+usa `testExeDir` en `totals.go`, no una interfaz nueva) queda para
+después — ver Pendientes.
+
 ## Offsets de memoria
 
 Todo se resuelve escaneando firmas AOB en el módulo del juego. Los tres
@@ -191,73 +266,16 @@ No son preferencias de estilo. Cada una costó un bug en producción.
 
 ## Pendientes
 
-- No hay `README.md` todavía. Escribir uno en inglés (instalación, modo
-  hub/peer, capturas) es lo único que falta del pendiente de idioma — el
-  código ya está en inglés de punta a punta, ver "Idioma del código"
-  arriba.
+- No hay `README.md` todavía. Escribir uno en inglés (instalación en
+  Windows y Linux/Proton — incluyendo el paso de `setcap`, modo hub/peer,
+  capturas) es lo único que falta del pendiente de idioma — el código ya
+  está en inglés de punta a punta, ver "Idioma del código" arriba.
 - La identificación por nombre (respaldo cuando no se lee el slot) mezcla
   personajes homónimos. Documentado, no resuelto.
-- **Compilar y correr en Linux (Proton).** Mucha gente juega Elden Ring
-  con Proton, y hoy el programa sólo existe para Windows. El juego sigue
-  siendo el mismo binario de Windows corriendo bajo Wine, así que **las
-  firmas AOB y todos los offsets valen igual**: lo único que cambia es
-  cómo se encuentra el proceso y cómo se lee su memoria.
-
-  **Refactor primero.** Hoy `main.go` mezcla lo específico de Windows con
-  lo que no lo es. Separar en `process_windows.go` y `process_linux.go`
-  detrás de unas pocas funciones — `findProcessID`, `openProcess`,
-  `readMemory`, `findModuleBase`, `productVersion`, `systemLang` — y dejar
-  el resto (escaneo de firmas, resolución de punteros, loop de polling,
-  lectura del nombre) en un archivo portable. Beneficio extra que vale por
-  sí solo: con eso el escaneo y el polling **se pueden testear con un
-  lector de memoria falso**, que es justo la parte que hoy no tiene tests.
-
-  Lo específico de Linux:
-
-  1. **Encontrar el proceso por sus mapeos, no por el nombre.** Proton
-     levanta varios procesos. Lo robusto es recorrer `/proc/*/maps` y
-     quedarse con el pid que tenga mapeado un archivo terminado en
-     `eldenring.exe`. De paso, esos mismos mapeos dan la base y el tamaño
-     del módulo, que es lo que `findModuleBase` necesita. Puede venir
-     partido en varios tramos (`.text`, `.rdata`, `.data`) con permisos
-     distintos: tomar el span completo alcanza, porque el escáner ya lee
-     por chunks y saltea los que no puede leer.
-  2. **Leer con `process_vm_readv(2)`**, que no necesita adjuntarse al
-     proceso. `/proc/<pid>/mem` sirve de alternativa.
-  3. **El obstáculo real es `ptrace_scope`.** En casi todas las distros
-     vale `1`, y con eso `process_vm_readv` sobre un proceso ajeno falla
-     con `EPERM`. La salida recomendada es darle la capacidad al binario:
-
-     ```bash
-     sudo setcap cap_sys_ptrace+ep ./deathwatch
-     ```
-
-     **No** recomendar `sysctl kernel.yama.ptrace_scope=0`, que baja la
-     defensa de todo el sistema, ni correrlo como root. Y que el mensaje
-     de error diga exactamente esto cuando falle: sin eso el programa
-     parece simplemente roto, y es el primer problema que va a tener
-     cualquiera que lo pruebe.
-  4. **`productVersion` no tiene equivalente** (usa `version.dll` sobre el
-     exe). En Linux devolver `ok=false` y listo: la versión es sólo una
-     corazonada para decidir qué offset de `PlayerIns` probar primero, y
-     el valor bueno se confirma leyendo memoria igual. Una decisión vieja
-     que acá se paga sola.
-  5. **`systemLang`** sale de `$LC_ALL` / `$LANG` en vez de
-     `GetUserDefaultLocaleName`.
-
-  No hay que tocar: los offsets, las firmas, `counter.go`, `names.go`,
-  `totals.go`, `duo.go`, `ws.go`, `tlscert.go`, `pin.go`, `invite.go`,
-  `config.go` ni el overlay. Ya son portables.
-
-  **Modo sólo-hub, de regalo.** Un hub que no lee memoria —que sólo junta
-  lo de los peers y sirve el overlay— no necesita `setcap` ni permiso
-  alguno. Es exactamente el caso de quien tiene el OBS en una PC con Linux
-  y juega en otra, y sale casi gratis una vez separado lo de arriba.
-
-  Referencias de gente que ya leyó memoria de juegos bajo Proton:
-  [pika](https://github.com/delfianto/pika),
-  [cheat-engine-linux](https://github.com/wleeaf/cheat-engine-linux),
-  [un trainer para D2R en Linux paso a paso](https://axiom0x0.sh/posts/d2r-memory-trainer-part2/).
-
+- Testear el escáner/poller (`process.go`) con un lector de memoria
+  falso. El split multiplataforma (ver "Multiplataforma" arriba) ya lo
+  habilita — falta escribir los tests en sí, con una variable de paquete
+  intercambiable por función de frontera (mismo patrón que `testExeDir`
+  en `totals.go`), no una interfaz nueva.
 - El README debería mencionar que meter todo adentro de una VPN sigue
   siendo una opción perfectamente válida, TLS+pinning aparte.
diff --git a/i18n.go b/i18n.go
index f47a982..c552ad8 100644
--- a/i18n.go
+++ b/i18n.go
@@ -1,5 +1,3 @@
-//go:build windows
-
 // i18n.go: interface languages.
 //
 // Translations live in locales/*.json and get embedded into the binary.
@@ -10,6 +8,11 @@
 // Console messages deliberately do NOT go through here: they're
 // diagnostics, and it helps if they're always in the same language so a
 // log pasted into an issue reads the same no matter where it came from.
+//
+// systemLang(), which this file calls to pick the default language, is
+// the one OS-specific piece — implemented in process_windows.go/
+// process_linux.go, not here, since this file's own job (embedding and
+// picking a dictionary) doesn't depend on the OS at all.
 package main
 
 import (
@@ -19,8 +22,6 @@ import (
 	"path"
 	"sort"
 	"strings"
-	"syscall"
-	"unsafe"
 )
 
 //go:embed locales/*.json
@@ -64,24 +65,8 @@ func availableLangs() []string {
 	return out
 }
 
-var procGetUserDefaultLocaleName = kernel32.NewProc("GetUserDefaultLocaleName")
-
-// systemLang returns Windows's language ("es-AR" -> "es").
-func systemLang() string {
-	buf := make([]uint16, 85) // LOCALE_NAME_MAX_LENGTH
-	r, _, _ := procGetUserDefaultLocaleName.Call(uintptr(unsafe.Pointer(&buf[0])), uintptr(len(buf)))
-	if r == 0 {
-		return ""
-	}
-	name := syscall.UTF16ToString(buf[:r])
-	if base, _, ok := strings.Cut(name, "-"); ok {
-		return strings.ToLower(base)
-	}
-	return strings.ToLower(name)
-}
-
-// resolveLang decides the final language. "auto" (or empty) uses
-// Windows's; if that language isn't translated, it falls back to English.
+// resolveLang decides the final language. "auto" (or empty) uses the
+// system's; if that language isn't translated, it falls back to English.
 func resolveLang(want string) string {
 	want = strings.ToLower(strings.TrimSpace(want))
 	if want == "" || want == "auto" {
diff --git a/main.go b/main.go
index beffff4..33df5d5 100644
--- a/main.go
+++ b/main.go
@@ -1,9 +1,10 @@
-//go:build windows
-
 // deathwatch: reads Elden Ring's death counter read-only, straight from
 // the process (the same byte pattern / offset used by LiveSplit's
-// "eldenring_boss_timer.asl" ASL script, verified by hand on this PC).
-// Writes nothing to the game's memory.
+// "eldenring_boss_timer.asl" ASL script, verified by hand). Writes
+// nothing to the game's memory. Runs on Windows natively and on Linux
+// against the game running under Proton — same signatures and offsets
+// either way, since it's the same Windows binary in memory; see
+// process.go/process_windows.go/process_linux.go for the split.
 //
 // Exposes:
 //
@@ -15,798 +16,16 @@ package main
 import (
 	"crypto/tls"
 	_ "embed"
-	"encoding/binary"
 	"encoding/json"
-	"fmt"
 	"log"
 	"net/http"
 	"os"
-	"strconv"
 	"strings"
-	"syscall"
-	"time"
-	"unicode/utf16"
-	"unsafe"
 )
 
 //go:embed overlay.html
 var overlayHTML []byte
 
-const (
-	processName = "eldenring.exe"
-
-	processQueryInformation = 0x0400
-	processVMRead           = 0x0010
-	th32csSnapProcess       = 0x00000002
-	th32csSnapModule        = 0x00000008
-	th32csSnapModule32      = 0x00000010
-	maxPath                 = 260
-)
-
-var (
-	kernel32                     = syscall.NewLazyDLL("kernel32.dll")
-	procOpenProcess              = kernel32.NewProc("OpenProcess")
-	procCloseHandle              = kernel32.NewProc("CloseHandle")
-	procReadProcessMemory        = kernel32.NewProc("ReadProcessMemory")
-	procCreateToolhelp32Snapshot = kernel32.NewProc("CreateToolhelp32Snapshot")
-	procModule32FirstW           = kernel32.NewProc("Module32FirstW")
-	procModule32NextW            = kernel32.NewProc("Module32NextW")
-	procProcess32FirstW          = kernel32.NewProc("Process32FirstW")
-	procProcess32NextW           = kernel32.NewProc("Process32NextW")
-
-	versionDLL                  = syscall.NewLazyDLL("version.dll")
-	procGetFileVersionInfoSizeW = versionDLL.NewProc("GetFileVersionInfoSizeW")
-	procGetFileVersionInfoW     = versionDLL.NewProc("GetFileVersionInfoW")
-	procVerQueryValueW          = versionDLL.NewProc("VerQueryValueW")
-)
-
-// vsFixedFileInfo is Windows's VS_FIXEDFILEINFO struct: used to pull the
-// game's version straight from eldenring.exe, same as SoulMemory (which
-// reads MainModule.FileVersionInfo.ProductVersion).
-type vsFixedFileInfo struct {
-	Signature        uint32
-	StrucVersion     uint32
-	FileVersionMS    uint32
-	FileVersionLS    uint32
-	ProductVersionMS uint32
-	ProductVersionLS uint32
-	FileFlagsMask    uint32
-	FileFlags        uint32
-	FileOS           uint32
-	FileType         uint32
-	FileSubtype      uint32
-	FileDateMS       uint32
-	FileDateLS       uint32
-}
-
-type processEntry32 struct {
-	Size              uint32
-	CntUsage          uint32
-	ProcessID         uint32
-	DefaultHeapID     uintptr
-	ModuleID          uint32
-	CntThreads        uint32
-	ParentProcessID   uint32
-	PriorityClassBase int32
-	Flags             uint32
-	ExeFile           [maxPath]uint16
-}
-
-type moduleEntry32 struct {
-	Size         uint32
-	ModuleID     uint32
-	ProcessID    uint32
-	GlblcntUsage uint32
-	ProccntUsage uint32
-	ModBaseAddr  uintptr
-	ModBaseSize  uint32
-	HModule      syscall.Handle
-	ModuleName   [256]uint16
-	ExePath      [maxPath]uint16
-}
-
-// ---------------------------- Windows API helpers ----------------------------
-
-func findProcessID(name string) (uint32, error) {
-	snap, _, _ := procCreateToolhelp32Snapshot.Call(uintptr(th32csSnapProcess), 0)
-	if snap == 0 || snap == uintptr(^uintptr(0)) {
-		return 0, fmt.Errorf("couldn't take a process snapshot")
-	}
-	defer procCloseHandle.Call(snap)
-
-	var pe processEntry32
-	pe.Size = uint32(unsafe.Sizeof(pe))
-	r, _, _ := procProcess32FirstW.Call(snap, uintptr(unsafe.Pointer(&pe)))
-	if r == 0 {
-		return 0, fmt.Errorf("Process32First failed")
-	}
-	for {
-		exe := syscall.UTF16ToString(pe.ExeFile[:])
-		if strings.EqualFold(exe, name) {
-			return pe.ProcessID, nil
-		}
-		r, _, _ := procProcess32NextW.Call(snap, uintptr(unsafe.Pointer(&pe)))
-		if r == 0 {
-			break
-		}
-	}
-	return 0, fmt.Errorf("process not found: %s", name)
-}
-
-func findModuleBase(pid uint32, name string) (uintptr, uint32, string, error) {
-	snap, _, _ := procCreateToolhelp32Snapshot.Call(uintptr(th32csSnapModule|th32csSnapModule32), uintptr(pid))
-	if snap == 0 || snap == uintptr(^uintptr(0)) {
-		return 0, 0, "", fmt.Errorf("couldn't take a module snapshot")
-	}
-	defer procCloseHandle.Call(snap)
-
-	var me moduleEntry32
-	me.Size = uint32(unsafe.Sizeof(me))
-	r, _, _ := procModule32FirstW.Call(snap, uintptr(unsafe.Pointer(&me)))
-	if r == 0 {
-		return 0, 0, "", fmt.Errorf("Module32First failed")
-	}
-	for {
-		mname := syscall.UTF16ToString(me.ModuleName[:])
-		if strings.EqualFold(mname, name) {
-			return me.ModBaseAddr, me.ModBaseSize, syscall.UTF16ToString(me.ExePath[:]), nil
-		}
-		r, _, _ := procModule32NextW.Call(snap, uintptr(unsafe.Pointer(&me)))
-		if r == 0 {
-			break
-		}
-	}
-	return 0, 0, "", fmt.Errorf("module not found: %s", name)
-}
-
-// productVersion reads the game executable's version. label carries both
-// full version numbers (product and file) because they don't always
-// match, which helps diagnose whether the PlayerIns offset ever needs
-// adjusting.
-func productVersion(path string) (major, minor uint16, label string, ok bool) {
-	if path == "" {
-		return 0, 0, "", false
-	}
-	p, err := syscall.UTF16PtrFromString(path)
-	if err != nil {
-		return 0, 0, "", false
-	}
-	size, _, _ := procGetFileVersionInfoSizeW.Call(uintptr(unsafe.Pointer(p)), 0)
-	if size == 0 {
-		return 0, 0, "", false
-	}
-	buf := make([]byte, size)
-	r, _, _ := procGetFileVersionInfoW.Call(uintptr(unsafe.Pointer(p)), 0, size, uintptr(unsafe.Pointer(&buf[0])))
-	if r == 0 {
-		return 0, 0, "", false
-	}
-	sub, err := syscall.UTF16PtrFromString(`\`)
-	if err != nil {
-		return 0, 0, "", false
-	}
-	var info *vsFixedFileInfo
-	var infoLen uint32
-	r, _, _ = procVerQueryValueW.Call(
-		uintptr(unsafe.Pointer(&buf[0])),
-		uintptr(unsafe.Pointer(sub)),
-		uintptr(unsafe.Pointer(&info)),
-		uintptr(unsafe.Pointer(&infoLen)),
-	)
-	if r == 0 || info == nil || infoLen == 0 {
-		return 0, 0, "", false
-	}
-	quad := func(ms, ls uint32) string {
-		return fmt.Sprintf("%d.%d.%d.%d", ms>>16, ms&0xFFFF, ls>>16, ls&0xFFFF)
-	}
-	label = fmt.Sprintf("product %s / file %s",
-		quad(info.ProductVersionMS, info.ProductVersionLS),
-		quad(info.FileVersionMS, info.FileVersionLS))
-	return uint16(info.ProductVersionMS >> 16), uint16(info.ProductVersionMS & 0xFFFF), label, true
-}
-
-// playerInsOffsetForVersion mirrors SoulMemory's table (InitializeOffsets):
-// up to 1.06 the PlayerIns offset inside WorldChrMan is 0x18468, from
-// 1.07 onward it's 0x1E508.
-//
-// WATCH OUT: the version the exe reports is NOT what the game shows on
-// screen (the exe can say 2.7.1.0 while the game says 1.17.1), and
-// SoulMemory's table is written with the game's own numbers. So this is
-// only a HUNCH for deciding which one to try first: what actually
-// decides is playerInsCandidates plus the in-memory verification.
-func playerInsOffsetForVersion(major, minor uint16, ok bool) uintptr {
-	if ok && major == 1 && minor <= 6 {
-		return 0x18468
-	}
-	return 0x1E508
-}
-
-// playerInsCandidates returns the known offsets to try, with the one the
-// version suggests listed first.
-func playerInsCandidates(major, minor uint16, ok bool) []uintptr {
-	if playerInsOffsetForVersion(major, minor, ok) == 0x18468 {
-		return []uintptr{0x18468, 0x1E508}
-	}
-	return []uintptr{0x1E508, 0x18468}
-}
-
-func openProcessHandle(pid uint32) (syscall.Handle, error) {
-	h, _, err := procOpenProcess.Call(uintptr(processQueryInformation|processVMRead), 0, uintptr(pid))
-	if h == 0 {
-		return 0, err
-	}
-	return syscall.Handle(h), nil
-}
-
-func readMemory(h syscall.Handle, addr uintptr, size int) ([]byte, bool) {
-	if addr == 0 {
-		return nil, false
-	}
-	buf := make([]byte, size)
-	var n uintptr
-	r, _, _ := procReadProcessMemory.Call(uintptr(h), addr, uintptr(unsafe.Pointer(&buf[0])), uintptr(size), uintptr(unsafe.Pointer(&n)))
-	if r == 0 || int(n) != size {
-		return nil, false
-	}
-	return buf, true
-}
-
-// ------------------------- signature scans -------------------------
-//
-// Two signatures, both the same shape: a 7-byte "mov reg,[rip+disp32]"
-// instruction, where the first 3 bytes are the opcode and the next 4 are
-// the displacement. The resolved static slot holds the pointer to the
-// object (one more dereference needed).
-//
-//	GameDataMan  -> holds the death counter (+0x94) and the boss-fight flag
-//	              (+0xC0). Same pattern eldenring_boss_timer.asl (LiveSplit)
-//	              uses.
-//	WorldChrMan  -> holds the pointer to PlayerIns (+playerInsOffset). If
-//	              that pointer is null, there's no character in the world:
-//	              you're at the main menu or on a loading screen. Exactly
-//	              what SoulMemory.IsPlayerLoaded() does.
-
-type patByte struct {
-	val      byte
-	wildcard bool
-}
-
-// parsePattern accepts "48 8B 05 ?? ?? ?? ??" (?? = wildcard).
-func parsePattern(s string) []patByte {
-	var out []patByte
-	for _, tok := range strings.Fields(s) {
-		if strings.HasPrefix(tok, "?") {
-			out = append(out, patByte{wildcard: true})
-			continue
-		}
-		v, err := strconv.ParseUint(tok, 16, 8)
-		if err != nil {
-			panic("invalid pattern: " + tok)
-		}
-		out = append(out, patByte{val: byte(v)})
-	}
-	return out
-}
-
-var (
-	// mov rax,[rip+disp32]; test rax,rax; jz +5; mov rax,[rax+58]; ret; ret
-	gameDataManPattern = parsePattern("48 8B 05 ?? ?? ?? ?? 48 85 C0 74 05 48 8B 40 58 C3 C3")
-	// mov rsi,[rip+disp32]; test rsi,rsi; ...  (WorldChrManImp, same as SoulMemory)
-	worldChrManPattern = parsePattern("48 8B 35 ?? ?? ?? ?? 48 85 F6 ?? ?? BB 01 00 00 00 89 5C 24 20 48 8B B6")
-	// mov rax,[rip+disp32]; cmp byte ptr [rax+disp32],0D; sete al; ret
-	// (GameMan: +0xAC0 holds the loaded character's save slot)
-	gameManPattern = parsePattern("48 8B 05 ?? ?? ?? ?? 80 B8 ?? ?? ?? ?? 0D 0F 94 C0 C3")
-)
-
-// saveSlotOffset: where GameMan stores the save-slot index (0-9) of the
-// loaded game. This is a character's REAL identity: it doesn't depend on
-// the name, so two characters sharing a name never mix.
-const saveSlotOffset = 0xAC0
-
-func matchAt(buf []byte, i int, pattern []patByte) bool {
-	if i+len(pattern) > len(buf) {
-		return false
-	}
-	for j, p := range pattern {
-		if !p.wildcard && buf[i+j] != p.val {
-			return false
-		}
-	}
-	return true
-}
-
-// scanModule looks for several patterns in a single pass over the module,
-// reading it in chunks (with overlap, in case a pattern straddles a chunk
-// boundary). Returns, for each pattern, the match address or 0.
-func scanModule(h syscall.Handle, base uintptr, size uint32, patterns [][]patByte) []uintptr {
-	const chunk = 1 << 20 // 1 MiB
-	const overlap = 64
-
-	found := make([]uintptr, len(patterns))
-	remaining := len(patterns)
-
-	var pos uint32
-	for pos < size && remaining > 0 {
-		readSize := chunk
-		if rem := int(size - pos); readSize > rem {
-			readSize = rem
-		}
-		buf, ok := readMemory(h, base+uintptr(pos), readSize)
-		if ok {
-			for i := 0; i < len(buf); i++ {
-				for p := range patterns {
-					if found[p] != 0 {
-						continue
-					}
-					if matchAt(buf, i, patterns[p]) {
-						found[p] = base + uintptr(pos) + uintptr(i)
-						remaining--
-					}
-				}
-			}
-		}
-		if uint32(readSize) <= overlap {
-			break
-		}
-		pos += uint32(readSize) - overlap
-	}
-	return found
-}
-
-// ripSlot turns the address of a 7-byte "mov reg,[rip+disp32]"
-// instruction into the address of the static slot it points to.
-func ripSlot(h syscall.Handle, matchAddr uintptr) (uintptr, error) {
-	if matchAddr == 0 {
-		return 0, fmt.Errorf("pattern not found")
-	}
-	codeLocation := matchAddr + 3 // the first 3 bytes are the opcode
-	dispBytes, ok := readMemory(h, codeLocation, 4)
-	if !ok {
-		return 0, fmt.Errorf("couldn't read the RIP-relative displacement")
-	}
-	disp := int32(binary.LittleEndian.Uint32(dispBytes))
-	return codeLocation + 4 + uintptr(int64(disp)), nil
-}
-
-// gamePointers gathers everything resolved just once per process session:
-// the static slots (which don't move) and the game's version.
-type gamePointers struct {
-	gameDataManSlot uintptr
-	worldChrManSlot uintptr // 0 if the pattern wasn't found (we keep going without the menu check)
-	gameManSlot     uintptr // 0 if not found: falls back to identifying by name
-	playerInsOffset uintptr // the one currently in use (or the preferred candidate)
-	playerInsTried  []uintptr
-	offsetConfirmed bool // true once verified by actually reading memory
-	versionLabel    string
-
-	nameChain     nameCandidate // how we got to the character's name
-	nameConfirmed bool
-	namePending   string // candidate waiting to repeat (see resolveCharName)
-	namePendingOf nameCandidate
-}
-
-// resolvePointers does the signature scans (expensive: walks the whole
-// module) just once per process session. Deliberately does NOT return
-// the resolved objects themselves: those pointers get re-read every tick,
-// because the game can destroy and recreate GameDataMan (e.g. going back
-// to the main menu and loading again). If we cached a stale address,
-// we'd keep reading it successfully (the memory page is still valid) but
-// its contents would belong to something else entirely — the most likely
-// cause of a counter that "goes up on its own" without an actual death.
-// SoulMemory does the same: its Pointer class resolves the whole chain on
-// every read, never caching the final address.
-func resolvePointers(h syscall.Handle, pid uint32) (gamePointers, error) {
-	var gp gamePointers
-
-	base, size, exePath, err := findModuleBase(pid, processName)
-	if err != nil {
-		return gp, err
-	}
-
-	major, minor, label, okVer := productVersion(exePath)
-	gp.playerInsTried = playerInsCandidates(major, minor, okVer)
-	gp.playerInsOffset = gp.playerInsTried[0]
-	gp.versionLabel = label
-	if !okVer {
-		gp.versionLabel = "unknown"
-	}
-
-	matches := scanModule(h, base, size, [][]patByte{gameDataManPattern, worldChrManPattern, gameManPattern})
-
-	gp.gameDataManSlot, err = ripSlot(h, matches[0])
-	if err != nil {
-		return gp, fmt.Errorf("GameDataMan's pattern wasn't found (did the game update?)")
-	}
-
-	// WorldChrMan is optional: if it's missing, we keep counting deaths,
-	// we just lose menu/loading-screen detection.
-	if slot, werr := ripSlot(h, matches[1]); werr == nil {
-		gp.worldChrManSlot = slot
-	}
-	// GameMan is optional too: without it, we identify by name.
-	if slot, gerr := ripSlot(h, matches[2]); gerr == nil {
-		gp.gameManSlot = slot
-	}
-	return gp, nil
-}
-
-// derefPointer reads a static slot (cheap: 8 bytes) and returns the
-// object's CURRENT address. Called every tick, not just once.
-func derefPointer(h syscall.Handle, slot uintptr) (uintptr, bool) {
-	if slot == 0 {
-		return 0, false
-	}
-	buf, ok := readMemory(h, slot, 8)
-	if !ok {
-		return 0, false
-	}
-	return uintptr(binary.LittleEndian.Uint64(buf)), true
-}
-
-// ------------------------- character name -------------------------
-//
-// Sourced from a Cheat Engine table: "GameDataMan +0C +9C, unicode,
-// length 19". That notation allows more than one reading (is 0x0C a
-// pointer to dereference, or do the two offsets just add up?), and on
-// top of that the community/the ASL use GameDataMan+0x08 to reach
-// PlayerGameData. So instead of picking one, all three get tried, and
-// whichever gives back something that looks like a real name wins.
-
-type nameCandidate struct {
-	ptrOffset  uintptr // offset where the pointer lives (0 = no dereference)
-	nameOffset uintptr // offset of the text within the object
-	label      string
-}
-
-var nameCandidates = []nameCandidate{
-	{0x08, 0x9C, "[GameDataMan+0x08]+0x9C (PlayerGameData)"},
-	{0x0C, 0x9C, "[GameDataMan+0x0C]+0x9C"},
-	{0x00, 0xA8, "GameDataMan+0xA8 (0x0C and 0x9C added together)"},
-}
-
-// looksLikeName (and charNameMaxChars) live in names.go: they don't
-// depend on Windows, so they're kept out of this file to be testable
-// without a PC with the game open.
-
-func readCharName(h syscall.Handle, gameDataMan uintptr, c nameCandidate) (string, bool) {
-	base := gameDataMan
-	if c.ptrOffset != 0 {
-		p, ok := derefPointer(h, gameDataMan+c.ptrOffset)
-		if !ok || p == 0 {
-			return "", false
-		}
-		base = p
-	}
-	buf, ok := readMemory(h, base+c.nameOffset, charNameMaxChars*2)
-	if !ok {
-		return "", false
-	}
-	u16 := make([]uint16, 0, charNameMaxChars)
-	for i := 0; i+1 < len(buf); i += 2 {
-		ch := binary.LittleEndian.Uint16(buf[i : i+2])
-		if ch == 0 {
-			break
-		}
-		u16 = append(u16, ch)
-	}
-	s := strings.TrimSpace(string(utf16.Decode(u16)))
-	if !looksLikeName(s) {
-		return "", false
-	}
-	return s, true
-}
-
-// resolveCharName returns the character's name. Before locking in a
-// variant, it requires seeing it give the SAME text on two readings in a
-// row: the real name doesn't change from one second to the next, but a
-// chunk of memory that happens to pass the filter is far less likely to
-// repeat.
-func resolveCharName(h syscall.Handle, gameDataMan uintptr, gp *gamePointers) (string, bool) {
-	if gp.nameConfirmed {
-		return readCharName(h, gameDataMan, gp.nameChain)
-	}
-	for _, c := range nameCandidates {
-		s, ok := readCharName(h, gameDataMan, c)
-		if !ok {
-			continue
-		}
-		if gp.namePending == s && gp.namePendingOf == c {
-			gp.nameChain = c
-			gp.nameConfirmed = true
-			gp.namePending = ""
-			log.Printf("character name: \"%s\" (read via %s)", s, c.label)
-			return s, true
-		}
-		gp.namePending = s
-		gp.namePendingOf = c
-		return "", false
-	}
-	gp.namePending = ""
-	return "", false
-}
-
-// readSaveSlot returns the loaded game's slot index (0-9), or -1 if it
-// couldn't be read. Elden Ring has 10 slots, so any other value is
-// garbage and gets discarded.
-func readSaveSlot(h syscall.Handle, gp gamePointers) int {
-	if gp.gameManSlot == 0 {
-		return -1
-	}
-	gameMan, ok := derefPointer(h, gp.gameManSlot)
-	if !ok || gameMan == 0 {
-		return -1
-	}
-	buf, ok := readMemory(h, gameMan+saveSlotOffset, 1)
-	if !ok {
-		return -1
-	}
-	slot := int(buf[0])
-	if slot < 0 || slot > 9 {
-		return -1
-	}
-	return slot
-}
-
-// isPlayerLoaded mirrors SoulMemory.IsPlayerLoaded(): resolves
-// WorldChrMan and reads the pointer to PlayerIns; if it's null, there's
-// no character in the world. The second return value says whether we
-// were able to evaluate it at all.
-//
-// Until the offset is confirmed, instead of trusting the version number
-// (which in Elden Ring doesn't match what the game displays), the known
-// offsets are tried and whichever one first points at genuinely readable
-// memory wins. That's decided by the machine, not by a table that can
-// age badly.
-func isPlayerLoaded(h syscall.Handle, gp *gamePointers) (loaded bool, known bool) {
-	if gp.worldChrManSlot == 0 {
-		return true, false
-	}
-	worldChrMan, ok := derefPointer(h, gp.worldChrManSlot)
-	if !ok {
-		return true, false
-	}
-	if worldChrMan == 0 {
-		return false, true
-	}
-
-	if gp.offsetConfirmed {
-		playerIns, ok := derefPointer(h, worldChrMan+gp.playerInsOffset)
-		if !ok {
-			return true, false
-		}
-		return playerIns != 0, true
-	}
-
-	for _, cand := range gp.playerInsTried {
-		playerIns, ok := derefPointer(h, worldChrMan+cand)
-		if !ok || playerIns == 0 {
-			continue
-		}
-		// A real pointer points at mapped memory; a garbage one almost
-		// never survives this read.
-		if _, ok := readMemory(h, playerIns, 8); !ok {
-			continue
-		}
-		gp.playerInsOffset = cand
-		gp.offsetConfirmed = true
-		log.Printf("PlayerIns confirmed at +0x%X (verified by reading the object, not by version number)", cand)
-		return true, true
-	}
-	return false, true
-}
-
-// -------------------------------- poller loop --------------------------------
-
-// maxPlausibleDeltaPerTick: between two readings ~1s apart, with the
-// character loaded the whole time, the real death counter can't go up by
-// more than this (and never goes down). A bigger jump almost always
-// means we grabbed memory that's no longer GameDataMan (a stale/invalid
-// address), not an actual death.
-const maxPlausibleDeltaPerTick = 3
-
-func pollLoop() {
-	var (
-		handle       syscall.Handle
-		pid          uint32
-		gp           gamePointers
-		resolved     bool
-		lastRaw      int32
-		haveLastRaw  bool
-		sawUnloaded  bool
-		warnedNoWCM  bool
-		lastBossRead bool
-
-		// Menu-detection watchdog: while we believe no character is
-		// loaded, we still peek at the death counter. If it climbs the
-		// way a real death does, our detection is lying (you were
-		// actually playing) and we turn it off.
-		unloadedRaw      int32
-		unloadedRawFirst int32
-		haveUnloadedRaw  bool
-	)
-
-	closeHandle := func() {
-		if handle != 0 {
-			procCloseHandle.Call(uintptr(handle))
-			handle = 0
-		}
-		pid = 0
-		gp = gamePointers{}
-		resolved = false
-		haveLastRaw = false
-		sawUnloaded = false
-		haveUnloadedRaw = false
-	}
-
-	for {
-		if handle == 0 {
-			newPid, err := findProcessID(processName)
-			if err != nil {
-				st.setDisconnected("waiting for eldenring.exe")
-				time.Sleep(3 * time.Second)
-				continue
-			}
-			h, err := openProcessHandle(newPid)
-			if err != nil {
-				st.setDisconnected("couldn't open the process (permissions?): " + err.Error())
-				time.Sleep(3 * time.Second)
-				continue
-			}
-			pid = newPid
-			handle = h
-			log.Printf("eldenring.exe found (PID %d), scanning signatures...", pid)
-		}
-
-		if !resolved {
-			p, err := resolvePointers(handle, pid)
-			if err != nil {
-				st.setDisconnected(err.Error())
-				time.Sleep(2 * time.Second)
-				// if the process died, release the handle to retry from scratch
-				if _, ferr := findProcessID(processName); ferr != nil {
-					closeHandle()
-				}
-				continue
-			}
-			gp = p
-			resolved = true
-			haveLastRaw = false
-			log.Printf("game version: %s | GameDataMan slot 0x%X", gp.versionLabel, gp.gameDataManSlot)
-			if gp.gameManSlot != 0 {
-				log.Printf("GameMan slot 0x%X (identifying characters by their save slot)", gp.gameManSlot)
-			} else {
-				log.Printf("warning: GameMan's pattern wasn't found; identifying characters by name")
-			}
-			if gp.worldChrManSlot != 0 {
-				log.Printf("WorldChrMan slot 0x%X | PlayerIns: trying +0x%X and confirming against memory", gp.worldChrManSlot, gp.playerInsOffset)
-			} else if !warnedNoWCM {
-				warnedNoWCM = true
-				log.Printf("warning: WorldChrMan's pattern wasn't found; still counting deaths but without menu/loading-screen detection")
-			}
-		}
-
-		// Same as LiveSplit's ASL, which does "if (!IsPlayerLoaded) return;":
-		// with no character in the world, nothing gets read. The total
-		// stays frozen on screen (no dash shown) so it doesn't flicker on
-		// every loading screen.
-		if loaded, known := isPlayerLoaded(handle, &gp); known && !loaded {
-			// Safety net. The PlayerIns offset depends on the game's
-			// version: if a patch ever moves it, we'd read null forever
-			// and the counter would freeze mid-stream.
-			//
-			// The only judge we can trust is the death counter itself:
-			// at the start menu it NEVER goes up. So the raw value gets
-			// peeked at without being used, and if it climbs the way a
-			// real death does (+1, +2, +3), you were actually playing
-			// and our detection was lying. Unlike a timeout, this can't
-			// fire just from leaving the game sitting at the menu a while.
-			if gdm, ok := derefPointer(handle, gp.gameDataManSlot); ok && gdm != 0 {
-				if buf, ok := readMemory(handle, gdm+0x94, 4); ok {
-					raw := int32(binary.LittleEndian.Uint32(buf))
-					if raw >= 0 && raw < 1_000_000 {
-						if haveUnloadedRaw {
-							if d := raw - unloadedRaw; d >= 1 && d <= maxPlausibleDeltaPerTick {
-								log.Printf("the death counter went from %d to %d while I thought no character was loaded: menu detection is wrong on this version, turning it off and continuing to count", unloadedRaw, raw)
-								gp.worldChrManSlot = 0
-								// Recover what happened during the confused
-								// stretch: keep the reference at that
-								// period's first reading so the "crossed a
-								// loading screen" logic can credit the
-								// deaths if there weren't many.
-								lastRaw = unloadedRawFirst
-								haveLastRaw = true
-								sawUnloaded = true
-								haveUnloadedRaw = false
-								continue
-							}
-						} else {
-							unloadedRawFirst = raw
-						}
-						unloadedRaw = raw
-						haveUnloadedRaw = true
-					}
-				}
-			}
-			// Careful: haveLastRaw/lastRaw are NOT touched, precisely so
-			// they can be compared against the last good reading once
-			// the world comes back.
-			st.setPlayerUnloaded("main menu or loading screen")
-			sawUnloaded = true
-			time.Sleep(1 * time.Second)
-			continue
-		}
-		haveUnloadedRaw = false
-
-		// Re-dereference the slot on EVERY tick (not just on connect) to
-		// never end up stuck with a stale GameDataMan address.
-		gameDataMan, ok := derefPointer(handle, gp.gameDataManSlot)
-		if !ok {
-			st.setDisconnected("lost the memory reading (the game closed or restarted)")
-			closeHandle()
-			time.Sleep(2 * time.Second)
-			continue
-		}
-		if gameDataMan == 0 {
-			st.setPlayerUnloaded("no game loaded")
-			sawUnloaded = true
-			time.Sleep(1 * time.Second)
-			continue
-		}
-
-		deathsBuf, ok1 := readMemory(handle, gameDataMan+0x94, 4)
-		bossBuf, ok2 := readMemory(handle, gameDataMan+0xC0, 1)
-		if !ok1 {
-			st.setDisconnected("lost the memory reading (the game closed or restarted)")
-			closeHandle()
-			time.Sleep(2 * time.Second)
-			continue
-		}
-		raw := int32(binary.LittleEndian.Uint32(deathsBuf))
-		boss := lastBossRead
-		if ok2 {
-			boss = bossBuf[0] != 0
-			lastBossRead = boss
-		}
-		if raw < 0 || raw > 1_000_000 {
-			log.Printf("discarding an impossible reading (raw %d) - rescanning signatures", raw)
-			resolved = false
-			haveLastRaw = false
-			time.Sleep(1 * time.Second)
-			continue
-		}
-
-		// Within the same save, the counter never goes down or jumps: if
-		// it does, it's memory that's no longer GameDataMan. Crossing a
-		// load, on the other hand, can change to anything, since it might
-		// be a different character — and setCharacter handles that case.
-		if haveLastRaw && !sawUnloaded {
-			delta := int64(raw) - int64(lastRaw)
-			if delta < 0 || delta > maxPlausibleDeltaPerTick {
-				log.Printf("discarding a suspicious reading (raw %d, previous %d) - rescanning signatures", raw, lastRaw)
-				resolved = false
-				haveLastRaw = false
-				time.Sleep(1 * time.Second)
-				continue
-			}
-		}
-
-		// Which character this is gets resolved BEFORE recording the
-		// reading: if you switched characters, the total jumps to theirs
-		// on this very pass, with no need to wait for a death.
-		name, _ := resolveCharName(handle, gameDataMan, &gp)
-		st.setCharacter(readSaveSlot(handle, gp), name, raw)
-
-		st.setReading(raw, boss)
-
-		lastRaw = raw
-		haveLastRaw = true
-		sawUnloaded = false
-
-		time.Sleep(1 * time.Second)
-	}
-}
-
-// ---------------------------------- HTTP ----------------------------------
-
 // portOf pulls the port out of an address like "0.0.0.0:47822", so your
 // partner can be told exactly what to put in their config.
 func portOf(addr string) string {
diff --git a/process.go b/process.go
new file mode 100644
index 0000000..d8a7363
--- /dev/null
+++ b/process.go
@@ -0,0 +1,616 @@
+// process.go: everything about finding the death counter in the game's
+// memory that does NOT depend on which OS is doing the reading — AOB
+// signature scanning, pointer resolution, character-name reading, and the
+// poll loop that ties it all together.
+//
+// The actual reading is behind procHandle and a handful of functions
+// (findProcessID, openProcess, closeProcessHandle, readMemory,
+// findModuleBase, productVersion, systemLang) implemented once per
+// platform in process_windows.go/process_linux.go. Elden Ring under
+// Proton on Linux is the exact same Windows binary Wine is running, so
+// every signature and offset below is identical on both platforms — only
+// how the process gets found and read differs.
+package main
+
+import (
+	"encoding/binary"
+	"fmt"
+	"log"
+	"strconv"
+	"strings"
+	"time"
+	"unicode/utf16"
+)
+
+const processName = "eldenring.exe"
+
+// procHandle is an opaque reference to an open process, produced by
+// openProcess and consumed by readMemory/closeProcessHandle. What it
+// actually holds is platform-specific: a real Windows HANDLE value, or
+// just a pid on Linux (which needs no persistent OS resource — see
+// process_linux.go).
+type procHandle uintptr
+
+// ------------------------- signature scans -------------------------
+//
+// Two signatures, both the same shape: a 7-byte "mov reg,[rip+disp32]"
+// instruction, where the first 3 bytes are the opcode and the next 4 are
+// the displacement. The resolved static slot holds the pointer to the
+// object (one more dereference needed).
+//
+//	GameDataMan  -> holds the death counter (+0x94) and the boss-fight flag
+//	              (+0xC0). Same pattern eldenring_boss_timer.asl (LiveSplit)
+//	              uses.
+//	WorldChrMan  -> holds the pointer to PlayerIns (+playerInsOffset). If
+//	              that pointer is null, there's no character in the world:
+//	              you're at the main menu or on a loading screen. Exactly
+//	              what SoulMemory.IsPlayerLoaded() does.
+
+type patByte struct {
+	val      byte
+	wildcard bool
+}
+
+// parsePattern accepts "48 8B 05 ?? ?? ?? ??" (?? = wildcard).
+func parsePattern(s string) []patByte {
+	var out []patByte
+	for _, tok := range strings.Fields(s) {
+		if strings.HasPrefix(tok, "?") {
+			out = append(out, patByte{wildcard: true})
+			continue
+		}
+		v, err := strconv.ParseUint(tok, 16, 8)
+		if err != nil {
+			panic("invalid pattern: " + tok)
+		}
+		out = append(out, patByte{val: byte(v)})
+	}
+	return out
+}
+
+var (
+	// mov rax,[rip+disp32]; test rax,rax; jz +5; mov rax,[rax+58]; ret; ret
+	gameDataManPattern = parsePattern("48 8B 05 ?? ?? ?? ?? 48 85 C0 74 05 48 8B 40 58 C3 C3")
+	// mov rsi,[rip+disp32]; test rsi,rsi; ...  (WorldChrManImp, same as SoulMemory)
+	worldChrManPattern = parsePattern("48 8B 35 ?? ?? ?? ?? 48 85 F6 ?? ?? BB 01 00 00 00 89 5C 24 20 48 8B B6")
+	// mov rax,[rip+disp32]; cmp byte ptr [rax+disp32],0D; sete al; ret
+	// (GameMan: +0xAC0 holds the loaded character's save slot)
+	gameManPattern = parsePattern("48 8B 05 ?? ?? ?? ?? 80 B8 ?? ?? ?? ?? 0D 0F 94 C0 C3")
+)
+
+// saveSlotOffset: where GameMan stores the save-slot index (0-9) of the
+// loaded game. This is a character's REAL identity: it doesn't depend on
+// the name, so two characters sharing a name never mix.
+const saveSlotOffset = 0xAC0
+
+func matchAt(buf []byte, i int, pattern []patByte) bool {
+	if i+len(pattern) > len(buf) {
+		return false
+	}
+	for j, p := range pattern {
+		if !p.wildcard && buf[i+j] != p.val {
+			return false
+		}
+	}
+	return true
+}
+
+// scanModule looks for several patterns in a single pass over the module,
+// reading it in chunks (with overlap, in case a pattern straddles a chunk
+// boundary). Returns, for each pattern, the match address or 0.
+func scanModule(h procHandle, base uintptr, size uint32, patterns [][]patByte) []uintptr {
+	const chunk = 1 << 20 // 1 MiB
+	const overlap = 64
+
+	found := make([]uintptr, len(patterns))
+	remaining := len(patterns)
+
+	var pos uint32
+	for pos < size && remaining > 0 {
+		readSize := chunk
+		if rem := int(size - pos); readSize > rem {
+			readSize = rem
+		}
+		buf, ok := readMemory(h, base+uintptr(pos), readSize)
+		if ok {
+			for i := 0; i < len(buf); i++ {
+				for p := range patterns {
+					if found[p] != 0 {
+						continue
+					}
+					if matchAt(buf, i, patterns[p]) {
+						found[p] = base + uintptr(pos) + uintptr(i)
+						remaining--
+					}
+				}
+			}
+		}
+		if uint32(readSize) <= overlap {
+			break
+		}
+		pos += uint32(readSize) - overlap
+	}
+	return found
+}
+
+// ripSlot turns the address of a 7-byte "mov reg,[rip+disp32]"
+// instruction into the address of the static slot it points to.
+func ripSlot(h procHandle, matchAddr uintptr) (uintptr, error) {
+	if matchAddr == 0 {
+		return 0, fmt.Errorf("pattern not found")
+	}
+	codeLocation := matchAddr + 3 // the first 3 bytes are the opcode
+	dispBytes, ok := readMemory(h, codeLocation, 4)
+	if !ok {
+		return 0, fmt.Errorf("couldn't read the RIP-relative displacement")
+	}
+	disp := int32(binary.LittleEndian.Uint32(dispBytes))
+	return codeLocation + 4 + uintptr(int64(disp)), nil
+}
+
+// gamePointers gathers everything resolved just once per process session:
+// the static slots (which don't move) and the game's version.
+type gamePointers struct {
+	gameDataManSlot uintptr
+	worldChrManSlot uintptr // 0 if the pattern wasn't found (we keep going without the menu check)
+	gameManSlot     uintptr // 0 if not found: falls back to identifying by name
+	playerInsOffset uintptr // the one currently in use (or the preferred candidate)
+	playerInsTried  []uintptr
+	offsetConfirmed bool // true once verified by actually reading memory
+	versionLabel    string
+
+	nameChain     nameCandidate // how we got to the character's name
+	nameConfirmed bool
+	namePending   string // candidate waiting to repeat (see resolveCharName)
+	namePendingOf nameCandidate
+}
+
+// resolvePointers does the signature scans (expensive: walks the whole
+// module) just once per process session. Deliberately does NOT return
+// the resolved objects themselves: those pointers get re-read every tick,
+// because the game can destroy and recreate GameDataMan (e.g. going back
+// to the main menu and loading again). If we cached a stale address,
+// we'd keep reading it successfully (the memory page is still valid) but
+// its contents would belong to something else entirely — the most likely
+// cause of a counter that "goes up on its own" without an actual death.
+// SoulMemory does the same: its Pointer class resolves the whole chain on
+// every read, never caching the final address.
+func resolvePointers(h procHandle, pid uint32) (gamePointers, error) {
+	var gp gamePointers
+
+	base, size, exePath, err := findModuleBase(pid, processName)
+	if err != nil {
+		return gp, err
+	}
+
+	major, minor, label, okVer := productVersion(exePath)
+	gp.playerInsTried = playerInsCandidates(major, minor, okVer)
+	gp.playerInsOffset = gp.playerInsTried[0]
+	gp.versionLabel = label
+	if !okVer {
+		gp.versionLabel = "unknown"
+	}
+
+	matches := scanModule(h, base, size, [][]patByte{gameDataManPattern, worldChrManPattern, gameManPattern})
+
+	gp.gameDataManSlot, err = ripSlot(h, matches[0])
+	if err != nil {
+		return gp, fmt.Errorf("GameDataMan's pattern wasn't found (did the game update?)")
+	}
+
+	// WorldChrMan is optional: if it's missing, we keep counting deaths,
+	// we just lose menu/loading-screen detection.
+	if slot, werr := ripSlot(h, matches[1]); werr == nil {
+		gp.worldChrManSlot = slot
+	}
+	// GameMan is optional too: without it, we identify by name.
+	if slot, gerr := ripSlot(h, matches[2]); gerr == nil {
+		gp.gameManSlot = slot
+	}
+	return gp, nil
+}
+
+// derefPointer reads a static slot (cheap: 8 bytes) and returns the
+// object's CURRENT address. Called every tick, not just once.
+func derefPointer(h procHandle, slot uintptr) (uintptr, bool) {
+	if slot == 0 {
+		return 0, false
+	}
+	buf, ok := readMemory(h, slot, 8)
+	if !ok {
+		return 0, false
+	}
+	return uintptr(binary.LittleEndian.Uint64(buf)), true
+}
+
+// ------------------------- character name -------------------------
+//
+// Sourced from a Cheat Engine table: "GameDataMan +0C +9C, unicode,
+// length 19". That notation allows more than one reading (is 0x0C a
+// pointer to dereference, or do the two offsets just add up?), and on
+// top of that the community/the ASL use GameDataMan+0x08 to reach
+// PlayerGameData. So instead of picking one, all three get tried, and
+// whichever gives back something that looks like a real name wins.
+
+type nameCandidate struct {
+	ptrOffset  uintptr // offset where the pointer lives (0 = no dereference)
+	nameOffset uintptr // offset of the text within the object
+	label      string
+}
+
+var nameCandidates = []nameCandidate{
+	{0x08, 0x9C, "[GameDataMan+0x08]+0x9C (PlayerGameData)"},
+	{0x0C, 0x9C, "[GameDataMan+0x0C]+0x9C"},
+	{0x00, 0xA8, "GameDataMan+0xA8 (0x0C and 0x9C added together)"},
+}
+
+// looksLikeName (and charNameMaxChars) live in names.go: they don't
+// depend on the OS at all, let alone Windows vs. Linux, so they're kept
+// out of this file to be testable without a PC with the game open.
+
+func readCharName(h procHandle, gameDataMan uintptr, c nameCandidate) (string, bool) {
+	base := gameDataMan
+	if c.ptrOffset != 0 {
+		p, ok := derefPointer(h, gameDataMan+c.ptrOffset)
+		if !ok || p == 0 {
+			return "", false
+		}
+		base = p
+	}
+	buf, ok := readMemory(h, base+c.nameOffset, charNameMaxChars*2)
+	if !ok {
+		return "", false
+	}
+	u16 := make([]uint16, 0, charNameMaxChars)
+	for i := 0; i+1 < len(buf); i += 2 {
+		ch := binary.LittleEndian.Uint16(buf[i : i+2])
+		if ch == 0 {
+			break
+		}
+		u16 = append(u16, ch)
+	}
+	s := strings.TrimSpace(string(utf16.Decode(u16)))
+	if !looksLikeName(s) {
+		return "", false
+	}
+	return s, true
+}
+
+// resolveCharName returns the character's name. Before locking in a
+// variant, it requires seeing it give the SAME text on two readings in a
+// row: the real name doesn't change from one second to the next, but a
+// chunk of memory that happens to pass the filter is far less likely to
+// repeat.
+func resolveCharName(h procHandle, gameDataMan uintptr, gp *gamePointers) (string, bool) {
+	if gp.nameConfirmed {
+		return readCharName(h, gameDataMan, gp.nameChain)
+	}
+	for _, c := range nameCandidates {
+		s, ok := readCharName(h, gameDataMan, c)
+		if !ok {
+			continue
+		}
+		if gp.namePending == s && gp.namePendingOf == c {
+			gp.nameChain = c
+			gp.nameConfirmed = true
+			gp.namePending = ""
+			log.Printf("character name: \"%s\" (read via %s)", s, c.label)
+			return s, true
+		}
+		gp.namePending = s
+		gp.namePendingOf = c
+		return "", false
+	}
+	gp.namePending = ""
+	return "", false
+}
+
+// readSaveSlot returns the loaded game's slot index (0-9), or -1 if it
+// couldn't be read. Elden Ring has 10 slots, so any other value is
+// garbage and gets discarded.
+func readSaveSlot(h procHandle, gp gamePointers) int {
+	if gp.gameManSlot == 0 {
+		return -1
+	}
+	gameMan, ok := derefPointer(h, gp.gameManSlot)
+	if !ok || gameMan == 0 {
+		return -1
+	}
+	buf, ok := readMemory(h, gameMan+saveSlotOffset, 1)
+	if !ok {
+		return -1
+	}
+	slot := int(buf[0])
+	if slot < 0 || slot > 9 {
+		return -1
+	}
+	return slot
+}
+
+// isPlayerLoaded mirrors SoulMemory.IsPlayerLoaded(): resolves
+// WorldChrMan and reads the pointer to PlayerIns; if it's null, there's
+// no character in the world. The second return value says whether we
+// were able to evaluate it at all.
+//
+// Until the offset is confirmed, instead of trusting the version number
+// (which in Elden Ring doesn't match what the game displays, and on
+// Linux isn't available at all — see productVersion), the known offsets
+// are tried and whichever one first points at genuinely readable memory
+// wins. That's decided by the machine, not by a table that can age badly
+// or an OS that can't report a version at all.
+func isPlayerLoaded(h procHandle, gp *gamePointers) (loaded bool, known bool) {
+	if gp.worldChrManSlot == 0 {
+		return true, false
+	}
+	worldChrMan, ok := derefPointer(h, gp.worldChrManSlot)
+	if !ok {
+		return true, false
+	}
+	if worldChrMan == 0 {
+		return false, true
+	}
+
+	if gp.offsetConfirmed {
+		playerIns, ok := derefPointer(h, worldChrMan+gp.playerInsOffset)
+		if !ok {
+			return true, false
+		}
+		return playerIns != 0, true
+	}
+
+	for _, cand := range gp.playerInsTried {
+		playerIns, ok := derefPointer(h, worldChrMan+cand)
+		if !ok || playerIns == 0 {
+			continue
+		}
+		// A real pointer points at mapped memory; a garbage one almost
+		// never survives this read.
+		if _, ok := readMemory(h, playerIns, 8); !ok {
+			continue
+		}
+		gp.playerInsOffset = cand
+		gp.offsetConfirmed = true
+		log.Printf("PlayerIns confirmed at +0x%X (verified by reading the object, not by version number)", cand)
+		return true, true
+	}
+	return false, true
+}
+
+// playerInsOffsetForVersion mirrors SoulMemory's table (InitializeOffsets):
+// up to 1.06 the PlayerIns offset inside WorldChrMan is 0x18468, from
+// 1.07 onward it's 0x1E508.
+//
+// WATCH OUT: the version the exe reports is NOT what the game shows on
+// screen (the exe can say 2.7.1.0 while the game says 1.17.1), and
+// SoulMemory's table is written with the game's own numbers. So this is
+// only a HUNCH for deciding which one to try first: what actually
+// decides is playerInsCandidates plus the in-memory verification in
+// isPlayerLoaded. On Linux, where productVersion always reports ok=false,
+// this hunch is simply skipped — the in-memory verification still nails
+// it down.
+func playerInsOffsetForVersion(major, minor uint16, ok bool) uintptr {
+	if ok && major == 1 && minor <= 6 {
+		return 0x18468
+	}
+	return 0x1E508
+}
+
+// playerInsCandidates returns the known offsets to try, with the one the
+// version suggests listed first.
+func playerInsCandidates(major, minor uint16, ok bool) []uintptr {
+	if playerInsOffsetForVersion(major, minor, ok) == 0x18468 {
+		return []uintptr{0x18468, 0x1E508}
+	}
+	return []uintptr{0x1E508, 0x18468}
+}
+
+// -------------------------------- poller loop --------------------------------
+
+// maxPlausibleDeltaPerTick: between two readings ~1s apart, with the
+// character loaded the whole time, the real death counter can't go up by
+// more than this (and never goes down). A bigger jump almost always
+// means we grabbed memory that's no longer GameDataMan (a stale/invalid
+// address), not an actual death.
+const maxPlausibleDeltaPerTick = 3
+
+func pollLoop() {
+	var (
+		handle       procHandle
+		pid          uint32
+		gp           gamePointers
+		resolved     bool
+		lastRaw      int32
+		haveLastRaw  bool
+		sawUnloaded  bool
+		warnedNoWCM  bool
+		lastBossRead bool
+
+		// Menu-detection watchdog: while we believe no character is
+		// loaded, we still peek at the death counter. If it climbs the
+		// way a real death does, our detection is lying (you were
+		// actually playing) and we turn it off.
+		unloadedRaw      int32
+		unloadedRawFirst int32
+		haveUnloadedRaw  bool
+	)
+
+	closeHandle := func() {
+		if handle != 0 {
+			closeProcessHandle(handle)
+			handle = 0
+		}
+		pid = 0
+		gp = gamePointers{}
+		resolved = false
+		haveLastRaw = false
+		sawUnloaded = false
+		haveUnloadedRaw = false
+	}
+
+	for {
+		if handle == 0 {
+			newPid, err := findProcessID(processName)
+			if err != nil {
+				st.setDisconnected("waiting for eldenring.exe")
+				time.Sleep(3 * time.Second)
+				continue
+			}
+			h, err := openProcess(newPid)
+			if err != nil {
+				st.setDisconnected("couldn't open the process (permissions?): " + err.Error())
+				time.Sleep(3 * time.Second)
+				continue
+			}
+			pid = newPid
+			handle = h
+			log.Printf("eldenring.exe found (PID %d), scanning signatures...", pid)
+		}
+
+		if !resolved {
+			p, err := resolvePointers(handle, pid)
+			if err != nil {
+				st.setDisconnected(err.Error())
+				time.Sleep(2 * time.Second)
+				// if the process died, release the handle to retry from scratch
+				if _, ferr := findProcessID(processName); ferr != nil {
+					closeHandle()
+				}
+				continue
+			}
+			gp = p
+			resolved = true
+			haveLastRaw = false
+			log.Printf("game version: %s | GameDataMan slot 0x%X", gp.versionLabel, gp.gameDataManSlot)
+			if gp.gameManSlot != 0 {
+				log.Printf("GameMan slot 0x%X (identifying characters by their save slot)", gp.gameManSlot)
+			} else {
+				log.Printf("warning: GameMan's pattern wasn't found; identifying characters by name")
+			}
+			if gp.worldChrManSlot != 0 {
+				log.Printf("WorldChrMan slot 0x%X | PlayerIns: trying +0x%X and confirming against memory", gp.worldChrManSlot, gp.playerInsOffset)
+			} else if !warnedNoWCM {
+				warnedNoWCM = true
+				log.Printf("warning: WorldChrMan's pattern wasn't found; still counting deaths but without menu/loading-screen detection")
+			}
+		}
+
+		// Same as LiveSplit's ASL, which does "if (!IsPlayerLoaded) return;":
+		// with no character in the world, nothing gets read. The total
+		// stays frozen on screen (no dash shown) so it doesn't flicker on
+		// every loading screen.
+		if loaded, known := isPlayerLoaded(handle, &gp); known && !loaded {
+			// Safety net. The PlayerIns offset depends on the game's
+			// version: if a patch ever moves it, we'd read null forever
+			// and the counter would freeze mid-stream.
+			//
+			// The only judge we can trust is the death counter itself:
+			// at the start menu it NEVER goes up. So the raw value gets
+			// peeked at without being used, and if it climbs the way a
+			// real death does (+1, +2, +3), you were actually playing
+			// and our detection was lying. Unlike a timeout, this can't
+			// fire just from leaving the game sitting at the menu a while.
+			if gdm, ok := derefPointer(handle, gp.gameDataManSlot); ok && gdm != 0 {
+				if buf, ok := readMemory(handle, gdm+0x94, 4); ok {
+					raw := int32(binary.LittleEndian.Uint32(buf))
+					if raw >= 0 && raw < 1_000_000 {
+						if haveUnloadedRaw {
+							if d := raw - unloadedRaw; d >= 1 && d <= maxPlausibleDeltaPerTick {
+								log.Printf("the death counter went from %d to %d while I thought no character was loaded: menu detection is wrong on this version, turning it off and continuing to count", unloadedRaw, raw)
+								gp.worldChrManSlot = 0
+								// Recover what happened during the confused
+								// stretch: keep the reference at that
+								// period's first reading so the "crossed a
+								// loading screen" logic can credit the
+								// deaths if there weren't many.
+								lastRaw = unloadedRawFirst
+								haveLastRaw = true
+								sawUnloaded = true
+								haveUnloadedRaw = false
+								continue
+							}
+						} else {
+							unloadedRawFirst = raw
+						}
+						unloadedRaw = raw
+						haveUnloadedRaw = true
+					}
+				}
+			}
+			// Careful: haveLastRaw/lastRaw are NOT touched, precisely so
+			// they can be compared against the last good reading once
+			// the world comes back.
+			st.setPlayerUnloaded("main menu or loading screen")
+			sawUnloaded = true
+			time.Sleep(1 * time.Second)
+			continue
+		}
+		haveUnloadedRaw = false
+
+		// Re-dereference the slot on EVERY tick (not just on connect) to
+		// never end up stuck with a stale GameDataMan address.
+		gameDataMan, ok := derefPointer(handle, gp.gameDataManSlot)
+		if !ok {
+			st.setDisconnected("lost the memory reading (the game closed or restarted)")
+			closeHandle()
+			time.Sleep(2 * time.Second)
+			continue
+		}
+		if gameDataMan == 0 {
+			st.setPlayerUnloaded("no game loaded")
+			sawUnloaded = true
+			time.Sleep(1 * time.Second)
+			continue
+		}
+
+		deathsBuf, ok1 := readMemory(handle, gameDataMan+0x94, 4)
+		bossBuf, ok2 := readMemory(handle, gameDataMan+0xC0, 1)
+		if !ok1 {
+			st.setDisconnected("lost the memory reading (the game closed or restarted)")
+			closeHandle()
+			time.Sleep(2 * time.Second)
+			continue
+		}
+		raw := int32(binary.LittleEndian.Uint32(deathsBuf))
+		boss := lastBossRead
+		if ok2 {
+			boss = bossBuf[0] != 0
+			lastBossRead = boss
+		}
+		if raw < 0 || raw > 1_000_000 {
+			log.Printf("discarding an impossible reading (raw %d) - rescanning signatures", raw)
+			resolved = false
+			haveLastRaw = false
+			time.Sleep(1 * time.Second)
+			continue
+		}
+
+		// Within the same save, the counter never goes down or jumps: if
+		// it does, it's memory that's no longer GameDataMan. Crossing a
+		// load, on the other hand, can change to anything, since it might
+		// be a different character — and setCharacter handles that case.
+		if haveLastRaw && !sawUnloaded {
+			delta := int64(raw) - int64(lastRaw)
+			if delta < 0 || delta > maxPlausibleDeltaPerTick {
+				log.Printf("discarding a suspicious reading (raw %d, previous %d) - rescanning signatures", raw, lastRaw)
+				resolved = false
+				haveLastRaw = false
+				time.Sleep(1 * time.Second)
+				continue
+			}
+		}
+
+		// Which character this is gets resolved BEFORE recording the
+		// reading: if you switched characters, the total jumps to theirs
+		// on this very pass, with no need to wait for a death.
+		name, _ := resolveCharName(handle, gameDataMan, &gp)
+		st.setCharacter(readSaveSlot(handle, gp), name, raw)
+
+		st.setReading(raw, boss)
+
+		lastRaw = raw
+		haveLastRaw = true
+		sawUnloaded = false
+
+		time.Sleep(1 * time.Second)
+	}
+}
diff --git a/process_linux.go b/process_linux.go
new file mode 100644
index 0000000..14ecd1f
--- /dev/null
+++ b/process_linux.go
@@ -0,0 +1,183 @@
+//go:build linux
+
+// process_linux.go: the Linux side of the portable boundary defined in
+// process.go — for players running Elden Ring through Proton. Proton
+// runs the exact same Windows binary under Wine, so every AOB signature
+// and memory offset in process.go is unchanged; only how the process
+// gets found and read differs.
+//
+// No external dependencies (matching the project's single-binary goal):
+// process memory is read via /proc/<pid>/mem instead of hand-rolling a
+// raw process_vm_readv(2) syscall, which CLAUDE.md explicitly allows as
+// an equivalent alternative.
+package main
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+)
+
+// findProcessID finds Elden Ring's pid by walking every process's memory
+// mappings, not by matching a process name: Proton runs several helper
+// processes, and the one that actually has eldenring.exe mapped is the
+// one we want.
+func findProcessID(name string) (uint32, error) {
+	entries, err := os.ReadDir("/proc")
+	if err != nil {
+		return 0, fmt.Errorf("couldn't list /proc: %w", err)
+	}
+	for _, e := range entries {
+		pid, err := strconv.ParseUint(e.Name(), 10, 32)
+		if err != nil {
+			continue // not a pid directory
+		}
+		if _, _, _, ok := scanMaps(uint32(pid), name); ok {
+			return uint32(pid), nil
+		}
+	}
+	return 0, fmt.Errorf("process not found: %s", name)
+}
+
+// scanMaps walks /proc/<pid>/maps looking for lines whose mapped file's
+// base name matches name (case-insensitively), and returns the full span
+// across every matching line: the module can be split into several
+// segments (.text/.rdata/.data with different permissions), and the
+// scanner in process.go already reads in chunks and tolerates unreadable
+// ones, so the min-start/max-end span across all of them is enough.
+func scanMaps(pid uint32, name string) (base, end uintptr, path string, ok bool) {
+	f, err := os.Open(fmt.Sprintf("/proc/%d/maps", pid))
+	if err != nil {
+		return 0, 0, "", false
+	}
+	defer f.Close()
+
+	sc := bufio.NewScanner(f)
+	for sc.Scan() {
+		// Format: "start-end perms offset dev inode [pathname]". The
+		// pathname (anonymous mappings don't have one) is everything
+		// after the first 5 fields, rejoined with single spaces — a
+		// pathname with unusual internal spacing could theoretically
+		// come out collapsed, but that's a cosmetic edge case that
+		// doesn't affect matching against a base filename like
+		// "eldenring.exe".
+		fields := strings.Fields(sc.Text())
+		if len(fields) < 6 {
+			continue
+		}
+		mapPath := strings.Join(fields[5:], " ")
+		if !strings.EqualFold(filepath.Base(mapPath), name) {
+			continue
+		}
+
+		startStr, endStr, cut := strings.Cut(fields[0], "-")
+		if !cut {
+			continue
+		}
+		start, err1 := strconv.ParseUint(startStr, 16, 64)
+		stop, err2 := strconv.ParseUint(endStr, 16, 64)
+		if err1 != nil || err2 != nil {
+			continue
+		}
+		if !ok || uintptr(start) < base {
+			base = uintptr(start)
+		}
+		if uintptr(stop) > end {
+			end = uintptr(stop)
+		}
+		path = mapPath
+		ok = true
+	}
+	return base, end, path, ok
+}
+
+func findModuleBase(pid uint32, name string) (uintptr, uint32, string, error) {
+	base, end, path, ok := scanMaps(pid, name)
+	if !ok {
+		return 0, 0, "", fmt.Errorf("module not found: %s", name)
+	}
+	return base, uint32(end - base), path, nil
+}
+
+// openProcess doesn't need to attach to anything (readMemory reads via
+// /proc/<pid>/mem per call, no persistent handle involved) — it just
+// probes that memory is actually readable now, so a permissions problem
+// surfaces here with a clear explanation instead of as a silent stream
+// of failed reads later.
+func openProcess(pid uint32) (procHandle, error) {
+	if _, err := os.Stat(fmt.Sprintf("/proc/%d", pid)); err != nil {
+		return 0, fmt.Errorf("process %d not found: %w", pid, err)
+	}
+	f, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", pid), os.O_RDONLY, 0)
+	if err != nil {
+		return 0, fmt.Errorf(
+			"can't read process %d's memory (%v).\n"+
+				"This is almost always ptrace_scope blocking it. Grant this binary the capability once with:\n\n"+
+				"    sudo setcap cap_sys_ptrace+ep %s\n\n"+
+				"(don't lower kernel.yama.ptrace_scope or run this as root instead — that weakens "+
+				"ptrace protection for your whole system, not just this program)",
+			pid, err, exePathForSetcap())
+	}
+	f.Close()
+	return procHandle(pid), nil
+}
+
+// closeProcessHandle has nothing to release: see openProcess.
+func closeProcessHandle(h procHandle) {}
+
+func readMemory(h procHandle, addr uintptr, size int) ([]byte, bool) {
+	if addr == 0 {
+		return nil, false
+	}
+	f, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", uint32(h)), os.O_RDONLY, 0)
+	if err != nil {
+		return nil, false
+	}
+	defer f.Close()
+	buf := make([]byte, size)
+	if _, err := f.ReadAt(buf, int64(addr)); err != nil {
+		return nil, false
+	}
+	return buf, true
+}
+
+// productVersion has no Linux equivalent: it reads version.dll's
+// resource off the exe. This was always only a hint for which PlayerIns
+// offset to try first — isPlayerLoaded (process.go) confirms the real
+// one by reading memory regardless, so ok=false just skips straight to
+// that confirmation.
+func productVersion(path string) (major, minor uint16, label string, ok bool) {
+	return 0, 0, "", false
+}
+
+// systemLang reads the Unix locale environment instead of calling a
+// Windows API. Matches the Windows implementation's contract: returns
+// just the short base language code ("es", not "es_AR.UTF-8" or
+// "es-AR"), since that's what resolveLang (i18n.go) expects.
+func systemLang() string {
+	for _, key := range []string{"LC_ALL", "LC_MESSAGES", "LANG"} {
+		v := os.Getenv(key)
+		if v == "" || v == "C" || v == "POSIX" {
+			continue
+		}
+		v = strings.ToLower(v)
+		cut := len(v)
+		for _, sep := range []byte{'_', '.', '@'} {
+			if i := strings.IndexByte(v, sep); i >= 0 && i < cut {
+				cut = i
+			}
+		}
+		return v[:cut]
+	}
+	return ""
+}
+
+func exePathForSetcap() string {
+	if exe, err := os.Executable(); err == nil {
+		return exe
+	}
+	return "./deathwatch"
+}
diff --git a/process_windows.go b/process_windows.go
new file mode 100644
index 0000000..2121702
--- /dev/null
+++ b/process_windows.go
@@ -0,0 +1,223 @@
+//go:build windows
+
+// process_windows.go: the Windows side of the portable boundary defined
+// in process.go — finding the game process, opening/closing it, reading
+// its memory, finding a loaded module, reading the exe's file version,
+// and the system's UI language. All via raw Windows API calls (no
+// external dependencies, per the project's single-.exe goal).
+package main
+
+import (
+	"fmt"
+	"strings"
+	"syscall"
+	"unsafe"
+)
+
+const (
+	processQueryInformation = 0x0400
+	processVMRead           = 0x0010
+	th32csSnapProcess       = 0x00000002
+	th32csSnapModule        = 0x00000008
+	th32csSnapModule32      = 0x00000010
+	maxPath                 = 260
+)
+
+var (
+	kernel32                     = syscall.NewLazyDLL("kernel32.dll")
+	procOpenProcess              = kernel32.NewProc("OpenProcess")
+	procCloseHandle              = kernel32.NewProc("CloseHandle")
+	procReadProcessMemory        = kernel32.NewProc("ReadProcessMemory")
+	procCreateToolhelp32Snapshot = kernel32.NewProc("CreateToolhelp32Snapshot")
+	procModule32FirstW           = kernel32.NewProc("Module32FirstW")
+	procModule32NextW            = kernel32.NewProc("Module32NextW")
+	procProcess32FirstW          = kernel32.NewProc("Process32FirstW")
+	procProcess32NextW           = kernel32.NewProc("Process32NextW")
+	procGetUserDefaultLocaleName = kernel32.NewProc("GetUserDefaultLocaleName")
+
+	versionDLL                  = syscall.NewLazyDLL("version.dll")
+	procGetFileVersionInfoSizeW = versionDLL.NewProc("GetFileVersionInfoSizeW")
+	procGetFileVersionInfoW     = versionDLL.NewProc("GetFileVersionInfoW")
+	procVerQueryValueW          = versionDLL.NewProc("VerQueryValueW")
+)
+
+// vsFixedFileInfo is Windows's VS_FIXEDFILEINFO struct: used to pull the
+// game's version straight from eldenring.exe, same as SoulMemory (which
+// reads MainModule.FileVersionInfo.ProductVersion).
+type vsFixedFileInfo struct {
+	Signature        uint32
+	StrucVersion     uint32
+	FileVersionMS    uint32
+	FileVersionLS    uint32
+	ProductVersionMS uint32
+	ProductVersionLS uint32
+	FileFlagsMask    uint32
+	FileFlags        uint32
+	FileOS           uint32
+	FileType         uint32
+	FileSubtype      uint32
+	FileDateMS       uint32
+	FileDateLS       uint32
+}
+
+type processEntry32 struct {
+	Size              uint32
+	CntUsage          uint32
+	ProcessID         uint32
+	DefaultHeapID     uintptr
+	ModuleID          uint32
+	CntThreads        uint32
+	ParentProcessID   uint32
+	PriorityClassBase int32
+	Flags             uint32
+	ExeFile           [maxPath]uint16
+}
+
+type moduleEntry32 struct {
+	Size         uint32
+	ModuleID     uint32
+	ProcessID    uint32
+	GlblcntUsage uint32
+	ProccntUsage uint32
+	ModBaseAddr  uintptr
+	ModBaseSize  uint32
+	HModule      syscall.Handle
+	ModuleName   [256]uint16
+	ExePath      [maxPath]uint16
+}
+
+func findProcessID(name string) (uint32, error) {
+	snap, _, _ := procCreateToolhelp32Snapshot.Call(uintptr(th32csSnapProcess), 0)
+	if snap == 0 || snap == uintptr(^uintptr(0)) {
+		return 0, fmt.Errorf("couldn't take a process snapshot")
+	}
+	defer procCloseHandle.Call(snap)
+
+	var pe processEntry32
+	pe.Size = uint32(unsafe.Sizeof(pe))
+	r, _, _ := procProcess32FirstW.Call(snap, uintptr(unsafe.Pointer(&pe)))
+	if r == 0 {
+		return 0, fmt.Errorf("Process32First failed")
+	}
+	for {
+		exe := syscall.UTF16ToString(pe.ExeFile[:])
+		if strings.EqualFold(exe, name) {
+			return pe.ProcessID, nil
+		}
+		r, _, _ := procProcess32NextW.Call(snap, uintptr(unsafe.Pointer(&pe)))
+		if r == 0 {
+			break
+		}
+	}
+	return 0, fmt.Errorf("process not found: %s", name)
+}
+
+func openProcess(pid uint32) (procHandle, error) {
+	h, _, err := procOpenProcess.Call(uintptr(processQueryInformation|processVMRead), 0, uintptr(pid))
+	if h == 0 {
+		return 0, err
+	}
+	return procHandle(h), nil
+}
+
+func closeProcessHandle(h procHandle) {
+	procCloseHandle.Call(uintptr(h))
+}
+
+func readMemory(h procHandle, addr uintptr, size int) ([]byte, bool) {
+	if addr == 0 {
+		return nil, false
+	}
+	buf := make([]byte, size)
+	var n uintptr
+	r, _, _ := procReadProcessMemory.Call(uintptr(h), addr, uintptr(unsafe.Pointer(&buf[0])), uintptr(size), uintptr(unsafe.Pointer(&n)))
+	if r == 0 || int(n) != size {
+		return nil, false
+	}
+	return buf, true
+}
+
+func findModuleBase(pid uint32, name string) (uintptr, uint32, string, error) {
+	snap, _, _ := procCreateToolhelp32Snapshot.Call(uintptr(th32csSnapModule|th32csSnapModule32), uintptr(pid))
+	if snap == 0 || snap == uintptr(^uintptr(0)) {
+		return 0, 0, "", fmt.Errorf("couldn't take a module snapshot")
+	}
+	defer procCloseHandle.Call(snap)
+
+	var me moduleEntry32
+	me.Size = uint32(unsafe.Sizeof(me))
+	r, _, _ := procModule32FirstW.Call(snap, uintptr(unsafe.Pointer(&me)))
+	if r == 0 {
+		return 0, 0, "", fmt.Errorf("Module32First failed")
+	}
+	for {
+		mname := syscall.UTF16ToString(me.ModuleName[:])
+		if strings.EqualFold(mname, name) {
+			return me.ModBaseAddr, me.ModBaseSize, syscall.UTF16ToString(me.ExePath[:]), nil
+		}
+		r, _, _ := procModule32NextW.Call(snap, uintptr(unsafe.Pointer(&me)))
+		if r == 0 {
+			break
+		}
+	}
+	return 0, 0, "", fmt.Errorf("module not found: %s", name)
+}
+
+// productVersion reads the game executable's version. label carries both
+// full version numbers (product and file) because they don't always
+// match, which helps diagnose whether the PlayerIns offset ever needs
+// adjusting.
+func productVersion(path string) (major, minor uint16, label string, ok bool) {
+	if path == "" {
+		return 0, 0, "", false
+	}
+	p, err := syscall.UTF16PtrFromString(path)
+	if err != nil {
+		return 0, 0, "", false
+	}
+	size, _, _ := procGetFileVersionInfoSizeW.Call(uintptr(unsafe.Pointer(p)), 0)
+	if size == 0 {
+		return 0, 0, "", false
+	}
+	buf := make([]byte, size)
+	r, _, _ := procGetFileVersionInfoW.Call(uintptr(unsafe.Pointer(p)), 0, size, uintptr(unsafe.Pointer(&buf[0])))
+	if r == 0 {
+		return 0, 0, "", false
+	}
+	sub, err := syscall.UTF16PtrFromString(`\`)
+	if err != nil {
+		return 0, 0, "", false
+	}
+	var info *vsFixedFileInfo
+	var infoLen uint32
+	r, _, _ = procVerQueryValueW.Call(
+		uintptr(unsafe.Pointer(&buf[0])),
+		uintptr(unsafe.Pointer(sub)),
+		uintptr(unsafe.Pointer(&info)),
+		uintptr(unsafe.Pointer(&infoLen)),
+	)
+	if r == 0 || info == nil || infoLen == 0 {
+		return 0, 0, "", false
+	}
+	quad := func(ms, ls uint32) string {
+		return fmt.Sprintf("%d.%d.%d.%d", ms>>16, ms&0xFFFF, ls>>16, ls&0xFFFF)
+	}
+	label = fmt.Sprintf("product %s / file %s",
+		quad(info.ProductVersionMS, info.ProductVersionLS),
+		quad(info.FileVersionMS, info.FileVersionLS))
+	return uint16(info.ProductVersionMS >> 16), uint16(info.ProductVersionMS & 0xFFFF), label, true
+}
+
+// systemLang returns Windows's language ("es-AR" -> "es").
+func systemLang() string {
+	buf := make([]uint16, 85) // LOCALE_NAME_MAX_LENGTH
+	r, _, _ := procGetUserDefaultLocaleName.Call(uintptr(unsafe.Pointer(&buf[0])), uintptr(len(buf)))
+	if r == 0 {
+		return ""
+	}
+	name := syscall.UTF16ToString(buf[:r])
+	if base, _, ok := strings.Cut(name, "-"); ok {
+		return strings.ToLower(base)
+	}
+	return strings.ToLower(name)
+}