Hook WriteFormattedLogMessage to inject session marker before first combat log write

Instead of hooking EnableChatLogging or SignalEvent (which fire too late
or before the log file is open), hook WriteFormattedLogMessage directly.
On the first write to the combat log, prepend COMBATLOG_SESSION with the
player name resolved via the name cache. This guarantees the session
marker is the very first line in the file.

Also add World_HandleLogoutCleanup (0x491180) hook to reset per-session
state on real logout/disconnect (not /reload), and move remove_on_shutdown
cleanup from shutdownDetour to logoutDetour.

src/combatlog/RESEARCH.md

@@ -269,17 +269,67 @@ On DLL init, scan `Logs\` directory for `WoWCombatLog_*.txt` files:
       .ebx=true })` barrier at top of detour forces Zig to push/pop in
       prologue/epilogue regardless of internal register allocation.
 
+## Logout Detection
+
+`CGGameUI_Shutdown` (0x490BD0) fires on `/reload` AND logout -- unsuitable for
+per-session reset. The correct logout-only function is:
+
+### World_HandleLogoutCleanup (0x491180) -- __stdcall(), no params
+
+Only called from `ShutdownClientSystems` (0x401ee0), which is called from:
+- `NetworkDisconnectHandler` (0x46c540)
+- `handleDisconnectWithReason` (0x5aad70)
+- `cleanupAfterDisconnect` (0x5aaeb0)
+
+These are all real disconnect/logout paths. NOT called on `/reload` or map change.
+
+`World_HandleLogoutCleanup` calls `World_HandlePlayerLogin` (0x490BD0) as its
+first action, which fires `SignalEvent(271)` (PLAYER_LOGOUT).
+
+Full cleanup sequence: World_HandlePlayerLogin (fires PLAYER_LOGOUT event),
+UI_ProcessDirtyFaces, release_minimap_textures, ShutdownChatSubsystem,
+SpellBookManager_Cleanup, CleanupUnitSystem, UnregisterPlayerStateHandlers, etc.
+
+Hooked in `main.zig` as `logout_hook` -- available to all modules via `onLogout()`.
+
+### Event IDs (verified via event table at 0xbe1198)
+
+- PLAYER_LOGIN = 270 (table slot 0xbe15d0, string at 0x852d50)
+- PLAYER_LOGOUT = 271 (table slot 0xbe15d4, string at 0x852d40)
+- PLAYER_ENTERING_WORLD = 272 (table slot 0xbe15d8, string at 0x852d28)
+- LOGOUT_CANCEL = event at table slot 0xbe15f0, string at 0x852cc4
+
+### SignalEvent callers for login/logout events
+
+- `SignalEvent(270)` PLAYER_LOGIN: from `CGGameUI::EnterWorld` (0x4908c0)
+- `SignalEvent(271)` PLAYER_LOGOUT: from `World_HandlePlayerLogin` (0x490bd0)
+- `SignalEvent(272)` PLAYER_ENTERING_WORLD: from `CGGameUI::EnterWorld` (0x4908c0)
+- `SignalEvent(273)`: from `World_HandleEnterWorldCleanup` (0x490a80)
+
+### Player Name Resolution
+
+`GetObjectPtr` (0x464870) -> `GetUnitName` (0x609210) does NOT work during early
+login -- player object not in object manager yet. Use name cache instead:
+`RetrieveNPCDataFromCache` (0x55f080, cache at 0xc0e228) is populated before the
+object manager and works for the local player GUID.
+
 ## Open Questions
 
 - [ ] Where does SuperWoW write `WoWRawCombatLog.txt`? Need to check
       SuperWoWhook.dll. Same redirect approach should work if it uses the
       same table or a similar one.
-- [ ] Exact prologue size of `EnableChatLogging` for hooking (need 5+ bytes).
+- [x] Exact prologue size of `EnableChatLogging` for hooking (need 5+ bytes).
       First 3 instructions = PUSH EBX; PUSH ESI; PUSH EDI = 3 bytes. Then
       MOV ESI,EDX = 2 bytes. Total = 5 bytes -- just enough for a jmp hook.
 - [x] **Markers permission check**: markers' `getNameFromGUID` uses
       `RetrieveNPCDataFromCache` (0x55f080) and works correctly for roster GUIDs.
       The combatlog crash (EAX=name bytes instead of char*) is specific to the
-      context — either the player's own GUID is handled differently, or the cache
-      state during `EnableChatLogging` is incomplete. Combatlog uses `GetObjectName`
-      (0x6264E0) instead, which avoids the issue. No change needed for markers.
+      context -- either the player's own GUID is handled differently, or the cache
+      state during `EnableChatLogging` is incomplete. Combatlog now uses
+      RetrieveNPCDataFromCache directly (same pattern as markers) and it works.
+- [x] **Player name at EnableChatLogging time**: name cache IS populated when
+      LoggingCombat(1) fires. The GetObjectPtr path failed because the object
+      manager isn't ready, not because the name isn't known.
+- [x] **Session marker ordering**: Hooking WriteFormattedLogMessage (0x65ac20)
+      directly guarantees our marker is the first line -- intercepts the first
+      write to the combat log handle and prepends the session marker.