- Remove complex ramp/edge filtering that was skipping building walls
- Simpler wall detection: any vertical geometry above step height
- Add intro camera pan on game start
- Use explicit loaded flag for heightmap instead of checking height values
- Increase player collision radius from 0.50 to 0.55 for better wall collision
Don't snap player down when detected floor is >2 units below current
position. This prevents falling through upper floors when the raycast
accidentally detects the ground floor through geometry gaps.
- Increase footprint sampling radius from 0.28 to 0.4 units
- Only update floor cache if found floor is close to query height
- Prevents caching wrong floor from different stories
Skip cached floor height if it's too far below query point (>4 units),
forcing a full raycast. This handles cases where the cache has a
different floor's height (e.g., ground floor cached while on upper floor).
Save/load floor cache per map (e.g., cache/wmo_floor_Azeroth.bin) instead
of a single global file. Saves current zone's cache before teleporting,
loads target zone's cache after terrain streaming completes.
Add precomputeFloorCache() method that samples all WMO bounds at load time
and populates the persistent floor height grid. Called after terrain
streaming completes (initial spawn and teleport) when cache is empty.
- Add texture-sorted batch merging with glMultiDrawElements to reduce draw calls
- Pre-compute merged batches at load time instead of per-frame
- Add persistent floor height cache with disk save/load (cache/wmo_floor_cache.bin)
- Reduce collision focus radius and sweep steps for faster collision checks
- Add floor cache size to performance HUD
- Reduce WMO group distance culling to 80 units
Renders bounding boxes in depth-only pre-pass and queries GPU for
visibility. Groups fully occluded in previous frame are skipped.
Significantly improves performance in dense areas like Stormwind.
Portal-based visibility culling for WMO rendering (disabled by default,
needs debugging for complex WMOs like Stormwind). Skip character creation
screen when characters already exist in single-player mode.
Save SHA1(UPPER(user):UPPER(pass)) hash to login.cfg instead of the
plaintext password. On subsequent logins, use the stored hash directly
with a new authenticateWithHash() method that bypasses password hashing.
The password field shows a placeholder when using a stored hash.
For Section 0 (body skin) and Section 4 (underwear), the skin color is
stored in colorIndex with variation=0, not the other way around. This
fixes the preview showing no texture for skin colors other than 0.
Render an animated M2 character model in the creation screen using a
dedicated CharacterRenderer with an offscreen FBO. Preview updates on
race/gender/appearance changes, supports mouse-drag rotation, and
composites skin, face, hair scalp, and underwear textures from
CharSections.dbc. Extracts getPlayerModelPath() to a shared free
function in game/character.
The charCreateCallback was firing synchronously during ImGui render,
causing a state transition before the create screen's ImGui::End() was
called. Defer the callback to GameHandler::update() and ensure update()
runs during CHARACTER_CREATION and CHARACTER_SELECTION states.
Implements a full character creation UI integrated into the existing flow.
In single-player mode, auth screen now goes to character creation before
entering the world. In online mode, a "Create Character" button on the
character selection screen sends CMSG_CHAR_CREATE to the server. Includes
WoW 3.3.5a race/class combo validation and appearance range limits.
Game name uses big-endian (not reversed), platform/os/locale use
little-endian with chars reversed and null at end (not beginning).
Protocol version is 8 for WoW 3.3.5a, not 3.
Account for PIN/matrix/authenticator extra data in packet size
calculation to prevent receive buffer corruption. Add hex dump
of raw auth packets and show actual server error codes.
TrinityCore/AzerothCore's UpdateBigNumbers uses BN_num_bytes (natural
size without padding) when hashing values for u and M1. Our code was
using fixed 32-byte padding which produces different hashes when any
value (salt, A, B, N) has leading zeros in big-endian representation.
Two bugs that caused the server to always reject our login proof:
- N was hashed as 256 bytes (2048 bits) instead of 32 bytes (256 bits),
producing completely wrong H(N)^H(g) and therefore wrong M1
- Session key computation B-k*g^x could go negative; OpenSSL's
BN_mod_exp doesn't handle negative bases. Add k*N before subtracting
(standard TrinityCore approach) to keep the value positive
