With GLFW_CURSOR_DISABLED the official GLFW backend keeps feeding
virtual unbounded cursor coords to ImGui each frame, so hover state
wanders across hidden panels while the physical cursor is invisible.
Append a "mouse left the window" pos event before ImGui::NewFrame
whenever the cursor is captured - ImGui treats the frame as mouseless.
UE's far clip is effectively unlimited and shadow distance is an
independent parameter. Split kCameraFar (main projection, 10km) from
kShadowDistance (cascade splits, unchanged 100m) so distant geometry
is no longer clipped while shadow quality stays bit-identical.
Main projection now references the shared constants instead of
duplicated literals.
Assimp flags every mesh-less scene as AI_SCENE_FLAGS_INCOMPLETE
(FBXConverter), so pure animation FBX never reached clip parsing -
the scene was discarded before ParseAnimations ran. Only treat null
scene / missing root node as hard failure; log INCOMPLETE as a
warning and continue. Final validity stays with ParsedModel::Valid().
Pure-animation FBX (no meshes, skeleton clips only) was rejected at
the cache layer: ParsedModel::Valid() required non-empty meshes, so
Model::FromParsed/LoadModel returned null and the asset never made it
into the cache. Animation assets are now first-class:
- ParsedModel::Valid() = meshes or clips non-empty; parse failures
(both empty) still rejected. Empty-mesh consumers already guarded
(rendering skips empty mesh loop, mesh collider falls back to Box)
- ParseFromDisk: fix null-deref when assimp returns a null scene
- AssetCache::Find(path) + AssetManager::GetModelByPath for panels
- Asset Browser hover card shows meshes/bones/clips for loaded models
- Inspector model line flags "animation-only" when meshes == 0
- Animation section: collapsible clip data table (name/duration/
channel count) to verify bone-name matching of attached clips
Editor toolbar gains Play/Stop (Ctrl+P) with UE semantics:
- Play: serialize current scene to an in-memory YAML snapshot via new
SceneSerializer::SerializeToString, switch display to kDisplayDirect
(primary camera entity renders), unfreeze time (timeScale 0 -> 1)
- Stop: DeserializeFromString atomically rebuilds the pre-play scene,
back to kDisplayViewport (editor camera), refreeze time; selection
cleared and gizmo drag reset since all entity handles dangle
- SceneSerializer refactored: SerializeDoc/ParseRoot shared by file
and memory variants; file Save/Load behavior unchanged
- Running state disables gizmo + editor hotkeys in the viewport (input
goes to game scripts), draws yellow PLAYING hint and a no-camera
warning; File->Save/Load disabled while playing (UE behavior)
Physics bodies/constraints/characters and script instances are
reconciled automatically by the existing per-frame sync on restore.
Jolt's GENERATE_DEBUG_SYMBOLS option defaults to ON and injects -g in
every config, so the Release binary carried ~44MB of DWARF sections
(.debug_info/.debug_line/...) from Jolt TUs alone -- 80% of the 55MB
exe. Force the option OFF for Release builds only (the vendored
submodule is untouched; Debug/RelWithDebInfo keep symbols).
Measured: 55.0MB -> 11.3MB, .text unchanged at 8.5MB, zero debug
sections left. The remaining ~1.6MB is the ELF symtab, useful for
crash stacks; plain strip takes it to 9.7MB if a distributable is
needed.
Dragging rotation hit a wall at the wrap point: each frame re-read
euler from the quat and wrapped to [0,360), so dragging past 360
snapped the display back to 0 (and backwards dragging wrapped up to
360) -- values "jumped around" instead of accumulating.
Port UE TransformDetails session semantics: the rotation row shows a
local mirror value while the user edits it (accumulating freely in
both directions, never wrapping), and re-syncs from the quat only on
foreign changes (gizmo drag, physics writeback, entity switch).
Detection is exact component-wise quat equality: the mirror quat comes
from the same QuatFromEuler used by SetRotation, so an untouched pair
compares bit-equal, while any foreign writeback -- including sub-0.05
degree gizmo steps that a dot-product threshold would miss -- differs
in at least one float. Edit frames skip the check (mirror leads the
quat by one drag delta), gated on the previous frame's item-active
state.
User pointed out the real gap: UE Details shows rotation in [0,360)
(-90 reads 270); our readback exposed [-180,180] with negative zeros.
GetRotation now wraps each axis to [0,360) via fmod -- strictly
equivalent math (each axis has 360-degree period), so matrix-rebuild
consumers like Camera::EulerRotMat are unaffected. The one consumer
needing a continuous domain is the free-look pitch clamp [-89,89] in
Camera::Update, which wraps back to [-180,180) before applying deltas.
Typing 90 degrees on the pitch axis showed 89.98. The stored quat is
exact (matrices differ by 0.0) -- the noise is display-side: quat's
sin(90) components square to 0.99999994 in float, and asin() diverges
at |sin|->1, amplifying 6e-8 into 0.02 degrees. Yaw/roll use atan2 and
were always exact, matching the user's report being X-axis only.
Two fixes, both aligned with how UE gets away with it (internally UE
is quat + equally pathological, its Details panel just shows one
decimal):
- GetRotation gimbal-lock threshold 1.0 -> 1-1e-6: pitch near +-90
now takes the composite-angle atan2 branch (ratio-stable) instead
of asin, so typing 90.0 reads back exactly 90.0. Probed: X=90 and
X=-90 now round-trip to 0 error; generic poses stay ~1e-6 deg.
- Rotation row shows one decimal ("%.1f", UE Details default), masking
the residual asin noise for angles near (not at) the pole; Ctrl+Click
input stays full precision. Negative zero from atan2 cleaned up.
Remaining asin noise for e.g. 89.9 (~9e-4 deg) is below the 1-decimal
display and unreachable in practice.
Animation-only FBX files (no mesh, clips only) could not be used:
clips are keyed by node index into the parsed model's node table, so
there was no path to get them onto a different model's skeleton.
Add AssetManager::AttachAnimationClips: parses the container with the
CPU-only path (no GPU upload -- empty meshes cannot build a Model),
remaps each clip's channels to the target model's node indices by
bone name, and appends to its clip list. Channels whose bone name is
absent are skipped and counted; clips with zero matching channels are
dropped. Same-source attach is idempotent via a new
AnimationClip::sourcePath marker; name clashes get the file stem
prefixed.
Inspector Animation block gains an attach slot (UE Details semantics)
listing model-container files plus a bullet list of attached sources;
already-attached paths are replayed idempotently when the entity is
selected, so clips survive restarts. The paths persist in the scene
(Animation.Attached); the clips themselves stay runtime state.
Limitations (documented, by design): same-skeleton retargeting only
-- bone names must match between source and target (same DCC export
or Mixamo-style per-animation downloads); cross-skeleton retargeting
is out of scope.
Mesh Renderer was missing from the Add Component menu and its model
reference was read-only, so a mesh could only be attached by loading
a saved scene. Add it to the menu (Unity Add Component semantics) and
give the block a UE-Details-style asset slot: a dropdown listing all
model files under assets/, selecting applies LoadModel (cache deduped).
Asset browser files now show a UE Content Browser style hover card:
kind, size, modified time, and loaded state (new AssetCache::Contains
query, read-only -- hovering never triggers loading).
Fix latent FATALs unlocked by manual assignment: AssetCache::Get
aborts on invalid handles, so rendering, animation sampling, and mesh
collider build now guard mr.model.Valid() before calling it.
The per-entity script list allowed picking the same script class
twice; the name-keyed runtime silently ignores duplicates, leaving
the extra rows dead weight with unusable enable bits. Enforce
DisallowMultipleComponent semantics: the dropdown of each row only
offers unattached names (own name kept for echo), "+ Add Script"
disables when every registered script is attached, and scene loading
dedupes as a safety net for hand-edited files.
The file list was hardwired to the root directory, so it was always
empty (all assets live in subdirectories). Directories are now
selectable from both the tree (click row) and the list (click [D]
row), with a breadcrumb + Root button. Prefix check done manually
(path::starts_with is missing in this libstdc++).
ScriptComponent now holds a list of {name, active} items (Unity
semantics: one entity, many scripts, each with its own enable bit;
component-level active stays as the master switch). ScriptRuntime
reconciles instances by name per entity -- rename/remove destroys
only the affected instances, untouched ones keep ticking. Scenes
persist Script.Items; old single-script key still loads.
AssetBrowserPanel (lightweight Unity Project window): fixed roots
(assets/scenes) as tabs, recursive directory tree, file list with
type filter (models/textures/scenes) and name substring filter,
double-click preloads models/HDRs into the asset caches (path-
deduped). Scene files point to the File menu for now; drag & drop
is a later iteration. Docked with Stats by default in fresh layouts.
Clearer accessor naming on ScriptBehavior (Get -> GetComponent,
TryGet -> TryGetComponent), WASDMove follows. Teleport marked const
(JPH Ref does not propagate constness). Spinner: auto-rotation
disabled for the moment, registration switched from the registration
macro to an explicit lambda to verify the mechanism.
CharacterVirtual is not a registered body, so PhysicsSystem::DrawBodies
never renders it (Jolt provides no CharacterVirtual::DebugDraw either).
Add CharacterController::DebugDraw (shape wireframe at foot-origin +
shapeOffset, green) and drive it from CollectDebugData under the
colliders bit.
UE calls it Trigger (Overlap Events / Trigger preset) and Unity's UI
says "Is Trigger"; "sensor" is the physics-backend term (Jolt sensor,
Box2D sensor). Rename the component/record flag and the Inspector
label. Scene files now persist "IsTrigger" with backward-compatible
read of the old "IsSensor" key.
ScriptBehavior gains App() host access (Application injected via
ScriptRuntime), TryGet<T> and EntityBits for physics facade addressing.
WASDMove maps WASD to the editor camera's horizontal heading (UE
gameplay convention): dynamic bodies go through the physics velocity
facade preserving the vertical component (zeroing Y would cancel
gravity mid-fall), bodiless entities translate their transform directly.
Editor-mode key conflict is naturally avoided: RMB-held flight consumes
WASD, release hands keys to the script.
ScriptComponent stores only a script name string (POD, serializable);
instances are reconciled by ScriptRuntime each frame (PhysicsWorld
pattern): orphan/rename cleanup with OnDestroy, per-name instantiation
from a file-level self-registering registry (ENGINE_REGISTER_SCRIPT),
OnCreate on bind, OnUpdate after physics step on scaled dt. Engine
framerate loop only sees the base interface, so future Lua/C# behaviors
plug in as foreign ScriptBehavior implementations with zero engine
changes. Ships a Spinner sample script and Inspector script picker.
SyncScene only reconciled static/kinematic transforms into Jolt bodies;
dragging a dynamic entity (esp. while paused, where Writeback never runs)
moved the visual but left the body -- and the collider debug wireframe --
at the old pose. Dynamic bodies now get the same drag reconciliation
(teleport + zero velocity + activate), and Writeback syncs
lastPos/lastRot to the body pose so physics motion is never misread as
editor dragging (which would teleport the body back).
Skinned meshes keep vertices in raw mesh space (bakeTransform=false);
rendering transforms them via bind palette (global(bone) x invBind, folds
FBX unit scale). Physics MeshShape/ConvexHull and the mesh bounding sphere
consumed the raw positions directly -> M1911 hull measured ~10x the visual
size. ParseFromDisk now bakes bind-palette-skinned positions per vertex
(same formula as the skinning shader) into ParsedMesh::cpuPositions; Mesh
uses them for m_CpuPositions and bounds. GPU vertex data and rendering are
untouched.
- JointComponent: type, target UUID, per-body local anchors, hinge axis
+ angle limits (data only, zero Jolt deps)
- PhysicsWorld::SyncConstraints: create/remove/rebuild on snapshot
change; body rebuild contagion via invalidated-bits list; target
resolved lazily by UUID each sync (no two-pass load needed)
- Jolt FixedConstraint/HingeConstraint/PointConstraint in
LocalToBodyCOM space (anchors minus shape COM); fixed frames lock
creation-time relative pose; hinge axis aligned across bodies
- Inspector: Joint section with target entity name dropdown
- Serialization: Joint map with UUID target, as<T> fallbacks
EditorCamera (UE FEditorViewportClient semantics) owned by SceneRenderer:
own TransformComponent + reused Camera controller. BuildFrame dispatches
render source by display mode — editor form always renders from the
editor camera (usable in empty scenes), game form renders from the
scene's primary camera entity. AlignTo copies an entity pose into the
editor camera ("Align Viewport to Camera" button in Inspector). F focus
and viewport gizmo now follow the same dispatch. main.cpp stripped of
dangling sandbox camera/spinner state; onSceneLoaded hook removed.
yaml-cpp master hardens zombie nodes: operator=, size(), Type() and
chained operator[] all throw InvalidNode instead of silently passing.
Storing per-component nodes via Node::operator= threw for every key
that was absent, so all 7 entities were skipped on load (0 loaded).
Reads are now level-by-level guarded (copy-init + IsDefined), which
only uses the safe surface (as<T>(fallback), begin(), IsDefined).
Save side switched from node-tree streaming to a manual Emitter:
vector/quat fields emit as single-line flow arrays ([0, 1.2, -6])
and model refs as inline maps, with 9-digit float precision for
lossless float32 round-trips.
Verified offline against the previously saved scene: 7/7 entities
parse with correct component counts; emitter output round-trips.
Engine-layer SceneSerializer (yaml-cpp submodule): entities/components
to versioned YAML, atomic clear-and-rebuild on load with per-entity
fault tolerance, missing-field defaults, unknown-component skip and
UUID dedup. Asset references are {Guid, Path} pairs resolved through
the new AssetManager GUID registry (assets/asset_registry.yaml) with
path fallback; procedural:// pseudo-paths rebuild from parameters.
Editor File menu gets Save/Load Scene modals; app re-resolves stale
entity handles via onSceneLoaded callback.
Key conflict fix (UE semantics): while RMB is held the keyboard belongs
to the camera, so WASD no longer collides with the gizmo W/E/R tool
hotkeys. Editor mode gates fly movement behind RMB; game mode keeps
always-on WASD. Viewport tool hotkeys (W/E/R/T/F) are suppressed while
RMB is held.
Focus orbit (pivot): MMB drag rotates the camera around a pivot keeping
distance; first orbit picks a pivot 5m ahead on the view ray. Shift+MMB
pans in screen plane (camera and pivot move together). Mouse wheel
dollies toward/away from the pivot with exponential steps (no pivot:
move along view ray). F focuses the selected entity (frame-selection):
pivot set to its position, camera slides along the view ray; orbit
distance persists.
Window gains a polled scroll-delta channel (ConsumeScrollDelta) that
respects ImGui mouse capture; Camera::Update mutates orbit state (no
longer const).
User feedback: the self-drawn scale handles clash visually with the
ImGuizmo translate/rotate rods (different size/style, clearly not one
widget). Keep the library widget for all three ops for visual
consistency and prevent mis-drags by masking degenerate axes via
SetAxisMask:
- translate: axis ~ parallel to camera-to-object line (sin < 0.25)
- scale: axis ~ perpendicular to camera-to-object line (|dot| < 0.1,
intersection numerator identically zero -> grab fails/values jump)
OR axis ~ parallel to view direction (|dot| > 0.98, zero screen
displacement -> value never changes)
- rotate: screen-space math, no degeneracy, no mask
Masked axes are neither drawn nor activatable; a slight view change
restores them. Center uniform-scale square (MT_SCALE_XYZ, pure screen
space) stays available in any view.
ImGuizmo scale hover/drag relies on ray-plane intersection with the
axis as plane normal; in front views (axis ⟂ camera-to-object) the
intersection numerator is identically zero, so handles can't be
grabbed and values jump randomly; axes parallel to the view have zero
screen displacement. Submodule is read-only and AxisMask can only hide
degenerate axes, not replace them.
Replace the library scale path entirely with a screen-space gizmo:
- three axis rods (line + end square) + center uniform-scale square,
constant screen size in any view
- axis drag: mouse delta projected onto the axis' screen direction
- axis pointing at camera: radial mapping (drag away = grow)
- center square: uniform scale (library MT_SCALE_XYZ semantics)
- Ctrl snapping quantizes the result; scale clamped >= 0.01
- pick suppressed while hovering/dragging the self-drawn handles
atdunbg