Adds PortalInfo struct and extends CellPhysics with CellBSP (third BSP
for point-in-cell tests, typed CellBSPTree from DatReaderWriter),
Portals (from envCell.CellPortals), PortalPolygons (resolved
cellStruct.Polygons — portals reference visible polys, not
PhysicsPolygons), and VisibleCellIds (populated for future use;
envCell.VisibleCells is List<UInt16>, not Dictionary).
Deletes CellPhysics.LocalAabbMin/Max and PhysicsDataCache.TryFindContainingCell
— Phase D's AABB shortcut is gone. CacheCellStruct's AABB compute
removed; the [cell-cache] diagnostic updated with portal/visible counts
instead.
CacheCellStruct signature gains an EnvCell parameter (one call site in
GameWindow.cs:5384 updated). ResolveOutdoorCellId drops the
TryFindContainingCell call; portal-graph CellTransit replaces it next.
ResolveOutdoorCellIdTests object initializers had the deleted AABB
properties stripped temporarily so the build stays green; the file gets
replaced wholesale in the next commit (CellTransit integration). Those
2 AABB-containment tests continue to fail (they were pre-broken on this
branch); no new failures introduced.
Spec: docs/superpowers/specs/2026-05-19-indoor-portal-cell-tracking-design.md
Plan: docs/superpowers/plans/2026-05-19-indoor-portal-cell-tracking.md
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The original Phase E [cell-cache] probe (fda6af7) only showed the BSP root
node's direct poly count, which was always 0 for non-trivial trees (internal
node root). Extending the probe to:
- Recursively walk the BSP tree and count total leaf polys
- Detect unmatched poly IDs (BSP leaves referencing IDs not in our resolved dict)
- Dump the BSP root bounding sphere (center + radius)
- Dump the cell's local AABB (min/max from poly vertices)
- Dump the cell's world origin (cellTransform * (0,0,0))
The extended data made the route-δ diagnosis definitive: Holtburg cells DO
have full physics polygons in their BSPs (e.g. 0xA9B40143 has 14 polys all
resolved, full Z range 0-2.8 m). The bug is upstream — AABB-based cell
containment is too tight to capture a standing player at most thresholds
between rooms, so the indoor cell-BSP branch fires only intermittently.
Retail uses portal traversal (CObjMaint::HandleObjectEnterCell + cell-side
portal data) which propagates CellId at door crossings. Our AABB-containment
shortcut is partial. This diagnostic stays in place as infrastructure for
the follow-up "Indoor portal-based cell tracking" phase.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Every [indoor-bsp] probe line reports result=OK poly=n/a, meaning
BSPQuery.FindCollisions never records a hit polygon. Four hypotheses:
(a) PhysicsPolygons.Count == 0 for all cached EnvCells (empty data),
(b) BSP leaf Polygons IDs don't match PhysicsPolygons dict keys,
(c) ResolvePolygons filters out all polygons (vertex lookups fail or
degenerate normals), or (d) sphere is too far from BSP leaf bounds.
Format analysis rules out (b): retail BSPLEAF::PackLeaf writes
poly_id (not array index) into the BSP leaf ushort list; CPolygon::Pack
writes poly_id as first field; DatReaderWriter reads it as dictionary
key. ACE DatLoader does the same. Keys are consistent end-to-end.
Add ProbeCellCacheEnabled (ACDREAM_PROBE_CELL_CACHE=1) to
PhysicsDiagnostics and a [cell-cache] log line at the end of
CacheCellStruct. One line per cached EnvCell:
[cell-cache] envCellId=0x... physicsPolyCount=N resolvedCount=M
bspRootPolyCount=K bspRootHasChildren=true|false
physicsPolyCount=0 -> hypothesis (a).
resolvedCount < physicsPolyCount -> hypothesis (c).
Non-zero counts + bspRootPolyCount=0 + bspRootHasChildren=true ->
expected (internal node, leaves hold poly refs); then investigate (d).
Non-zero counts + bspRootPolyCount=0 + bspRootHasChildren=false ->
leaf with empty Polygons list, deeper investigation needed.
Cross-referencing cell-cache lines with indoor-bsp lines (same
envCellId) will pin the root cause in the next launch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ResolveOutdoorCellId only resolved outdoor terrain landcells. A player
geometrically inside an EnvCell stayed in outdoor-landcell range, so
FindEnvCollisions' indoor cell-BSP branch (gated on cellLow >= 0x0100)
never fired. Both #84 (blocked by air indoors) and #85 (pass through
walls outside→in) are downstream of this — without indoor cell-BSP
collision the player gets stuck against outdoor-stab back-faces of the
building shell, and walls only block from one side.
Adds an indoor-cell-containment check via PhysicsDataCache: at
CacheCellStruct time, compute each cell's local AABB from its resolved
polygon vertices; at ResolveOutdoorCellId time, transform the world
position into each cached cell's local space and return the matched
cell's full id when contained. Falls through to the existing outdoor
terrain logic when no EnvCell contains the position.
Also fixes a pre-existing prefix-preservation bug in the outdoor branch:
the function now always applies the matched landblock's high-16 prefix
even when the input fallbackCellId arrived bare-low-byte (the L.2e
finding from CLAUDE.md). Updated two existing PhysicsEngineTests that
encoded the old bare-low-byte output.
Evidence: launch-cluster-a-capture.log @ 2026-05-19 — player at
worldPos (155.376, 14.010, 94.000) geometrically inside cottage cell
0xA9B40172, but sp.CheckCellId stuck at 0x00000031 (outdoor landcell)
across 454 [resolve] lines; zero [indoor-bsp] lines because the gate
never opened.
Closes#84.
Closes#85.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
WorldPicker.Pick previously had no occlusion test — any entity along
the click ray within maxDistance was a candidate, including ones
behind walls. Adds the CellBspRayOccluder static helper that
Möller-Trumbore-tests the click ray against every polygon in every
currently-cached EnvCell BSP, returning the nearest wall-hit `t`.
Both Pick overloads gate candidate selection by that wall-t (legacy
ray-sphere via world-space `t`, screen-rect via camera-space clip.W
depth — matching ScreenProjection.TryProjectSphereToScreenRect's
convention).
PhysicsDataCache exposes a new CellStructIds snapshot accessor so the
caller can iterate without needing the private cache dictionary.
CellPhysics.BSP/PhysicsPolygons/Vertices relaxed from required to
nullable so test fixtures can construct a CellPhysics from Resolved
alone without a real DAT BSP object. GameWindow snapshots the loaded
cell physics on each Pick call and passes the occluder callback.
Closes#86.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the [indoor-bsp] probe + ProbeIndoorBspEnabled toggle for the
Indoor walking Phase 1 BSP-cluster investigation. Mirrors the existing
[resolve] / [cell-transit] / [indoor-*] pattern: one log line per
BSPQuery.FindCollisions call from FindEnvCollisions' cell branch,
capturing cell id, sphere local-pos, result TransitionState, and the
hit poly's normal + side-type via the LastBspHitPoly side-channel
(already wired for ProbeBuildingEnabled, now also fires for the indoor
flag).
Toggle via ACDREAM_PROBE_INDOOR_BSP=1 env var or DebugPanel checkbox.
Zero-cost when off.
Predecessor for the three fix commits that will close ISSUES.md
#84/#85/#86 after the capture session.
Spec: docs/superpowers/specs/2026-05-19-indoor-walking-phase1-bsp-cluster-design.md
Plan: docs/superpowers/plans/2026-05-19-indoor-walking-phase1-bsp-cluster.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 diagnostic chain identified the EXACT cause of 26/123 Holtburg
cells silently failing in WB's PrepareEnvCellMeshData:
ArgumentOutOfRangeException thrown from Setup.Unpack inside
DatReaderWriter when WB calls TryGet<Setup>(stab.Id, ...) on a stab id
whose prefix is GfxObj (0x01xxxxxx), not Setup (0x02xxxxxx).
DatReaderWriter finds the file in Portal's tree (GfxObjs and Setups
share tree-lookups), attempts to parse GfxObj bytes as Setup format,
throws OOR. Exception bubbles to PrepareMeshData's outer try/catch
which silently swallows + returns null. Entire cell fails to upload.
This commit lands the diagnostic infrastructure that surfaced the bug:
- WbMeshAdapter: replaced NullLogger<ObjectMeshManager> with a small
Console-backed ConsoleErrorLogger<T> private class. Filters to
LogLevel.Error+. WB's existing _logger.LogError(ex, ...) at the
swallow site now writes [wb-error] lines with type + message + top 5
stack frames. Bridges WB's intentional log point to acdream's console.
- WbMeshAdapter: extended [indoor-upload] NULL_RESULT probe with
reader-divergence diagnostic (ourCellDb.TryGet, wbResolveId.Count,
wbSelectedType, wbDbIsPortal, wbDbTryGet<EnvCell>, hadRenderData).
Made it possible to rule out cache-hits and reader-divergence as
causes before identifying the real one.
- Cause report at docs/research/2026-05-19-indoor-cell-rendering-cause.md
documents the full chain: 55 ArgumentOutOfRangeException stack traces
captured in one launch, all from PrepareEnvCellMeshData line 1223.
The fix itself (1-line guard at WB's TryGet<Setup> call site) is applied
to references/WorldBuilder/.../ObjectMeshManager.cs — which is a git
submodule. Will be committed separately to the WB submodule after
visual verification.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 Task 1's continuation logged [indoor-upload] NULL_RESULT
when WB's PrepareMeshDataAsync returned null. Extend the line to
include two cross-checks:
ourCellDb.TryGet=<bool> — acdream's DatCollection.Cell.TryGet<EnvCell>
wbResolveId.Count=<int> — WB's DefaultDatReaderWriter.ResolveId().Count
This narrows the cause among WB's null-return paths (ResolveId empty
vs TryGet<EnvCell> failure vs wrong type). Best-effort: both calls
wrapped in try/catch so diagnostic failures don't propagate.
Capture: 55 NULL_RESULTs across multiple landblocks ALL show
ourCellDb.TryGet=True + wbResolveId.Count=1. Both readers find the
cells in their indices, but WB's downstream PrepareMeshData logic
still returns null. Divergence is downstream of ResolveId.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 1 confirmed 26/123 Holtburg cells silently fail in WB's
PrepareEnvCellMeshData / PrepareMeshData. WB's catch block at
ObjectMeshManager.cs:589 calls _logger.LogError(ex, ...) — but we
construct ObjectMeshManager with NullLogger, so the log is dropped.
Capture the Task from PrepareMeshDataAsync (previously fire-and-forget)
and attach a ContinueWith that, for EnvCell ids only when the probe
is on, logs:
[indoor-upload] FAILED cellId=0x... exception=<Type>: <Message>
stack=[<top 3 frames>]
[indoor-upload] NULL_RESULT cellId=0x...
Runs on ThreadPool — non-blocking. Zero cost when ProbeIndoorUploadEnabled
is off. AggregateException is unwrapped to InnerException for readability.
Stack truncated to top 3 frames.
Next: capture procedure, identify cause, target the fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Instruments the per-MeshRef draw loop in WbDrawDispatcher:
- [indoor-lookup]: per cell entity, dumps render-data hit/miss,
IsSetup, parts count, and a partsHit/partsMiss tally over the
SetupParts. Disambiguates hypothesis H2 (WB produces empty
ObjectRenderData with zero parts) and H6 (dispatcher fails to
traverse Setup).
- [indoor-xform]: only fires for the cell's synthetic geometry part
(the SetupPart whose GfxObjId has bit 32 set, per WB's
PrepareEnvCellMeshData cellGeomId convention). Logs the three
composed transform translations: entityWorld, meshRef.PartTransform,
partTransform, and the final composed matrix translation. Disambiguates
hypothesis H5 (transform double-apply — composedT lands at 2 ×
cellOrigin).
Rate-limited via the ShouldEmitIndoorProbe instance helper added in
Task 6 (now consumed — no longer dead code).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Instruments WalkVisibleEntities to identify whether cell entities (first
MeshRef.GfxObjId low-16-bits >= 0x0100) pass all visibility filters or
get culled. Three emission paths:
- [indoor-cull] reason=visibleCellIds-miss -- when the ParentCellId
filter rejects the entity.
- [indoor-cull] reason=frustum -- when AABB frustum cull rejects.
- [indoor-walk] -- when the entity passes all filters and reaches the
draw list.
Rate-limited to once per cellId per ~1 sec (30 frames at 30 Hz) via
IndoorProbeState, a nested class wrapping _lastIndoorProbeFrame dictionary
and _indoorProbeFrameCounter (bumped at top of Draw()). WalkEntitiesInto
accepts a new optional IndoorProbeState? parameter (null = probes off,
default) so the test-friendly WalkEntities overload is unaffected. The
ShouldEmitIndoorProbe instance helper is also retained for Task 7 use.
Disambiguates hypothesis H3 (cull bug -- cell entity dropped before draw).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Instruments WbMeshAdapter at two sites:
- IncrementRefCount: on first call for an EnvCell id (low 16 bits >=
0x0100), tag the id in _pendingEnvCellRequests and log
[indoor-upload] requested.
- Tick: when WB's StagedMeshData drains an ObjectMeshData whose
ObjectId matches a pending EnvCell, log [indoor-upload] completed
with parts count, EnvCellGeometry vertex count, and upload result.
Missing "completed" lines after "requested" identify hypothesis H1
(WB silently returns null from PrepareEnvCellMeshData).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Six checkboxes (ALL master + five individual probes) in the existing
DrawDiagnostics block. Toggling flips the corresponding
RenderingDiagnostics.Probe* flag live via DebugVM forwarding.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live-toggle wrappers for the five indoor-rendering probe flags plus the
ProbeIndoorAll master cascade. Pattern matches existing ProbeResolve /
ProbeCell / ProbeBuilding / ProbeAutoWalk mirrors so a checkbox flip in
the DebugPanel takes effect on the next frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User report: third-person chase camera enters interiors before the
player body does, so the camera-based cameraInsideCell flag was
flipping the scene to indoor lighting prematurely (ambient drops to
0.2 white before the player has actually crossed the doorway).
Retail keys lighting off the PLAYER's cell. CellManager::ChangePosition
@ 0x004559B0 reads CObjCell::seen_outside on the player's current
cell — never on the camera. Match that semantics.
- CellVisibility.IsInsideAnyCell(Vector3): new non-caching brute-force
scan that's safe to call alongside ComputeVisibility(cameraPos)
without thrashing the camera cell cache.
- GameWindow render loop: derive playerInsideCell from the player's
Position when in player mode, otherwise fall back to cameraInsideCell
(orbit/fly debug camera).
- UpdateSunFromSky now takes playerInsideCell. The sky-render and
depth-buffer-clear decisions still use cameraInsideCell — those are
legitimately camera-POV concerns.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Indoor cells rendered "almost black" because the hardcoded ambient at
GameWindow.cs:8342-8345 was an early-2026 guess (0.10, 0.09, 0.08 — half
retail brightness, warm-tinted) rather than the retail value. The named
retail decomp (acclient.pdb, Sept 2013 EoR build) shows
CellManager::ChangePosition @ 0x004559B0 calls
SmartBox::SetWorldAmbientLight(0.2f, 0xFFFFFFFF) whenever the player's
CObjCell::seen_outside flag is 0 — a flat 0.20 white floor, not a
dungeon-tone warm color.
Investigation also confirmed:
- EnvCell.dat does NOT carry inline lights — CEnvCell::UnPack reads
numVisibleCells where Binary Ninja's heuristic decomp inferred
"num_lights". Retail's CObjCell.light_list is populated at runtime via
add_light() calls from neighbouring cell light registrations + per-cell
static-object Setup.Lights, NOT from the dat byte stream.
- Setup.Lights from indoor static objects (entity.SourceGfxObjOrSetupId
prefix 0x02xxxxxx) DO flow through LightInfoLoader.Load (line 5765)
and reach LightManager via LightingHookSink. The wire is intact; the
per-frame Tick + UBO upload chain (line 6865-6867) is intact.
- Retail's particle system does NOT emit lights from particles themselves.
The light comes from the owning Setup's LightInfo records.
Pre-existing failures in DispatcherToMovementIntegrationTests, BSPStepUpTests,
and MotionInterpreterTests are on the branch already and unrelated to this
change (verified by stashing + retesting).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After visual verification 2026-05-18 (turn lag, coast-and-settle,
slope-tilt, jump tracking with contact-plane projection all working),
make the retail chase camera the default. Legacy ChaseCamera stays
available via the DebugPanel toggle (ACDREAM_RETAIL_CHASE=0 or the
checkbox) pending a follow-up deletion commit.
Env var polarity now matches AlignToSlope: default-on if unset, off
only when explicitly "0".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Original symptom: jumping made the camera swing around the player
vertically — the basis tilted up/down with the player's Z velocity.
Root cause: ComputeHeading used the raw 3D velocity vector as the
heading direction. During a jump, velocity has a substantial Z
component (vy ≈ jump speed), and `normalize((vx, vy, vz))` produced
a heading pointing up. The basis tilted accordingly and the camera
went under/over the player.
Retail's actual ALIGN_WITH_PLANE algorithm (decomp at
acclient_2013_pseudo_c.txt:95644-95795) is different:
1. Velocity is only used as a gate. If |vx| AND |vy| > epsilon
(player is moving in XY), proceed; otherwise fall back to the
LOOK_IN_DIRECTION path (player's facing direction unchanged).
2. The base heading is `localtoglobalvec(player, (0, 1, 0))` —
the player's local +Y axis in world space, which in our
convention is `(cos yaw, sin yaw, 0)`.
3. Pick a surface normal:
grounded: contact_plane.N
airborne: (0, 0, 1) [world up]
4. Project the base heading onto the plane perpendicular to that
normal: projected = forward - normal * dot(forward, normal).
5. Normalize. Fall back to the base if projection collapses.
Behaviorally:
* Standing jump (vx≈0, vy≈0): gate fails → base heading. Camera
doesn't move with the jump.
* Running jump (vx, vy, vz all nonzero, airborne): projects onto
world up → no-op since base is already horizontal. Camera basis
stays horizontal; player visibly rises in frame.
* Walking uphill (grounded, slope normal tilted): projection
adds a Z component matching the slope angle. Camera basis tilts
with the terrain.
* Walking on flat ground: projection is a no-op. Camera basis
horizontal.
Surface changes:
* RetailChaseCamera.ComputeHeading gains `isOnGround` and
`contactPlaneNormal` parameters.
* RetailChaseCamera.Update gains the same two parameters and
threads them through.
* GameWindow's two Update call sites pass `result.IsOnGround` and
`_playerController.ContactPlane.Normal` (already exposed on
PlayerMovementController — no plumbing change there).
* Tests: 2 existing heading tests reshaped (Moving* and Uphill);
2 new tests added (AirborneJumping straight-up + running-jump);
1 renamed (SlopeAlignDisabled). Net 25 → 27 tests in
RetailChaseCameraTests; full AcDream.App.Tests: 39 → 41.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
GameWindow now constructs both ChaseCamera + RetailChaseCamera at
player-mode entry, updates both per frame (legacy with isOnGround,
retail with BodyVelocity), and routes mouse/wheel/held-key input to
whichever the CameraDiagnostics flag selects. Mouse-Y goes through
RetailChaseCamera.FilterMouseDelta before AdjustPitch when retail is
active; legacy path is unchanged. Held-key bindings (CameraZoomIn/Out,
CameraRaise/Lower; default-unbound) integrate distance/pitch at
CameraDiagnostics.CameraAdjustmentSpeed per second.
Default behavior: ACDREAM_RETAIL_CHASE unset -> legacy camera as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New CollapsingHeader between Player Info and Performance: toggle +
slope-align checkbox + four sliders (translation stiffness, rotation
stiffness, mouse low-pass window, adjustment speed). All controls
write through DebugVM mirror properties to CameraDiagnostics statics;
changes take effect on the next frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Four new InputAction entries for held-key offset integration
(CameraZoomIn/Out, CameraRaise/Lower; default unbound). Six new
DebugVM mirror properties forwarding to CameraDiagnostics so the
upcoming "Chase camera" DebugPanel section can drive them live.
Also folds in four small cleanups from the Task 4 code review:
- Both CameraDiagnostics-mutating tests in CameraControllerTests now
use try/finally save/restore (consistency with Task-3 follow-up B)
- Drop unused `using System.Numerics` from CameraControllerTests
- Reword the XML doc on CameraController.Active to explain WHY both
cameras are held simultaneously (flag flip takes effect on the
next Active access without re-entry) rather than restating the
getter logic
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EnterChaseMode now takes (ChaseCamera, RetailChaseCamera); Active
consults CameraDiagnostics.UseRetailChaseCamera to pick which to
expose. Flag flip at runtime swaps cameras instantly (both are kept
warm). GameWindow's two EnterChaseMode call sites get a temporary
stub RetailChaseCamera; Task 7 wires proper construction +
per-frame updates.
Also folds in two minor cleanups from the Task 3 code review:
- Update() discards the unused `right` axis from BuildBasis (no
caller in the chase-cam math; viewer_offset.X is always 0)
- The three CameraDiagnostics-mutating integration tests now
save and restore the static state in try/finally to avoid
ordering-dependent contamination
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the per-frame Update(playerPos, yaw, velocity, dt) entrypoint
that composes the math primitives into a renderable View matrix +
PlayerTranslucency. State: 5-frame velocity ring, damped eye + forward
unit vector, first-frame snap flag, mouse-filter shared state.
Public surface: Distance/Pitch/YawOffset/PivotHeight tunables,
AdjustDistance/Pitch (with clamps), FilterMouseDelta entry, View +
Position + PlayerTranslucency outputs. 5 new integration tests, all
pass; total RetailChaseCamera test count 25.
Also folds in two minor cleanups from the Task 2 code review:
- AverageVelocity uses ring.Length instead of hardcoded 5
- Basis_NearVerticalHeading test asserts orthogonality of right & up
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Seven pure-math helpers in the new RetailChaseCamera class:
ComputeHeading (slope-align with flat fallback), BuildBasis (heading
→ orthonormal frame, near-vertical fallback), PushVelocity +
AverageVelocity (5-entry FIFO ring), ComputeDampingAlpha (retail's
stiffness*dt*10), FilterMouseAxis (0.25s low-pass), ComputeTranslucency
(linear ramp 0.20..0.45 m). 20 tests, all pass. State machine + Update()
land in the next commit.
Per spec docs/superpowers/specs/2026-05-18-retail-chase-camera-design.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two related AnimationSequencer fixes for visible animation glitches at
motion-cycle boundaries.
1. Link-tail blend hold (closes#61). BuildBlendedFrame was wrapping
nextIdx unconditionally to rangeLo at the high-frame boundary —
correct for looping cyclic nodes (idle/run/walk loops), wrong for
one-shot links and action overlays. During the ~30 ms fractional
tail before the sequencer transitions to the next queue node, the
blend mixed frame[end] with frame[0], producing a one-frame flash
through the anim's starting pose. Symptoms: door swing-open flap
(frame 0 = closed pose) and player run-stop twitch (frame 0 =
mid-stride). Fix: gate the wrap on curr.IsLooping; non-looping
nodes hold the boundary frame until AdvanceToNextAnimation fires.
2. Stop-anim direction fallback. Stopping from WalkBackward /
SideStepLeft / TurnLeft hit a null linkData from GetLink (the dat
authors a single forward/right stop link and reuses it for both
directions). SetCycle then enqueued only the Ready cycle, snapping
straight to idle with no leg-settle blend. Fix: when the primary
GetLink lookup is null, retry with the substate's low byte remapped
to its forward/right peer (0x06→0x05, 0x10→0x0F, 0x0E→0x0D).
Both fixes are pinned by new regression tests in
AnimationSequencerTests that fail against the prior code (Y=5.02 for
the link tail wrap → frame 0 blend; Y=0 for the backward stop snapping
to Ready cycle).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two related close-range bugs reported in #77 share a root in
PlayerMovementController.DriveServerAutoWalk + BeginServerAutoWalk:
1. **Walk-vs-run misclassification.** BeginServerAutoWalk decided
`_autoWalkInitiallyRunning = (initialDist - distanceToObject) >= 1.0f`,
forcing run at any chase past ~1.6 m. ACE's wire-level walk-vs-run
answer is the MovementParameters CanCharge bit (0x10), which
Creature.SetWalkRunThreshold sets when server-side player→target
distance >= WalkRunThreshold/2 (= 7.5 m default). Retail's
MovementParameters::get_command (decomp 0x0052aa00) gates the run
path on CanCharge first; the inner walk_run_threshold check
practically always walks given ACE's 15 m default. The hardcoded
1.0 m threshold pushed run into the 3-5 m walk-range the user
reported should walk.
2. **Velocity leak in turn-in-place phase.** When the auto-walked body
crossed the destination, desiredYaw flipped ~180°, walkAligned
dropped to false, and the `if (!moveForward) return true;` branch
returned without zeroing body velocity. The body kept the prior
frame's running velocity (RunAnimSpeed × runRate ≈ 11 m/s) and
slid 4-5 m past the target before the turn-around rotation
completed — the "runs and slides away, runs back, picks up"
symptom in #77 bug B.
Changes:
- `CreateObject.ServerMotionState.CanCharge`: new bool prop reading
bit 0x10 of MoveToParameters. Cross-ref ACE
MovementParams.CanCharge = 0x10.
- `PlayerMovementController.BeginServerAutoWalk`: replaces the unused
`walkRunThreshold` parameter with `bool canCharge`; sets
`_autoWalkInitiallyRunning = canCharge`.
- `PlayerMovementController.DriveServerAutoWalk` turn-in-place branch:
calls `_motion.DoMotion(Ready, 1.0)` and zeros body horizontal
velocity (preserving Z for gravity). No-op for case (a) initial-turn
with stationary body; fixes (b) overshoot recovery and (c) settling
cases.
- `GameWindow.OnLiveMotionUpdated`: passes
`update.MotionState.CanCharge` through; [autowalk-begin] trace
shows `canCharge=` instead of `walkRunThresh=`.
- `GameWindow.InstallSpeculativeTurnToTarget`: predicts ACE's
CanCharge from local distance using ACE's exact 7.5 m rule, so the
speculative install agrees with the wire-triggered overwrite that
arrives moments later.
Visual-verified at Holtburg 2026-05-18: walk-range NPC click walks +
fires Use, walk-range F-key pickup walks + no overshoot, far-range
(8-10 m) pickup still runs. Test baseline unchanged (8 Core pre-existing
failures, 0 net-new failures across Core/Net/UI/App suites).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Step 2 of the extraction sequence in docs/architecture/code-structure.md
§4. Lifts DNS resolution + WorldSession instantiation + wireEvents
callback + per-frame Tick + Dispose out of GameWindow.TryStartLiveSession
into a dedicated AcDream.App.Net.LiveSessionController.
What moves:
- DNS resolution (IPAddress.TryParse + Dns.GetHostAddresses fallback,
IPv4 preferred) → LiveSessionController.ResolveEndpoint
- WorldSession instantiation → LiveSessionController.CreateAndWire
- "live: connecting to ..." console line → CreateAndWire
- Try/catch around the setup phase → CreateAndWire (separate from the
Connect/EnterWorld try block that stays in GameWindow)
- Per-frame _liveSession?.Tick() → _liveSessionController?.Tick()
- OnClosing _liveSession?.Dispose() → _liveSessionController?.Dispose()
What stays in GameWindow:
- The 25+ event subscriptions (extracted into a new private
WireLiveSessionEvents method that the controller invokes via callback)
- The Connect → CharacterList → EnterWorld → post-EnterWorld setup dance
(touches Chat, _playerServerGuid, _vitalsVm, _worldState, _settingsStore,
_settingsVm, _playerModeAutoEntry; moving these would balloon Step 2's
scope and risk surface)
- All 60+ outbound _liveSession.Send* call sites (touch the field by
name; LiveSessionController.Session is the controller-side mirror)
The _liveSession field remains as a convenience handle synced with
_liveSessionController.Session; it tracks the same WorldSession instance.
Behavior preservation:
- Same DNS-resolution sequence, same "live: connecting to ..." line,
same wiring-vs-Connect ordering as pre-refactor.
- Same 25+ event subscriptions in the same order, byte-for-byte.
- Same Connect/EnterWorld error handling (the catch block stays in
GameWindow because it disposes _combatChatTranslator which is also
a GameWindow field).
Closes#76.
One subtle nullable-flow fix the compiler required: the chat-bus
lambda's `var liveSession = _liveSession;` capture became
`var liveSession = session;` (the non-null parameter) so the compiler
can prove non-null inside the lambda body. Both pointed to the same
WorldSession instance; only the static analysis changed.
Verification:
- dotnet build green
- dotnet test: AcDream.App.Tests 10/10, Core.Net.Tests 294/294,
UI.Abstractions.Tests 419/419 — all green. Core.Tests 1073/1081
(same 8 pre-existing physics failures as baseline; unrelated).
- Live ACE session against +Acdream verified end-to-end:
* Connection + handshake + EnterWorld
* Door double-click → OnLiveMotionUpdated round-trip
(cmd=0x000B open / cmd=0x000C close)
* NPC double-click → outbound Use
* Ground item F-key pickup × 4 successful (Amaranth, Comfrey,
Damiana, Dragonsblood)
* Spawn stream + chat channels + remote-entity motion all healthy
* Clean OnClosing → controller.Dispose
Walking-range auto-walk + pickup-overshoot bugs observed during
verification are pre-existing (filed as #77); they live in
PlayerMovementController.DriveServerAutoWalk / threshold logic
which this refactor did not touch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First application of CLAUDE.md's new Code Structure Rules §5
("Runtime probes belong in diagnostic owner classes"). Migrates the
four ACDREAM_DUMP_STEEP_ROOF call-site env reads into a single
PhysicsDiagnostics.DumpSteepRoofEnabled property initialized from the
env var at type init, with a runtime setter that follows the existing
ProbeResolveEnabled / ProbeCellEnabled / ProbeBuildingEnabled pattern.
Sites migrated:
- AcDream.Core/Physics/PhysicsEngine.cs:637 (KILL-VELOCITY-APPLIED log)
- AcDream.Core/Physics/TransitionTypes.cs:718 (PHASE3-RESET log)
- AcDream.App/Input/PlayerMovementController.cs:1117 (FRAME log)
- AcDream.App/Input/PlayerMovementController.cs:1199 (per-frame bounce log)
Behavior-preservation only. ACDREAM_DUMP_STEEP_ROOF=1 still produces
identical [steep-roof] log output. The class-comment in
PhysicsDiagnostics already anticipated this migration
("Future slices may fold the older ACDREAM_DUMP_* env vars into this
class for unified runtime toggling").
Not yet wired to a DebugVM checkbox — runtime toggling is available
via the property setter for future debugging sessions, but exposing
it on the panel is a 30-second future cut, not in scope here.
Build: green.
Tests: same pass/fail profile as before this commit (8 pre-existing
Core failures unrelated to physics-diagnostics; App.Tests / Core.Net.Tests
/ UI.Abstractions.Tests all green).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lifts 13 startup-time environment variables out of GameWindow.cs into a
single typed AcDream.App.RuntimeOptions record read once in Program.cs.
Behavior-preservation only — no live behavior change, no visual change.
Verified end-to-end against ACE on 127.0.0.1:9000: full M1 demo loop
(walk Holtburg, click door, click NPC, portal entry) plus DEVTOOLS
ImGui panels load cleanly.
Why: GameWindow.cs is 10,304 LOC and scattered Environment.GetEnvironmentVariable
calls were one of the structural smells called out in the new
"Code Structure Rules" doc. Typed options is the safest cut to make
first because the substitution is mechanical and parsing semantics
get pinned by unit tests.
What lands:
- CLAUDE.md: removed stale R1→R8 execution-phases line, replaced with
pointers to the milestones doc + strategic roadmap (the actual
source of truth). Tightened the "check ALL FOUR references"
section to describe WB as the production rendering base, not
just a reference. New "Code Structure Rules" section (6 rules)
captures the discipline we're committing to.
- docs/architecture/acdream-architecture.md: removed dangling link
to the deleted memory/project_ui_architecture.md.
- docs/architecture/code-structure.md (NEW, 376 LOC): rationale for
the 6 rules + 6-step extraction sequence
(RuntimeOptions → LiveSessionController → LiveEntityRuntime →
SelectionInteractionController → RenderFrameOrchestrator →
GameEntity aggregation). This PR is Step 1.
- src/AcDream.App/RuntimeOptions.cs (NEW, 100 LOC): typed record
with FromEnvironment(string) factory and Parse(datDir, env)
overload for testability. Covers ACDREAM_LIVE, _TEST_HOST/PORT/
USER/PASS, _DEVTOOLS, _DUMP_MOVE_TRUTH, _NO_AUDIO,
_ENABLE_SKY_PES, _HIDE_PART, _RETAIL_CLOSE_DEGRADES,
_DUMP_SCENERY_Z, _STREAM_RADIUS.
- src/AcDream.App/Program.cs: builds RuntimeOptions once, passes
to GameWindow.
- src/AcDream.App/Rendering/GameWindow.cs: ctor takes RuntimeOptions;
7 startup-cached env-var fields become expression-bodied
properties or direct _options.X reads; TryStartLiveSession,
audio init, legacy stream-radius branch all route through
_options.
- tests/AcDream.App.Tests/ (NEW project, 10 unit tests + csproj):
pins parser semantics — default-off bools, the literal "0"
gate for RETAIL_CLOSE_DEGRADES, the >=0 guard for
STREAM_RADIUS, null-vs-empty for user/pass, exact-"1" check
for diagnostic flags. Registered in AcDream.slnx.
Out of scope (per code-structure.md §4):
- Per-call-site ACDREAM_DUMP_* / _REMOTE_VEL_DIAG diagnostic reads
sprinkled through GameWindow (~40 sites). Rule 5 in CLAUDE.md
commits us to migrating these opportunistically as larger
extractions land, not in a bulk pass.
- AcDream.Core's project-reference to Chorizite.OpenGLSDLBackend.
Only the stateless .Lib namespace is used; tightening the project
reference is documented as future work in code-structure.md §2.
Build: green.
Tests: AcDream.App.Tests 10/10 ✓, Core.Net.Tests 294/294 ✓,
UI.Abstractions.Tests 419/419 ✓,
AcDream.Core.Tests 1073/1081 (8 pre-existing failures verified
against pre-refactor baseline by stash-and-rerun).
Visual verification: full M1 demo loop against ACE +Acdream login
including DEVTOOLS panel host load.
Next: Step 2 — extract LiveSessionController per code-structure.md §4.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes#63, #69, #74, #75. Replaces the chain of Commit-B workarounds
that compensated for ACE's MoveToChain getting cancelled by a leaked
user-MoveToState packet during inbound auto-walk. The fix is
architectural — auto-walk drives the body directly from the
server-supplied path data, no player-input synthesis, no spurious
wire-packet transitions, no grace-period band-aid.
Architectural change (closes#75):
PlayerMovementController.ApplyAutoWalkOverlay → DriveServerAutoWalk.
- Steps Yaw toward target at retail-faithful turn rates.
- Computes desired forward velocity from path runRate.
- Calls _motion.DoMotion(WalkForward, speed) directly for the
motion-interpreter state (drives animation cycle).
- Sets _body.set_local_velocity directly when grounded.
- Returns true to gate the user-input motion + velocity section
in Update so user-input flow doesn't overwrite auto-walk
velocity or motion state.
Mirrors retail's MovementManager::PerformMovement case 6 (decomp
0x00524440) which never touches the user-input pipeline during
server-controlled auto-walk.
Wire-layer guard at GameWindow.cs:6419 retained as a SEMANTIC
statement (`if (result.MotionStateChanged && !IsServerAutoWalking)`):
user-MoveToState packets are for user-driven motion intent. During
server-controlled auto-walk, the motion-state transitions caused by
the animation override (RunForward / WalkForward / TurnLeft /
TurnRight cycles) must not leak as user-cancellation packets. This
is NOT the deleted 500ms grace-period band-aid; it's the wire-layer
expressing the user-vs-server motion split.
Animation plumbed for auto-walk phases (closes#69):
- Moving forward → WalkForward (speed=1.0) / RunForward (speed=runRate)
- Turn-first phase → TurnLeft / TurnRight (sign of yawStep)
- Aligned-but-pre-step / arrival → no override (idle)
Driven via _autoWalkMovingForwardThisFrame + _autoWalkTurnDirectionThisFrame
fields set in DriveServerAutoWalk and read in the MovementResult
construction at the bottom of Update. UpdatePlayerAnimation picks up
the localAnimCmd as the highest-priority animation source.
Walk/run threshold = 1.0m, retail-observed. ACE's wire-default of
15.0f is too generous; ACE's own physics layer uses 1.0f at
MovementParameters.cs:50 (with the 15.0f line commented out) and
Creature.cs:312 notes "default 15 distance seems too far". The
formula matches retail's MovementParameters::get_command at decomp
0x0052aa00: running = (initialDist - distance_to_object) >=
threshold, evaluated ONCE at chain start and held for the rest of
the auto-walk (matches retail "runs all the way / walks all the way"
behaviour). Wire-supplied threshold is ignored.
Pickup gate (IsPickupableTarget) now uses BF_STUCK
(acclient.h:6435, bit 0x4) to discriminate immovable scenery from
real pickup items that share a Misc ItemType. Sign (pwd=0x14 with
BF_STUCK) → blocked; spell component (pwd=0x10, no BF_STUCK) →
allowed. ACE's PutItemInContainer (Player_Inventory.cs:831-836)
responds with WeenieError.Stuck (0x29) on stuck items so the gate
prevents wasted wire packets + a UX dead-end.
R-key dispatch by target type. UseCurrentSelection's top-level
IsUseableTarget gate was wrong (blocked USEABLE_NO=1 items that
ARE pickupable). Reordered:
1. Creature → SendUse
2. Pickupable → SendPickUp
3. Useable → SendUse
4. Otherwise → "cannot be used" toast
Each handler keeps its own gate. Matches retail's per-action
server-side validation.
AP cadence revert (closes#74). With the MoveToChain race fixed,
the per-frame "send while moving" cadence is no longer load-bearing.
Reverted to retail's two-branch ShouldSendPositionEvent gate
(acclient_2013_pseudo_c.txt:700233-700285):
Interval NOT elapsed (< 1 sec): send if cell or contact-plane changed.
Interval elapsed (>= 1 sec): send if cell or position frame changed.
Adds _lastSentContactPlane field + ApproxPlaneEqual helper +
PlayerMovementController.ContactPlane public accessor. Extended
NotePositionSent(Vector3, uint, Plane, float) — both outbound sites
(MoveToState + AP) pass _playerController.ContactPlane.
Effective rates: 0 Hz idle, ~1 Hz smooth motion, per-event on
cell/plane changes, 0 Hz airborne.
CLAUDE.md updated with no-workarounds rule (commit `da126f9` on
the worktree branch). Saved as feedback memory at
memory/feedback_no_workarounds.md.
Tests: build green; Core.Net 294/294; Core 1073/1081 (baseline,
8 pre-existing Physics failures unchanged). Visual-verified
end-to-end on 2026-05-16 for far/near Use + PickUp on NPCs,
doors, items, spell components, signs (correctly blocked), corpses,
turn-first animation, run/walk thresholds, idle quiet, smooth-
motion 1Hz.
Spec: docs/superpowers/specs/2026-05-16-phase-b6-suppress-movetostate-during-inbound-autowalk-design.md
Plan: docs/superpowers/plans/2026-05-16-phase-b6-suppress-movetostate-during-inbound-autowalk.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Retires divergences flagged in the 2026-05-16 faithfulness audit:
1. AP cadence. Replaces the 1 Hz idle / 10 Hz active flat heartbeat
with a diff-driven model gated on `Contact && OnWalkable`
(acclient_2013_pseudo_c.txt:700327 SendPositionEvent). Sends on
position or cell change while grounded on walkable, plus a 1 sec
heartbeat; suppressed entirely airborne. PlayerMovementController
exposes `NotePositionSent(pos, cellId, now)` which GameWindow stamps
after each AutonomousPosition / MoveToState send — mirrors retail's
shared `last_sent_position_time` between SendPositionEvent
(0x006b4770) and SendMovementEvent (0x006b4680). Known divergence
from retail: ours is per-frame-while-moving, retail's effective rate
is ~1 Hz during smooth motion (cell/plane checks). Filed as #74,
blocked by #63 — when #63 lands we revert to retail's narrower gate.
2. Workaround retirement. Removes TinyMargin (0.05 m inside arrival)
and the AP-flush before re-send (`SendAutonomousPositionNow`). The
diff-driven cadence makes both obsolete. Close-range turn-first
deferred Use is kept (it IS retail — ACE Player_Move.cs:66-87
mirrors retail's CreateMoveToChain pre-callback rotation), renamed
`OnAutoWalkArrivedSendDeferredAction` to clarify it's a FIRST send.
`isRetryAfterArrival` parameter dropped.
3. Far-range Use/PickUp retry. Restored — was load-bearing, not the
"redundant cleanup" the Group 2 audit thought. Issue #63 means ACE
drops the first Use as too-far without re-polling on subsequent APs;
the arrival re-send is what makes far-range Use complete. Logs
include `(queued for arrival re-send pending #63)` to make this
explicit. Removes when #63 closes.
4. Screen-rect picker. New `AcDream.Core.Selection.ScreenProjection`
helper shared by `WorldPicker` and `TargetIndicatorPanel`. The
`Setup.SelectionSphere` projects to a screen-space square (retail
anchor `SmartBox::GetObjectBoundingBox` 0x00452e20); picker
hit-tests the mouse pixel against the same rect the indicator draws,
inflated by 8 px (`TriangleSize`). Guarantees what-you-see is
what-you-click — including rect corners that were dead zones under
the old ray-sphere picker. Per-type radius (1.0/1.6/2.0 m) and
vertical-offset (0.2/0.9/1.0/1.5 m) heuristic lambdas retired;
`IsTallSceneryGuid` deleted; `EntityHeightFor` trimmed to 1.5 m × scale
defensive default. No defensive sphere synth — entities without a
baked `SelectionSphere` are skipped, matching retail's
`GfxObjUnderSelectionRay` (0x0054c740).
5. Rotation rate run multiplier (Commit A precursor). `TurnRateFor(running)`
helper applies retail's `run_turn_factor = 1.5f` (PDB-named
0x007c8914) under HoldKey.Run, matching `apply_run_to_command` at
0x00527be0 (line 305098). Effective: walking ≈ 90°/s, running ≈ 135°/s.
Keyboard A/D + ApplyAutoWalkOverlay both use it.
6. Useability gate (Commit A precursor). `IsUseableTarget` corrected to
`useability != 0` per `ItemUses::IsUseable` at 256455 — ANY non-zero
passes (USEABLE_NO=1, USEABLE_CONTAINED=8, etc.), not just the
USEABLE_REMOTE bit. Cross-checked against 4 call sites in retail
(ItemHolder::UseObject 0x00588a80, DetermineUseResult 0x402697,
UsingItem 0x367638, disable-button-state 0x198826). Added
`ProbeUseabilityFallbackEnabled` diagnostic
(`ACDREAM_PROBE_USEABILITY_FALLBACK=1`) to measure how often the
creature/BF_DOOR fallback fires for ACE-seed-DB entities with
null useability.
CLAUDE.md updated with the graceful-shutdown rule for relaunch:
Stop-Process bypasses the logout packet, leaving ACE's session marked
logged-in for ~3+ min. CloseMainWindow() sends WM_CLOSE so the
shutdown hook runs and the logout packet reaches ACE.
Tests: +3 ScreenProjectionTests + 6 WorldPickerRectOverloadTests = +9.
Core.Net 294/294 pass; Core 1073/1081 (8 pre-existing Physics failures
unchanged). Visual-verified 2026-05-16: rotation rate, useability,
screen-rect click area, double-click + R-key + F-key Use/PickUp at
short and long range — dialogue/door/pickup fire on arrival.
Filed follow-ups #70 (triangle apex/size DAT sprite), #71 (picker
Stage B polygon refine), #72 (cdb omega.z probe), #73 (retail-message
sweep pattern), #74 (per-frame AP chattier than retail — blocked by
#63). Old ray-sphere `WorldPicker.Pick(origin, direction, ...)`
overload kept for back-compat; no callers in acdream proper.
Plan: docs/superpowers/plans/2026-05-16-retail-faithfulness-fixes.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two retail divergences fixed from the 2026-05-16 faithfulness audit
(Commit A of the plan at docs/superpowers/plans/2026-05-16-retail-faithfulness-fixes.md).
1. Rotation rate ignored HoldKey.Run. Retail's CMotionInterp::
apply_run_to_command (decomp 0x00527be0 line 305098) multiplies
turn_speed by run_turn_factor (1.5, PDB-named symbol at 0x007c8914)
when input is TurnRight/TurnLeft under HoldKey.Run. Effective
running rotation is 50% faster (~135°/s vs walking ~90°/s).
Our keyboard A/D and ApplyAutoWalkOverlay used a fixed walking
rate.
New: RemoteMoveToDriver.TurnRateFor(running) helper. Keyboard
path passes input.Run; auto-walk overlay passes
_autoWalkInitiallyRunning. The walking-rate base
(BaseTurnRateRadPerSec = π/2) is unchanged; TurnRateRadPerSec
constant is preserved as the walking-rate alias for callers
that don't have run/walk state (NPC remotes).
2. IsUseableTarget gated on `useability & USEABLE_REMOTE (0x20)`,
which was stricter than retail. Per ItemUses::IsUseable
(acclient_2013_pseudo_c.txt:256455) cross-referenced with 4
call sites, retail's IsUseable() semantic is `_useability != 0`.
But visually retail's USEABLE_NO (1) entities don't approach
either, because ACE never broadcasts MovementType=6 for them.
Our client installs a speculative auto-walk BEFORE the server
responds, so we'd visibly approach + face signs before the
wire packet was rejected.
Pragmatic fix: block USEABLE_UNDEF (0) AND USEABLE_NO (1) in
IsUseableTarget — slightly stricter than retail's
IsUseable but matches retail's user-visible behaviour
("R on sign does nothing"). Documented in the doc-comment so
a future implementer knows the gap.
3. New IsPickupableTarget gate for F-key path — requires
USEABLE_REMOTE (0x20) bit. Null-useability fallback for
BF_CORPSE + small-item ItemTypes (preserves M1 ground-item
pickup flow when ACE seed DB doesn't publish useability).
4. R-key (UseCurrentSelection) upfront gate now ALWAYS uses
IsUseableTarget. R is conceptually "use" with smart-routing
to pickup as a downstream optimization. F-key (SendPickUp)
uses IsPickupableTarget directly.
5. Retail toast strings on block, centralised in new
src/AcDream.Core/Ui/RetailMessages.cs:
- "The X cannot be used" (data 0x007e2a70, sprintf 0x00588ea4)
fires on UseCurrentSelection / SendUse gate block.
- "The X can't be picked up!" (sprintf 0x00587353) fires on
SendPickUp non-pickupable block.
- "You cannot pick up creatures!" (data 0x007e22b4) fires on
SendPickUp creature block (was previously silent).
- Plus 4 inactive retail strings ready for future call sites:
CannotBeUsedWith (two-target Use), CannotBePickedUp (formal
pickup variant), CannotBeUsedWhileOnHook_HooksOff +
CannotBeUsedWhileOnHook_NotOwner (housing). All cite their
retail data addresses + runtime sprintf addresses.
6. ProbeUseabilityFallbackEnabled diagnostic (env var
ACDREAM_PROBE_USEABILITY_FALLBACK=1) logs every time the
null-useability fallback fires. Settles whether the
fallback for creature + BF_DOOR/LIFESTONE/PORTAL/CORPSE
entries in ACE's seed DB without useability is hot code
or theoretical defense.
Test coverage:
- +3 RemoteMoveToDriverTests cover TurnRateFor walking/running/back-compat.
- +7 RetailMessagesTests cover each retail string with retail anchor.
- +1 CreateObjectTests TryParse_WeenieFlagsUsable_ReadsUseableNoValue
pins parser correctness for USEABLE_NO=1.
- 294/294 Core.Net pass; 24/24 new+touched Core tests pass.
- Pre-existing baseline of 8 Physics test failures unchanged
(BSPStepUp + MotionInterpreter regression noise from prior
sessions; out of scope here).
Deferred to a separate session per user direction:
- Click area = indicator-rect retail fidelity. Retail's picker
uses per-part CGfxObj.drawing_sphere + polygon refine
(0x0054c740); ours uses single Setup.SelectionSphere ray-
intersect. The rect corners are dead zones today. Three fix
options analyzed: screen-space rectangle hit-test, sqrt(2)
sphere inflation, polygon refine Stage B.
Plan: docs/superpowers/plans/2026-05-16-retail-faithfulness-fixes.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the mesh-AABB approximation with retail's actual selection
mechanism. The user observed the indicator was too small and didn't
scale with the object the way retail does — root cause was using the
wrong source data.
Retail trace (decomp anchors in named-retail/acclient_2013_pseudo_c.txt):
- VividTargetIndicator::Draw at 0x004f6c30 is registered as the
SmartBox targetting callback (0x004f6df6).
- SmartBox::DoTargettingChecks at 0x00453bb4 calls
SmartBox::GetObjectBoundingBox (0x00452e20) to compute the rect.
- GetObjectBoundingBox uses CPhysicsObj::GetSelectionSphere
(0x0050ea40) → CPartArray::GetSelectionSphere (0x00518b80) which
reads setup->selection_sphere from the DAT, applies part-array
scale (component-wise on center, Z-scale on radius), then calls
Render::GetViewerBBox (0x0054b400) to project the sphere as a
screen-space camera-aligned BBox.
- VividTargetIndicator::OnDraw at 0x004f62b0 inflates that rect by
one triangle width/height on every side before drawing (eax_21 /
eax_23 in 0x004f6a0b–0x004f6a99), so the corner triangles sit
outside the projected sphere with a small gap.
Implementation:
- GameWindow.TryGetEntitySelectionSphere reads setup.SelectionSphere
from the DAT (Setup type already exposes Origin + Radius),
applies entity scale, rotates center via entity orientation, and
produces a world-space sphere.
- TargetIndicatorPanel.TryComputeScreenRectFromSphere projects the
sphere center via the view-projection matrix and computes
screenRadius = worldRadius * projection.M22 * viewport.Y /
(2 * clip.W). M22 = cot(fovY/2) for a standard right-handed
perspective. Mathematically equivalent to retail's
Render::GetViewerBBox followed by 2-corner xformPointInternal,
faster (no double projection).
- TargetInfo carries WorldSphereCenter + WorldSphereRadius (replaces
the previous WorldAabbMin/Max). Fallback to per-type height
heuristic still in place if Setup has no baked selection_sphere
(rare; Radius <= 1e-4f short-circuits).
- Inflate by TriangleSize on every side matches retail's eax_21 +
eax_23 offsets exactly.
- Triangle right-angle apex flipped to point INWARD toward the
target (per user feedback) — apex at corner + (±t, ±t),
hypotenuse along the outer diagonal of the corner.
- TriangleSize 10 → 14 → 8 (retail sprite is small).
Also fixes a parser bug in CreateObject.cs introduced in 58e1556:
BF_INCLUDES_SECOND_HEADER is 0x04000000 per acclient.h:6458 (ACE
ObjectDescriptionFlag.IncludesSecondHeader matches), NOT 0x80000000.
The wrong bit meant the weenieFlags2 4-byte skip never fired for
entities that had the bit set, potentially shifting Useability /
UseRadius reads by 4 bytes. Now correct.
Visual verification (2026-05-16):
- Holtburg town sign — indicator traces the visible sign + pole at
the right size (matches retail screenshot proportions).
- Sign R-key still silent no-op (B.8 useability gate intact).
- NPCs / doors / items still get correctly-sized indicators.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two retail divergences fixed end-to-end:
1. R-key Use on non-useable entities (signs, banners, decorative
scenery) was silently sending Use/PickUp to ACE, triggering
auto-walk + NPC-style chat fallback. Retail's client checks
ITEM_USEABLE (acclient.h:6478) and silently ignores Use when
the USEABLE_REMOTE (0x20) bit isn't set. Now ports that gate.
2. Holtburg town sign indicator + click sphere only covered the
base of the pole because the "everything else" default in
EntityHeightFor was 1.5 m and the picker's vertical offset
for default class was 0.2 m. A 3 m sign on a pole was almost
entirely outside both shapes.
Wire change:
- CreateObject parser now walks the WeenieHeader optional tail
(per ACE WorldObject_Networking.cs:87-114) up through Useability
+ UseRadius. Captures weenieFlags upfront, then conditionally
skips PluralName, ItemCapacity, ContainerCapacity, AmmoType,
Value before reading Useability (u32) and UseRadius (f32).
- CreateObject.Parsed + WorldSession.EntitySpawn record append two
new optional fields (Useability uint?, UseRadius float?), both
defaulting to null. Existing call sites unchanged.
- 3 new tests cover: no weenieFlags → null, weenieFlags=0x10 alone
→ useability read, weenieFlags=0x8|0x10|0x20 → walker skips Value
then reads Useability + UseRadius in correct order.
Behaviour change:
- GameWindow.IsUseableTarget(guid) — authoritative path uses spawn
.Useability when present (REMOTE bit gate); fallback when null
permits Use on creatures + BF_DOOR/LIFESTONE/PORTAL/CORPSE for
M1 flow continuity.
- UseCurrentSelection (R-key dispatcher) and SendUse + SendPickUp
(double-click + F-key direct paths) gate on IsUseableTarget,
silent early-return matching retail. isRetryAfterArrival skips
the gate (re-fires only previously-gated actions).
- TargetIndicatorPanel.EntityHeightFor default branch 1.5 m → 3 m
for non-creature non-flat non-small-item entities (sign-class).
Scale > 1 still grows proportionally.
- WorldPicker callbacks: new IsTallSceneryGuid branch lifts sphere
centre to 1.5 m with 1.6 m radius for sign-class entities,
mirroring the indicator's 3 m default so click sphere matches
the visible box.
Tests: 293/293 pass in AcDream.Core.Net.Tests (+3 new walker
tests). dotnet build clean.
Retail anchors:
- acclient.h:6478 — ITEM_USEABLE enum (USEABLE_REMOTE = 0x20)
- acclient.h:6431-6463 — PWD bitfield (BF_DOOR etc.)
- ACE WorldObject_Networking.cs:87-114 — wire field order
- ACE WeenieHeaderFlag — Usable = 0x10, UseRadius = 0x20
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User report: 'No we only walk, not running from the correct threshold.
regression?'
Cause: InstallSpeculativeTurnToTarget passed walkRunThreshold=9999,
which made BeginServerAutoWalk evaluate the initial-distance run/walk
decision as walk-mode (no distance > 9999). ACE's MovementType=6
arrives ~100 ms later with the real wire threshold (15.0) and
overwrites, but the body had already started walking by then; for
far targets near the 15 m boundary, the speculative walking shortened
the distance enough that ACE's overwrite re-evaluated to walk also.
Fix: pass 15.0 in the speculative install — matches ACE's default
MoveToParameters.SetDefaults() for non-combat Use/PickUp.
Effect: a >15 m target now correctly enters run-mode at the
speculative install, and the ACE overwrite preserves that decision.
The body runs all the way, stopping at the target as before.
User report: 'quick snap at like 30 degrees to the last position. Not
a smooth turn. Did you verify with retail?'
Verified against retail decomp at MoveToManager::HandleTurnToHeading
(0x0052a0c0). Retail's pattern:
- Body rotates continuously via a TurnLeft/TurnRight motion cycle.
- The ONLY snap is set_heading(target, 1) after heading_greater()
detects we've passed the target (overshoot protection).
- No 'snap when close to target' tolerance band — that's purely
a sparse-update fudge in RemoteMoveToDriver (the remote-creature
path with ~1Hz UpdatePosition broadcasts).
I'd copied the snap-on-approach tolerance from RemoteMoveToDriver to
ApplyAutoWalkOverlay. Wrong: local player rotates at per-tick
resolution, no sparse-update problem to compensate for. Removed.
The MathF.Min(|delta|, maxStep) clamp naturally lands the body on
the target heading without overshoot in the final partial tick, so
no separate snap-on-overshoot branch is needed for our integrator
either.
Visible effect: 1.8m humanoid rotating ~180° at π/2 rad/s takes ~2 s
of smooth turn now, instead of ~1.3 s of turn + instant 20° snap at
the end.
User report: 'You should be face to face with the NPC before sending
use. So first is rotation, when you are facing, then using.' and
'it does not face it completely.'
Two changes:
1. Split alignment thresholds in ApplyAutoWalkOverlay:
walkAligned (30°) — gate for synthesised Forward+Run motion
during far-range approach; body walks
while finishing residual turn within 30°.
aligned (5°) — gate for arrival-fire. Final facing
before the auto-walk ends and the action
re-sends. Matches retail's tight pre-Use
rotation tolerance.
Within-arrival check still requires alignment; without alignment
the body holds in turn-only mode regardless of distance.
2. Defer wire Use/PickUp packet for CLOSE-range targets. SendUse
and SendPickUp now check IsCloseRangeTarget(guid): if the player
is already within the target's use-radius, we install the
speculative overlay, set _pendingPostArrivalAction, and RETURN
without sending the wire packet. AutoWalkArrived fires after the
local rotation completes (alignment within 5°); the existing
re-send handler then fires SendUse with isRetryAfterArrival=true,
sending the wire packet at that moment. Effect: rotate first,
THEN Use — the NPC/door/item only sees the action after the
character has turned to face it.
Far-range path unchanged: send immediately, ACE auto-walks,
arrival re-sends.
Filed #69: turn animation (leg/arm cycle while pivoting). The body
now rotates but doesn't play the TurnLeft/TurnRight cycle the user
wants to see. Separate scope — needs motion-interpreter integration.
User report: 'It should always face the NPC. When I'm close I'm not
facing. But now it turns if I'm far away.'
Cause: ACE skips MoveToChain when the player is already within
WithinUseRadius (Player_Move.cs:66 shortcut) — it rotates the body
server-side via Rotate(target) but doesn't broadcast a MovementType=6
to us, so our auto-walk overlay never installs. The local body never
turns; the player remains facing wherever the camera/mouse last left
them.
Fix has two pieces:
1. PlayerMovementController.ApplyAutoWalkOverlay: arrival is now
gated on BOTH within-radius AND aligned. Previously a body that
started already in-range ended the overlay before turning. Now
it turns in place, then ends once facing.
Also: forward motion stays suppressed while withinArrival (we
just need to finish the turn, no point stepping forward into a
target we're already touching).
2. GameWindow.SendUse / SendPickUp: install a speculative auto-walk
overlay at send time via new InstallSpeculativeTurnToTarget
helper. For far targets ACE's MovementType=6 arrives moments
later and overwrites with its wire-supplied use-radius. For
close targets our overlay is the only thing that runs — body
turns, then ends.
The per-type use-radius mirrors the picker's heuristic (3 m
creature / 2 m large flat / 0.6 m item).
10 Hz heartbeat (301281d) made ACE see us in-radius before its
MoveToChain timeout; user confirmed doors work now. Closing #67 — root
cause was 1 Hz position outbound on our side, not anything door-
specific. Same fix unblocked door + NPC.
Two visible refinements:
1. Turn-first gate. User report: 'when I use from far range, I should
face that object and then start moving. Now it starts running
before facing is complete.'
ApplyAutoWalkOverlay now suppresses Forward motion when the
heading delta to the target is > 30°. Body turns IN PLACE first,
then walks forward once roughly aligned. Within the 30° band the
body walks while finishing the residual turn. Matches the user-
observed retail rhythm.
2. Arrival margin shrunk 0.2 m → 0.05 m. User report: 'NPC dialogue
fires, but still a bit too close. In retail it fires from a longer
range.' With the 10 Hz heartbeat the server-side Player.Location
tracks us within ~100 ms, so the bigger safety margin is no longer
needed — only a tiny epsilon to absorb the sub-tick race between
local arrival fire and the next outbound packet.
Filed #68: remote players' running animation doesn't transition to
Ready on auto-walk arrival when observed from acdream. Separate
visual bug — server-side action completes correctly; just the cycle
on the dead-reckoned remote body doesn't flip back to idle.
User correctly called out: 'why workarounds? Nothing wrong with ACE,
our client is wrong.' Three of the four workarounds in B.6/B.7
(arrival margin, re-send-on-arrival, AP flush on arrival) exist
because our client sends 1Hz position heartbeat. Retail sent every
tick. ACE's CreateMoveToChain polls WithinUseRadius every ~0.1 s
using the latest Player.Location — at 1Hz we leave up to 1 s of
stale position data on the server, so ACE rejects re-sent actions
as still-out-of-range.
Fix: bump heartbeat to ~10 Hz when the body is actively moving
(auto-walk OR user pressing W/A/S/D). Idle still 1 Hz.
ACE sees us approach in near-real-time; server-side MoveToChain
converges normally; CreateMoveToChain's own callback fires the
action when in radius — no client-side re-send needed.
This SHOULD make the existing workarounds redundant:
* Arrival margin (0.2m) — can shrink toward 0 since position
drift is bounded by 100ms instead of 1s
* Re-send on arrival — ACE's chain completes on its own
* AP flush on arrival — included in normal heartbeat
Plan to retire them in a follow-up commit once we verify the
heartbeat bump alone is enough.
Margin trim:
Previous: min(0.5, threshold * 0.4) — for 3 m NPC arrived at 2.5 m
New: min(0.2, threshold * 0.2) — for 3 m NPC arrives at 2.8 m
User feedback: 'compared to retail, it fires too close. In retail
it fires from a longer range.' Smaller margin matches that — still
safely inside ACE's strict WithinUseRadius but closer to the boundary.
Tight pickup radii (0.6 m item) now arrive at 0.48 m (was 0.36 m).
Filed issues:
#67 Door Use action doesn't complete after auto-walk arrival.
NPC dialogue fires correctly post-flush-AP+re-send, but
doors still go silent — need to investigate door-specific
state requirements in ACE's Door.ActOnUse or our wire
payload differences.
#66 Rotation: local player flips back after auto-walk arrival
(observed from retail observer); NPCs don't turn to face
the player when used. Both rooted in missing MovementType=8
TurnToObject handling. Supersedes #65 (which was local-only)
with a unified rotation-handling phase scope.
Trace showed local arrival landing at 3.025 m from a target with
objDist=3.00. The previous arrival used ArrivalEpsilon=0.05 to
EXPAND the threshold (dist <= 3.05), so the body stopped right at
the boundary. ACE's server-side WithinUseRadius is strict
(dist <= radius), so 3.025 > 3.00 — ACE rejected the re-sent
action and looped back to MoveToObject. User had to manually
re-press R because auto-arrival kept landing just-outside-range.
Fix: walk INSIDE ACE's radius by a safety margin (0.3–0.5 m, capped
at 40 % of threshold so tight pickup radii like 0.6 m stay
reachable).
arrivalThreshold = wire's distanceToObject or minDistance
safetyMargin = min(0.5, arrivalThreshold * 0.4)
effectiveArrival = max(arrivalThreshold - safetyMargin, 0.1)
Examples:
objDist=3.00 (NPC) → walk to ≤2.50 m (ACE happy at 3.0)
objDist=2.00 (door) → walk to ≤1.50 m
objDist=0.60 (item) → walk to ≤0.36 m
objDist=0.50 (small) → walk to ≤0.30 m
Flee case (moveTowards=false) keeps its original predicate with
+ArrivalEpsilon — that's the boundary check semantics for fleeing
to a min-distance, not a max-distance use radius.
Previous re-send-on-arrival didn't actually unstick the action. Trace
showed ACE replying to the re-sent Use with another MoveToObject —
i.e. ACE's Player.Location was still the pre-walk position, so the
'I'm in range' fast-path didn't fire.
Cause: packet ordering. OnAutoWalkArrivedReSendAction was firing the
re-send immediately (sub-frame), BEFORE the next per-frame
AutonomousPosition heartbeat. ACE processed the Use against stale
location data.
Fix: SendAutonomousPositionNow() — an out-of-frame AutonomousPosition
build using the controller's current authoritative position +
rotation + cell. Called from OnAutoWalkArrivedReSendAction BEFORE the
re-send. ACE now processes 'I'm here at (target_pos)' then 'Use'
in order; CreateMoveToChain's WithinUseRadius shortcut
(Player_Move.cs:66) fires immediately and the action completes.
[autowalk-flush-ap] trace line under ACDREAM_PROBE_AUTOWALK so the
sequence is visible end-to-end:
autowalk-end → autowalk-flush-ap → autowalk-arrived-resend → autowalk-out
User report: 'It still however just approach it and does not use it.'
Root cause: local auto-walk arrives at the target visually, but ACE's
server-side MoveToChain may have timed out before our position was
recognised as in-range (we don't echo authoritative position back to
ACE during the walk yet). The action never fires.
Fix (re-send on arrival):
* PlayerMovementController.AutoWalkArrived event fires once when
EndServerAutoWalk(reason='arrived') is called.
* GameWindow tracks _pendingPostArrivalAction = (guid, isPickup)
on each SendUse / SendPickUp.
* OnAutoWalkArrivedReSendAction (subscribed at EnterPlayerModeNow)
re-sends the action with isRetryAfterArrival=true. The retry
flag prevents the re-sent action from itself setting a new
pending action — breaks any potential re-fire loop.
* The re-sent action is close-range from the local body's
perspective, so ACE's CreateMoveToChain hits the WithinUseRadius
shortcut (Player_Move.cs:66) and completes immediately —
dialogue opens, item picks up.
User report: 'items dropped on the ground now have a smaller triangle
box, perhaps too small. Also now other stuff like signs also have a
very small triangle box, should not have it should scale to the size
of the object.'
Fix (scale-aware indicator height):
* TargetIndicatorPanel.TargetInfo now carries entity Scale.
* EntityHeightFor multiplies the per-type base by Scale so an
upscaled NPC / sign / lifestone gets a proportionally larger box.
* Per-type table refined:
Creature : 1.8 m * scale
Door/Lifestone/Portal : 2.4 m * scale
Small carry items (weapon/armor/clothing/jewelry/food/money/
misc/missile-weapon/container/gem/spellcomp/writable/key/
caster — most pickup-able): 0.8 m * scale (up from 0.5 m)
Everything else (signs / scenery interactables / untyped):
1.5 m * scale (up from 0.5 m default)
Deferred to follow-up: exact mesh-AABB-derived box (need to read
each entity's actual rendered bounds at registration time).
Three user-reported fixes:
1. (B.6) Run-vs-walk decision lifted out of the per-frame overlay
into BeginServerAutoWalk. Once set at auto-walk start, the
character runs (or walks) the full way to the target instead of
transitioning. Matches user-observed retail behaviour:
'if its far away it should run all the way to the object and
then stop'.
_autoWalkWalkRunThreshold → _autoWalkInitiallyRunning (bool,
sampled once from initial distance vs the wire's WalkRunThreshold).
2. (B.7) TargetIndicatorPanel now picks EntityHeight per-type:
Creature (NPC/player) → 1.8 m
Door / Lifestone / Portal (tall structures) → 2.4 m
Default (small ground item) → 0.5 m
Items now get a small box hugging the silhouette instead of a
humanoid-tall rectangle floating around them.
3. (Interact) R-key (UseCurrentSelection) now dispatches by target
type:
Item (no Creature flag, no BF_DOOR|LIFESTONE|PORTAL|CORPSE)
→ SendPickUp (PutItemInContainer 0x0019)
Everything else → SendUse (0x0036)
Single hotkey to interact with whatever's selected.
Deferred (separate phase): turn-to-face on close-range use. ACE
server-side does Rotate(target) before the close-range pickup
callback (Player_Move.cs:71), but our local body doesn't echo
the turn yet — needs a synthesized client-side rotation or
MovementType=8 TurnToObject handling. Filing as follow-up.
User reported intermittent selection — 'sometimes can be selected,
sometimes not'. Cause: WorldEntity.Position is at FEET level (Z=ground
for standing humanoids), so a 0.7m sphere centred there only covered
the lower legs. Clicks on chest (Z≈1.2m) or head (Z≈1.7m) missed
because the closest-approach distance from the cursor ray to the
feet-centered sphere exceeded the radius.
Fix:
- Sphere centre now defaults to position.Z + 0.9 m (humanoid
mid-body). New optional verticalOffsetForGuid callback overrides
per entity.
- Default radius bumped 0.7 → 1.0 m to match the new sphere
placement (1.0 m at 0.9 m height covers a 1.8 m humanoid from
shin to top-of-head).
GameWindow.PickAndStoreSelection wires the callback:
- Creatures (ItemType.Creature flag): vz = 0.9 m (humanoid centre)
- Large flat objects (BF_DOOR | BF_LIFESTONE | BF_PORTAL |
BF_CORPSE): vz = 1.0 m + radius 2.0 m (mid-door/lifestone)
- Everything else (ground items): vz = 0.2 m (just above feet)
Existing 9 WorldPicker tests still pass — their head-on ray geometry
doesn't depend on the vertical offset.
Visual test surfaced three follow-ups:
1. Square box, not 1:2 rectangle.
WidthHeightRatio: 0.5 → 1.0. Retail's Vivid Target Indicator draws
a square; the earlier humanoid-aspect ratio looked wrong for
non-humanoids and didn't match retail screenshots.
2. Large flat objects (doors / lifestones / portals / corpses)
weren't selectable with the new tight 0.7 m pick sphere.
WorldPicker.Pick now takes an optional radiusForGuid callback so
the host can per-entity decide a larger radius. GameWindow's pick
site supplies a lambda that bumps to 2.0 m for any entity with
BF_DOOR (0x1000), BF_LIFESTONE (0x4000), BF_PORTAL (0x40000), or
BF_CORPSE (0x2000) set in ObjectDescriptionFlags. Default stays
at 0.7 m for humanoids and items.
3. New [B.7] pick-info diagnostic on each successful pick:
[B.7] pick-info guid=0x... itemType=0x... pwd=0x... color=(r,g,b)
Lets us verify e.g. whether a 'green NPC' really is server-side
flagged as Vendor (BF_VENDOR=0x200, retail-defined green) vs a
bug in our colour lookup. The pwd bit table is acclient.h:6431-
6463 — same flags retail's gmRadarUI::GetBlipColor branches on.
Note: textured retail-sprite corner triangles remain a B.7 follow-up
deferred per the spec. MVP uses procedural fills.
User-observed bug: 'I selected a retail player once, now I cant select
anything else.' Cause: the 5 m fixed pick sphere covered most of the
visible area around an entity, so once the cursor was anywhere near an
NPC or player, every subsequent click resolved to that same NPC/player
instead of the actual cursor target.
0.7 m roughly matches the actual hitbox radius of humanoid bodies and
most pickable items. Clicking on the entity's silhouette still hits;
clicking next to it or through it to a closer target now correctly
picks the closer target.
Existing 9 WorldPicker unit tests all pass — they tested geometric
behaviour at picked test radii, not the literal 5 m constant.
Follow-ups (deferred to a future picker phase):
- Per-itemType radius (tighter for tapers, looser for chests).
- Priority sorting at equal hit-distance (items beat NPCs).
- Ray-vs-actual-mesh test instead of bounding sphere.
Together with B.7's target indicator (corner triangles, c7e5f9f /
4bc95ec) this gives the user both 'I can hit what I'm aiming at'
AND 'I can see what I just hit' — fixes the over-pick at the source
plus surfaces it visually when it does still happen.
