acdream/docs/architecture/code-structure.md

acdream — code structure & extraction sequence

Status: Living document. Created 2026-05-16 as the companion to the "Code Structure Rules" section in CLAUDE.md. Purpose: Describe the desired structural state of the App layer, explain the rules we've adopted, and lay out the safe extraction sequence from today's reality (one 14,557-line GameWindow.cs at the 2026-07-20 audit) to the target (thin GameWindow, small focused collaborators). Companion to: acdream-architecture.md (the layered architecture) and worldbuilder-inventory.md (what we take from WB vs port ourselves).

---

1. The structural problem we're solving

The layered architecture works: AcDream.Core is GL-free, the network layer is wire-compatible, the UI has a stable contract, plugins load. The structural debt is concentrated in one file:

src/AcDream.App/Rendering/GameWindow.cs   14,557 lines (2026-07-20)

GameWindow is the single object that:

• Owns the GL context, the window, input, and shaders.
• Reads ~40 different environment variables across its lifetime.
• Hosts the live network session (WorldSession) and the offline pre-login state.
• Owns parallel dictionaries for entity lookup (_entitiesByServerGuid, the per-landblock entity lists in GpuWorldState, plus the player controller's own state).
• Drives selection / interaction (WorldPicker, SendUse, SendPickUp).
• Drives per-frame render orchestration (sky → terrain → opaque mesh → transparent mesh → particles → debug lines → UI).
• Wires up every plugin hook sink, every diagnostic, every panel.

Almost every M1 / M2 bug touches this file. Every new feature adds a field plus a method plus a wiring call. It is not getting better on its own.

The fix is not "rewrite GameWindow in one pass" — that's a high-risk change that would block M2. The fix is to extract one collaborator at a time, verify behavior is unchanged, ship, and move on. This document defines that sequence.

---

2. Code Structure Rules — the discipline

Recap of the rules from CLAUDE.md with the rationale:

Rule 1: No new substantial feature bodies in GameWindow.cs

Why: Every line we add to GameWindow makes the eventual decomposition harder. New features that "live in" GameWindow instead of being extracted are the reason the file is 10k lines.

How to apply: A new feature gets its own class under src/AcDream.App/<Subsystem>/ (or deeper in AcDream.Core if it's pure logic). GameWindow owns a field and a wiring call, nothing more. If you find yourself adding a 200-line method to GameWindow, stop and extract.

Exemption: Trivial wiring that must stay in GameWindow because it touches GL state during OnLoad is acceptable, but should still delegate to a collaborator for the substance.

Rule 2: AcDream.Core must not depend on window / GL / backend projects

Why: Core is the GL-free, testable layer. The moment Core imports a GL or windowing namespace, we've lost the ability to test it without a graphics context, and the layer split becomes fiction.

How to apply: Phase O removed both external WorldBuilder/backend project references. The only currently allowed seams are the GL-free helpers owned in our tree under src/AcDream.Core/Rendering/Wb/: TerrainUtils, TerrainEntry, RegionInfo, SceneryHelpers, and TextureHelpers. ObjectMeshManager and every GL resource owner remain in App. If Core needs a new capability, define a narrow Core interface and implement it in App; adding a new project reference requires an inventory-doc update explaining why.

Rule 3: UI panels target AcDream.UI.Abstractions only

Why: This is the one rule that keeps D.2b (the future retail-look backend) viable. Every panel that imports ImGuiNET directly is a panel we'd have to rewrite when the backend swaps.

How to apply: A panel's using block must mention AcDream.UI.Abstractions.* and nothing from AcDream.UI.ImGui. The panel writes against IPanelRenderer. The ImGuiPanelRenderer translates those calls to ImGui at runtime. Plugin-facing UI follows the same rule.

Rule 4: Startup env vars enter through RuntimeOptions

Why: Environment variables are global mutable state. Reading them at random call sites means (a) duplicated Environment.GetEnvironmentVariable boilerplate, (b) no single place to see "what flags does the client respond to?", (c) impossible to unit-test parsing.

How to apply: src/AcDream.App/RuntimeOptions.cs is the typed options object. Program.cs builds it once from args + env and passes it to GameWindow. New startup flags add a field to RuntimeOptions and a parser in RuntimeOptions.FromEnvironment. They don't add Environment.GetEnvironmentVariable reads.

Scope: RuntimeOptions is for startup-time configuration — things that don't change once the window is up. Runtime diagnostic toggles are Rule 5's domain.

Rule 5: Runtime diagnostic toggles live in diagnostic owner classes

Why: Diagnostic flags (ACDREAM_DUMP_MOTION, ACDREAM_PROBE_*, etc.) need to be both env-readable at startup and runtime-toggleable from the DebugPanel. Per-call-site env reads can't be runtime-toggled.

How to apply: Today's template is src/AcDream.Core/Physics/PhysicsDiagnostics.cs — one static class with typed Probe* properties read from env vars once at startup, plus runtime setters that the DebugPanel binds. New diagnostic flags follow this shape, not the per-call-site pattern that dominates GameWindow.cs.

Cleanup direction: The dozens of existing ACDREAM_DUMP_* reads inside GameWindow.cs are tech debt. We do NOT bulk-migrate them as part of this refactor — they're working, they're scattered, and moving them carries risk without a current acceptor. We migrate them opportunistically: when a GameWindow extraction lands and a diagnostic moves with it, route it through the new owner's diagnostic class.

Rule 6: Tests live in the project matching the layer

Why: Test discoverability + dependency hygiene. A test for a Core class belongs next to other Core tests; a test for an App class belongs in an App test project. Co-locating tests across layers makes the dependency graph dishonest.

How to apply: One test project per source project that has tests. Today:

• tests/AcDream.Core.Tests/ ← src/AcDream.Core/
• tests/AcDream.Core.Net.Tests/ ← src/AcDream.Core.Net/
• tests/AcDream.UI.Abstractions.Tests/ ← src/AcDream.UI.Abstractions/
• tests/AcDream.App.Tests/ ← src/AcDream.App/

tests/AcDream.App.Tests/ now exists and owns App-layer controller, streaming, render-resource lifetime, retained-UI, and RuntimeOptions tests. New App tests belong there; do not place GL-free Core behavior in that project merely because App currently wires it.

---

3. Target structure of the App layer

The end state — not what we're shipping in one pass, but the shape we're aiming at.

src/AcDream.App/
├── Program.cs                          # parse args + env → RuntimeOptions, build GameWindow
├── RuntimeOptions.cs                   # typed startup options (Rule 4)
├── Rendering/
│   ├── GameWindow.cs                   # thin: GL/window lifecycle + delegates per-frame to RenderFrameOrchestrator
│   ├── RenderFrameOrchestrator.cs      # per-frame draw order (sky → terrain → opaque → trans → particles → debug → UI)
│   ├── LiveEntityAnimationScheduler.cs # shipped: ordinary live-object update workset
│   ├── RetailStaticAnimatingObjectScheduler.cs # shipped: separate static-animation workset
│   ├── StaticLiveRootCommitter.cs       # live static root → pose + collision boundary
│   ├── TerrainModernRenderer.cs        # (already exists)
│   ├── TextureCache.cs                 # (already exists)
│   ├── ParticleRenderer.cs             # (already exists)
│   ├── Sky/                            # (already exists)
│   ├── Wb/                             # WB seam + EnvCellLandblockBuild transaction
│   └── Vfx/                            # (already exists)
├── Net/
│   └── LiveSessionController.cs        # owns WorldSession lifecycle, login/handshake, reconnect
├── Physics/
│   ├── ProjectileController.cs          # canonical live-record projectile orchestration
│   ├── RemotePhysicsUpdater.cs          # ordinary/Hidden remote narrow-tick integration
│   ├── LiveEntityOrdinaryPhysicsUpdater.cs # manager-less canonical body Transition path
│   ├── LiveEntityShadowPublisher.cs     # authoritative exact-owner/residency collision gate
│   ├── RemoteInboundMotionDispatcher.cs # shared animated/headless UpdateMotion funnel
│   ├── RemoteTeleportController.cs      # loaded/pending teleport placement ownership
│   ├── RemoteTeleportHook.cs            # ordered retail teleport teardown actions
│   └── RemoteTeleportPlacement.cs       # collision-seated SetPosition transition commit
├── World/
│   ├── InboundPhysicsStateController.cs # timestamps + accepted spawn snapshots
│   ├── LiveEntityRuntime.cs             # shipped: logical lifetime + ServerGuid↔entity.Id translation
│   ├── RetailInboundEventDispatcher.cs  # update-thread packet/frame FIFO barrier
│   ├── LiveEntityPresentationController.cs # ordered Hidden/NoDraw/effect/collision side effects
│   ├── LiveEntityTeardown.cs            # failure-isolated multi-owner lifecycle drain
│   └── ParentAttachmentState.cs         # parent generations + pending ParentEvent relations
├── Interaction/
│   └── SelectionInteractionController.cs  # owns WorldPicker, selection state, Use/PickUp dispatch
├── Streaming/                          # LandblockStreamer + immutable LandblockBuild completion
├── Input/                              # (already exists)
├── Audio/                              # (already exists)
└── Plugins/                            # (already exists)

What GameWindow keeps:

• IWindow / GL / IInputContext lifecycle (constructor + OnLoad + Run + OnClosing).
• RuntimeOptions reference (the typed startup config).
• One field per collaborator (_liveSessionController, _liveEntityRuntime, _selectionInteraction, _renderFrameOrchestrator).
• The Silk.NET event-handler stubs that delegate to collaborators.

What GameWindow loses:

• The 7 startup-time env var fields → moved into RuntimeOptions.
• TryStartLiveSession + the post-login network drain → moved into LiveSessionController.
• _entitiesByServerGuid + per-entity dictionaries + ServerGuid↔Id translation → moved into LiveEntityRuntime.
• WorldPicker + selection-driven SendUse / SendPickUp orchestration → moved into SelectionInteractionController; Core SelectionState remains the already-shipped session owner and is injected into that controller.
• Per-frame draw orchestration → moved into RenderFrameOrchestrator.

The eventual GameEntity aggregation (target state described in acdream-architecture.md §"GameEntity: The Unified Entity") happens after LiveEntityRuntime is the single owner of entity state. Until then, the parallel-dicts problem is bounded inside one class instead of spread across GameWindow.

LiveEntityRuntime is now that single boundary. It composes InboundPhysicsStateController for the nine-channel retail timestamp gates and latest accepted immutable CreateObject snapshot, owns the canonical local ID and optional runtime components, and separates logical registration from spatial projection. It also retains raw PhysicsState separately from the final state produced by retail's ordered side effects; LiveEntityPresentationController projects those transitions into draw, collision, effect, child-NoDraw, and target visibility without becoming a second lifetime or GUID owner. ParentAttachmentState is runtime-owned and keys unresolved relations by child and parent generation. Rendering/Vfx/EntityEffectController owns the focused mixed F754/F755 pending FIFO, effect profiles, typed-table resolution, and a readiness set; canonical ServerGuid-to-local-ID translation always stays in LiveEntityRuntime. EntityScriptActivator uses the same canonical WorldEntity.Id as rendering and physics; the disjoint static ID allocators fail fast instead of wrapping into another landblock's namespace. All other non-Parent packet families still need the future general queue tracked by divergence AD-32.

The per-frame object body is no longer an animation-dictionary loop inside GameWindow. LiveEntityAnimationScheduler snapshots canonical spatial root records and advances the incarnation-stable object clock, PartArray, hooks, one selected movement owner, and manager tail in retail order. Manager-less bodies delegate their candidate/Transition/cell/shadow commit to LiveEntityOrdinaryPhysicsUpdater; retained projectile bodies and remote MovementManagers remain mutually exclusive movement owners. Static animation is deliberately separate: RetailStaticAnimatingObjectScheduler owns the CPhysics::static_animating_objects workset for DAT and live PhysicsState- Static owners with Setup DefaultAnimation. Both schedulers are wired by GameWindow, but neither owns GUID identity or logical resources. The typed animation view fills its reusable snapshot and render-ID set from LiveEntityRuntime's concrete spatial dictionary, avoiding interface-enumerator boxing on both update and render hot paths. RemoteInboundMotionDispatcher similarly keeps UpdateMotion protocol behavior outside GameWindow: GameWindow resolves the canonical record/body and the optional PartArray sink, while one dispatcher owns retail's interrupt, style, MoveTo/type-0, sticky, and standing-long-jump order. Static root projection is bounded by StaticLiveRootCommitter, which synchronizes changed roots to effects and collision without rebuilding zero-omega shadows or resurrecting a Hidden/withdrawn registration.

Synchronous network and lifecycle callbacks are bounded by RetailInboundEventDispatcher. It owns no wire state and no identity; it only serializes nested live-object operations until the current packet or full object-frame tail completes. State-bearing direct dispatch is allocation-free; only a genuinely nested operation allocates its retained queue wrapper. LiveEntityRecord then supplies exact-incarnation and per-channel authority versions at callback boundaries. Position, State, Vector, and Movement remain independent, while a separate velocity version invalidates only an older operation that would overwrite a newer velocity installed by Position, Vector, or Movement. This prevents re-entrant App observers from creating call-stack ordering that retail's update-thread packet FIFO cannot produce.

Pose-dependent hook deferral is similarly incarnation-scoped rather than GUID- or local-ID-scoped. EntityEffectPoseRegistry publishes a monotonic pose-owner lifetime, and AnimationHookFrameQueue captures it before semantic callbacks and rechecks it before each semantic AnimationDone and each routed hook. Static animation retains process_hooks until its root, live parts, and children are published; withdrawal invalidates both its prepared pose and pending hook tail.

Resolved ordinary motion commits its body/root/contact state before the canonical full-cell setter enters LiveEntityRuntime.RebucketLiveEntity. Because that setter may synchronously move the projection to pending or replace the GUID, both RemotePhysicsUpdater and LiveEntityOrdinaryPhysicsUpdater revalidate the exact incarnation before collision or manager-tail publication. Collision residency itself is projection-owned, not updater-owned: LiveEntityPresentationController suspends retained non-projectile shadows on every unavailable-projection edge, restores them on hydration, and reconciles the pending-first case at OnLiveEntityReady after collision registration. Local projection and both authoritative remote UpdatePosition tails commit the complete root before rebucketing, then publish collision only through an exact-record/spatial-residency gate. Remote reflood tracks translation, sign-invariant complete orientation, and cell changes so an in-place turn or a same-pose EnvCell crossing cannot leave offset Setup shapes stale. The projectile controller retains its separate body/InWorld/shadow edge owner.

Remote teleport placement is bounded in Physics/RemoteTeleportController, not GameWindow: it retains at most one pending request per materialized incarnation, scopes it by the live generation and accepted PositionSequence, and asks RemoteTeleportPlacement to collision-seat the current body when the destination projection is available. GameWindow supplies lifecycle and shadow-sync callbacks only; canonical identity remains in LiveEntityRuntime. Failed hydration restores the captured source and delegates an incarnation-scoped shadow restore to LiveEntityPresentationController while that source is unloaded, so Hidden/UnHide and teleport never become competing restore owners. A newer placement transfers that restore into an explicit generation-scoped active-placement state before its rebucket visibility edge even while Hidden. All intervening Hidden/UnHide and projection edges defer to that owner until stable success or rollback completes; only then can it restore, re-defer the source, or hand a Hidden result back for UnHide. The ILiveEntityRemotePlacementRuntime seam keeps the complete cell/contact handoff available across same-body runtime-wrapper replacement; replacing the canonical body or dropping the placement contract within one incarnation is rejected even after an operational component clear. RemoteMotion.Body is constructor-owned; hydration compares pending/current wrappers directly to the record body rather than trusting wrapper-to-wrapper equality. Binding reads an interface Body getter once and reuses that snapshot. GpuWorldState performs remove+place as one spatial rebucket, then commits and serially drains visibility edges; LiveEntityRuntime filters delayed duplicates. A rollback inside an observer cannot race the outer destination-visible notification or expose an intermediate false pulse. LiveEntityTeardown executes those independent owner callbacks to completion and aggregates failures afterwards, so a throwing effect/plugin sink cannot strand teleport, movement, shadow, light, or GUID-scoped state.

---

4. Extraction sequence — safest first

Each step is one PR-sized refactor. Each must build clean, all tests pass, and visual verification at Holtburg looks identical to the previous step. Don't bundle two steps.

Step 1 — RuntimeOptions — SHIPPED (eda936dc)

Scope: Replace startup-time env var reads with a typed options object built once in Program.cs.

Behavior change: None. Same env vars, same defaults, same effects.

Risk: Low. Mechanical substitution at ~10-15 call sites in GameWindow.cs + one constructor signature change.

Test: Unit tests for RuntimeOptions.FromEnvironment parsing (the new tests/AcDream.App.Tests/ project).

Verification: dotnet build + dotnet test green. Visual launch verifies live mode + dat dir resolution still work.

Step 2 — LiveSessionController — SHIPPED (0b25df53)

Scope: Extract TryStartLiveSession + the WorldSession ownership + the post-EnterWorld drain (OnLiveStateUpdated, OnLiveEntityDeleted, etc.) into a controller class.

Behavior change: None. Same wire behavior, same handshake.

Risk: Medium. WorldSession lifecycle is load-bearing — every session-state crash would surface here. The change is a class extraction with the same event subscriptions, not a rewrite.

Test: Existing AcDream.Core.Net.Tests already cover the wire layer. The controller itself gets a smoke test that verifies it can be constructed without a live socket (offline mode).

Verification: Visual login + Holtburg traversal + door interaction identical to pre-extraction.

Step 3 — LiveEntityRuntime — SHIPPED 2026-07-14

Shipped scope: One LiveEntityRecord per server-object incarnation now owns ServerGuid↔local-ID translation, accepted state, runtime components, parent relations, logical resource activation, exact teardown, and spatial projection state. RegisterLiveEntity, RebucketLiveEntity, and UnregisterLiveEntity make the lifetime boundary explicit. Landblock unload/reload moves the same WorldEntity; it cannot reconstruct from a stale CreateObject or replay setup scripts. Equipped children use an attached projection and never enter the top-level target/radar/status view. Canonical materialized lookup remains available while a projection is pending; the separate visible view is the only surface radar, picking, status, and targeting consume. Pickup/parent leave-world clears cell membership and pauses root movement/animation without destroying the retained owners. Top-level spawn publication is one-shot per incarnation, so leave/re-entry restores presentation without duplicating plugin event replay.

Remaining target: the player-specific controller is still a separate aggregation. The focused effect queue is shipped; AD-32 now tracks only the future non-effect, non-Parent packet queue.

Behavior change: Spatial withdrawal and re-entry now preserve logical identity and active resources instead of replaying create-time effects.

Risk: Medium-high. Entity lookup is in every hot path. The change is structural (one owner instead of three) but the lookup semantics must be byte-identical.

Test: LiveEntityRuntimeTests cover duplicate CreateObject, generation replacement, appearance mutation, loaded/pending rebucketing, attached projection, pickup leave/re-entry, canonical-versus-visible lookup, resource rollback, GUID reuse, and idempotent session teardown.

Verification: Walk Holtburg, click NPC, open door, pick up item. All four M1 demo targets must still work.

Step 4 — SelectionInteractionController

Scope: Extract WorldPicker, the Core SelectionState consumers, SendUse, SendPickUp, and the InputAction.Select* / UseSelected / SelectionPickUp switch cases into one controller. Depends on Step 3 (uses LiveEntityRuntime).

Behavior change: None.

Risk: Low-medium. Selection state is local to interactions; the network outbound side is well-defined (InteractRequests.BuildUse / BuildPickUp).

Test: Selection state machine tests in tests/AcDream.App.Tests/.

Verification: Click-to-select, double-click-to-Use, F-key pickup all still work.

Step 5 — RenderFrameOrchestrator

Scope: Extract the per-frame draw sequence (sky → terrain → opaque mesh → translucent mesh → particles → debug → UI) into a dedicated orchestrator that GameWindow.OnRender delegates to.

Behavior change: None. Same draw order, same GL state.

Risk: Medium. GL state management is touchy; the orchestrator must hand the GL context to the same renderers in the same order with the same per-pass state setup.

Test: Visual verification only. Render orchestration is hard to unit-test without a GL context.

Verification: Holtburg at radius 4, radius 8, radius 12 looks identical across all four quality presets.

Step 6 — GameEntity aggregation (the big one)

Scope: Consolidate WorldEntity + AnimatedEntity + the per-entity state in LiveEntityRuntime into one GameEntity class (the target described in acdream-architecture.md). Every entity in the world — player, NPC, monster, door, item — becomes a single GameEntity.

Behavior change: None at the wire / visual level; substantial at the call-site level (everyone moves to the new entity API).

Risk: High. Touches every system that reads entity state.

Test: All existing tests + the new AcDream.App.Tests suite. Visual verification at every M1 / M2 scenario.

Verification: Full M2 demo loop (equip sword, kill drudge, pick up loot, open inventory) works identically.

---

5. Rules of the road during the extraction

1. One step at a time. A PR that ships Step 1 ships only Step 1. Bundling steps makes failures hard to isolate.
2. Behavior preservation is the acceptance criterion. Every step must build clean, all tests pass, and visual verification at the appropriate M1 / M2 scenarios must succeed. We're moving code, not changing it.
3. No new features during an extraction step. If you spot a real bug while extracting, file it in docs/ISSUES.md and address it in a separate commit (before or after the extraction, not folded into it).
4. Diagnostic toggle migrations are opportunistic. When a method moves to a new owner, the diagnostic flag inside it can move to a diagnostic class as part of the same commit. We do not do a bulk diagnostic-cleanup pass.
5. Update this document when the plan changes. If Step 3 turns out to need a different shape than described above, update §4 in the same session you discover the divergence.

---

6. What this document is not

• Not a full rewrite plan. The point is the opposite — small steps, verified at each boundary.
• Not blocking M2. Step 1 is small enough to ship without disrupting M2 work. Later steps interleave with M2 / M3 phases as the corresponding code paths come into focus.
• Not a substitute for the milestones / roadmap. Those drive the feature work. This drives the structural work that runs underneath.