Date: 2026-02-28
Status: Canonical interaction contract (graph-surface baseline)
Priority: Pre-renderer/WGPU required

Related:

• PLANNING_REGISTER.md
• 2026-02-28_ux_contract_register.md
• ../subsystem_ux_semantics/2026-03-01_ux_execution_control_plane.md
• ../workbench/workbench_frame_tile_interaction_spec.md
• GRAPH.md
• ../workbench/WORKBENCH.md
• ../aspect_command/command_surface_interaction_spec.md
• ../subsystem_focus/focus_and_region_navigation_spec.md
• ../viewer/viewer_presentation_and_fallback_spec.md
• ../subsystem_ux_semantics/2026-03-04_model_boundary_control_matrix.md
• layout_behaviors_and_physics_spec.md
• ../design/KEYBINDINGS.md
• ../../TERMINOLOGY.md
• 2026-03-14_graph_relation_families.md — relation family vocabulary; Navigator replaces "file tree" as the hierarchical projection surface
• 2026-03-14_edge_visual_encoding_spec.md — per-family visual encoding and interaction affordances
• 2026-03-14_edge_operability_matrix.md — per-family operability contract
• 2026-04-03_node_glyph_spec.md — node visual form authority; consumes §4.8 LOD tiers

Adopted standards (see standards report §§3.3, 3.5, 3.6):

• WCAG 2.2 Level AA — SC 2.5.8 (minimum target size for nodes/edges/affordances), SC 2.4.3 (focus order), SC 2.4.11 (focus appearance)
• Fruchterman-Reingold 1991 — physics preset semantics (Liquid / Gas / Solid) documented against this model
• OpenTelemetry Semantic Conventions — diagnostics channels for blocked/degraded graph states

• GraphId = truth boundary.
• GraphViewId = scoped view state.
• Navigator = the section-structured projection of graph relations rendered through one or more Navigator hosts (replaces "file tree" — see 2026-03-14_graph_relation_families.md §5).
• workbench = arrangement boundary.

This spec defines graph-surface semantics and must not redefine workbench arrangement ownership.

Normative graph contracts use: intent, trigger, preconditions, semantic result, focus result, visual result, degradation result, owner, verification.

• Tile/frame arrangement is not content hierarchy.
• Navigator, regardless of host edge or form factor, is not content truth authority — it is a read-only projection over graph relation families.
• Physics presets are not camera modes.
• "File tree" is a legacy alias — use Navigator in new code and docs.

---

This spec defines the interaction contract for the graph surface of Graphshell.

Within the architecture, this surface is the user-facing interaction layer of the Graph subsystem.

It explains:

• what the graph pane is for,
• what node, edge, and canvas actions mean semantically,
• who owns each behavior,
• what state transitions those interactions imply,
• what visual feedback must accompany them,
• what fallback behavior must happen when the ideal path is unavailable,
• which controls are core, planned, and exploratory.

This is not the workbench arrangement contract.

This spec governs:

• Graph Pane
• Node
• Edge
• Canvas
• Navigator (hierarchical projection over graph relation families rendered through Navigator hosts) — see 2026-03-14_graph_relation_families.md §5
• camera and selection semantics inside the graph pane
• the user-facing contract of the graph structure subsystem

For workbench/frame/tile semantics, see:

• ../workbench/workbench_frame_tile_interaction_spec.md

For command invocation, focus ownership, and viewer-state clarity, see:

• ../aspect_command/command_surface_interaction_spec.md
• ../subsystem_focus/focus_and_region_navigation_spec.md
• ../viewer/viewer_presentation_and_fallback_spec.md

---

The graph surface is composed of four primary interactive layers:

1. Graph Pane
   • the active graph viewport and command context
2. Canvas
   • the navigable graph space inside the pane
3. Node
   • the primary graph entity users inspect, select, move, and open
4. Edge
   • the relationship or traversal surface between nodes

When the Navigator is active in any host, it is a section-structured hierarchical projection over graph relation families — not a separate content-truth authority. It reads graph edges and renders them as a tree; it does not own graph identity or topology.

Selectable graph-surface objects are the interactable objects rendered on the canvas: nodes, edges, arrangement objects (for example frames or minimized tile objects), and minimap markers when they expose direct actions. Purely informational hover-only UI (labels, passive badges, transient tooltips) is not selectable.

• Graph Pane: the semantic navigation and manipulation surface for graph work.
• Canvas: the continuous space users pan, zoom, lasso, and inspect.
• Node: the primary content entity users act on.
• Edge: the relationship surface users inspect and, when defined, traverse. Edges have families (Semantic, Traversal, Containment, Arrangement, Imported) — see 2026-03-14_graph_relation_families.md.
• Navigator: a section-structured hierarchical projection over graph relation families rendered through one or more hosts. Sections may include Workbench (arrangement), Folders (containment/user-folder), Domain (containment/derived), Unrelated, Recent (traversal), and Imported.

• Graphshell owns camera policy, selection truth, activation semantics, routing, and graph interaction meaning.
• The framework layer may provide hit surfaces, event capture, and paint execution.
• The framework must not be the semantic owner of:
  • camera state,
  • selection truth,
  • lasso meaning,
  • node activation policy,
  • graph-to-workbench routing.
• The Navigator, when active in any host, is a read-only projection over graph relation families and must not become the owner of graph identity.

• The Graph subsystem owns graph truth:
  • nodes,
  • edges,
  • graph-backed hierarchical containment relations where defined,
  • graphlets / connected groups,
  • graph selection,
  • graph camera target semantics,
  • GraphViewId scope and view-state semantics.
• The Workbench subsystem does not own graph truth.
• The Workbench subsystem may host graph presentation surfaces, but it only does so after Graphshell routing bridges graph content into a workbench destination.
• A GraphViewId is a scoped view instance within a GraphId; panes may host a GraphViewId, but they do not define its identity.

---

Graph interactions fall into five semantic categories:

1. Selection
   • choose graph targets without committing a structural action
2. Activation
   • invoke the primary action of the target
3. Manipulation
   • move nodes or the viewport
4. Inspection
   • reveal information without changing semantic ownership
5. Routing
   • open graph content into workbench structures

• Single click selects the target under the pointer.
• Single click without a selection modifier replaces the current selection set with the target under the pointer.
• Ctrl-click (or configured additive-selection modifier) toggles the target's membership in the current selection set.
• Double click activates the target's primary action.
• If a target has no defined activation behavior, double click is a no-op beyond maintaining selection or inspection state.
• Hierarchical navigation actions, when invoked through the Navigator projection, must resolve through the same graph identity and routing rules as canvas-originated actions.
• Selection sets may mix any visible interactable graph-surface objects. Graphshell must not invent a hidden primary target inside a mixed selection.

The graph surface must make these user expectations reliable:

• camera controls affect the active graph pane only,
• selection is explicit and visible,
• node activation routes through Graphshell open policy,
• graph gestures do not silently fight each other,
• blocked or degraded behavior is explicit rather than silent,
• Navigator navigation, regardless of host, and graph-canvas navigation resolve to the same underlying graph-backed identities.

---

This section defines the stable target behavior for the current graph surface.

The following guardrails are mandatory for camera/navigation changes:

1. Metadata/layout ID parity must be explicit.
   • Any key used to read/write graph metadata or layout state must match the key shape used by egui_graphs for the same surface.
   • Partial identity migrations (some callsites None, others custom id) are forbidden.
2. No per-frame implicit camera override.
   • Continuous fit/recenter loops must not silently overwrite manual pan/zoom every frame.
   • First-frame fit and explicit fit commands are allowed; implicit perpetual fit is not.
3. Camera command ownership remains explicit.
   • CameraCommand::Fit (and related fit-family commands) are Graphshell-owned semantic actions.
   • Framework helpers may execute drawing but do not define camera truth.
4. Coordinate-space invariant must hold.
   • MetadataFrame.pan and related camera state writes are in widget-local space.
   • Screen-space conversion is render-path-only.
5. Physics/camera boundaries must not blur.
   • Physics presets affect node simulation only.
   • Physics profile changes must not mutate camera lock policy or zoom ownership.
6. Multi-view identity policy must be all-or-nothing.
   • Shared-slot behavior is acceptable only when explicitly declared.
   • Per-view isolation requires consistent GraphView id usage across graph view creation, layout state, and metadata state.
7. Regression checks are mandatory before merge.
   • Pan, wheel zoom, fit command, and lock toggles must all be verified in the same patch lane.
   • Any change touching metadata/layout IDs must include targeted tests or scenario updates proving no dead-slot write path.

What this domain is for

• Let the user move through graph space predictably and keep the active graph pane under user control.
• Define camera behavior as a Graphshell-owned policy surface, not a fixed widget behavior.

Core controls

• Drag selected node: move graph content.
• Wheel over graph pane: zoom at pointer.
• Pointer drag on empty canvas: pan the active graph pane.
• Camera Fit: fit relevant graph content.
• Graphlet Fit: fit the current connected cluster or explicitly targeted graph subset.
• Focus Selection: fit selected nodes only.
• Zoom Reset: restore canonical default zoom baseline.
• C: toggle position-fit lock (camera position follows fit target when enabled).
• Z: toggle zoom-fit lock (camera zoom follows fit target when enabled).
• Keyboard pan (Arrow Keys / WASD) is a camera control and must be configurable when enabled.
• Manual camera is the default policy: new graph views start with fit locks disabled and no startup auto-fit command.
• Physics presets (Liquid, Gas, Solid) affect node dynamics only and must not implicitly change camera behavior.

Who owns it

• Graphshell camera controller owns camera state, command targeting, zoom policy, fit-lock policy, and fit-strength policy.
• Graphshell graph/layout controller owns physics preset selection (Liquid, Gas, Solid) and node-dynamics behavior.
• The framework may only provide raw pointer and wheel events plus drawing surfaces.

State transitions

• Camera settings independently configure position-fit lock, zoom-fit lock, and fit strength for the active graph pane.
• C toggles position-fit lock for the active graph pane.
• Z toggles zoom-fit lock for the active graph pane.
• Panning changes viewport translation in the active graph pane.
• After manual pan input, camera translation may decay with slight inertia damping before settling.
• Wheel zoom changes viewport scale around the pointer target in the active graph pane.
• Camera Fit updates camera state to show relevant graph bounds.
• Graphlet Fit updates camera state to show the chosen graph subset bounds.
• Focus Selection updates camera state to show selected-node bounds.
• When position-fit lock is enabled, fit-family commands and graph-bounds changes may reconverge camera translation toward the current fit target.
• When zoom-fit lock is enabled, fit-family commands and graph-bounds changes may reconverge camera scale toward the current fit target.
• Node dragging changes graph bounds and may therefore update the fit target, but it must not silently disable manual pan or zoom.
• Fit Strength changes how strongly the camera converges toward the fit target over time.
• Physics preset changes update node simulation behavior only; they must not mutate camera locks or current camera state.

Visual feedback

• Camera movement must be immediate and legible.
• Zoom must visually anchor around the pointer, not drift arbitrarily.
• Fit operations must visibly land on the intended target set.
• If either fit lock is active, the user must be able to perceive that the camera is following graph bounds rather than being in a fully manual state.

Camera invariants (pre-renderer/WGPU closure)

• Zoom ownership is per active graph pane; a zoom request for one view must not mutate camera state in another view.
• Wheel zoom is pointer-relative when an anchor is available; missing anchor/metadata paths must defer safely or emit diagnostics rather than silently mutating camera state.
• Zoom Reset and fit-family commands resolve to deterministic camera targets under the active camera policy.
• Manual pan and manual zoom remain available regardless of active physics preset.
• Physics presets must never be used as implicit camera-mode selectors.

Pointer-anchor and passive-input invariants (UX migration §5.2 closure)

• Pointer-relative zoom anchor is resolved at wheel-sequence start from this order: hovered node anchor -> hovered edge midpoint anchor -> canvas pointer world position.
• If no valid world-space anchor can be resolved, wheel input is treated as passive for graph zoom (no camera mutation) and a diagnostics warning is emitted.
• Target lock for a wheel sequence is sticky: once anchor is chosen, subsequent wheel deltas in the same sequence must use the same anchor until sequence end.
• Graph zoom must not claim wheel input when a higher-priority input owner is active (Modal, CommandPalette, TextEntry, or non-active graph pane).
• Pointer-relative zoom is scoped to the active graph pane only; non-active panes may inspect hover but must not mutate camera state.

Zoom diagnostics assertions (normative)

Flow	Required diagnostics assertions
Anchor resolved and zoom applied	ux:navigation_transition with operation=zoom_pointer_relative, graph_view_id, anchor_kind, and scale delta
Missing anchor / passive no-op	ux:navigation_violation (Warn) with operation=zoom_pointer_relative and reason=anchor_unresolved
Wheel denied due to input ownership	ux:contract_warning with reason=zoom_input_owner_mismatch and owning context

Fallback / degraded behavior

• If input cannot be claimed, Graphshell must emit diagnostics and preserve existing camera state.
• Silent camera no-op behavior is forbidden.
• If a fit lock cannot be honored, Graphshell must preserve manual camera control and make the degraded fit-follow state explicit.

What this domain is for

• Support selecting, inspecting, manipulating, and opening graph content.

Core controls

• Single click node: select node.
• Modifier-click node: add or remove the node from the current selection set.
• Double click node: activate the node's primary open path.
• Drag selected node: move node.
• Drag one selected node while multiple are selected: move the selected group when group-move mode is active.
• Hover node: inspect node metadata and show hover affordance.
• Context action on node: expose node-scoped commands.

Who owns it

• Graphshell selection and graph interaction controllers own node selection truth, drag semantics, and activation meaning.
• The framework may only report hits and render hover or selected visuals.

State transitions

• Single selection replaces the current selection set with the target node.
• Additive selection mutates the current selection set without changing graph identity.
• Node activation routes through the workbench open policy.
• Node drag changes node position and, when enabled, group position.

Node manipulation command map (UX migration §5.3 closure)

Default bindings are profile-configurable, but semantic action mapping is fixed:

Action semantic	Default command/binding	Required behavior
Create node	New Node command (default N)	Create node in active graph context; selection moves to new node
Delete selected nodes	Delete Selected command (default Delete/Backspace)	Remove selected nodes via reducer-owned delete intent
Pin selected nodes	Pin Selected command (default P)	Set selected nodes to pinned state without changing selection scope
Unpin selected nodes	Unpin Selected command (Shift+P)	Clear pinned state for selected nodes
Group-move mode toggle	Toggle Group Move command (G)	While active, dragging any selected node moves all selected nodes as a cohort

Group-move invariants

• Group-move applies only when selected set cardinality is greater than 1.
• Group-move drag preserves pairwise offsets among selected nodes for the duration of one drag sequence.
• Group-move must not mutate non-selected nodes except through physics side effects after drag commit.

Visual feedback

• Selected nodes must look selected.
• Hovered nodes must look inspectable.
• Dragging must clearly show that movement is in progress.

Fallback / degraded behavior

• If a node cannot be opened, Graphshell must surface an explicit blocked-state or fallback target.
• Failed drag or selection claims must not leave hidden partial state.

What this domain is for

• Make relationships visible, inspectable, and traversable where defined.

Core controls

• Hover edge: inspect relationship or traversal information.
• Single click edge: select the edge and make it the active inspection target.
• Double click edge: open the edge's family/category in History Manager and reveal the matching recency-ordered entry.
• Context action on edge: expose edge-scoped commands through the canonical palette shell.

Who owns it

• Graphshell graph interaction controller owns edge inspection meaning, traversal policy, and any edge-selection semantics.
• The framework may only report edge hits and paint edge state.

State transitions

• Edge hover changes inspection context only.
• Edge single-click replaces the current selection set with the target edge unless modifier selection is active.
• Edge double-click opens History Manager scoped to the relevant edge family and edge/event entry; it does not append traversal history by itself.
• Dismissing an edge updates the current GraphViewId's EdgePolicy for that edge instance only, removing its presentation and graph effect from the current view without deleting the underlying relation/event truth.

Edge-focus and traversal invariants

• Edge focus (SetHighlightedEdge / ClearHighlightedEdge) is inspection-only state and must not append traversal history by itself.
• Traversal history mutation is owned by the history/reducer traversal append path, not by edge hover or single-click inspection.
• Clearing edge focus is explicit and deterministic when the active inspection target changes to none.
• Edge dismissal is view-local. Suppressing one edge in the current graph view must not hide other edges of the same family or delete provenance truth.

Edge-management interaction parity (UX migration §5.4 closure)

Interaction	Canvas semantic result	History semantic result
Edge hover	Update inspection context only	No traversal append
Edge single-click	Select edge and set highlighted edge for inspection	No traversal append
Edge double-click	Open matching family/category entry in History Manager	No traversal append
Edge dismiss	Suppress this edge instance in current GraphViewId via EdgePolicy	Underlying truth remains visible to History Manager

Edge-context command map (C2.2 closure)

Right-clicking an edge summons the canonical contextual shell with the Edge category first. The minimum edge-context command set is:

Action semantic	Default command/binding	Required behavior
Dismiss edge in this view	Dismiss Edge command (contextual)	Suppress only this edge instance in the current GraphViewId
Open edge family in history	Open Edge History command (default Enter on selected edge if double-click unavailable)	Open History Manager to the edge's family/category and reveal matching entry
Remove user edge	Remove User Edge command (default Alt+G for selected pair)	Remove user-grouped edge semantics for the active/selected pair
Connect source -> target	Connect Pair command (default G for selected pair)	Create one directed user-grouped edge from ordered pair source to target
Connect both directions	Connect Both Directions command (default Shift+G for selected pair)	Create user-grouped edges in both directions for the active/selected pair

Edge-context invariants:

• Edge context is command-capable only when the resolved edge target maps to a valid pair context.
• Inspection-only traversal edges must still expose disabled edge commands with explicit reasons rather than silently hiding the category.
• Edge-context command invocation must route through the same ActionRegistry semantics used by keyboard and other palette modes.

Edge diagnostics assertions (normative)

Flow	Required diagnostics assertions
Inspection-only edge focus update	ux:navigation_transition with operation=edge_inspection and history_append=false
Traversal-eligible edge activation	ux:navigation_transition with operation=edge_traversal_activate and resolver target
Edge activation blocked/fallback	ux:navigation_violation (Warn) plus ux:contract_warning with reason

Visual feedback

• Hovered edges must visibly read as inspectable.
• If an edge is the active inspection target, that state must be legible.
• Traversal activation must visibly change focus or destination context.

Fallback / degraded behavior

• If an edge has no primary action, double click is a no-op beyond preserving inspection state.
• Edges must not pretend to be activatable when they are inspection-only in the current context.

What this domain is for

• Provide predictable behavior for interacting with graph space rather than specific graph entities.

Core controls

• Empty-canvas click: clear selection when no other mode is active.
• Lasso gesture: select visible interactable graph-surface objects by region using the declared lasso binding.
• Empty-canvas context action: expose graph-scope commands.

Who owns it

• Graphshell graph interaction controller owns empty-canvas meaning, lasso semantics, and gesture precedence.
• The framework may only provide pointer events and paint transient lasso visuals.

State transitions

• Empty-canvas click clears the current selection set.
• Lasso updates the app-owned mixed selection set based on the chosen region and selection mode.
• Context commands do not directly mutate graph truth until a command is executed.

Lasso invariants (pre-renderer/WGPU closure)

• Lasso selection mode is deterministic from binding + modifiers: Alt => Toggle; otherwise Add when Ctrl is active or when the binding is RightDrag with Shift; otherwise Replace.
• Lasso candidate sets are canonicalized before intent dispatch (stable sort + deduplicate by selection-object key).
• Lasso, pan, and node drag are mutually exclusive gesture owners for a pointer sequence; ambiguity must be canceled explicitly with diagnostics.

Lasso boundary semantics (UX migration §5.1 closure)

• Boundary inclusion policy is intersection-inclusive: an interactable object is in the lasso candidate set when its hit bounds intersect the lasso polygon/rectangle by any positive area.
• Replace/Add/Toggle semantics apply over the candidate set after canonicalization (stable sort + dedupe).
• Lasso region sampling is deterministic: evaluation uses pointer-up finalized region, not intermediate hover jitter.
• If lasso starts on an eligible node-handle region, node drag owns the sequence and lasso does not arm.
• If lasso starts on empty canvas under declared lasso binding, lasso owns the sequence and pan is suppressed.

Lasso diagnostics assertions (normative)

Flow	Required diagnostics assertions
Lasso commit success	ux:navigation_transition with operation=lasso_commit, selection_mode, and candidate_count
Lasso canceled due to owner ambiguity	ux:navigation_violation (Warn) with reason=gesture_owner_ambiguous
Lasso denied by higher-priority context	ux:contract_warning with owning context and denied operation

Gesture precedence

• Pointer-down on a node enters node interaction first.
• Pointer-down on empty canvas enters pan or lasso based on the declared binding.
• Pan, lasso, and node-drag must not silently claim the same gesture at the same time.

Visual feedback

• Lasso region must be visible while active.
• Empty-canvas clear must visibly remove prior selection state.

Fallback / degraded behavior

• If the app cannot determine a valid gesture owner, Graphshell must cancel safely and emit diagnostics.
• Ambiguous gesture fights are forbidden.

What this domain is for

• Bridge graph activation semantics to the workbench without bypassing Graphshell ownership.

Core controls

• Node activation uses the workbench spec's open-tile-first behavior.
• Explicit alternate destinations are command-driven, not hidden pointer-only gestures.

Who owns it

• Graphshell routing authority owns destination policy, tile reuse rules, and fallback behavior.
• The graph surface may request an open action, but it does not choose layout structure independently.

State transitions

• Default node activation requests open or focus for the node in the workbench.
• Existing tiles are focused before new tiles are created.
• Explicit alternate routes may create new tile or frame structure according to workbench rules.

Visual feedback

• Activation must visibly land in the destination context.
• Cross-frame jumps must be legible.

Fallback / degraded behavior

• Silent duplicate tile creation is forbidden.
• If Graphshell chooses a fallback destination, the fallback must be explicit and deterministic.

What this domain is for

• Ensure graph actions are semantic commands first and bindings second.

Core controls

• Zoom In
• Zoom Out
• Zoom Reset
• Camera Fit
• Focus Selection
• Delete Selected where deletion is valid
• New Node
• graph-context command palette invocation

Who owns it

• Graphshell action registry and command dispatcher own semantic meaning and target resolution.
• The framework may render command surfaces and capture input, but it does not define command meaning.

State transitions

• Commands act on the active graph pane and current graph selection context.
• Commands may mutate camera state, selection state, or graph content according to their meaning.

Visual feedback

• Command execution must visibly change the graph surface or show a clear result surface.
• Failed commands must show explicit blocked-state feedback.

Fallback / degraded behavior

• If a command cannot run in the current context, the failure must be explicit.
• Hidden command no-op behavior is forbidden.

What this domain is for

• Make graph interactions visible, trustworthy, and usable under normal and degraded conditions.

Visual feedback

• Selected nodes must look selected.
• Hovered nodes and edges must look inspectable.
• Focused graph pane state must be legible.
• Lasso region must be visible while active.
• Active camera changes must be visually obvious.

Diagnostics

• Blocked states, degraded states, and fallback reasons must be observable.
• Missing or reduced interaction capability must be explicit rather than silent.

Accessibility

• Every critical graph action must eventually have a keyboard-accessible equivalent.
• Region focus and return paths must be deterministic.
• If a pointer-first action lacks a keyboard equivalent, it must be called out as incomplete rather than implied.

Performance

• Camera motion, node drag, and selection feedback must remain responsive under quick smoke workloads.
• Degradation may reduce fidelity, but it must not hide interaction ownership or current selection truth.

What this domain is for

• Keep semantic zoom behavior predictable while preserving interaction clarity and UxTree signal quality.

LOD tiers and thresholds (canonical defaults)

LOD tier	Scale range (camera.scale)	Node rendering contract	Semantic interaction contract
Point	< 0.55	Node points/minimal marks only	Node-level interactions suppressed; graph-level navigation remains active
Compact	0.55 .. < 1.10	Compact node glyph + key badge	Node selection/inspection available; reduced detail fields
Expanded	>= 1.10	Full node affordances and labels	Full node interaction and contextual commands

Transition and hysteresis policy

• LOD transitions are threshold-based with hysteresis band ±0.05 around each threshold.
• A tier transition commits only when camera.scale exits the current tier's hysteresis band.
• For one zoom sequence, LOD recalculation runs after final camera-scale update for that event tick.

UxTree cross-link (normative)

• GraphNode semantic emission in UxTree must follow ux_tree_and_probe_spec.md §3.1 C5.
• At Point LOD, individual GraphNode children are omitted from UxTree and GraphView carries a StatusIndicator child labeled "Zoom in to interact with nodes.".
• At Compact and Expanded LOD, UxTree must emit node-semantic children for interactable nodes.

Diagnostics assertions (normative)

Flow	Required diagnostics assertions
LOD tier transition committed	ux:navigation_transition with operation=lod_tier_change, from_tier, to_tier, graph_view_id
LOD transition suppressed by hysteresis	ux:contract_warning with operation=lod_tier_change and reason=hysteresis_hold
UxTree emission mismatch for active LOD	ux:navigation_violation (Warn) with expected vs observed semantic node-emission mode

---

These are intended near-term behaviors. They are not yet required for baseline closure, but they are part of the immediate design direction.

• Explicit "open in split" and "open in frame" actions from node context
• Pin and unpin actions directly from the graph surface
• More explicit group-drag affordances

• Edge traversal previews
• Edge filtering or highlighting by relationship type (UserGrouped, TraversalDerived, AgentDerived)
• Better relationship-specific context actions
• Multi-kind edge visual controls (hide or emphasize specific kinds)

When tag-based categorical relationships are implemented, edges of kind Categorical will route through a Mediator Node (a system-managed node representing the shared tag, e.g. #Rust). The graph interaction surface must handle:

• Mediator Node creation on first tag assignment and garbage collection when its degree drops to zero.
• Categorical edge rendering distinct from UserGrouped and TraversalDerived (e.g. dashed line).
• Mediator Node presentation as a non-content node type (non-activatable, distinct visual treatment).
• The interaction spec for Categorical edge kind and Mediator Node lifecycle belongs in this spec; the EdgeKind data-model extension belongs in edge_traversal_spec.md.

This is a prospective extension — do not implement until a dedicated design doc for categorical/tag-based edges is written.

• Camera bookmarks
• Search-targeting affordances in graph space
• More explicit selection-to-navigation handoffs
• Configurable physics presets:
  • Liquid for more fluid node motion
  • Gas for more energetic / free-moving node motion
  • Solid for more damped / rigid node motion
• Configurable Fit Strength control for auto-fit behavior, from aggressive framing to gentle framing assistance
• Keyboard camera pan bindings (Arrow Keys and/or WASD)
• Explicit user-facing toggles for:
  • position-fit lock (C)
  • zoom-fit lock (Z)
  • cursor-driven camera behavior
  • auto-fit to graph
  • auto-fit to graphlet
  • follow-selection framing assistance

• node-dynamics preset selector (Liquid/Gas/Solid), physics engine coefficient controls, position-fit lock toggle, zoom-fit lock toggle, keyboard pan bindings, pan inertia controls
• exposed via the Graph settings category in aspect_control/settings_and_control_surfaces_spec.md §4.2

---

These are exploratory design directions. They are informative only and should not be treated as current implementation requirements.

• 2.5D projection modes
• isometric views
• perspective or full 3D graph presentation

• Path-selection modes across chained edges
• Semantic subgraph collapse and expand
• Cluster-level operations and interaction modes

• Multiple layout modes beyond the default graph layout
• Graph-space lenses and semantic overlays
• Command-driven multi-node structural opening workflows

Logical self-loops (edges where source == target) are currently forbidden by the traversal skip rule and must be excluded from physics simulation regardless of how they arise (see edge_traversal_spec.md §2.6). A future node audit log design may permit self-loop edges to carry MetadataChange or workbench-action events as a localized per-node event log. If implemented:

• Self-loop edges must remain headless (logical only) and excluded from physics.
• They must not render as circular lines on the canvas.
• The audit log design belongs in a dedicated node_audit_log_spec.md (see subsystem_history/ deferred stub); it is not part of the edge traversal model.

---

1. Camera semantics are defined for pan, wheel zoom, fit, and focus-selection.
2. Node single-click, modifier-click, double-click, hover, and drag semantics are explicit.
3. Edge hover, inspection, and conditional activation semantics are explicit.
4. Canvas gesture precedence (pan vs lasso vs node drag) is defined and deterministic.
5. Graph-to-workbench routing matches the workbench spec rather than bypassing it.
6. Command semantics are defined as app actions, not only as keybindings.
7. Visual feedback and degraded-state requirements are explicit.
8. Accessibility expectations and current incompleteness rules are explicit.
9. Planned and exploratory controls are separated from the canonical core.
10. Lasso modifier-mode resolution and key normalization rules are explicit and deterministic.
11. Edge-focus inspection is explicitly separated from traversal-history mutation semantics.
12. Pointer-anchor and passive-input zoom invariants are explicit and diagnostics-backed.
13. Node create/delete/pin/group-move semantic mapping is explicit and deterministic.
14. Lasso boundary inclusion policy and gesture-owner precedence are explicit and deterministic.
15. Edge management interactions are explicitly aligned with history append rules and diagnostics.
16. LOD semantic zoom thresholds and hysteresis rules are explicit and deterministic.
17. LOD-to-UxTree semantic emission behavior is explicitly cross-linked and diagnostics-backed.