Status: Deferred (blocked) — awaiting basic LOD/culling baseline from 2026-02-24_performance_tuning_plan.md Phases 1–2 Blocking Prerequisites: Viewport culling (Phase 1) and Node + Edge LOD (Phase 2) from 2026-02-24_performance_tuning_plan.md must be in place. The canonical zoom-adaptive LOD tiers (Point / Compact / Expanded) and their camera.scale thresholds are defined in graph_node_edge_interaction_spec.md §4.8; DOI rendering must build on top of those tiers, not redefine them. Context: Post-LOD readability improvement for dense graphs; preserves mental map while surfacing relevance. Relates to: 2026-02-24_performance_tuning_plan.md (LOD/culling primitives this plan builds on), graph_node_edge_interaction_spec.md §4.8 (canonical LOD tier thresholds and hysteresis — DOI is an overlay, not a replacement), 2026-02-18_graph_ux_research_report.md §§13.2, 14.8, 14.9 (research basis), multi_view_pane_spec.md (per-pane isolation rule)

Introduce DOI-driven rendering and semantic fisheye as a concrete post-LOD roadmap item for graph UX. Focus+context without geometric distortion: keep the graph topology undistorted while scaling visual emphasis by relevance.

Not a replacement for LOD/culling. DOI and semantic fisheye are rendering-emphasis layers that build on top of stable viewport culling and zoom-adaptive LOD primitives. They do not duplicate or replace those systems.

DOI(n) = α·Recency(n) + β·Frequency(n) + γ·ExplicitInterest(n) - δ·DistanceFromFocus(n)

Default weights: α = 0.30, β = 0.20, γ = 0.30, δ = 0.20. All weights must sum to 1.0. Weights are configurable via CanvasRegistry policy; the defaults are the starting point, not a fixed contract.

Ghost tier nodes are never hidden entirely — they provide the structural context that makes semantic fisheye meaningful. Hiding Ghost nodes is the responsibility of the search/filter system, not DOI.

DOI scores are cached in MetadataFrame per graph pane alongside the existing LOD zoom oracle:

pub struct DoiScore {
    pub value: f32,           // [0.0, 1.0]
    pub tier: DoiTier,        // High / Medium / Low / Ghost
    pub computed_at: Instant, // for staleness check
}

pub enum DoiTier {
    High,
    Medium,
    Low,
    Ghost,
}

The cache (HashMap<NodeIndex, DoiScore>) is write-owned by the DOI calculator worker and read-only from the render path. Reads require no lock (snapshot swap pattern) to avoid stalling the render thread.

DOI scores are recomputed on a background worker, not per-frame. The cadence targets:

The worker holds a read-only snapshot of node metadata. It does not hold the graph write lock during calculation.

Key design rule: DOI drives rendering emphasis only. It does not hide, remove, or filter nodes from the graph, and it does not redefine the canonical zoom-adaptive LOD tiers (Point / Compact / Expanded) defined in graph_node_edge_interaction_spec.md §4.8. DOI emphasis operates as an overlay: it may promote a node's label or size upward from the zoom-LOD baseline (e.g., show a full label at Compact LOD when DOI is High), but it must never suppress interactions that the active LOD tier makes available. Node visibility for filtering is governed by the separate search/filter system. These concerns must not be entangled.

rendered_radius = base_radius * size_scale(tier)

Semantic fisheye applies cursor-distance scaling on top of DOI tier rendering. It is a separate, independently toggleable pass.

Invariant: node (x, y) positions are never modified. Only draw-size changes. Graph topology remains undistorted.

For each visible node in the current frame:

dist  = length(mouse_pos - node_pos)
scale = max(1.0, 3.0 * (1.0 - dist / fisheye_radius))

where fisheye_radius is a configurable canvas-space radius (default: 300.0 canvas units, stored in CanvasRegistry).

Apply the fisheye scale on top of the DOI-tier size:

final_radius = doi_radius * scale

Render pass z-order: sort visible nodes by final_radius ascending (paint smallest first; largest — most focused — on top).

• CanvasRegistry: policy toggles and parameters:
  • doi_rendering_enabled: bool
  • semantic_fisheye_enabled: bool
  • doi_update_cadence_ms: u32 (default: 100)
  • doi_max_hops: u8 (default: 5)
  • fisheye_radius: f32 (default: 300.0)
  • doi_weights: DoiWeights (α, β, γ, δ; must sum to 1.0)
• MetadataFrame: DOI score cache (HashMap<NodeIndex, DoiScore>) stored alongside the LOD zoom oracle; snapshot-swap updated by worker, read-only to render.
• DOI calculator worker: background task owned by graph-pane context; reads graph snapshot + node visit metadata; writes snapshot to MetadataFrame cache.
• Render path (render/graph_node_shape.rs or equivalent): reads cached DoiScore per node; applies size/opacity/LOD tier overrides; applies fisheye scale when semantic_fisheye_enabled.
• Multi-pane rule: DOI scores and fisheye calculations are computed per graph pane, consistent with existing culling/LOD per-pane isolation (multi_view_pane_spec.md).

The following must be in place before implementation begins:

1. Viewport culling (2026-02-24_performance_tuning_plan.md Phase 1) — DOI rendering should only iterate over the culled visible node set to stay within the 2 ms per-recalculation budget.
2. Node Label LOD (2026-02-24_performance_tuning_plan.md Phase 2.1) — DOI LOD tier overrides zoom-adaptive LOD upward only; the zoom-adaptive LOD path must be stable before DOI can layer on top of it.
3. MetadataFrame available as a stable per-pane cache oracle — DOI cache fields are added here; the frame struct must be settled before new fields are introduced.

• DOI score recomputation completes in ≤ 2 ms for a 1 000-node graph on reference hardware.
• DOI cache hit rate ≥ 99 % during interactive hover/selection at 100 ms cadence.
• doi_rendering_enabled = false produces rendering identical to pre-DOI baseline (no visual regression).
• semantic_fisheye_enabled = false does not affect DOI tier-based size/opacity/LOD behavior.
• Node (x, y) positions are unchanged by the fisheye pass — verified by snapshot comparison of node position map before and after.
• High-DOI nodes render on top of low-DOI nodes within the same render pass.
• Ghost tier nodes remain visible as faint dots; they are not hidden by the DOI layer.
• All DOI + fisheye policy toggles are governed by CanvasRegistry and not hardcoded in render callsites.
• DOI rendering does not mutate graph state; it is a read-only render pass over cached scores.