Status: Implementation-Ready
Supersedes: 2026-02-11_performance_optimization_plan.md
Context: Post-Registry migration cleanup and scaling. Relates to: ../graph/multi_view_pane_spec.md (per-graph-pane culling/LOD semantics and graph-view isolation)

Target performance envelopes on reference hardware:

• 500 nodes @ 60 FPS
• 1000 nodes @ 30+ FPS

---

• CanvasRegistry: runtime toggles/policies (viewport_culling_enabled, label_culling_enabled, edge LOD mode).
• DiagnosticRegistry: frame-time/entity metrics with bounded sampling.
• MetadataFrame: per-graph-pane previous-frame zoom/pan oracle used for LOD and viewport decisions.
• render/spatial_index.rs: viewport candidate lookup using spatial index (avoid full O(N) scans).
• Multi-pane rule: culling and LOD decisions are computed per graph pane, not globally.

---

1. Compute visible world rect from previous-frame camera metadata.
2. Query spatial index for candidate visible nodes.
3. Submit only visible nodes to egui_graphs.

• Validate behavior when edge endpoints are culled.
• If renderer requires both endpoints present, add explicit edge filtering/ghost endpoint policy.

• Add/confirm viewport_culling_enabled in CanvasRegistry.

---

1. Zoom < 0.5: hide labels.
2. Zoom 0.5..=1.5: domain-only label.
3. Zoom > 1.5: full title.

• Rank visible nodes by importance (selection + graph importance).
• Greedy rect packing: skip intersecting lower-priority labels.

1. Zoom < 0.3: hide non-critical edges or all edges by policy.
2. Zoom < 0.8: reduced alpha/width, no arrowheads.
3. Zoom >= 0.8: full edge styling.

• Add/confirm label_culling_enabled and edge LOD policy in CanvasRegistry.

---

1. Cap animated badges per frame (initial cap: 20).
2. Disable animation for nodes beyond view-center distance threshold.
3. Fallback to static badge rendering when budget is exhausted.

---

• If average displacement remains below epsilon for N frames, pause simulation.
• Resume on structural/interaction intent.

• Bound physics update work per frame (initial budget target: ~5ms).
• Carry remaining simulation work to subsequent frames.
• Favor frame responsiveness over instant convergence.

• Keep Barnes-Hut as deferred path for sustained >1000 node regimes.

---

To avoid observer-effect distortion:

• Keep diagnostic aggregation/render sampling bounded (10 Hz target for diagnostic updates).
• Keep main UI/render loop unconstrained by diagnostic refresh cadence.

---

• 500-node benchmark at zoomed-out view meets 60 FPS target band.
• 1000-node benchmark meets 30+ FPS target band.
• Viewport culling produces expected visible-set reduction and frame-time drop.
• Label and edge LOD transitions occur at configured zoom thresholds.
• Physics budget cap prevents long-frame spikes during active simulation.
• Diagnostics enabled vs disabled does not materially skew benchmark readings.