perf(web): comprehensive graph rendering and UI performance overhaul

[QusayNaserFHS]

Summary

• Progressive node loading: For graphs >2k nodes, only Folders/Files/Classes visible initially. Functions/Methods/Variables hidden — reveal via Filters tab. Reduces visible nodes from 55k→10k.
• Edge optimization: Max 8-15 edges per node, skip structural edges (CONTAINS/DEFINES/IMPORTS) for 5k+, straight lines instead of curves for 5k+. Reduces edges from 186k→15k.
• Skip O(n²) layout: For 10k+ nodes, skip ForceAtlas2 entirely — hierarchical positions from graph-adapter are sufficient. Cached positions in sessionStorage for instant reload.
• Context splitting: New CodeRefsContext isolates AI citation updates so GraphCanvas/FileTree/Header don't re-render during chat streaming.
• O(1) lookups: Pre-built nodeMap, nodeIdSet, filePathIndex, fileNodeIndex replace all graph.nodes.find() O(n) calls in hot paths.
• Reducer fast-paths: For 2k+ graphs, skip dimming of non-active nodes — only modify selected/highlighted nodes. Turns O(N) allocations into O(k).
• Lazy mermaid: ~480KB loaded on first diagram render via import('mermaid') + SVG caching.
• Sigma tuning: 30fps animation throttle, color caching, disable hover/edge events for large graphs, adaptive label density.
• Bundle: 8 vendor chunks (added vendor-git, vendor-ml, vendor-markdown), debounced search, useTransition for filters, code viewer windowed to ~70 lines.
• FPS counter: Live overlay top-right for monitoring.

Test plan

• Load small repo zip — verify graph renders normally with all features
• Load agent_playground.zip (55k nodes) — verify FPS counter shows >30, interactions are responsive
• Click node — verify code panel opens, no lag
• Toggle Filters — verify node types show/hide (Functions initially hidden for large graphs)
• Use AI chat — verify no graph lag during streaming
• Refresh page after loading — verify cached layout restores instantly
• Check Mermaid diagrams render correctly in chat (lazy loaded)

🤖 Generated with Claude Code