perf(web): reduce startup bundle pressure for optional runtimes

[limityan]

Summary

This PR completes the P2 and P3 Web UI bundle optimization plan without changing chunkSizeWarningLimit, without lazy-loading the main session/chat path, and without removing product-facing capabilities. The work narrows accidental startup dependencies first, then applies guarded vendor chunking once the dependency graph is cleaner.

The PR is intentionally capped at 3 commits:

1. Document the staged bundle-splitting plan and measurement scripts.
2. Remove accidental startup dependencies for tools, Lucide namespace icons, Mermaid, and Prism.
3. Add P3 guarded vendor chunks plus targeted component-library hot-path import cleanup.

Optimized Scenarios

• App startup no longer imports the public ./tools barrel just to run post-render tool initialization. Startup now uses a narrow ./tools/initializeTools entry while Git/LSP/tool barrels remain compatible.
• Normal chat messages and plain Markdown no longer prepay Mermaid or Prism runtime cost. Mermaid loads only for Mermaid render/validate/export paths, and Prism enhances code blocks asynchronously after readable fallback text is already available.
• Context menu, MiniApp, and smart recommendation icons now use explicit Lucide icon maps instead of namespace lookup, preserving fallback behavior for unknown names while improving tree-shaking.
• P3 narrows @/component-library imports in chat input, pending queue, nav/session rows, model round header, and tool-card status paths so hot UI code does not go through the root component barrel unnecessarily.
• P3 adds guarded manualChunks for React, Monaco, terminal, Markdown, Tiptap, and Lucide icons, improving cache boundaries while keeping Mermaid/Prism out of index.html preloads.

Measured Benefit

Baseline before P2:

• Largest entry JS: 9953.0 KiB raw / 2802.0 KiB gzip
• JS/CSS total: 14332.4 KiB raw / 3905.1 KiB gzip
• Main entry static graph: 1067 local modules

After P2:

• Largest entry JS: 8242.6 KiB raw / 2286.1 KiB gzip
• JS/CSS total: 13952.4 KiB raw / 3930.0 KiB gzip
• Main entry static graph: 1013 local modules
• mermaid and react-syntax-highlighter are no longer statically reachable from src/web-ui/src/main.tsx

After P3:

• Largest entry asset: vendor-monaco-DdXR92sU.js at 3771.6 KiB raw / 959.3 KiB gzip
• App entry chunk: index-Nu1_3coB.js at 3226.0 KiB raw / 933.2 KiB gzip
• JS/CSS total: 13952.1 KiB raw / 3880.2 KiB gzip
• Main entry static graph: 1013 local modules
• dist/index.html does not preload Mermaid, Prism, syntax-highlight, or react-syntax-highlighter assets

Compared with P2, P3 reduces the largest single entry JS by about 4471.0 KiB raw / 1326.8 KiB gzip, and reduces the app entry itself by about 5016.6 KiB raw / 1352.9 KiB gzip. The remaining large warnings now point at real eager hotspots such as Monaco/editor/terminal rather than one monolithic app entry.

Verification

• pnpm run lint:web
• pnpm run type-check:web
• pnpm --dir src/web-ui run test:run
• pnpm --dir src/web-ui build
• pnpm run verify:monaco-assets
• node scripts/report-web-bundle-size.cjs --top=26
• node scripts/report-web-main-static-graph.cjs --assert-external-unreachable=react-syntax-highlighter --assert-external-unreachable=mermaid --assert-no-direct-import=src/web-ui/src/main.tsx:./tools --top=12
• Select-String -Path dist\index.html -Pattern mermaid|syntax|prism|react-syntax produced no output
• git diff --check

Note: pnpm run e2e:test:l0 was attempted during the P2 investigation, but the app shell did not become ready within 30000ms, so it is not counted as a passing verification signal.

[limityan]

Closed in favor of #746, which now contains the stage 0-4 implementation squashed into two commits and excludes the design document from the submitted diff.