fix: Graph DB should persist file content for real depth filtering

[VictorGjn]

Problem

The graph packer (#135) now calls depthFilter + renderFilteredMarkdown for real depth-filtered content. But FileNode.content is usually undefined because the in-memory graph DB doesn't store raw file content after scanning.

This means the packer falls back to stub generation in most cases, negating the bridge fix.

Success criteria

• After POST /api/graph/scan, file content is stored (or retrievable) in the graph DB
• packContext() produces real depth-filtered content for scanned files
• Memory-bounded: content is lazy-loaded or evicted based on token budget

Implementation

1. Store content in FileNode during scan (or in a parallel content cache)
2. Lazy-load from disk when packer needs it (for large repos)
3. LRU eviction based on total cached tokens