Problem

The pack reader currently loads the entire packfile into memory before parsing. Repositories with packfiles >500MB will cause excessive memory usage or OOM.

Impact

• Cannot serve or consume large repositories
• Memory usage is proportional to largest packfile, not working set
• Limits production viability

Acceptance Criteria

• Pack reader supports offset-based reads (seek + read range)
• Memory usage stays bounded regardless of packfile size
• Storage behaviour extended with stream_pack / random-access read support
• Works with all three backends (filesystem, S3, memory)
• Existing pack reader tests continue to pass
• Added benchmark comparing memory usage for large packs