mirror of
https://github.com/hanzoai/semantic.git
synced 2026-10-30 15:04:19 +00:00
The 256 MiB cap was per part, so a file of a few KB could name one sheet, slide or shared string a hundred times, or decode a part into a tree 33 times its size, and reach gigabytes. A read now spends from one budget: each part's stated size each time it is read, a node per element decoded, and the text each table writes. Every charge checks the caller's context. A part claiming a size past int64 was let through as negative; it is refused. Style levels resolve in one walk, not one per style. Co-authored-by: Hanzo Dev <[email protected]>