Transcript JSONL duplicates tool-result image payloads (message.content + toolUseResult), doubling image-turn size

In the session transcript JSONL, each tool-result image (and file) payload is written twice within the same line:

• once inside message.content[] as the {"type":"image","source":{"type":"base64","data":"…"}} block the model is shown, and
• again in the top-level toolUseResult field (the raw tool return), which carries an identical copy of the base64 data.

For image-heavy sessions this roughly doubles transcript size on the image bytes. Concrete example from our corpus: a single 18 MB transcript contained 5.3 MB of pure image duplication. Consequences:

• transcript files bloat on disk and over the wire (upload/resume cost);
• naive size heuristics (stat().st_size) over-count "information content" by ~2x on image turns, skewing clipping/retention/cost logic downstream;
• slower parse/scan of the JSONL.

Proposals:

1. Store each blob once and have the second site reference it (content hash / pointer) rather than re-embedding the base64.
2. Or omit the toolUseResult image data when it is byte-identical to the message.content copy (keep the metadata, drop the redundant bytes).

We've shipped a local read-side correction (per-entry blob dedup when sizing transcripts), but the on-disk duplication is upstream of us.