## Why
Serialized response items include message IDs, metadata, and JSON escaping that inflate token estimates without adding model-visible content.
## What changed
- Estimate each response item from its content, retaining JSON syntax for structured tool payloads.
- Apply image estimates to all image inputs, including non-base64 URLs, and count audio and encrypted content through their modality-specific estimates.
- Exclude plaintext reasoning and bookkeeping-only items from replay accounting.
## Testing
Update unit expectations for text, images, audio, and encrypted content. Add a remote compaction regression test showing that equal-length text produces identical token usage estimates despite different message IDs, metadata, and JSON escaping, while preserving the submitted messages.
GitOrigin-RevId: 1c43d5abbcc1668b4ab413901a2b4b6511ca5892
b04a2c2645