Auto-injected `※ recap:` session-state messages appear without opt-in

Claude Code is injecting short "※ recap:" session-state summary messages
into the conversation. They look like:

※ recap: <one-line summary of what the user is doing, followed by a
suggested next action>

These appear to come from the harness itself (not generated by the model)
and I can't find a way to disable them:

• No recap, summary, or statusline-related key in ~/.claude/settings.json
• No project-level override in .claude/settings.local.json
• Nothing in /help that I can identify as the toggle

Request: please add a setting (e.g. autoRecap: false) or document the
existing one if it already exists. They're noisy and I'd like to turn them off.

Environment:

• OS: Linux 6.8.0-106-generic
• Shell: bash
• Claude Code model: Claude Opus 4.6 (1M context)