Align cbox bootstrap/health docs with actual commands

Problem

The cbox documentation across canonical docs (docs/agent/cbox.md) and wrapper files (.cursor/rules/cbox.mdc, CLAUDE.md, CLAUDE_MD_TEMPLATE.md) shows bootstrap and git auth health-check commands that don't match what the entrypoint and health probes actually run. Operators following the documented commands get different behavior than the runtime, leading to confusion when diagnosing bootstrap failures or verifying sandbox health. Since cbox is the primary sandbox environment for AI agent workflows, stale command examples erode trust in the documentation and slow down incident triage.

Context

Possible Solutions

Plan

Implementation Progress

• Update docs to reflect actual bootstrap command (uv sync --frozen --extra dev with fallback and timeout behavior).
• Update docs to reflect actual git auth probe (GIT_TERMINAL_PROMPT=0 git ls-remote --exit-code -q origin HEAD).

• Keep canonical-first sync across wrappers.

• docs/agent/cbox.md matches runtime command behavior.

• Wrappers are in sync with canonical docs.

• No stale command examples remain for these checks.

• Source issue: host review flagged command mismatch causing operator confusion.

Review Feedback

• [ ] Review cleared