feat: full Clone MCP session integration (start/record/feedback/stop)

[Turtle-Hwan]

Warning

Review carefully before merge. This PR adds Clone MCP session memory recording in addition to prediction: session tracking, prompt/response recording, feedback submission, and session stop calls. Prediction and local safety gates still follow the latest main behavior.

Summary

• Add scripts/clone-mcp.mjs as the shared Clone MCP action layer for predict_next_prompt, start_session, stop_session, record_agent_prompt, record_agent_response, and submit_feedback, built on the existing shared MCP client/auth helpers from main.
• Start a Clone session during /clone:loop, record the initial user prompt, and persist clone_session_id, mcp_session_id, and last_prompt_event_id in local loop state when available.
• Record what the agent actually did:
  • Stop hook records the current iteration assistant response before prediction.
  • PostToolUse records Write, Edit, MultiEdit, and apply_patch mutation summaries.
  • Accepted Clone-predicted continuation prompts and AskUserQuestion auto-answers are recorded as prompt events.
• Submit accepted/rejected feedback for predictions and stop the Clone session on terminal loop paths or explicit /clone:cancel-loop.
• Keep storage-side MCP failures non-fatal: record/bootstrap/feedback/stop failures are logged to local history and do not break the existing loop behavior.

Alignment with merged direction

• Preserves the stale-loop safety model from PR Prevent stale Clone Loop auto-resume #15 / CLO-46: Stop, AskUserQuestion, and PostToolUse still share validateActiveLoopState, strict session matching, loop-start history validation, and TTL handling.
• Preserves the shared client/auth/data split from PR Connect API key setup to dashboard session ping #20 and PR Strengthen Clone Interview goal-to-plan flow #21: MCP transport/auth/version identity stay in shared helpers, while clone-mcp.mjs centralizes Clone-specific actions.
• Preserves the AskUserQuestion threshold/status gating from PR Fix Clone Loop AskUserQuestion threshold gating #22 / CLO-36: low-confidence, non-auto, unmapped, or MCP-failed question predictions defer to the user instead of choosing a fallback answer.
• Extends the multi-turn/rich-context direction from earlier merged PRs: prior work still feeds prediction context, and this PR adds persistent session memory writes so Clone can learn from prompts, responses, tool use, and feedback.
• Keeps Codex plugin direction from PR [codex] add Codex Clone Loop hooks #17 intact; this PR focuses on shared Clone Loop memory/session behavior used by the Claude hooks and compatible shared scripts.

Failure behavior

• start_session / initial record_agent_prompt failure: loop still starts, with a warning.
• record_agent_response, record_agent_prompt, submit_feedback, or stop_session failure: hook continues its existing control flow and appends an error record to .claude/clone-loop.history.local.jsonl.
• predict_next_prompt failure: loop does not fabricate a continuation. Stop hook escalates/stops; AskUserQuestion defers to the real user. This matches the latest safe gating behavior on main.

Validation

• npm test -> 72/72 pass
• claude plugin validate . -> pass
• git diff --check -> pass

Manual follow-ups before undrafting

• Run /clone:loop "..." in a real Claude Code session and confirm state frontmatter receives clone_session_id.
• Let the loop run for 2+ iterations and confirm record-response, record-prompt, and feedback-sent local history events.
• Trigger low-confidence escalation and confirm rejected feedback plus session-stopped are logged without prompt injection.
• Run /clone:cancel-loop and confirm the state file is removed and stop-session errors, if any, are non-fatal.