fix(loop): repair review-diff/verdict-parsing bugs, deduplicate session lookups

Applies the P0-P3 fixes from the multi-agent-mux-loop audit
(.mam/jobs/ab686e47/claude-reports/report-final.md):

- P0-1: capture BASE_COMMIT before Phase 2 and diff against it, so reviewer
  diffs stay non-empty and cumulative even after the Creator commits per the
  documented DoD (bare `git diff` alone showed nothing once committed).
- P0-2: has_verdict now matches only the report's last non-blank line, so a
  stray [VERDICT: ...] token quoted mid-report as a formatting example can no
  longer flip the outcome.
- P1-1: replace the English-only refactor/complex/design/architect keyword
  sniff (dead code against Korean-language reviewer reports) with an explicit
  [ESCALATE: PLANNER] tag the reviewer prompt now asks for.
- P1-2: resolve_all_reviewers/resolve_agent_type/resolve_planner_session now
  read through lib.sh's load_state_json single source of truth instead of
  each hand-rolling its own SQLite+YAML lookup; resolve_agent_type's name
  fallback matches exact hyphen segments instead of a substring `in` check.
- P2-1: warn when --all-reviewer and --reviewer are both given, since the
  latter is silently discarded.
- P2-2: correct the SKILL.md CLI-mapping table row that overstated an
  automated lint gate and an unconditional Planner feedback loop.
- P3: fix lib.sh shellcheck SC2164 (unguarded cd in start_watchdog) and
  annotate the intentional SC2317 dual source/exec guard.

Verified: shellcheck clean on both scripts, bash -n syntax OK, and the
rewritten has_verdict/resolve_* functions were unit-tested against this
repo's live .mam/agent-sessions state.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

.agents/skills/multi-agent-mux-loop/SKILL.md

@@ -143,7 +143,7 @@ sequenceDiagram
 | **Phase 2: Execution** | (기본값) | `--target-agent`로 명시한 주 작업 세션에 코딩 태스크를 주입합니다. |
 | **Phase 3: Review** | `--reviewer "A,B"` | 지정된 리뷰어 세션 리스트(`A`, `B` 등)에 교차 Peer Review를 위임합니다. |
 | **Phase 3: Consensus** | `--all-reviewer` | 레지스트리에 등록된 모든 active 리뷰어 세션을 자동으로 수집하여 리뷰를 돌립니다. (지정/수집된 모든 리뷰어의 PASS 만장일치가 항상 필요합니다.) |
-| **Iterative Loop** | `--max-loop M` | NOT PASS 또는 린트 실패 시 최대 `M`회까지 Planner와 Creator 간 피드백 루프를 반복합니다. |
+| **Iterative Loop** | `--max-loop M` | NOT PASS 판정 시 최대 `M`회까지 Creator가 자체 수정합니다. `--plan` 모드에서 리뷰어가 리포트에 `[ESCALATE: PLANNER]` 태그를 남기면 설계 변경 수준으로 판단하여 Planner에게 계획 갱신을 위임합니다 (린트는 리뷰어가 검토 관점 중 하나로 확인할 뿐, 별도의 자동 게이트는 아닙니다). |
 
 ---