Merge pull request 'docs: add live-operation issue tracking design and plan (#80)' (#395) from task/80-implementation-audit-closeout into base/ouroboros-workflow-20260302

Reviewed-on: #395

docs/commands.md

@@ -59,6 +59,18 @@ scripts/tea_comment.sh 374 /tmp/comment.md
 - `scripts/tea_comment.sh` accepts stdin with `-` as body source.
 - The helper fails fast when body looks like escaped-newline text only.
 
+#### PR Body Post-Check (Mandatory)
+
+PR 생성 직후 본문이 `\n` 문자열로 깨지지 않았는지 반드시 확인한다.
+
+```bash
+python3 scripts/validate_pr_body.py --pr <PR_NUMBER>
+```
+
+검증 실패 시:
+- PR 본문을 API patch 또는 파일 기반 본문으로 즉시 수정
+- 같은 명령으로 재검증 통과 후에만 리뷰/머지 진행
+
 #### ❌ TTY Error - Interactive Confirmation Fails
 ```bash
 ~/bin/tea issues create --repo X --title "Y" --description "Z"

docs/ouroboros/01_requirements_registry.md

@@ -1,6 +1,6 @@
 <!--
 Doc-ID: DOC-REQ-001
-Version: 1.0.3
+Version: 1.0.12
 Status: active
 Owner: strategy
 Updated: 2026-03-02
@@ -19,14 +19,14 @@ Updated: 2026-03-02
 - `REQ-V2-005`: 라벨링은 Triple Barrier(Upper/Lower/Time) 방식이어야 한다.
 - `REQ-V2-006`: 검증은 Walk-forward + Purge/Embargo를 강제한다.
 - `REQ-V2-007`: 백테스트는 비용/슬리피지/체결실패를 반영하지 않으면 채택 불가다.
-- `REQ-V2-008`: Kill Switch는 신규주문차단 -> 미체결취소 -> 재조회 -> 리스크축소 -> 스냅샷 순서다.
+- `REQ-V2-008`: Kill Switch는 신규주문차단 -> 미체결취소 -> 재조회(실패 시 최대 3회, 1s/2s backoff 재시도, 성공 시 즉시 중단) -> 리스크축소 -> 스냅샷 순서다.
 
 ## v3 핵심 요구사항
 
 - `REQ-V3-001`: 모든 신호/주문/로그는 `session_id`를 포함해야 한다.
 - `REQ-V3-002`: 세션 전환 시 리스크 파라미터 재로딩이 수행되어야 한다.
 - `REQ-V3-003`: 브로커 블랙아웃 시간대에는 신규 주문이 금지되어야 한다.
-- `REQ-V3-004`: 블랙아웃 중 신호는 Queue에 적재되고, 복구 후 유효성 재검증을 거친다.
+- `REQ-V3-004`: 블랙아웃 중 신호는 bounded Queue에 적재되며, 포화 시 oldest-drop 정책으로 최신 intent를 보존하고 복구 후 유효성 재검증을 거친다.
 - `REQ-V3-005`: 저유동 세션(`NXT_AFTER`, `US_PRE`, `US_DAY`, `US_AFTER`)은 시장가 주문 금지다.
 - `REQ-V3-006`: 백테스트 체결가는 불리한 방향 체결 가정을 기본으로 한다.
 - `REQ-V3-007`: US 운용은 환율 손익 분리 추적과 통화 버퍼 정책을 포함해야 한다.
@@ -38,3 +38,9 @@ Updated: 2026-03-02
 - `REQ-OPS-002`: 문서의 수치 정책은 원장에서만 변경한다.
 - `REQ-OPS-003`: 구현 태스크는 반드시 테스트 태스크를 동반한다.
 - `REQ-OPS-004`: 원본 계획 문서(`v2`, `v3`)는 `docs/ouroboros/source/` 경로를 단일 기준으로 사용한다.
+
+## 변경 이력
+
+- 2026-03-02: `v1.0.12` 문서 검증 게이트 강화(#390) 반영에 따라 정책 문서 동기화 체크를 수행했다. (`REQ-OPS-002`)
+
+- 2026-03-04: 86_peak_probability_model_plan.md 초안 추가에 따른 정책 문서 동기화 체크를 수행했다. (REQ-OPS-002)

docs/ouroboros/80_implementation_audit.md

@@ -9,7 +9,7 @@ Updated: 2026-03-02
 # v2/v3 구현 감사 및 수익률 분석 보고서
 
 작성일: 2026-02-28
-최종 업데이트: 2026-03-02 (#373 상태표 정합화 반영)
+최종 업데이트: 2026-03-02 (#377 kill-switch refresh 재시도 정책 반영)
 대상 기간: 2026-02-25 ~ 2026-02-28 (실거래)
 분석 브랜치: `feature/v3-session-policy-stream`
 
@@ -32,23 +32,23 @@ Updated: 2026-03-02
 | REQ-V2-001 | 4-상태 매도 상태기계 (HOLDING→BE_LOCK→ARMED→EXITED) | `src/strategy/position_state_machine.py` | ✅ 완료 |
 | REQ-V2-002 | 즉시 최상위 상태 승격 (갭 대응) | `position_state_machine.py:51-70` | ✅ 완료 |
 | REQ-V2-003 | EXITED 우선 평가 | `position_state_machine.py:38-48` | ✅ 완료 |
-| REQ-V2-004 | 4중 청산 로직 (Hard/BE/ATR Trailing/Model) | `src/strategy/exit_rules.py` | ⚠️ 부분 (`#369`) |
+| REQ-V2-004 | 4중 청산 로직 (Hard/BE/ATR Trailing/Model assist-only, 직접 EXIT 미트리거) | `src/strategy/exit_rules.py` | ✅ 완료 |
 | REQ-V2-005 | Triple Barrier 라벨링 | `src/analysis/triple_barrier.py` | ✅ 완료 |
 | REQ-V2-006 | Walk-Forward + Purge/Embargo 검증 | `src/analysis/walk_forward_split.py` | ✅ 완료 |
-| REQ-V2-007 | 비용/슬리피지/체결실패 모델 필수 | `src/analysis/backtest_cost_guard.py` | ⚠️ 부분 (`#368`) |
-| REQ-V2-008 | Kill Switch 실행 순서 (Block→Cancel→Refresh→Reduce→Snapshot) | `src/core/kill_switch.py` | ⚠️ 부분 (`#377`) |
+| REQ-V2-007 | 비용/슬리피지/체결실패 모델 필수 | `src/analysis/backtest_cost_guard.py`, `src/analysis/backtest_pipeline.py` | ✅ 완료 |
+| REQ-V2-008 | Kill Switch 실행 순서 (Block→Cancel→Refresh(retry)→Reduce→Snapshot) | `src/core/kill_switch.py` | ✅ 완료 |
 
 ### 1.3 v3 구현 상태: 부분 완료 (2026-03-02 기준)
 
 | REQ-ID | 요구사항 | 상태 | 비고 |
 |--------|----------|------|------|
 | REQ-V3-001 | 모든 신호/주문/로그에 session_id 포함 | ⚠️ 부분 | 큐 intent에 `session_id` 누락 (`#375`) |
-| REQ-V3-002 | 세션 전환 훅 + 리스크 파라미터 재로딩 | ⚠️ 부분 | 구현 존재, 세션 경계 E2E 회귀 보강 필요 (`#376`) |
+| REQ-V3-002 | 세션 전환 훅 + 리스크 파라미터 재로딩 | ✅ 완료 | 세션 경계 E2E 회귀(override 적용/해제 + 재로딩 실패 폴백) 보강 (`#376`) |
 | REQ-V3-003 | 블랙아웃 윈도우 정책 | ✅ 완료 | `src/core/blackout_manager.py` |
-| REQ-V3-004 | 블랙아웃 큐 + 복구 시 재검증 | ⚠️ 부분 | 큐 포화 시 intent 유실 경로 존재 (`#371`), 재검증 강화를 `#328`에서 추적 |
+| REQ-V3-004 | 블랙아웃 큐 + 복구 시 재검증 | ✅ 완료 | DB 기록(`#324`), 재검증 강화(`#328`), 큐 포화 oldest-drop(`#371`) 반영 |
 | REQ-V3-005 | 저유동 세션 시장가 금지 | ✅ 완료 | `src/core/order_policy.py` |
 | REQ-V3-006 | 보수적 백테스트 체결 (불리 방향) | ✅ 완료 | `src/analysis/backtest_execution_model.py` |
-| REQ-V3-007 | FX 손익 분리 (전략 PnL vs 환율 PnL) | ⚠️ 부분 | 스키마 존재, 런타임 분리 계산/전달 미적용 (`#370`) |
+| REQ-V3-007 | FX 손익 분리 (전략 PnL vs 환율 PnL) | ⚠️ 부분 | 런타임 분리 계산/전달 적용 (`#370`), buy-side `fx_rate` 미관측 시 `fx_pnl=0` fallback |
 | REQ-V3-008 | 오버나잇 예외 vs Kill Switch 우선순위 | ✅ 완료 | `src/main.py` — `_should_force_exit_for_overnight()`, `_apply_staged_exit_override_for_hold()` |
 
 ### 1.4 운영 거버넌스: 부분 완료 (2026-03-02 재평가)
@@ -80,22 +80,22 @@ Updated: 2026-03-02
 - **해소**: #326 머지 — `log_trade()` 호출 시 런타임 `session_id` 명시적 전달
 - **요구사항**: REQ-V3-001
 
-### GAP-3: 세션 전환 시 리스크 파라미터 재로딩 없음 → ⚠️ 부분 해소 (#327)
+### GAP-3: 세션 전환 시 리스크 파라미터 재로딩 없음 → ✅ 해소 (#327, #376)
 
 - **위치**: `src/main.py`, `src/config.py`
 - **해소 내용**: #327 머지 — `SESSION_RISK_PROFILES_JSON` 기반 세션별 파라미터 재로딩 메커니즘 구현
   - `SESSION_RISK_RELOAD_ENABLED=true` 시 세션 경계에서 파라미터 재로딩
   - 재로딩 실패 시 기존 파라미터 유지 (안전 폴백)
-- **잔여 갭**: 세션 경계 실시간 전환 E2E 통합 테스트 보강 필요 (`test_main.py`에 설정 오버라이드/폴백 단위 테스트는 존재)
+- **해소**: 세션 경계 E2E 회귀 테스트를 추가해 override 적용/해제, 재로딩 실패 시 폴백 유지를 검증함 (`#376`)
 - **요구사항**: REQ-V3-002
 
-### GAP-4: 블랙아웃 복구 DB 기록 + 재검증 → ⚠️ 부분 해소 (#324, #328, #371)
+### GAP-4: 블랙아웃 복구 DB 기록 + 재검증 → ✅ 해소 (#324, #328, #371)
 
 - **위치**: `src/core/blackout_manager.py`, `src/main.py`
 - **현 상태**:
-  - #324 추적 범위(DB 기록)는 구현 경로가 존재
-  - #328 범위(가격/세션 재검증 강화)는 추적 이슈 오픈 상태
-  - #371: 큐 포화 시 intent 유실 경로가 남아 있어 `REQ-V3-004`를 완료로 보기 어려움
+  - #324: 복구 주문 DB 기록 구현 및 테스트 반영
+  - #328: 가격/세션 재검증 강화 구현 및 머지 완료
+  - #371: 큐 포화 정책을 oldest-drop으로 명시/구현해 최신 intent 유실 경로 제거
 - **요구사항**: REQ-V3-004
 
 ### GAP-5: 시간장벽이 봉 개수 고정 → ✅ 해소 (#329)
@@ -107,10 +107,12 @@ Updated: 2026-03-02
   - `max_holding_bars` deprecated 경고 유지 (하위 호환)
 - **요구사항**: REQ-V2-005 / v3 확장
 
-### GAP-6 (신규): FX PnL 분리 미완료 (MEDIUM — 부분 구현)
+### GAP-6 (신규): FX PnL 분리 부분 해소 (MEDIUM)
 
 - **위치**: `src/db.py` (`fx_pnl`, `strategy_pnl` 컬럼 존재)
-- **문제**: 스키마와 함수는 존재하지만 런타임 경로에서 `strategy_pnl`/`fx_pnl` 분리 계산 전달이 누락됨 (`#370`)
+- **현 상태**: 런타임 SELL 경로에서 `strategy_pnl`/`fx_pnl` 분리 계산 및 전달을 적용함 (`#370`).
+- **운영 메모**: `trading_cycle`은 scanner 기반 `selection_context`에 `fx_rate`를 추가하고, `run_daily_session`은 scanner 컨텍스트 없이 `fx_rate` 스냅샷만 기록한다.
+- **잔여**: 과거 BUY 레코드에 `fx_rate`가 없으면 해외 구간도 `fx_pnl=0` fallback으로 기록됨.
 - **영향**: USD 거래에서 환율 손익과 전략 손익이 분리되지 않아 성과 분석 부정확
 - **요구사항**: REQ-V3-007
 
@@ -326,7 +328,7 @@ Updated: 2026-03-02
 | 블랙아웃 복구 주문 `log_trade()` 추가 (GAP-4) | #324 | ✅ 머지 |
 | 세션 전환 리스크 파라미터 동적 재로딩 (GAP-3) | #327 | ✅ 머지 |
 | session_id 거래/의사결정 로그 명시 전달 (GAP-1, GAP-2) | #326 | ✅ 머지 |
-| 블랙아웃 복구 가격/세션 재검증 강화 (GAP-4 잔여) | #328 | ✅ 머지 |
+| 블랙아웃 복구 가격/세션 재검증 강화 (GAP-4) | #328 | ✅ 머지 |
 
 **잔여 개선 항목:**
 
@@ -335,7 +337,6 @@ Updated: 2026-03-02
 | P1 | US 시장 ATR 공급 경로 완성 (ROOT-5 잔여) | 중간 |
 | P1 | FX PnL 운영 활성화 (REQ-V3-007) | 낮음 |
 | P2 | pred_down_prob ML 모델 대체 (ROOT-5 잔여) | 높음 |
-| P2 | 세션 경계 E2E 통합 테스트 보강 (GAP-3 잔여) | 낮음 |
 
 ### 5.3 권장 실행 순서
 
@@ -392,8 +393,7 @@ Phase 3 (중기): v3 세션 최적화
 
 ### 테스트 미존재 (잔여)
 
-- ❌ 세션 전환 훅 콜백 (GAP-3 잔여)
-- ❌ 세션 경계 리스크 파라미터 재로딩 단위 테스트 (GAP-3 잔여)
+- ✅ 세션 전환 훅 콜백/세션 경계 리스크 재로딩 E2E 회귀 (`#376`)
 - ❌ 실거래 경로 ↔ v2 상태기계 통합 테스트 (피처 공급 포함)
 - ❌ FX PnL 운영 활성화 검증 (GAP-6)
 

docs/ouroboros/86_peak_probability_model_plan.md

@@ -0,0 +1,168 @@
+<!--
+Doc-ID: DOC-PLAN-086
+Version: 0.1.0
+Status: draft
+Owner: strategy
+Updated: 2026-03-02
+-->
+
+# 고점 확률 모델 도입 계획 (초안)
+
+작성일: 2026-03-02  
+목적: 기존 계획(v2)의 "모델 보조 신호"를, 실운용 가능한 단순한 형태로 재정의한다.
+
+---
+
+## 1. 한 줄 요약
+
+- 목표는 "지금이 고점(또는 고점 근처)일 확률"을 예측해 익절/축소/오버나잇 리스크 결정을 보조하는 것이다.
+- 1차는 정규장 중심으로 시작하고, NXT/US_PRE 등 확장세션은 이후 분리 대응한다.
+
+---
+
+## 2. 왜 바꾸는가
+
+- 기존 논의는 범위가 넓어 구현/검증 비용이 컸다.
+- 현재 로그 구조에서도 빠르게 시작 가능한 타깃이 필요하다.
+- 실거래 효용 기준에서 "고점 근처 감지"는 청산 품질 개선에 직접 연결된다.
+
+---
+
+## 3. 이번 계획의 핵심 결정
+
+### 3.1 예측 타깃 (2개)
+
+- `p_top_d0`: 현재 시점이 "오늘 정규장 고점(근접 포함)"일 확률
+- `p_top_d1`: 현재 시점이 "내일 정규장 마감까지 포함한 고점(근접 포함)"일 확률
+
+### 3.2 모델 역할
+
+- 단독 매도 트리거가 아니라 보조 신호로 사용한다.
+- Hard stop / ATR trailing / Kill Switch 우선 원칙은 유지한다.
+
+### 3.3 적용 순서
+
+1. KR/US 정규장만 대상으로 MVP 구축
+2. 성능/운영 안정화 후 NXT, US_PRE, US_AFTER 분리 확장
+
+---
+
+## 4. 데이터/라벨 정의 (MVP)
+
+### 4.1 샘플 단위
+
+- 단위: "시점 t의 관측값 1개"
+- 대상: 주문 가능 시점 중 정책 게이트를 통과하는 시점
+
+### 4.2 라벨 정의
+
+- `y_d0 = 1` if `max_high(t ~ 오늘 정규장 마감) <= price_t * (1 + eps_d0)` else `0`
+- `y_d1 = 1` if `max_high(t ~ 내일 정규장 마감) <= price_t * (1 + eps_d1)` else `0`
+
+### 4.3 필수 피처 (초기)
+
+- 가격/수익률: `current_price`, `price_change_pct`
+- 미시지표: `rsi`, `volume_ratio`, `signal`, `score`
+- 포트폴리오: `portfolio_pnl_pct`, `total_cash`(선택)
+- 상태: `market`, `session_id`, `minutes_to_close`, `is_tradeable_now`
+
+### 4.4 데이터 제외 규칙 (초기)
+
+- `session_id in ('', 'UNKNOWN')` 제외
+- `price <= 0` 또는 `quantity <= 0` 제외
+- `rationale LIKE '[startup-sync]%'` 제외
+- 라벨 윈도우 불완전(휴장/데이터 누락) 샘플 제외
+
+---
+
+## 5. 추론/실행 규칙 (초안)
+
+### 5.1 추론 출력
+
+```json
+{
+  "p_top_d0": 0.00,
+  "p_top_d1": 0.00,
+  "model_version": "TBD"
+}
+```
+
+### 5.2 실행 반영 원칙
+
+- `p_top_d0`가 높으면: 당일 익절/트레일링 강화
+- `p_top_d1`가 높으면: 오버나잇 보유 축소/청산 우선
+- 확장세션(NXT/US_PRE)은 초기엔 보수 임계치 또는 실행 제한
+
+---
+
+## 6. 세션 이슈 대응 원칙
+
+### 6.1 거래 가능 세션 기준
+
+- 달력 시간이 아니라 "종목이 실제 거래 가능한 세션" 기준으로 라벨/추론을 구성한다.
+
+### 6.2 정규장 vs 확장세션 분리
+
+- MVP: 정규장 전용 라벨/모델
+- 확장: 세션별 헤드 또는 세션별 모델 분리 검토
+
+### 6.3 고점 정의 분리
+
+- `reg_high`: 정규장 고점 (주 라벨)
+- `ext_high`: 확장세션 포함 고점 (보조 분석)
+
+---
+
+## 7. 단계별 구현 계획
+
+### Phase A: 계약 고정
+
+- 학습용 canonical dataset 스키마 확정
+- 시점 정렬(as-of) 규칙 문서화
+- 라벨 생성 함수/검증 테스트 추가
+
+### Phase B: 오프라인 검증
+
+- Walk-forward + Purge/Embargo 평가
+- `d0`, `d1` 분리 리포트
+- B1(모델 없음) 대비 개선 여부 확인
+
+### Phase C: 런타임 통합
+
+- 추론 결과를 `pred_top_d0/pred_top_d1`로 주입
+- 매도/축소 보조 규칙 연결
+- Shadow 모드로 관측 후 제한적 활성화
+
+### Phase D: 세션 확장
+
+- NXT/US_PRE/US_AFTER 정책 분리
+- 실행 게이트/임계치 세션별 튜닝
+
+---
+
+## 8. 지금 당장 결정할 최소 항목 (TBD)
+
+- [ ] `eps_d0`, `eps_d1` 기본값 (`TBD`)
+- [ ] `p_top_d0`, `p_top_d1` 실행 임계치 (`TBD`)
+- [ ] 정규장/확장세션 대상 시장 범위 (`TBD`)
+- [ ] 최소 학습 샘플 수 기준 (`TBD`)
+- [ ] Shadow 관측 기간 (`TBD`)
+
+---
+
+## 9. 리스크 및 완화
+
+- 라벨 불균형: class weight, threshold 튜닝
+- 시간대 편향(장마감 근처): 시간 피처 추가 + 구간별 성능 보고
+- 데이터 무결성: 제외 규칙 강제 + CI 검증
+- 복잡도 증가: 정규장 MVP를 먼저 잠그고 확장세션은 후속 분리
+
+---
+
+## 10. 후속 문서 연결 (예정)
+
+- requirements registry: 신규 REQ/TASK/TEST 매핑 추가
+- code-level work orders: 구현 작업 분해
+- acceptance/test plan: `p_top_d0/d1` 검증 항목 추가
+- implementation audit: 기존 "모델 보조 신호" 항목 업데이트
+

docs/plans/2026-03-02-dual-track-workflow-execution-plan.md

@@ -0,0 +1,211 @@
+# Dual-Track Workflow Execution Plan (Revised)
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Re-seed workflow planning from a clean base branch (not the temporary branch), then execute isolated `86` and `80` tracks with real-operation-confirmed closeout and related Gitea issue closure.
+
+**Architecture:** Start from `feature/v3-session-policy-stream` and create `base/ouroboros-workflow-20260302`. Bring only two workflow plan docs as one squashed seed commit. Branch `task/86...` and `task/80...` from that base with strict file-scope separation. For `80`, update audit only after real-operation confirmation, then close related Gitea issues (children before parent). Final promotion happens by merging the validated stream to `main`.
+
+**Tech Stack:** Git CLI, Gitea (`tea`) issue/PR workflow, Markdown docs (`docs/ouroboros`, `docs/plans`)
+
+---
+
+### Task 1: Freeze temporary branch and identify export scope
+
+**Files:**
+- Modify: none
+- Create: none
+- Test: none
+
+**Step 1: Inspect temporary branch state**
+
+Run: `git status --short --branch`
+Expected: current temporary branch with local changes listed.
+
+**Step 2: Preserve changes using stash (recommended)**
+
+Run: `git stash push -u -m "tmp-before-workflow-reseed-2026-03-02"`
+Expected: stash created with tracked/untracked changes.
+
+**Step 3: Define transferable artifacts**
+
+Run: keep only these for reseed:  
+- `docs/plans/2026-03-02-ouroboros-dual-track-workflow-design.md`  
+- `docs/plans/2026-03-02-dual-track-workflow-execution-plan.md`
+Expected: transfer scope fixed to two files.
+
+**Step 4: Verify clean working tree**
+
+Run: `git status --short --branch`
+Expected: no local changes.
+
+**Step 5: Commit**
+
+```bash
+# no commit in this task (stash-only preservation)
+```
+
+### Task 2: Create clean official base and seed one commit
+
+**Files:**
+- Modify: `docs/plans/2026-03-02-ouroboros-dual-track-workflow-design.md`
+- Modify: `docs/plans/2026-03-02-dual-track-workflow-execution-plan.md`
+- Test: none
+
+**Step 1: Switch to canonical base**
+
+Run: `git checkout feature/v3-session-policy-stream`
+Expected: branch switch successful.
+
+**Step 2: Update local base**
+
+Run: `git pull --ff-only origin feature/v3-session-policy-stream`
+Expected: up-to-date or fast-forward update.
+
+**Step 3: Create staging base branch**
+
+Run: `git checkout -b base/ouroboros-workflow-20260302`
+Expected: new branch created from updated feature base.
+
+**Step 4: Restore only two plan docs from temporary branch commits**
+
+Run (example): `git checkout <temp-branch> -- docs/plans/2026-03-02-ouroboros-dual-track-workflow-design.md docs/plans/2026-03-02-dual-track-workflow-execution-plan.md`
+Expected: only two files restored in working tree.
+
+**Step 5: Create one squashed seed commit**
+
+Run: `git add docs/plans/2026-03-02-ouroboros-dual-track-workflow-design.md docs/plans/2026-03-02-dual-track-workflow-execution-plan.md && git commit -m "docs: seed revised dual-track workflow plan"`
+Expected: single commit with two files only.
+
+**Step 6: Verify seed scope**
+
+Run: `git show --name-only --oneline HEAD`
+Expected: exactly the two plan docs listed.
+
+**Step 7: Push official base**
+
+Run: `git push -u origin base/ouroboros-workflow-20260302`
+Expected: remote tracking set.
+
+### Task 3: Open Track A (`86`) branch and initial PR
+
+**Files:**
+- Modify: `docs/ouroboros/86_peak_probability_model_plan.md`
+- Test: none
+
+**Step 1: Create Track A branch**
+
+Run: `git checkout -b task/86-peak-probability-plan`
+Expected: branch created from official base.
+
+**Step 2: Stage only `86` document**
+
+Run: `git add docs/ouroboros/86_peak_probability_model_plan.md`
+Expected: only `86` doc staged.
+
+**Step 3: Commit initial PR seed**
+
+Run: `git commit -m "docs: seed peak probability model plan (#86)"`
+Expected: one-document commit.
+
+**Step 4: Push branch**
+
+Run: `git push -u origin task/86-peak-probability-plan`
+Expected: remote branch created.
+
+**Step 5: Open PR-A**
+
+Run: `tea pr create --base base/ouroboros-workflow-20260302 --head task/86-peak-probability-plan --title "docs: seed peak probability model plan (#86)" --description "Initial planning PR for iterative concretization."`
+Expected: PR URL returned.
+
+### Task 4: Open Track B (`80`) branch for operation-confirmed closeout
+
+**Files:**
+- Modify: `docs/ouroboros/80_implementation_audit.md`
+- Test: operation observation evidence (outside unit-test scope)
+
+**Step 1: Return to staging base**
+
+Run: `git checkout base/ouroboros-workflow-20260302`
+Expected: branch switch successful.
+
+**Step 2: Create Track B branch**
+
+Run: `git checkout -b task/80-implementation-audit-closeout`
+Expected: branch created.
+
+**Step 3: Apply closeout edits only after real-operation confirmation**
+
+Run: edit `docs/ouroboros/80_implementation_audit.md`
+Expected: status/evidence sections updated with observed behavior, confirmation timestamp, and operational proof summary.
+
+**Step 4: Commit closeout update**
+
+Run: `git add docs/ouroboros/80_implementation_audit.md && git commit -m "docs: close out implementation audit items after operation confirmation (#80)"`
+Expected: single-scope commit.
+
+**Step 5: Push and open PR-B**
+
+Run: `git push -u origin task/80-implementation-audit-closeout`
+Expected: remote branch created.
+
+Run: `tea pr create --base base/ouroboros-workflow-20260302 --head task/80-implementation-audit-closeout --title "docs: close out implementation audit items (#80)" --description "Closeout based on real-operation confirmation."`
+Expected: PR URL returned.
+
+### Task 5: Close related Gitea issues after `80` confirmation
+
+**Files:**
+- Modify: `docs/ouroboros/80_implementation_audit.md` (reference linkage only)
+- Test: issue parent-child completeness check
+
+**Step 1: Build related issue set from updated `80` audit**
+
+Run: parse issue references from `80` updated sections.
+Expected: target open issue list prepared.
+
+**Step 2: Close child issues first**
+
+Run: `tea issues close <child-issue-number>` for each confirmed child.
+Expected: all qualifying child issues closed.
+
+**Step 3: Close parent issues only when all children are closed**
+
+Run: `tea issues close <parent-issue-number>`
+Expected: parent closes only after dependency check passes.
+
+**Step 4: Record closure summary in PR comment**
+
+Run: add PR comment with closed issue IDs and parent-child resolution note.
+Expected: auditable trace for closeout.
+
+### Task 6: Validate branch hygiene and merge path
+
+**Files:**
+- Modify: none
+- Test: git diff/log checks
+
+**Step 1: Verify PR-A scope**
+
+Run: `git diff --name-only base/ouroboros-workflow-20260302...task/86-peak-probability-plan`
+Expected: only `docs/ouroboros/86_peak_probability_model_plan.md` (and intentional related plan files if any).
+
+**Step 2: Verify PR-B scope**
+
+Run: `git diff --name-only base/ouroboros-workflow-20260302...task/80-implementation-audit-closeout`
+Expected: only `docs/ouroboros/80_implementation_audit.md` (and intentional related plan files if any).
+
+**Step 3: Merge both PRs into staging base**
+
+Run: GitHub merge flow for PR-A and PR-B
+Expected: staging base includes both tracks without cross-file contamination.
+
+**Step 4: Promote validated stream to `main`**
+
+Run: open PR from `feature/v3-session-policy-stream` (or its validated promotion branch) to `main`
+Expected: merge-based history preservation with no cherry-pick/rebase surgery.
+
+**Step 5: Commit**
+
+```bash
+# no additional commit required in this task
+```

docs/plans/2026-03-02-live-operation-issue-tracking-design.md

@@ -0,0 +1,100 @@
+# Live Operation Issue Tracking Design
+
+**Date:** 2026-03-02  
+**Scope:** Real-operation confirmation tracking for all open Gitea issues, tied to `task/80-implementation-audit-closeout`.
+
+---
+
+## 1. Goal
+
+Define an issue-by-issue tracking method for real-operation confirmation so that:
+
+- every open issue is tracked with explicit lifecycle states,
+- parent-child close order is enforced,
+- and final closeout is reflected in `80` audit work without loss.
+
+---
+
+## 2. Target Set
+
+- Coverage target: all current open Gitea issues (`318~381`, and future open issues during this closeout window).
+- Tracking starts from open set snapshot and stays synced until closeout completion.
+
+---
+
+## 3. Tracking Architecture
+
+- Primary log: direct comments on each issue ticket.
+- Aggregate log: single summary comment on the `80` PR (`task/80-implementation-audit-closeout`).
+- No separate tracking doc as source of truth during execution.
+
+---
+
+## 4. State Model and Criteria
+
+State machine:
+
+- `NOT_STARTED`
+- `OBSERVING`
+- `CONFIRMED`
+- `CLOSED`
+
+Rules:
+
+- `CONFIRMED` requires one real-operation observation evidence item.
+- `CLOSED` requires close conditions satisfied and ticket close action executed.
+- Parent issue cannot become `CLOSED` until all child issues are `CLOSED`.
+
+---
+
+## 5. Issue Comment Template
+
+Each issue update comment uses:
+
+- `Status: <STATE>`
+- `Observed At: <KST>, <UTC>`
+- `Environment: live`
+- `Evidence: <log/query/order-execution identifier>`
+- `Next: <single next action>`
+
+This template is reused for every transition for consistent auditability.
+
+---
+
+## 6. Parent-Child Policy
+
+- Parent issue comment must list child issue IDs.
+- Child issues transition independently.
+- Parent close gate:
+  - if any child is not `CLOSED`, parent remains at most `CONFIRMED`;
+  - when all children are `CLOSED`, parent can transition to `CLOSED` and be closed.
+
+---
+
+## 7. Operational Loop
+
+1. Initialize all open issues with `NOT_STARTED` comment.
+2. Move active issues to `OBSERVING` when live confirmation begins.
+3. Move to `CONFIRMED` after one real-operation evidence capture.
+4. Close child issues first, then close parent issue(s).
+5. Update the single `80` PR summary comment whenever issue state changes.
+6. Reflect final confirmed/closed outcomes in `docs/ouroboros/80_implementation_audit.md`.
+
+---
+
+## 8. Evidence and Time Rules
+
+- Evidence must be replayable/referenceable by identifier (not vague narrative only).
+- Every update includes both KST and UTC timestamps.
+- Use absolute dates (example: `2026-03-02`) to avoid ambiguity in session handoff.
+
+---
+
+## 9. Completion Conditions
+
+Tracking design is complete when:
+
+- all open issues are represented in issue comments with valid lifecycle state,
+- parent-child closure constraints are respected,
+- `80` PR summary comment reflects current global status,
+- and final `80` audit document update is ready for closeout commit.

docs/plans/2026-03-02-live-operation-issue-tracking-implementation-plan.md

@@ -0,0 +1,203 @@
+# Live Operation Issue Tracking Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Track real-operation confirmation for all open Gitea issues with per-issue comments and a synchronized summary on the `80` PR, then close issues in dependency-safe order.
+
+**Architecture:** Use issue tickets as the primary state log with a strict four-state lifecycle. Maintain one aggregate status comment on `task/80-implementation-audit-closeout` PR. Enforce child-first closure and parent close gating.
+
+**Tech Stack:** Git CLI, Gitea CLI (`tea`), Markdown (`docs/ouroboros/80_implementation_audit.md`)
+
+---
+
+### Task 1: Snapshot open issues and classify parent-child graph
+
+**Files:**
+- Create: none
+- Modify: none
+- Test: `tea issues ls` output inspection
+
+**Step 1: Capture open issue list**
+
+Run: `tea issues ls --state open --limit 200`
+Expected: list of all open issues (including `318~381`).
+
+**Step 2: Fetch details for each issue**
+
+Run: `for n in $(tea issues ls --state open --limit 200 --output json | jq -r '.[].number'); do tea issue view "$n"; done`
+Expected: issue bodies/comments available for relation parsing.
+
+**Step 3: Extract parent-child references**
+
+Run: parse issue text for relation keywords (`parent`, `child`, `depends on`, `blocks`).
+Expected: dependency map ready.
+
+**Step 4: Validate unresolved dependency edges**
+
+Run: verify each parent references only existing issue numbers.
+Expected: no dangling child references.
+
+**Step 5: Commit**
+
+```bash
+# no commit (discovery task)
+```
+
+### Task 2: Initialize per-issue tracking comments (`NOT_STARTED`)
+
+**Files:**
+- Create: none
+- Modify: none
+- Test: issue comment presence check
+
+**Step 1: Prepare standard template text**
+
+```text
+Status: NOT_STARTED
+Observed At: <KST>, <UTC>
+Environment: live
+Evidence: N/A
+Next: Begin live observation for this issue.
+```
+
+**Step 2: Post initialization comment to each open issue**
+
+Run: `tea issues comment <issue-number> --message "<template>"`
+Expected: each open issue has one initialization comment.
+
+**Step 3: Verify comment coverage**
+
+Run: sample-check each issue (`tea issue view <issue-number>`) for latest status comment.
+Expected: 100% coverage across open issues.
+
+**Step 4: Commit**
+
+```bash
+# no commit (ticket operations)
+```
+
+### Task 3: Create and maintain `80` PR aggregate status comment
+
+**Files:**
+- Modify: `docs/ouroboros/80_implementation_audit.md` (final sync only, not this task)
+- Test: PR comment contains full issue table
+
+**Step 1: Identify PR number for `task/80-implementation-audit-closeout`**
+
+Run: `tea pr ls --state open`
+Expected: PR number for `task/80...` identified.
+
+**Step 2: Post initial summary comment**
+
+Run: `tea pr comment <pr-number> --message "<issue-state-summary-table>"`
+Expected: one aggregate comment listing `issue | state | last observed at | parent/child`.
+
+**Step 3: Define update protocol**
+
+Run: update same summary comment on every issue state change.
+Expected: summary remains single-source aggregate view.
+
+**Step 4: Commit**
+
+```bash
+# no commit (ticket/PR operations)
+```
+
+### Task 4: Execute live observation loop and mark `OBSERVING` -> `CONFIRMED`
+
+**Files:**
+- Create: none
+- Modify: none
+- Test: evidence-backed state transitions present in issue comments
+
+**Step 1: Mark active issue as `OBSERVING`**
+
+Run: `tea issues comment <issue-number> --message "Status: OBSERVING ..."`
+Expected: issue history shows observation started.
+
+**Step 2: Capture one real-operation evidence item**
+
+Run: collect log/query/order identifier tied to issue behavior.
+Expected: one concrete evidence reference captured.
+
+**Step 3: Mark issue `CONFIRMED`**
+
+Run: `tea issues comment <issue-number> --message "Status: CONFIRMED ... Evidence: <id> ..."`
+Expected: issue has explicit confirmed state.
+
+**Step 4: Sync PR aggregate summary**
+
+Run: update `80` PR summary comment row for that issue.
+Expected: PR summary and issue status aligned.
+
+**Step 5: Commit**
+
+```bash
+# no commit (ticket operations)
+```
+
+### Task 5: Close issues with child-first dependency enforcement
+
+**Files:**
+- Create: none
+- Modify: none
+- Test: parent close attempted only after all children closed
+
+**Step 1: Close confirmed child issues first**
+
+Run: `tea issues close <child-issue-number>`
+Expected: child issue state becomes closed.
+
+**Step 2: Verify all children for each parent**
+
+Run: evaluate dependency map for remaining non-closed children.
+Expected: parent close gate decision available.
+
+**Step 3: Close eligible parent issues**
+
+Run: `tea issues close <parent-issue-number>`
+Expected: parent is closed only when all children are closed.
+
+**Step 4: Post closure comment and sync PR summary**
+
+Run: comment `Status: CLOSED` on closed issue and update PR summary row.
+Expected: closure trace exists in both issue and PR aggregate.
+
+**Step 5: Commit**
+
+```bash
+# no commit (ticket operations)
+```
+
+### Task 6: Reflect final tracked outcomes in `80` audit and commit
+
+**Files:**
+- Modify: `docs/ouroboros/80_implementation_audit.md`
+- Test: markdown consistency + traceability check
+
+**Step 1: Update `80` with final live-confirmed statuses**
+
+Run: edit status lines and evidence summary sections.
+Expected: document reflects final issue states and confirmation notes.
+
+**Step 2: Add issue closure summary block**
+
+Run: include closed child/parent sequence summary.
+Expected: audit reader can verify closure logic quickly.
+
+**Step 3: Validate references**
+
+Run: ensure issue numbers in doc match actual closed/open states.
+Expected: no mismatch between doc and Gitea.
+
+**Step 4: Commit closeout doc update**
+
+```bash
+git add docs/ouroboros/80_implementation_audit.md
+git commit -m "docs: finalize 80 audit with live-operation issue confirmations"
+```
+
+**Step 5: Push branch**
+
+Run: `git push`
+Expected: closeout commit available on `task/80-implementation-audit-closeout`.

docs/plans/2026-03-02-ouroboros-dual-track-workflow-design.md

@@ -0,0 +1,99 @@
+# Ouroboros Dual-Track Workflow Design (Revised)
+
+**Date:** 2026-03-02  
+**Scope:** `docs/ouroboros/86_peak_probability_model_plan.md` and `docs/ouroboros/80_implementation_audit.md`
+
+---
+
+## 1. Goal
+
+Prevent work loss and history contamination while running two parallel tracks, while correcting an invalid starting condition (design commits created on a disposable temporary branch):
+
+- Track A: open and iterate the `86` plan as a dedicated PR.
+- Track B: finalize `80` only after real-operation behavior is confirmed, then reflect completion and close related Gitea open issues.
+
+---
+
+## 2. Constraints and Current Context
+
+- Current temporary branch is disposable and must not remain the authoritative base.
+- Canonical base line for current work is `feature/v3-session-policy-stream`.
+- Final integration strategy is to merge the full `feature/v3-session-policy-stream` line into `main` after operation-level validation is complete.
+
+---
+
+## 3. Approaches Considered
+
+1. Keep using temporary branch and continue from existing design commits.
+2. Create a clean base from `feature/v3-session-policy-stream` and move only plan artifacts as one commit (recommended).
+3. Split into two fresh bases (`86` and `80` separately) immediately.
+
+Recommendation rationale:
+
+- Approach 2 removes temporary-branch contamination risk.
+- It preserves clean lineage with a single "official starting point" commit.
+- It keeps downstream split for `86` and `80` straightforward.
+
+---
+
+## 4. Approved Branch and PR Topology
+
+- Long-lived source baseline: `feature/v3-session-policy-stream`
+- New official base: `base/ouroboros-workflow-20260302` (from `feature/v3-session-policy-stream`)
+- Official base contains only two workflow plan docs squashed into one commit.
+- Track A branch: `task/86-peak-probability-plan` (from official base)
+- Track B branch: `task/80-implementation-audit-closeout` (from official base)
+
+PR direction:
+
+- PR-A (`86`): `task/86-peak-probability-plan -> base/ouroboros-workflow-20260302`
+- PR-B (`80`): `task/80-implementation-audit-closeout -> base/ouroboros-workflow-20260302`
+- Final integration: `feature/v3-session-policy-stream` line merged into `main`
+
+---
+
+## 5. Completion Criteria per Track
+
+### Track A (`86`)
+
+- Open initial PR early with the planning document.
+- Continue iterative concretization in follow-up commits on the same branch.
+
+### Track B (`80`)
+
+- Do not close based on test-command artifacts alone.
+- Close only when modified behavior is confirmed in real operation.
+- Update `80_implementation_audit.md` with:
+  - observed behavior changes,
+  - confirmation timing,
+  - operation-side evidence summary.
+- After `80` real-operation confirmation and doc update, close related Gitea open issues immediately (same phase).
+- Parent issues close only when all child issues are confirmed closed.
+
+---
+
+## 6. Loss-Prevention Rules
+
+- Preserve temporary-branch unmerged changes first (stash or archival commit).
+- Exclude `workflow/session-handover.md` and `docs/ouroboros/86_peak_probability_model_plan.md` from official base-seeding commit.
+- Enforce file-scope separation:
+  - PR-A should avoid `80` edits.
+  - PR-B should avoid `86` edits.
+- Keep commit messages scoped to one track per commit.
+
+---
+
+## 7. Merge Strategy Decision
+
+- For this workflow, prefer merge-based lineage preservation.
+- Because the final promotion unit is the full `feature/v3-session-policy-stream` stream, avoid unnecessary cherry-pick/rebase-surgery unless a contamination incident occurs.
+
+---
+
+## 8. Immediate Next Actions
+
+1. Seed `base/ouroboros-workflow-20260302` from `feature/v3-session-policy-stream` with one squashed workflow-doc commit.
+2. Verify seed commit file scope contains only two plan docs.
+3. Create and open PR-A for `86`.
+4. Complete real-operation confirmation for `80`, update audit doc, then close related Gitea issues (children first, then parent).
+5. Merge validated `feature/v3-session-policy-stream` line into `main`.

docs/workflow.md

@@ -128,6 +128,16 @@ tea pr create \
   --description "$PR_BODY"
 ```
 
+PR 생성 직후 본문 무결성 검증(필수):
+
+```bash
+python3 scripts/validate_pr_body.py --pr <PR_NUMBER>
+```
+
+강제 규칙:
+- 검증 실패(`\n` 리터럴, 코드펜스 불균형, 헤더/리스트 누락) 상태에서는 리뷰/머지 금지
+- 본문 수정 후 같은 명령으로 재검증 통과 필요
+
 금지 패턴:
 
 - `-d "line1\nline2"` (웹 UI에 `\n` 문자 그대로 노출될 수 있음)

scripts/validate_docs_sync.py

@@ -92,6 +92,25 @@ def validate_testing_doc_has_dynamic_count_guidance(errors: list[str]) -> None:
         )
 
 
+def validate_pr_body_postcheck_guidance(errors: list[str]) -> None:
+    required_tokens = {
+        "commands": (
+            "PR Body Post-Check (Mandatory)",
+            "python3 scripts/validate_pr_body.py --pr <PR_NUMBER>",
+        ),
+        "workflow": (
+            "PR 생성 직후 본문 무결성 검증(필수)",
+            "python3 scripts/validate_pr_body.py --pr <PR_NUMBER>",
+        ),
+    }
+    for key, tokens in required_tokens.items():
+        path = REQUIRED_FILES[key]
+        text = _read(path)
+        for token in tokens:
+            if token not in text:
+                errors.append(f"{path}: missing PR body post-check guidance token -> {token}")
+
+
 def main() -> int:
     errors: list[str] = []
 
@@ -117,6 +136,7 @@ def main() -> int:
     validate_summary_docs_reference_core_docs(errors)
     validate_commands_endpoint_duplicates(errors)
     validate_testing_doc_has_dynamic_count_guidance(errors)
+    validate_pr_body_postcheck_guidance(errors)
 
     if errors:
         print("[FAIL] docs sync validation failed")
@@ -128,6 +148,7 @@ def main() -> int:
     print("[OK] summary docs link to core docs and links resolve")
     print("[OK] commands endpoint rows have no duplicates")
     print("[OK] testing doc includes dynamic count guidance")
+    print("[OK] PR body post-check guidance exists in commands/workflow docs")
     return 0
 
 

scripts/validate_ouroboros_docs.py

@@ -33,6 +33,9 @@ ALLOWED_PLAN_TARGETS = {
     "2": (DOC_DIR / "source" / "ouroboros_plan_v2.txt").resolve(),
     "3": (DOC_DIR / "source" / "ouroboros_plan_v3.txt").resolve(),
 }
+ISSUE_REF_PATTERN = re.compile(r"#(?P<issue>\d+)")
+ISSUE_DONE_PATTERN = re.compile(r"(?:✅|머지|해소|완료)")
+ISSUE_PENDING_PATTERN = re.compile(r"(?:잔여|오픈 상태|추적 이슈)")
 
 
 def iter_docs() -> list[Path]:
@@ -119,6 +122,38 @@ def collect_req_traceability(
                 req_to_test.setdefault(req_id, set()).add(item_id)
 
 
+def validate_issue_status_consistency(path: Path, text: str, errors: list[str]) -> None:
+    issue_done_lines: dict[str, list[int]] = {}
+    issue_pending_lines: dict[str, list[int]] = {}
+
+    for line_no, raw_line in enumerate(text.splitlines(), start=1):
+        line = raw_line.strip()
+        if not line:
+            continue
+        issue_ids = [m.group("issue") for m in ISSUE_REF_PATTERN.finditer(line)]
+        if not issue_ids:
+            continue
+
+        is_pending = bool(ISSUE_PENDING_PATTERN.search(line))
+        is_done = bool(ISSUE_DONE_PATTERN.search(line)) and not is_pending
+        if not is_pending and not is_done:
+            continue
+
+        for issue_id in issue_ids:
+            if is_done:
+                issue_done_lines.setdefault(issue_id, []).append(line_no)
+            if is_pending:
+                issue_pending_lines.setdefault(issue_id, []).append(line_no)
+
+    conflicted_issues = sorted(set(issue_done_lines) & set(issue_pending_lines))
+    for issue_id in conflicted_issues:
+        errors.append(
+            f"{path}: conflicting status for issue #{issue_id} "
+            f"(done at lines {issue_done_lines[issue_id]}, "
+            f"pending at lines {issue_pending_lines[issue_id]})"
+        )
+
+
 def main() -> int:
     if not DOC_DIR.exists():
         print(f"ERROR: missing directory {DOC_DIR}")
@@ -140,6 +175,8 @@ def main() -> int:
         text = path.read_text(encoding="utf-8")
         validate_metadata(path, text, errors, doc_ids)
         validate_links(path, text, errors)
+        if path.name == "80_implementation_audit.md":
+            validate_issue_status_consistency(path, text, errors)
         collect_ids(path, text, defs, refs)
         collect_req_traceability(text, req_to_task, req_to_test)
 

scripts/validate_pr_body.py

@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+"""Validate PR body formatting to prevent escaped-newline artifacts."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+HEADER_PATTERN = re.compile(r"^##\s+\S+", re.MULTILINE)
+LIST_ITEM_PATTERN = re.compile(r"^\s*(?:-|\*|\d+\.)\s+\S+", re.MULTILINE)
+
+
+def validate_pr_body_text(text: str) -> list[str]:
+    errors: list[str] = []
+    if "\\n" in text and "\n" not in text:
+        errors.append("body contains escaped newline sequence (\\n)")
+    if text.count("```") % 2 != 0:
+        errors.append("body has unbalanced fenced code blocks (``` count is odd)")
+    if not HEADER_PATTERN.search(text):
+        errors.append("body is missing markdown section headers (e.g. '## Summary')")
+    if not LIST_ITEM_PATTERN.search(text):
+        errors.append("body is missing markdown list items")
+    return errors
+
+
+def fetch_pr_body(pr_number: int) -> str:
+    try:
+        completed = subprocess.run(
+            [
+                "tea",
+                "api",
+                "-R",
+                "origin",
+                f"repos/{{owner}}/{{repo}}/pulls/{pr_number}",
+            ],
+            check=True,
+            capture_output=True,
+            text=True,
+        )
+    except (subprocess.CalledProcessError, FileNotFoundError) as exc:
+        raise RuntimeError(f"failed to fetch PR #{pr_number}: {exc}") from exc
+
+    try:
+        payload = json.loads(completed.stdout)
+    except json.JSONDecodeError as exc:
+        raise RuntimeError(f"failed to parse PR payload for #{pr_number}: {exc}") from exc
+
+    body = payload.get("body", "")
+    if not isinstance(body, str):
+        raise RuntimeError(f"unexpected PR body type for #{pr_number}: {type(body).__name__}")
+    return body
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Validate PR body markdown formatting and escaped-newline artifacts."
+    )
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument("--pr", type=int, help="PR number to fetch via `tea api`")
+    group.add_argument("--body-file", type=Path, help="Path to markdown body file")
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    if args.body_file is not None:
+        if not args.body_file.exists():
+            print(f"[FAIL] body file not found: {args.body_file}")
+            return 1
+        body = args.body_file.read_text(encoding="utf-8")
+        source = f"file:{args.body_file}"
+    else:
+        body = fetch_pr_body(args.pr)
+        source = f"pr:{args.pr}"
+
+    errors = validate_pr_body_text(body)
+    if errors:
+        print("[FAIL] PR body validation failed")
+        print(f"- source: {source}")
+        for err in errors:
+            print(f"- {err}")
+        return 1
+
+    print("[OK] PR body validation passed")
+    print(f"- source: {source}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/analysis/backtest_cost_guard.py

@@ -11,6 +11,7 @@ class BacktestCostModel:
     commission_bps: float | None = None
     slippage_bps_by_session: dict[str, float] | None = None
     failure_rate_by_session: dict[str, float] | None = None
+    partial_fill_rate_by_session: dict[str, float] | None = None
     unfavorable_fill_required: bool = True
 
 
@@ -31,6 +32,7 @@ def validate_backtest_cost_model(
 
     slippage = model.slippage_bps_by_session or {}
     failure = model.failure_rate_by_session or {}
+    partial_fill = model.partial_fill_rate_by_session or {}
 
     missing_slippage = [s for s in required_sessions if s not in slippage]
     if missing_slippage:
@@ -44,9 +46,19 @@ def validate_backtest_cost_model(
             f"missing failure_rate_by_session for sessions: {', '.join(missing_failure)}"
         )
 
+    missing_partial_fill = [s for s in required_sessions if s not in partial_fill]
+    if missing_partial_fill:
+        raise ValueError(
+            "missing partial_fill_rate_by_session for sessions: "
+            f"{', '.join(missing_partial_fill)}"
+        )
+
     for sess, bps in slippage.items():
         if not math.isfinite(bps) or bps < 0:
             raise ValueError(f"slippage bps must be >= 0 for session={sess}")
     for sess, rate in failure.items():
         if not math.isfinite(rate) or rate < 0 or rate > 1:
             raise ValueError(f"failure rate must be within [0,1] for session={sess}")
+    for sess, rate in partial_fill.items():
+        if not math.isfinite(rate) or rate < 0 or rate > 1:
+            raise ValueError(f"partial fill rate must be within [0,1] for session={sess}")

src/analysis/backtest_pipeline.py

@@ -13,6 +13,11 @@ from statistics import mean
 from typing import Literal, cast
 
 from src.analysis.backtest_cost_guard import BacktestCostModel, validate_backtest_cost_model
+from src.analysis.backtest_execution_model import (
+    BacktestExecutionModel,
+    ExecutionAssumptions,
+    ExecutionRequest,
+)
 from src.analysis.triple_barrier import TripleBarrierSpec, label_with_triple_barrier
 from src.analysis.walk_forward_split import WalkForwardFold, generate_walk_forward_splits
 
@@ -40,6 +45,7 @@ class WalkForwardConfig:
 class BaselineScore:
     name: Literal["B0", "B1", "M1"]
     accuracy: float
+    cost_adjusted_accuracy: float
 
 
 @dataclass(frozen=True)
@@ -50,6 +56,10 @@ class BacktestFoldResult:
     train_label_distribution: dict[int, int]
     test_label_distribution: dict[int, int]
     baseline_scores: list[BaselineScore]
+    execution_adjusted_avg_return_bps: float
+    execution_adjusted_trade_count: int
+    execution_rejected_count: int
+    execution_partial_count: int
 
 
 @dataclass(frozen=True)
@@ -115,6 +125,8 @@ def run_v2_backtest_pipeline(
         ).label
 
     ordered_labels = [labels_by_bar_index[idx] for idx in normalized_entries]
+    ordered_sessions = [bars[idx].session_id for idx in normalized_entries]
+    ordered_prices = [bars[idx].close for idx in normalized_entries]
     folds = generate_walk_forward_splits(
         n_samples=len(normalized_entries),
         train_size=walk_forward.train_size,
@@ -129,8 +141,37 @@ def run_v2_backtest_pipeline(
     for fold_idx, fold in enumerate(folds):
         train_labels = [ordered_labels[i] for i in fold.train_indices]
         test_labels = [ordered_labels[i] for i in fold.test_indices]
+        test_sessions = [ordered_sessions[i] for i in fold.test_indices]
+        test_prices = [ordered_prices[i] for i in fold.test_indices]
         if not test_labels:
             continue
+        execution_model = _build_execution_model(cost_model=cost_model, fold_seed=fold_idx)
+        execution_return_model = _build_execution_model(
+            cost_model=cost_model,
+            fold_seed=fold_idx + 1000,
+        )
+        b0_pred = _baseline_b0_pred(train_labels)
+        m1_pred = _m1_pred(train_labels)
+        execution_returns_bps: list[float] = []
+        execution_rejected = 0
+        execution_partial = 0
+        for rel_idx in fold.test_indices:
+            entry_bar_index = normalized_entries[rel_idx]
+            bar = bars[entry_bar_index]
+            trade = _simulate_execution_adjusted_return_bps(
+                execution_model=execution_return_model,
+                bar=bar,
+                label=ordered_labels[rel_idx],
+                side=side,
+                spec=triple_barrier_spec,
+                commission_bps=float(cost_model.commission_bps or 0.0),
+            )
+            if trade["status"] == "REJECTED":
+                execution_rejected += 1
+                continue
+            execution_returns_bps.append(float(trade["return_bps"]))
+            if trade["status"] == "PARTIAL":
+                execution_partial += 1
         fold_results.append(
             BacktestFoldResult(
                 fold_index=fold_idx,
@@ -139,13 +180,49 @@ def run_v2_backtest_pipeline(
                 train_label_distribution=_label_dist(train_labels),
                 test_label_distribution=_label_dist(test_labels),
                 baseline_scores=[
-                    BaselineScore(name="B0", accuracy=_baseline_b0(train_labels, test_labels)),
-                    BaselineScore(name="B1", accuracy=_score_constant(1, test_labels)),
+                    BaselineScore(
+                        name="B0",
+                        accuracy=_score_constant(b0_pred, test_labels),
+                        cost_adjusted_accuracy=_score_with_execution(
+                            prediction=b0_pred,
+                            actual=test_labels,
+                            sessions=test_sessions,
+                            reference_prices=test_prices,
+                            execution_model=execution_model,
+                            commission_bps=float(cost_model.commission_bps or 0.0),
+                        ),
+                    ),
+                    BaselineScore(
+                        name="B1",
+                        accuracy=_score_constant(1, test_labels),
+                        cost_adjusted_accuracy=_score_with_execution(
+                            prediction=1,
+                            actual=test_labels,
+                            sessions=test_sessions,
+                            reference_prices=test_prices,
+                            execution_model=execution_model,
+                            commission_bps=float(cost_model.commission_bps or 0.0),
+                        ),
+                    ),
                     BaselineScore(
                         name="M1",
-                        accuracy=_score_constant(_m1_pred(train_labels), test_labels),
+                        accuracy=_score_constant(m1_pred, test_labels),
+                        cost_adjusted_accuracy=_score_with_execution(
+                            prediction=m1_pred,
+                            actual=test_labels,
+                            sessions=test_sessions,
+                            reference_prices=test_prices,
+                            execution_model=execution_model,
+                            commission_bps=float(cost_model.commission_bps or 0.0),
+                        ),
                     ),
                 ],
+                execution_adjusted_avg_return_bps=(
+                    mean(execution_returns_bps) if execution_returns_bps else 0.0
+                ),
+                execution_adjusted_trade_count=len(execution_returns_bps),
+                execution_rejected_count=execution_rejected,
+                execution_partial_count=execution_partial,
             )
         )
 
@@ -176,12 +253,15 @@ def _score_constant(pred: int, actual: Sequence[int]) -> float:
 
 
 def _baseline_b0(train_labels: Sequence[int], test_labels: Sequence[int]) -> float:
+    return _score_constant(_baseline_b0_pred(train_labels), test_labels)
+
+
+def _baseline_b0_pred(train_labels: Sequence[int]) -> int:
     if not train_labels:
-        return _score_constant(0, test_labels)
+        return 0
     # Majority-class baseline from training fold.
     choices = (-1, 0, 1)
-    pred = max(choices, key=lambda c: train_labels.count(c))
-    return _score_constant(pred, test_labels)
+    return max(choices, key=lambda c: train_labels.count(c))
 
 
 def _m1_pred(train_labels: Sequence[int]) -> int:
@@ -190,6 +270,56 @@ def _m1_pred(train_labels: Sequence[int]) -> int:
     return train_labels[-1]
 
 
+def _build_execution_model(
+    *,
+    cost_model: BacktestCostModel,
+    fold_seed: int,
+) -> BacktestExecutionModel:
+    return BacktestExecutionModel(
+        ExecutionAssumptions(
+            slippage_bps_by_session=dict(cost_model.slippage_bps_by_session or {}),
+            failure_rate_by_session=dict(cost_model.failure_rate_by_session or {}),
+            partial_fill_rate_by_session=dict(cost_model.partial_fill_rate_by_session or {}),
+            seed=fold_seed,
+        )
+    )
+
+
+def _score_with_execution(
+    *,
+    prediction: int,
+    actual: Sequence[int],
+    sessions: Sequence[str],
+    reference_prices: Sequence[float],
+    execution_model: BacktestExecutionModel,
+    commission_bps: float,
+) -> float:
+    if not actual:
+        return 0.0
+    contributions: list[float] = []
+    for label, session_id, reference_price in zip(actual, sessions, reference_prices, strict=True):
+        if prediction == 0:
+            contributions.append(1.0 if label == 0 else 0.0)
+            continue
+        side = "BUY" if prediction > 0 else "SELL"
+        execution = execution_model.simulate(
+            ExecutionRequest(
+                side=side,
+                session_id=session_id,
+                qty=100,
+                reference_price=reference_price,
+            )
+        )
+        if execution.status == "REJECTED":
+            contributions.append(0.0)
+            continue
+        fill_ratio = execution.filled_qty / 100.0
+        cost_penalty = min(0.99, (commission_bps + execution.slippage_bps) / 10000.0)
+        correctness = 1.0 if prediction == label else 0.0
+        contributions.append(correctness * fill_ratio * (1.0 - cost_penalty))
+    return mean(contributions)
+
+
 def _build_run_id(*, n_entries: int, n_folds: int, sessions: Sequence[str]) -> str:
     sess_key = "_".join(sessions)
     return f"v2p-e{n_entries}-f{n_folds}-s{sess_key}"
@@ -198,3 +328,58 @@ def _build_run_id(*, n_entries: int, n_folds: int, sessions: Sequence[str]) -> s
 def fold_has_leakage(fold: WalkForwardFold) -> bool:
     """Utility for tests/verification: True when train/test overlap exists."""
     return bool(set(fold.train_indices).intersection(fold.test_indices))
+
+
+def _simulate_execution_adjusted_return_bps(
+    *,
+    execution_model: BacktestExecutionModel,
+    bar: BacktestBar,
+    label: int,
+    side: int,
+    spec: TripleBarrierSpec,
+    commission_bps: float,
+) -> dict[str, float | str]:
+    qty = 100
+    entry_req = ExecutionRequest(
+        side="BUY" if side == 1 else "SELL",
+        session_id=bar.session_id,
+        qty=qty,
+        reference_price=float(bar.close),
+    )
+    entry_fill = execution_model.simulate(entry_req)
+    if entry_fill.status == "REJECTED":
+        return {"status": "REJECTED", "return_bps": 0.0}
+
+    exit_qty = entry_fill.filled_qty
+    if label == 1:
+        gross_return_bps = spec.take_profit_pct * 10000.0
+    elif label == -1:
+        gross_return_bps = -spec.stop_loss_pct * 10000.0
+    else:
+        gross_return_bps = 0.0
+
+    if side == 1:
+        exit_price = float(bar.close) * (1.0 + gross_return_bps / 10000.0)
+    else:
+        exit_price = float(bar.close) * (1.0 - gross_return_bps / 10000.0)
+
+    exit_req = ExecutionRequest(
+        side="SELL" if side == 1 else "BUY",
+        session_id=bar.session_id,
+        qty=exit_qty,
+        reference_price=max(0.01, exit_price),
+    )
+    exit_fill = execution_model.simulate(exit_req)
+    if exit_fill.status == "REJECTED":
+        return {"status": "REJECTED", "return_bps": 0.0}
+
+    fill_ratio = min(entry_fill.filled_qty, exit_fill.filled_qty) / qty
+    cost_bps = (
+        float(entry_fill.slippage_bps)
+        + float(exit_fill.slippage_bps)
+        + (2.0 * float(commission_bps))
+    )
+    net_return_bps = (gross_return_bps * fill_ratio) - cost_bps
+    is_partial = entry_fill.status == "PARTIAL" or exit_fill.status == "PARTIAL"
+    status = "PARTIAL" if is_partial else "FILLED"
+    return {"status": status, "return_bps": net_return_bps}

src/core/blackout_manager.py

@@ -23,6 +23,7 @@ class BlackoutWindow:
 class QueuedOrderIntent:
     market_code: str
     exchange_code: str
+    session_id: str
     stock_code: str
     order_type: str
     quantity: int
@@ -68,11 +69,16 @@ class BlackoutOrderManager:
         self._queue: deque[QueuedOrderIntent] = deque()
         self._was_blackout = False
         self._max_queue_size = max_queue_size
+        self._overflow_drop_count = 0
 
     @property
     def pending_count(self) -> int:
         return len(self._queue)
 
+    @property
+    def overflow_drop_count(self) -> int:
+        return self._overflow_drop_count
+
     def in_blackout(self, now: datetime | None = None) -> bool:
         if not self.enabled or not self._windows:
             return False
@@ -81,8 +87,11 @@ class BlackoutOrderManager:
         return any(window.contains(kst_now) for window in self._windows)
 
     def enqueue(self, intent: QueuedOrderIntent) -> bool:
-        if len(self._queue) >= self._max_queue_size:
+        if self._max_queue_size <= 0:
             return False
+        if len(self._queue) >= self._max_queue_size:
+            self._queue.popleft()
+            self._overflow_drop_count += 1
         self._queue.append(intent)
         return True
 

src/core/kill_switch.py

@@ -3,13 +3,14 @@
 Order is fixed:
 1) block new orders
 2) cancel pending orders
-3) refresh order state
+3) refresh order state (retry up to 3 attempts with exponential backoff)
 4) reduce risk
 5) snapshot and notify
 """
 
 from __future__ import annotations
 
+import asyncio
 import inspect
 from collections.abc import Awaitable, Callable
 from dataclasses import dataclass, field
@@ -34,16 +35,55 @@ class KillSwitchOrchestrator:
         report: KillSwitchReport,
         name: str,
         fn: StepCallable | None,
-    ) -> None:
+    ) -> bool:
         report.steps.append(name)
         if fn is None:
-            return
+            return True
         try:
             result = fn()
             if inspect.isawaitable(result):
                 await result
+            if result is False:
+                raise RuntimeError("step returned False")
+            return True
         except Exception as exc:  # pragma: no cover - intentionally resilient
             report.errors.append(f"{name}: {exc}")
+            return False
+
+    async def _run_refresh_with_retry(
+        self,
+        report: KillSwitchReport,
+        fn: StepCallable | None,
+        *,
+        max_attempts: int,
+        base_delay_sec: float,
+    ) -> None:
+        report.steps.append("refresh_order_state")
+        if fn is None:
+            return
+
+        attempts = max(1, max_attempts)
+        delay = max(0.0, base_delay_sec)
+        last_exc: Exception | None = None
+        for attempt in range(1, attempts + 1):
+            try:
+                result = fn()
+                if inspect.isawaitable(result):
+                    await result
+                if result is False:
+                    raise RuntimeError("step returned False")
+                return
+            except Exception as exc:
+                last_exc = exc
+                if attempt >= attempts:
+                    break
+                if delay > 0:
+                    await asyncio.sleep(delay * (2 ** (attempt - 1)))
+        if last_exc is not None:
+            report.errors.append(
+                "refresh_order_state: failed after "
+                f"{attempts} attempts ({last_exc})"
+            )
 
     async def trigger(
         self,
@@ -54,6 +94,8 @@ class KillSwitchOrchestrator:
         reduce_risk: StepCallable | None = None,
         snapshot_state: StepCallable | None = None,
         notify: StepCallable | None = None,
+        refresh_retry_attempts: int = 3,
+        refresh_retry_base_delay_sec: float = 1.0,
     ) -> KillSwitchReport:
         report = KillSwitchReport(reason=reason)
 
@@ -61,7 +103,12 @@ class KillSwitchOrchestrator:
         report.steps.append("block_new_orders")
 
         await self._run_step(report, "cancel_pending_orders", cancel_pending_orders)
-        await self._run_step(report, "refresh_order_state", refresh_order_state)
+        await self._run_refresh_with_retry(
+            report,
+            refresh_order_state,
+            max_attempts=refresh_retry_attempts,
+            base_delay_sec=refresh_retry_base_delay_sec,
+        )
         await self._run_step(report, "reduce_risk", reduce_risk)
         await self._run_step(report, "snapshot_state", snapshot_state)
         await self._run_step(report, "notify", notify)

src/db.py

@@ -318,7 +318,7 @@ def get_latest_buy_trade(
     if exchange_code:
         cursor = conn.execute(
             """
-            SELECT decision_id, price, quantity
+            SELECT decision_id, price, quantity, selection_context
             FROM trades
             WHERE stock_code = ?
               AND market = ?
@@ -339,7 +339,7 @@ def get_latest_buy_trade(
     else:
         cursor = conn.execute(
             """
-            SELECT decision_id, price, quantity
+            SELECT decision_id, price, quantity, selection_context
             FROM trades
             WHERE stock_code = ?
               AND market = ?

src/main.py

@@ -128,6 +128,84 @@ def _resolve_sell_qty_for_pnl(*, sell_qty: int | None, buy_qty: int | None) -> i
     return max(0, int(buy_qty or 0))
 
 
+def _extract_fx_rate_from_sources(*sources: dict[str, Any] | None) -> float | None:
+    """Best-effort FX rate extraction from broker payloads."""
+    # KIS overseas payloads expose exchange-rate fields with varying key names
+    # across endpoints/responses (price, balance, buying power). Keep this list
+    # centralised so schema drifts can be patched in one place.
+    rate_keys = (
+        "frst_bltn_exrt",
+        "bass_exrt",
+        "ovrs_exrt",
+        "aply_xchg_rt",
+        "xchg_rt",
+        "exchange_rate",
+        "fx_rate",
+    )
+    for source in sources:
+        if not isinstance(source, dict):
+            continue
+        for key in rate_keys:
+            rate = safe_float(source.get(key), 0.0)
+            if rate > 0:
+                return rate
+    return None
+
+
+def _split_trade_pnl_components(
+    *,
+    market: MarketInfo,
+    trade_pnl: float,
+    buy_price: float,
+    sell_price: float,
+    quantity: int,
+    buy_fx_rate: float | None = None,
+    sell_fx_rate: float | None = None,
+) -> tuple[float, float]:
+    """Split total trade pnl into strategy/fx components.
+
+    For overseas symbols, use buy/sell FX rates when both are available.
+    Otherwise preserve backward-compatible behaviour (all strategy pnl).
+    """
+    if trade_pnl == 0.0:
+        return 0.0, 0.0
+    if market.is_domestic:
+        return trade_pnl, 0.0
+
+    if (
+        buy_fx_rate is not None
+        and sell_fx_rate is not None
+        and buy_fx_rate > 0
+        and sell_fx_rate > 0
+        and quantity > 0
+        and buy_price > 0
+        and sell_price > 0
+    ):
+        buy_notional = buy_price * quantity
+        fx_return = (sell_fx_rate - buy_fx_rate) / buy_fx_rate
+        fx_pnl = buy_notional * fx_return
+        strategy_pnl = trade_pnl - fx_pnl
+        return strategy_pnl, fx_pnl
+
+    return trade_pnl, 0.0
+
+
+def _extract_buy_fx_rate(buy_trade: dict[str, Any] | None) -> float | None:
+    if not buy_trade:
+        return None
+    raw_ctx = buy_trade.get("selection_context")
+    if not isinstance(raw_ctx, str) or not raw_ctx.strip():
+        return None
+    try:
+        decoded = json.loads(raw_ctx)
+    except (TypeError, ValueError):
+        return None
+    if not isinstance(decoded, dict):
+        return None
+    rate = safe_float(decoded.get("fx_rate"), 0.0)
+    return rate if rate > 0 else None
+
+
 def _compute_kr_dynamic_stop_loss_pct(
     *,
     market: MarketInfo | None = None,
@@ -926,6 +1004,7 @@ async def build_overseas_symbol_universe(
 def _build_queued_order_intent(
     *,
     market: MarketInfo,
+    session_id: str,
     stock_code: str,
     order_type: str,
     quantity: int,
@@ -935,6 +1014,7 @@ def _build_queued_order_intent(
     return QueuedOrderIntent(
         market_code=market.code,
         exchange_code=market.exchange_code,
+        session_id=session_id,
         stock_code=stock_code,
         order_type=order_type,
         quantity=quantity,
@@ -947,6 +1027,7 @@ def _build_queued_order_intent(
 def _maybe_queue_order_intent(
     *,
     market: MarketInfo,
+    session_id: str,
     stock_code: str,
     order_type: str,
     quantity: int,
@@ -956,9 +1037,11 @@ def _maybe_queue_order_intent(
     if not BLACKOUT_ORDER_MANAGER.in_blackout():
         return False
 
+    before_overflow_drops = BLACKOUT_ORDER_MANAGER.overflow_drop_count
     queued = BLACKOUT_ORDER_MANAGER.enqueue(
         _build_queued_order_intent(
             market=market,
+            session_id=session_id,
             stock_code=stock_code,
             order_type=order_type,
             quantity=quantity,
@@ -967,6 +1050,7 @@ def _maybe_queue_order_intent(
         )
     )
     if queued:
+        after_overflow_drops = BLACKOUT_ORDER_MANAGER.overflow_drop_count
         logger.warning(
             (
                 "Blackout active: queued order intent %s %s (%s) "
@@ -980,9 +1064,22 @@ def _maybe_queue_order_intent(
             source,
             BLACKOUT_ORDER_MANAGER.pending_count,
         )
+        if after_overflow_drops > before_overflow_drops:
+            logger.error(
+                (
+                    "Blackout queue overflow policy applied: evicted oldest intent "
+                    "to keep latest %s %s (%s) source=%s pending=%d total_evicted=%d"
+                ),
+                order_type,
+                stock_code,
+                market.code,
+                source,
+                BLACKOUT_ORDER_MANAGER.pending_count,
+                after_overflow_drops,
+            )
     else:
         logger.error(
-            "Blackout queue full: dropped order intent %s %s (%s) qty=%d source=%s",
+            "Blackout queue unavailable: could not queue order intent %s %s (%s) qty=%d source=%s",
             order_type,
             stock_code,
             market.code,
@@ -1115,7 +1212,6 @@ async def process_blackout_recovery_orders(
 
             accepted = result.get("rt_cd", "0") == "0"
             if accepted:
-                runtime_session_id = get_session_info(market).session_id
                 log_trade(
                     conn=db_conn,
                     stock_code=intent.stock_code,
@@ -1127,7 +1223,7 @@ async def process_blackout_recovery_orders(
                     pnl=0.0,
                     market=market.code,
                     exchange_code=market.exchange_code,
-                    session_id=runtime_session_id,
+                    session_id=intent.session_id,
                 )
                 logger.info(
                     "Recovered queued order executed: %s %s (%s) qty=%d price=%.4f source=%s",
@@ -1279,7 +1375,10 @@ async def _cancel_pending_orders_for_kill_switch(
                 )
 
     if failures:
-        raise RuntimeError("; ".join(failures[:3]))
+        summary = "; ".join(failures[:3])
+        if len(failures) > 3:
+            summary = f"{summary} (+{len(failures) - 3} more)"
+        raise RuntimeError(summary)
 
 
 async def _refresh_order_state_for_kill_switch(
@@ -1288,6 +1387,7 @@ async def _refresh_order_state_for_kill_switch(
     overseas_broker: OverseasBroker,
     markets: list[MarketInfo],
 ) -> None:
+    failures: list[str] = []
     seen_overseas: set[str] = set()
     for market in markets:
         try:
@@ -1303,6 +1403,12 @@ async def _refresh_order_state_for_kill_switch(
                 market.exchange_code,
                 exc,
             )
+            failures.append(f"{market.code}/{market.exchange_code}: {exc}")
+    if failures:
+        summary = "; ".join(failures[:3])
+        if len(failures) > 3:
+            summary = f"{summary} (+{len(failures) - 3} more)"
+        raise RuntimeError(summary)
 
 
 def _reduce_risk_for_kill_switch() -> None:
@@ -1372,6 +1478,7 @@ async def trading_cycle(
     _session_risk_overrides(market=market, settings=settings)
 
     # 1. Fetch market data
+    balance_info: dict[str, Any] = {}
     price_output: dict[str, Any] = {}  # Populated for overseas markets; used for fallback metrics
     if market.is_domestic:
         current_price, price_change_pct, foreigner_net = await broker.get_current_price(stock_code)
@@ -1394,8 +1501,6 @@ async def trading_cycle(
             balance_info = output2[0]
         elif isinstance(output2, dict):
             balance_info = output2
-        else:
-            balance_info = {}
 
         total_eval = safe_float(balance_info.get("frcr_evlu_tota", "0") or "0")
         purchase_total = safe_float(balance_info.get("frcr_buy_amt_smtl", "0") or "0")
@@ -1815,6 +1920,9 @@ async def trading_cycle(
     quantity = 0
     trade_price = current_price
     trade_pnl = 0.0
+    buy_trade: dict[str, Any] | None = None
+    buy_price = 0.0
+    sell_qty = 0
     if decision.action in ("BUY", "SELL"):
         if KILL_SWITCH.new_orders_blocked and decision.action == "BUY":
             logger.critical(
@@ -1962,6 +2070,7 @@ async def trading_cycle(
                 return
             if _maybe_queue_order_intent(
                 market=market,
+                session_id=runtime_session_id,
                 stock_code=stock_code,
                 order_type=decision.action,
                 quantity=quantity,
@@ -2009,6 +2118,7 @@ async def trading_cycle(
                 return
             if _maybe_queue_order_intent(
                 market=market,
+                session_id=runtime_session_id,
                 stock_code=stock_code,
                 order_type=decision.action,
                 quantity=quantity,
@@ -2129,6 +2239,26 @@ async def trading_cycle(
             "signal": candidate.signal,
             "score": candidate.score,
         }
+    sell_fx_rate = _extract_fx_rate_from_sources(price_output, balance_info)
+    if sell_fx_rate is not None and not market.is_domestic:
+        if selection_context is None:
+            selection_context = {"fx_rate": sell_fx_rate}
+        else:
+            selection_context["fx_rate"] = sell_fx_rate
+
+    strategy_pnl: float | None = None
+    fx_pnl: float | None = None
+    if decision.action == "SELL" and order_succeeded:
+        buy_fx_rate = _extract_buy_fx_rate(buy_trade)
+        strategy_pnl, fx_pnl = _split_trade_pnl_components(
+            market=market,
+            trade_pnl=trade_pnl,
+            buy_price=buy_price,
+            sell_price=trade_price,
+            quantity=sell_qty or quantity,
+            buy_fx_rate=buy_fx_rate,
+            sell_fx_rate=sell_fx_rate,
+        )
 
     log_trade(
         conn=db_conn,
@@ -2139,6 +2269,8 @@ async def trading_cycle(
         quantity=quantity,
         price=trade_price,
         pnl=trade_pnl,
+        strategy_pnl=strategy_pnl,
+        fx_pnl=fx_pnl,
         market=market.code,
         exchange_code=market.exchange_code,
         session_id=runtime_session_id,
@@ -2737,6 +2869,7 @@ async def run_daily_session(
             )
             continue
 
+        balance_info: dict[str, Any] = {}
         if market.is_domestic:
             output2 = balance_data.get("output2", [{}])
             total_eval = safe_float(output2[0].get("tot_evlu_amt", "0")) if output2 else 0
@@ -2991,6 +3124,9 @@ async def run_daily_session(
             quantity = 0
             trade_price = stock_data["current_price"]
             trade_pnl = 0.0
+            buy_trade: dict[str, Any] | None = None
+            buy_price = 0.0
+            sell_qty = 0
             order_succeeded = True
             if decision.action in ("BUY", "SELL"):
                 if KILL_SWITCH.new_orders_blocked and decision.action == "BUY":
@@ -3143,6 +3279,7 @@ async def run_daily_session(
                             continue
                         if _maybe_queue_order_intent(
                             market=market,
+                            session_id=runtime_session_id,
                             stock_code=stock_code,
                             order_type=decision.action,
                             quantity=quantity,
@@ -3180,6 +3317,7 @@ async def run_daily_session(
                             continue
                         if _maybe_queue_order_intent(
                             market=market,
+                            session_id=runtime_session_id,
                             stock_code=stock_code,
                             order_type=decision.action,
                             quantity=quantity,
@@ -3273,6 +3411,30 @@ async def run_daily_session(
             # Log trade (skip if order was rejected by API)
             if decision.action in ("BUY", "SELL") and not order_succeeded:
                 continue
+            strategy_pnl: float | None = None
+            fx_pnl: float | None = None
+            selection_context: dict[str, Any] | None = None
+            if decision.action == "SELL" and order_succeeded:
+                buy_fx_rate = _extract_buy_fx_rate(buy_trade)
+                sell_fx_rate = _extract_fx_rate_from_sources(balance_info, stock_data)
+                strategy_pnl, fx_pnl = _split_trade_pnl_components(
+                    market=market,
+                    trade_pnl=trade_pnl,
+                    buy_price=buy_price,
+                    sell_price=trade_price,
+                    quantity=sell_qty or quantity,
+                    buy_fx_rate=buy_fx_rate,
+                    sell_fx_rate=sell_fx_rate,
+                )
+                if sell_fx_rate is not None and not market.is_domestic:
+                    # Daily path does not carry scanner candidate metrics, so this
+                    # context intentionally stores FX snapshot only.
+                    selection_context = {"fx_rate": sell_fx_rate}
+            elif not market.is_domestic:
+                snapshot_fx_rate = _extract_fx_rate_from_sources(balance_info, stock_data)
+                if snapshot_fx_rate is not None:
+                    # BUY/HOLD in daily path: persist FX snapshot for later SELL split.
+                    selection_context = {"fx_rate": snapshot_fx_rate}
             log_trade(
                 conn=db_conn,
                 stock_code=stock_code,
@@ -3282,9 +3444,12 @@ async def run_daily_session(
                 quantity=quantity,
                 price=trade_price,
                 pnl=trade_pnl,
+                strategy_pnl=strategy_pnl,
+                fx_pnl=fx_pnl,
                 market=market.code,
                 exchange_code=market.exchange_code,
                 session_id=runtime_session_id,
+                selection_context=selection_context,
                 decision_id=decision_id,
                 mode=settings.MODE,
             )

src/strategy/exit_rules.py

@@ -85,8 +85,8 @@ def evaluate_exit(
         reason = "atr_trailing_stop"
     elif be_lock_threat:
         reason = "be_lock_threat"
-    elif model_exit_signal:
-        reason = "model_liquidity_exit"
+    elif model_exit_signal and next_state == PositionState.BE_LOCK:
+        reason = "model_assist_be_lock"
     elif take_profit_hit:
         # Backward-compatible immediate profit-taking path.
         reason = "arm_take_profit"

src/strategy/position_state_machine.py

@@ -40,7 +40,8 @@ def evaluate_exit_first(inp: StateTransitionInput) -> bool:
 
     EXITED must be evaluated before any promotion.
     """
-    return inp.hard_stop_hit or inp.trailing_stop_hit or inp.model_exit_signal or inp.be_lock_threat
+    # model_exit_signal is assist-only and must not trigger EXIT directly.
+    return inp.hard_stop_hit or inp.trailing_stop_hit or inp.be_lock_threat
 
 
 def promote_state(current: PositionState, inp: StateTransitionInput) -> PositionState:
@@ -61,5 +62,8 @@ def promote_state(current: PositionState, inp: StateTransitionInput) -> Position
         target = PositionState.ARMED
     elif inp.unrealized_pnl_pct >= inp.be_arm_pct:
         target = PositionState.BE_LOCK
+    elif inp.model_exit_signal:
+        # Model signal assists risk posture by tightening to BE_LOCK.
+        target = PositionState.BE_LOCK
 
     return target if _STATE_RANK[target] > _STATE_RANK[current] else current

tests/test_backtest_cost_guard.py

@@ -10,6 +10,7 @@ def test_valid_backtest_cost_model_passes() -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": 10.0, "US_PRE": 50.0},
         failure_rate_by_session={"KRX_REG": 0.01, "US_PRE": 0.08},
+        partial_fill_rate_by_session={"KRX_REG": 0.1, "US_PRE": 0.2},
         unfavorable_fill_required=True,
     )
     validate_backtest_cost_model(model=model, required_sessions=["KRX_REG", "US_PRE"])
@@ -20,6 +21,7 @@ def test_missing_required_slippage_session_raises() -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": 10.0},
         failure_rate_by_session={"KRX_REG": 0.01, "US_PRE": 0.08},
+        partial_fill_rate_by_session={"KRX_REG": 0.1, "US_PRE": 0.2},
         unfavorable_fill_required=True,
     )
     with pytest.raises(ValueError, match="missing slippage_bps_by_session.*US_PRE"):
@@ -31,6 +33,7 @@ def test_missing_required_failure_rate_session_raises() -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": 10.0, "US_PRE": 50.0},
         failure_rate_by_session={"KRX_REG": 0.01},
+        partial_fill_rate_by_session={"KRX_REG": 0.1, "US_PRE": 0.2},
         unfavorable_fill_required=True,
     )
     with pytest.raises(ValueError, match="missing failure_rate_by_session.*US_PRE"):
@@ -42,6 +45,7 @@ def test_invalid_failure_rate_range_raises() -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": 10.0},
         failure_rate_by_session={"KRX_REG": 1.2},
+        partial_fill_rate_by_session={"KRX_REG": 0.2},
         unfavorable_fill_required=True,
     )
     with pytest.raises(ValueError, match="failure rate must be within"):
@@ -53,6 +57,7 @@ def test_unfavorable_fill_requirement_cannot_be_disabled() -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": 10.0},
         failure_rate_by_session={"KRX_REG": 0.02},
+        partial_fill_rate_by_session={"KRX_REG": 0.2},
         unfavorable_fill_required=False,
     )
     with pytest.raises(ValueError, match="unfavorable_fill_required must be True"):
@@ -65,6 +70,7 @@ def test_non_finite_commission_rejected(bad_commission: float) -> None:
         commission_bps=bad_commission,
         slippage_bps_by_session={"KRX_REG": 10.0},
         failure_rate_by_session={"KRX_REG": 0.02},
+        partial_fill_rate_by_session={"KRX_REG": 0.2},
         unfavorable_fill_required=True,
     )
     with pytest.raises(ValueError, match="commission_bps"):
@@ -77,7 +83,33 @@ def test_non_finite_slippage_rejected(bad_slippage: float) -> None:
         commission_bps=5.0,
         slippage_bps_by_session={"KRX_REG": bad_slippage},
         failure_rate_by_session={"KRX_REG": 0.02},
+        partial_fill_rate_by_session={"KRX_REG": 0.2},
         unfavorable_fill_required=True,
     )
     with pytest.raises(ValueError, match="slippage bps"):
         validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])
+
+
+def test_missing_required_partial_fill_session_raises() -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": 10.0, "US_PRE": 50.0},
+        failure_rate_by_session={"KRX_REG": 0.01, "US_PRE": 0.08},
+        partial_fill_rate_by_session={"KRX_REG": 0.1},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="missing partial_fill_rate_by_session.*US_PRE"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG", "US_PRE"])
+
+
+@pytest.mark.parametrize("bad_partial_fill", [float("nan"), float("inf"), float("-inf"), -0.1, 1.1])
+def test_invalid_partial_fill_rate_rejected(bad_partial_fill: float) -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": 10.0},
+        failure_rate_by_session={"KRX_REG": 0.02},
+        partial_fill_rate_by_session={"KRX_REG": bad_partial_fill},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="partial fill rate must be within"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])

tests/test_backtest_pipeline_integration.py

@@ -35,6 +35,7 @@ def _cost_model() -> BacktestCostModel:
         commission_bps=3.0,
         slippage_bps_by_session={"KRX_REG": 10.0, "US_PRE": 50.0},
         failure_rate_by_session={"KRX_REG": 0.01, "US_PRE": 0.08},
+        partial_fill_rate_by_session={"KRX_REG": 0.05, "US_PRE": 0.2},
         unfavorable_fill_required=True,
     )
 
@@ -71,6 +72,10 @@ def test_pipeline_happy_path_returns_fold_and_artifact_contract() -> None:
         assert names == {"B0", "B1", "M1"}
         for score in fold.baseline_scores:
             assert 0.0 <= score.accuracy <= 1.0
+            assert 0.0 <= score.cost_adjusted_accuracy <= 1.0
+        assert fold.execution_adjusted_trade_count >= 0
+        assert fold.execution_rejected_count >= 0
+        assert fold.execution_partial_count >= 0
 
 
 def test_pipeline_cost_guard_fail_fast() -> None:
@@ -78,6 +83,7 @@ def test_pipeline_cost_guard_fail_fast() -> None:
         commission_bps=3.0,
         slippage_bps_by_session={"KRX_REG": 10.0},
         failure_rate_by_session={"KRX_REG": 0.01},
+        partial_fill_rate_by_session={"KRX_REG": 0.05},
         unfavorable_fill_required=True,
     )
     try:
@@ -166,3 +172,49 @@ def test_pipeline_rejects_minutes_spec_when_timestamp_missing() -> None:
         assert "BacktestBar.timestamp is required" in str(exc)
     else:
         raise AssertionError("expected timestamp validation error")
+
+
+def test_pipeline_fold_scores_reflect_cost_and_execution_effects() -> None:
+    cfg = dict(
+        bars=_bars(),
+        entry_indices=[0, 1, 2, 3, 4, 5, 6, 7],
+        side=1,
+        triple_barrier_spec=TripleBarrierSpec(
+            take_profit_pct=0.02,
+            stop_loss_pct=0.01,
+            max_holding_minutes=3,
+        ),
+        walk_forward=WalkForwardConfig(
+            train_size=4,
+            test_size=2,
+            step_size=2,
+            purge_size=1,
+            embargo_size=1,
+            min_train_size=3,
+        ),
+    )
+    optimistic = BacktestCostModel(
+        commission_bps=0.0,
+        slippage_bps_by_session={"KRX_REG": 0.0, "US_PRE": 0.0},
+        failure_rate_by_session={"KRX_REG": 0.0, "US_PRE": 0.0},
+        partial_fill_rate_by_session={"KRX_REG": 0.0, "US_PRE": 0.0},
+        unfavorable_fill_required=True,
+    )
+    conservative = BacktestCostModel(
+        commission_bps=10.0,
+        slippage_bps_by_session={"KRX_REG": 30.0, "US_PRE": 80.0},
+        failure_rate_by_session={"KRX_REG": 0.2, "US_PRE": 0.4},
+        partial_fill_rate_by_session={"KRX_REG": 0.5, "US_PRE": 0.7},
+        unfavorable_fill_required=True,
+    )
+    optimistic_out = run_v2_backtest_pipeline(cost_model=optimistic, **cfg)
+    conservative_out = run_v2_backtest_pipeline(cost_model=conservative, **cfg)
+
+    assert optimistic_out.folds and conservative_out.folds
+    optimistic_score = optimistic_out.folds[0].baseline_scores[1].cost_adjusted_accuracy
+    conservative_score = conservative_out.folds[0].baseline_scores[1].cost_adjusted_accuracy
+    assert conservative_score < optimistic_score
+
+    optimistic_avg_return = optimistic_out.folds[0].execution_adjusted_avg_return_bps
+    conservative_avg_return = conservative_out.folds[0].execution_adjusted_avg_return_bps
+    assert conservative_avg_return < optimistic_avg_return

tests/test_blackout_manager.py

@@ -35,6 +35,7 @@ def test_recovery_batch_only_after_blackout_exit() -> None:
     intent = QueuedOrderIntent(
         market_code="KR",
         exchange_code="KRX",
+        session_id="KRX_REG",
         stock_code="005930",
         order_type="BUY",
         quantity=1,
@@ -64,6 +65,7 @@ def test_requeued_intent_is_processed_next_non_blackout_cycle() -> None:
     intent = QueuedOrderIntent(
         market_code="KR",
         exchange_code="KRX",
+        session_id="KRX_REG",
         stock_code="005930",
         order_type="BUY",
         quantity=1,
@@ -79,3 +81,54 @@ def test_requeued_intent_is_processed_next_non_blackout_cycle() -> None:
     manager.requeue(first_batch[0])
     second_batch = manager.pop_recovery_batch(outside_blackout)
     assert len(second_batch) == 1
+
+
+def test_queue_overflow_drops_oldest_and_keeps_latest() -> None:
+    manager = BlackoutOrderManager(
+        enabled=True,
+        windows=parse_blackout_windows_kst("23:30-00:10"),
+        max_queue_size=2,
+    )
+    first = QueuedOrderIntent(
+        market_code="KR",
+        exchange_code="KRX",
+        session_id="KRX_REG",
+        stock_code="000001",
+        order_type="BUY",
+        quantity=1,
+        price=100.0,
+        source="first",
+        queued_at=datetime.now(UTC),
+    )
+    second = QueuedOrderIntent(
+        market_code="KR",
+        exchange_code="KRX",
+        session_id="KRX_REG",
+        stock_code="000002",
+        order_type="BUY",
+        quantity=1,
+        price=101.0,
+        source="second",
+        queued_at=datetime.now(UTC),
+    )
+    third = QueuedOrderIntent(
+        market_code="KR",
+        exchange_code="KRX",
+        session_id="KRX_REG",
+        stock_code="000003",
+        order_type="SELL",
+        quantity=2,
+        price=102.0,
+        source="third",
+        queued_at=datetime.now(UTC),
+    )
+
+    assert manager.enqueue(first)
+    assert manager.enqueue(second)
+    assert manager.enqueue(third)
+    assert manager.pending_count == 2
+    assert manager.overflow_drop_count == 1
+
+    outside_blackout = datetime(2026, 1, 1, 15, 20, tzinfo=UTC)
+    batch = manager.pop_recovery_batch(outside_blackout)
+    assert [intent.stock_code for intent in batch] == ["000002", "000003"]

tests/test_kill_switch.py

@@ -53,3 +53,52 @@ async def test_kill_switch_collects_step_errors() -> None:
 
     report = await ks.trigger(reason="test", cancel_pending_orders=_boom)
     assert any(err.startswith("cancel_pending_orders:") for err in report.errors)
+
+
+@pytest.mark.asyncio
+async def test_kill_switch_refresh_retries_then_succeeds() -> None:
+    ks = KillSwitchOrchestrator()
+    refresh_calls = {"count": 0}
+
+    def _flaky_refresh() -> None:
+        refresh_calls["count"] += 1
+        if refresh_calls["count"] < 3:
+            raise RuntimeError("temporary refresh failure")
+
+    report = await ks.trigger(
+        reason="test",
+        refresh_order_state=_flaky_refresh,
+        refresh_retry_attempts=3,
+        refresh_retry_base_delay_sec=0.0,
+    )
+    assert refresh_calls["count"] == 3
+    assert report.errors == []
+
+
+@pytest.mark.asyncio
+async def test_kill_switch_refresh_retry_exhausted_records_error_and_continues() -> None:
+    ks = KillSwitchOrchestrator()
+    calls: list[str] = []
+
+    def _refresh_fail() -> None:
+        raise RuntimeError("persistent refresh failure")
+
+    def _reduce() -> None:
+        calls.append("reduce")
+
+    def _snapshot() -> None:
+        calls.append("snapshot")
+
+    report = await ks.trigger(
+        reason="test",
+        refresh_order_state=_refresh_fail,
+        reduce_risk=_reduce,
+        snapshot_state=_snapshot,
+        refresh_retry_attempts=2,
+        refresh_retry_base_delay_sec=0.0,
+    )
+    assert any(
+        err.startswith("refresh_order_state: failed after 2 attempts")
+        for err in report.errors
+    )
+    assert calls == ["reduce", "snapshot"]

tests/test_main.py

@@ -1,6 +1,7 @@
 """Tests for main trading loop integration."""
 
 from datetime import UTC, date, datetime
+from typing import Any
 from unittest.mock import ANY, AsyncMock, MagicMock, patch
 
 import pytest
@@ -9,6 +10,7 @@ import src.main as main_module
 from src.config import Settings
 from src.context.layer import ContextLayer
 from src.context.scheduler import ScheduleResult
+from src.core.blackout_manager import BlackoutOrderManager
 from src.core.order_policy import OrderPolicyRejected, get_session_info
 from src.core.risk_manager import CircuitBreakerTripped, FatFingerRejected
 from src.db import init_db, log_trade
@@ -33,6 +35,7 @@ from src.main import (
     _extract_held_qty_from_balance,
     _handle_market_close,
     _inject_staged_exit_features,
+    _maybe_queue_order_intent,
     _resolve_market_setting,
     _resolve_sell_qty_for_pnl,
     _retry_connection,
@@ -40,6 +43,7 @@ from src.main import (
     _run_evolution_loop,
     _should_block_overseas_buy_for_fx_buffer,
     _should_force_exit_for_overnight,
+    _split_trade_pnl_components,
     _start_dashboard_server,
     _stoploss_cooldown_minutes,
     _trigger_emergency_kill_switch,
@@ -102,22 +106,22 @@ def _make_sell_match(stock_code: str = "005930") -> ScenarioMatch:
 @pytest.fixture(autouse=True)
 def _reset_kill_switch_state() -> None:
     """Prevent cross-test leakage from global kill-switch state."""
+    def _reset_session_risk_globals() -> None:
+        _SESSION_RISK_LAST_BY_MARKET.clear()
+        _SESSION_RISK_OVERRIDES_BY_MARKET.clear()
+        _SESSION_RISK_PROFILES_MAP.clear()
+        main_module._SESSION_RISK_PROFILES_RAW = "{}"
+
     KILL_SWITCH.clear_block()
     _RUNTIME_EXIT_STATES.clear()
     _RUNTIME_EXIT_PEAKS.clear()
-    _SESSION_RISK_LAST_BY_MARKET.clear()
-    _SESSION_RISK_OVERRIDES_BY_MARKET.clear()
-    _SESSION_RISK_PROFILES_MAP.clear()
-    main_module._SESSION_RISK_PROFILES_RAW = "__reset__"
+    _reset_session_risk_globals()
     _STOPLOSS_REENTRY_COOLDOWN_UNTIL.clear()
     yield
     KILL_SWITCH.clear_block()
     _RUNTIME_EXIT_STATES.clear()
     _RUNTIME_EXIT_PEAKS.clear()
-    _SESSION_RISK_LAST_BY_MARKET.clear()
-    _SESSION_RISK_OVERRIDES_BY_MARKET.clear()
-    _SESSION_RISK_PROFILES_MAP.clear()
-    main_module._SESSION_RISK_PROFILES_RAW = "__reset__"
+    _reset_session_risk_globals()
     _STOPLOSS_REENTRY_COOLDOWN_UNTIL.clear()
 
 
@@ -3181,6 +3185,13 @@ async def test_sell_order_uses_broker_balance_qty_not_db() -> None:
     updated_buy = decision_logger.get_decision_by_id(buy_decision_id)
     assert updated_buy is not None
     assert updated_buy.outcome_pnl == -25.0
+    sell_row = db_conn.execute(
+        "SELECT pnl, strategy_pnl, fx_pnl FROM trades WHERE action='SELL' ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    assert sell_row is not None
+    assert sell_row[0] == -25.0
+    assert sell_row[1] == -25.0
+    assert sell_row[2] == 0.0
 
 
 @pytest.mark.asyncio
@@ -4598,6 +4609,23 @@ def test_fx_buffer_guard_applies_only_to_us_and_respects_boundary() -> None:
     assert required_jp == 0.0
 
 
+def test_split_trade_pnl_components_overseas_fx_split_preserves_total() -> None:
+    market = MagicMock()
+    market.is_domestic = False
+    strategy_pnl, fx_pnl = _split_trade_pnl_components(
+        market=market,
+        trade_pnl=20.0,
+        buy_price=100.0,
+        sell_price=110.0,
+        quantity=2,
+        buy_fx_rate=1200.0,
+        sell_fx_rate=1260.0,
+    )
+    assert strategy_pnl == 10.0
+    assert fx_pnl == 10.0
+    assert strategy_pnl + fx_pnl == pytest.approx(20.0)
+
+
 # run_daily_session — daily CB baseline (daily_start_eval) tests (issue #207)
 # ---------------------------------------------------------------------------
 
@@ -6351,6 +6379,225 @@ async def test_us_min_price_filter_not_applied_to_kr_market() -> None:
     broker.send_order.assert_called_once()
 
 
+@pytest.mark.asyncio
+async def test_session_boundary_reloads_us_min_price_override_in_trading_cycle() -> None:
+    db_conn = init_db(":memory:")
+    decision_logger = DecisionLogger(db_conn)
+
+    broker = MagicMock()
+    broker.get_balance = AsyncMock(return_value={"output1": [], "output2": [{}]})
+
+    overseas_broker = MagicMock()
+    overseas_broker.get_overseas_price = AsyncMock(
+        return_value={"output": {"last": "7.0", "rate": "0.0"}}
+    )
+    overseas_broker.get_overseas_balance = AsyncMock(
+        return_value={
+            "output1": [],
+            "output2": [{"frcr_evlu_tota": "10000", "frcr_buy_amt_smtl": "0"}],
+        }
+    )
+    overseas_broker.get_overseas_buying_power = AsyncMock(
+        return_value={"output": {"ovrs_ord_psbl_amt": "10000"}}
+    )
+    overseas_broker.send_overseas_order = AsyncMock(return_value={"rt_cd": "0", "msg1": "OK"})
+
+    market = MagicMock()
+    market.name = "NASDAQ"
+    market.code = "US_NASDAQ"
+    market.exchange_code = "NASD"
+    market.is_domestic = False
+
+    telegram = MagicMock()
+    telegram.notify_trade_execution = AsyncMock()
+    telegram.notify_fat_finger = AsyncMock()
+    telegram.notify_circuit_breaker = AsyncMock()
+    telegram.notify_scenario_matched = AsyncMock()
+
+    settings = Settings(
+        KIS_APP_KEY="k",
+        KIS_APP_SECRET="s",
+        KIS_ACCOUNT_NO="12345678-01",
+        GEMINI_API_KEY="g",
+        MODE="paper",
+        PAPER_OVERSEAS_CASH=50000.0,
+        US_MIN_PRICE=5.0,
+        USD_BUFFER_MIN=1000.0,
+        SESSION_RISK_RELOAD_ENABLED=True,
+        SESSION_RISK_PROFILES_JSON=(
+            '{"US_PRE": {"US_MIN_PRICE": 8.0}, "US_DAY": {"US_MIN_PRICE": 5.0}}'
+        ),
+    )
+
+    current_session = {"id": "US_PRE"}
+
+    def _session_info(_: Any) -> MagicMock:
+        return MagicMock(session_id=current_session["id"])
+
+    with (
+        patch("src.main.get_open_position", return_value=None),
+        patch("src.main.get_session_info", side_effect=_session_info),
+    ):
+        await trading_cycle(
+            broker=broker,
+            overseas_broker=overseas_broker,
+            scenario_engine=MagicMock(evaluate=MagicMock(return_value=_make_buy_match("AAPL"))),
+            playbook=_make_playbook("US_NASDAQ"),
+            risk=MagicMock(validate_order=MagicMock(), check_circuit_breaker=MagicMock()),
+            db_conn=db_conn,
+            decision_logger=decision_logger,
+            context_store=MagicMock(
+                get_latest_timeframe=MagicMock(return_value=None),
+                set_context=MagicMock(),
+            ),
+            criticality_assessor=MagicMock(
+                assess_market_conditions=MagicMock(return_value=MagicMock(value="NORMAL")),
+                get_timeout=MagicMock(return_value=5.0),
+            ),
+            telegram=telegram,
+            market=market,
+            stock_code="AAPL",
+            scan_candidates={},
+            settings=settings,
+        )
+        assert overseas_broker.send_overseas_order.call_count == 0
+
+        current_session["id"] = "US_DAY"
+        await trading_cycle(
+            broker=broker,
+            overseas_broker=overseas_broker,
+            scenario_engine=MagicMock(evaluate=MagicMock(return_value=_make_buy_match("AAPL"))),
+            playbook=_make_playbook("US_NASDAQ"),
+            risk=MagicMock(validate_order=MagicMock(), check_circuit_breaker=MagicMock()),
+            db_conn=db_conn,
+            decision_logger=decision_logger,
+            context_store=MagicMock(
+                get_latest_timeframe=MagicMock(return_value=None),
+                set_context=MagicMock(),
+            ),
+            criticality_assessor=MagicMock(
+                assess_market_conditions=MagicMock(return_value=MagicMock(value="NORMAL")),
+                get_timeout=MagicMock(return_value=5.0),
+            ),
+            telegram=telegram,
+            market=market,
+            stock_code="AAPL",
+            scan_candidates={},
+            settings=settings,
+        )
+
+    assert overseas_broker.send_overseas_order.call_count == 1
+
+
+@pytest.mark.asyncio
+async def test_session_boundary_falls_back_when_profile_reload_fails() -> None:
+    db_conn = init_db(":memory:")
+    decision_logger = DecisionLogger(db_conn)
+
+    broker = MagicMock()
+    broker.get_balance = AsyncMock(return_value={"output1": [], "output2": [{}]})
+
+    overseas_broker = MagicMock()
+    overseas_broker.get_overseas_price = AsyncMock(
+        return_value={"output": {"last": "7.0", "rate": "0.0"}}
+    )
+    overseas_broker.get_overseas_balance = AsyncMock(
+        return_value={
+            "output1": [],
+            "output2": [{"frcr_evlu_tota": "10000", "frcr_buy_amt_smtl": "0"}],
+        }
+    )
+    overseas_broker.get_overseas_buying_power = AsyncMock(
+        return_value={"output": {"ovrs_ord_psbl_amt": "10000"}}
+    )
+    overseas_broker.send_overseas_order = AsyncMock(return_value={"rt_cd": "0", "msg1": "OK"})
+
+    market = MagicMock()
+    market.name = "NASDAQ"
+    market.code = "US_NASDAQ"
+    market.exchange_code = "NASD"
+    market.is_domestic = False
+
+    telegram = MagicMock()
+    telegram.notify_trade_execution = AsyncMock()
+    telegram.notify_fat_finger = AsyncMock()
+    telegram.notify_circuit_breaker = AsyncMock()
+    telegram.notify_scenario_matched = AsyncMock()
+
+    settings = Settings(
+        KIS_APP_KEY="k",
+        KIS_APP_SECRET="s",
+        KIS_ACCOUNT_NO="12345678-01",
+        GEMINI_API_KEY="g",
+        MODE="paper",
+        PAPER_OVERSEAS_CASH=50000.0,
+        US_MIN_PRICE=5.0,
+        USD_BUFFER_MIN=1000.0,
+        SESSION_RISK_RELOAD_ENABLED=True,
+        SESSION_RISK_PROFILES_JSON='{"US_PRE": {"US_MIN_PRICE": 8.0}}',
+    )
+
+    current_session = {"id": "US_PRE"}
+
+    def _session_info(_: Any) -> MagicMock:
+        return MagicMock(session_id=current_session["id"])
+
+    with (
+        patch("src.main.get_open_position", return_value=None),
+        patch("src.main.get_session_info", side_effect=_session_info),
+    ):
+        await trading_cycle(
+            broker=broker,
+            overseas_broker=overseas_broker,
+            scenario_engine=MagicMock(evaluate=MagicMock(return_value=_make_buy_match("AAPL"))),
+            playbook=_make_playbook("US_NASDAQ"),
+            risk=MagicMock(validate_order=MagicMock(), check_circuit_breaker=MagicMock()),
+            db_conn=db_conn,
+            decision_logger=decision_logger,
+            context_store=MagicMock(
+                get_latest_timeframe=MagicMock(return_value=None),
+                set_context=MagicMock(),
+            ),
+            criticality_assessor=MagicMock(
+                assess_market_conditions=MagicMock(return_value=MagicMock(value="NORMAL")),
+                get_timeout=MagicMock(return_value=5.0),
+            ),
+            telegram=telegram,
+            market=market,
+            stock_code="AAPL",
+            scan_candidates={},
+            settings=settings,
+        )
+        assert overseas_broker.send_overseas_order.call_count == 0
+
+        settings.SESSION_RISK_PROFILES_JSON = "{invalid-json"
+        current_session["id"] = "US_DAY"
+        await trading_cycle(
+            broker=broker,
+            overseas_broker=overseas_broker,
+            scenario_engine=MagicMock(evaluate=MagicMock(return_value=_make_buy_match("AAPL"))),
+            playbook=_make_playbook("US_NASDAQ"),
+            risk=MagicMock(validate_order=MagicMock(), check_circuit_breaker=MagicMock()),
+            db_conn=db_conn,
+            decision_logger=decision_logger,
+            context_store=MagicMock(
+                get_latest_timeframe=MagicMock(return_value=None),
+                set_context=MagicMock(),
+            ),
+            criticality_assessor=MagicMock(
+                assess_market_conditions=MagicMock(return_value=MagicMock(value="NORMAL")),
+                get_timeout=MagicMock(return_value=5.0),
+            ),
+            telegram=telegram,
+            market=market,
+            stock_code="AAPL",
+            scan_candidates={},
+            settings=settings,
+        )
+
+    assert overseas_broker.send_overseas_order.call_count == 1
+
+
 def test_overnight_policy_prioritizes_killswitch_over_exception() -> None:
     market = MagicMock()
     with patch("src.main.get_session_info", return_value=MagicMock(session_id="US_AFTER")):
@@ -6475,6 +6722,7 @@ async def test_blackout_queues_order_and_skips_submission() -> None:
     blackout_manager.in_blackout.return_value = True
     blackout_manager.enqueue.return_value = True
     blackout_manager.pending_count = 1
+    blackout_manager.overflow_drop_count = 0
 
     with patch("src.main.BLACKOUT_ORDER_MANAGER", blackout_manager):
         await trading_cycle(
@@ -6504,6 +6752,43 @@ async def test_blackout_queues_order_and_skips_submission() -> None:
     blackout_manager.enqueue.assert_called_once()
 
 
+def test_blackout_queue_overflow_keeps_latest_intent() -> None:
+    manager = BlackoutOrderManager(enabled=True, windows=[], max_queue_size=1)
+    manager.in_blackout = lambda now=None: True  # type: ignore[method-assign]
+
+    market = MagicMock()
+    market.code = "KR"
+    market.exchange_code = "KRX"
+
+    with patch("src.main.BLACKOUT_ORDER_MANAGER", manager):
+        assert _maybe_queue_order_intent(
+            market=market,
+            session_id="KRX_REG",
+            stock_code="005930",
+            order_type="BUY",
+            quantity=1,
+            price=100.0,
+            source="test-first",
+        )
+        assert _maybe_queue_order_intent(
+            market=market,
+            session_id="KRX_REG",
+            stock_code="000660",
+            order_type="BUY",
+            quantity=2,
+            price=200.0,
+            source="test-second",
+        )
+
+    assert manager.pending_count == 1
+    assert manager.overflow_drop_count == 1
+    manager.in_blackout = lambda now=None: False  # type: ignore[method-assign]
+    batch = manager.pop_recovery_batch()
+    assert len(batch) == 1
+    assert batch[0].stock_code == "000660"
+    assert batch[0].session_id == "KRX_REG"
+
+
 @pytest.mark.asyncio
 async def test_process_blackout_recovery_executes_valid_intents() -> None:
     """Recovery must execute queued intents that pass revalidation."""
@@ -6525,6 +6810,7 @@ async def test_process_blackout_recovery_executes_valid_intents() -> None:
     intent.quantity = 1
     intent.price = 100.0
     intent.source = "test"
+    intent.session_id = "NXT_AFTER"
     intent.attempts = 0
 
     blackout_manager = MagicMock()
@@ -6555,7 +6841,7 @@ async def test_process_blackout_recovery_executes_valid_intents() -> None:
     assert row is not None
     assert row[0] == "BUY"
     assert row[1] == 1
-    assert row[2] == "KRX_REG"
+    assert row[2] == "NXT_AFTER"
     assert row[3].startswith("[blackout-recovery]")
 
 
@@ -6580,6 +6866,7 @@ async def test_process_blackout_recovery_drops_policy_rejected_intent() -> None:
     intent.quantity = 1
     intent.price = 100.0
     intent.source = "test"
+    intent.session_id = "KRX_REG"
     intent.attempts = 0
 
     blackout_manager = MagicMock()
@@ -6629,6 +6916,7 @@ async def test_process_blackout_recovery_drops_intent_on_excessive_price_drift()
     intent.quantity = 1
     intent.price = 100.0
     intent.source = "test"
+    intent.session_id = "US_PRE"
     intent.attempts = 0
 
     blackout_manager = MagicMock()
@@ -6679,6 +6967,7 @@ async def test_process_blackout_recovery_drops_overseas_intent_on_excessive_pric
     intent.quantity = 1
     intent.price = 100.0
     intent.source = "test"
+    intent.session_id = "KRX_REG"
     intent.attempts = 0
 
     blackout_manager = MagicMock()
@@ -6728,6 +7017,7 @@ async def test_process_blackout_recovery_requeues_intent_when_price_lookup_fails
     intent.quantity = 1
     intent.price = 100.0
     intent.source = "test"
+    intent.session_id = "KRX_REG"
     intent.attempts = 0
 
     blackout_manager = MagicMock()
@@ -6864,3 +7154,27 @@ async def test_trigger_emergency_kill_switch_records_cancel_failure() -> None:
         )
 
     assert any(err.startswith("cancel_pending_orders:") for err in report.errors)
+
+
+@pytest.mark.asyncio
+async def test_refresh_order_state_failure_summary_includes_more_count() -> None:
+    broker = MagicMock()
+    broker.get_balance = AsyncMock(side_effect=RuntimeError("domestic down"))
+    overseas_broker = MagicMock()
+    overseas_broker.get_overseas_balance = AsyncMock(side_effect=RuntimeError("overseas down"))
+
+    markets = []
+    for code, exchange in [("KR", "KRX"), ("US_PRE", "NASD"), ("US_DAY", "NYSE"), ("JP", "TKSE")]:
+        market = MagicMock()
+        market.code = code
+        market.exchange_code = exchange
+        market.is_domestic = code == "KR"
+        markets.append(market)
+
+    with pytest.raises(RuntimeError, match=r"\(\+1 more\)$") as exc_info:
+        await main_module._refresh_order_state_for_kill_switch(
+            broker=broker,
+            overseas_broker=overseas_broker,
+            markets=markets,
+        )
+    assert "KR/KRX" in str(exc_info.value)

tests/test_strategy_exit_rules.py

@@ -22,17 +22,18 @@ def test_take_profit_exit_for_backward_compatibility() -> None:
     assert out.reason == "arm_take_profit"
 
 
-def test_model_assist_exit_signal() -> None:
+def test_model_assist_signal_promotes_be_lock_without_direct_exit() -> None:
     out = evaluate_exit(
-        current_state=PositionState.ARMED,
-        config=ExitRuleConfig(model_prob_threshold=0.62, arm_pct=10.0),
+        current_state=PositionState.HOLDING,
+        config=ExitRuleConfig(model_prob_threshold=0.62, be_arm_pct=1.2, arm_pct=10.0),
         inp=ExitRuleInput(
-            current_price=101.0,
+            current_price=100.5,
             entry_price=100.0,
             peak_price=105.0,
             pred_down_prob=0.8,
             liquidity_weak=True,
         ),
     )
-    assert out.should_exit is True
-    assert out.reason == "model_liquidity_exit"
+    assert out.should_exit is False
+    assert out.state == PositionState.BE_LOCK
+    assert out.reason == "model_assist_be_lock"

tests/test_strategy_state_machine.py

@@ -28,3 +28,29 @@ def test_exited_has_priority_over_promotion() -> None:
         ),
     )
     assert state == PositionState.EXITED
+
+
+def test_model_signal_promotes_be_lock_as_assist() -> None:
+    state = promote_state(
+        PositionState.HOLDING,
+        StateTransitionInput(
+            unrealized_pnl_pct=0.5,
+            be_arm_pct=1.2,
+            arm_pct=2.8,
+            model_exit_signal=True,
+        ),
+    )
+    assert state == PositionState.BE_LOCK
+
+
+def test_model_signal_does_not_force_exit_directly() -> None:
+    state = promote_state(
+        PositionState.ARMED,
+        StateTransitionInput(
+            unrealized_pnl_pct=1.0,
+            be_arm_pct=1.2,
+            arm_pct=2.8,
+            model_exit_signal=True,
+        ),
+    )
+    assert state == PositionState.ARMED

tests/test_validate_docs_sync.py

@@ -121,3 +121,44 @@ def test_validate_testing_doc_has_dynamic_count_guidance(monkeypatch) -> None:
     monkeypatch.setattr(module, "_read", fake_read)
     module.validate_testing_doc_has_dynamic_count_guidance(errors)
     assert errors == []
+
+
+def test_validate_pr_body_postcheck_guidance_passes(monkeypatch) -> None:
+    module = _load_module()
+    errors: list[str] = []
+    fake_docs = {
+        str(module.REQUIRED_FILES["commands"]): (
+            "PR Body Post-Check (Mandatory)\n"
+            "python3 scripts/validate_pr_body.py --pr <PR_NUMBER>\n"
+        ),
+        str(module.REQUIRED_FILES["workflow"]): (
+            "PR 생성 직후 본문 무결성 검증(필수)\n"
+            "python3 scripts/validate_pr_body.py --pr <PR_NUMBER>\n"
+        ),
+    }
+
+    def fake_read(path: Path) -> str:
+        return fake_docs[str(path)]
+
+    monkeypatch.setattr(module, "_read", fake_read)
+    module.validate_pr_body_postcheck_guidance(errors)
+    assert errors == []
+
+
+def test_validate_pr_body_postcheck_guidance_reports_missing_tokens(
+    monkeypatch,
+) -> None:
+    module = _load_module()
+    errors: list[str] = []
+    fake_docs = {
+        str(module.REQUIRED_FILES["commands"]): "PR Body Post-Check (Mandatory)\n",
+        str(module.REQUIRED_FILES["workflow"]): "PR Body Post-Check\n",
+    }
+
+    def fake_read(path: Path) -> str:
+        return fake_docs[str(path)]
+
+    monkeypatch.setattr(module, "_read", fake_read)
+    module.validate_pr_body_postcheck_guidance(errors)
+    assert any("commands.md" in err for err in errors)
+    assert any("workflow.md" in err for err in errors)

tests/test_validate_ouroboros_docs.py

@@ -79,3 +79,42 @@ def test_validate_links_avoids_duplicate_error_for_invalid_plan_link(tmp_path) -
 
     assert len(errors) == 1
     assert "invalid plan link path" in errors[0]
+
+
+def test_validate_issue_status_consistency_reports_conflicts() -> None:
+    module = _load_module()
+    errors: list[str] = []
+    path = Path("docs/ouroboros/80_implementation_audit.md").resolve()
+    text = "\n".join(
+        [
+            "| REQ-V3-004 | 상태 | 부분 | `#328` 잔여 |",
+            "| 항목 | 상태 | ✅ 완료 | `#328` 머지 |",
+        ]
+    )
+
+    module.validate_issue_status_consistency(path, text, errors)
+
+    assert len(errors) == 1
+    assert "conflicting status for issue #328" in errors[0]
+
+
+def test_validate_issue_status_consistency_allows_done_only() -> None:
+    module = _load_module()
+    errors: list[str] = []
+    path = Path("docs/ouroboros/80_implementation_audit.md").resolve()
+    text = "| 항목 | 상태 | ✅ 완료 | `#371` 머지 |"
+
+    module.validate_issue_status_consistency(path, text, errors)
+
+    assert errors == []
+
+
+def test_validate_issue_status_consistency_allows_pending_only() -> None:
+    module = _load_module()
+    errors: list[str] = []
+    path = Path("docs/ouroboros/80_implementation_audit.md").resolve()
+    text = "| 항목 | 상태 | 부분 | `#390` 추적 이슈 |"
+
+    module.validate_issue_status_consistency(path, text, errors)
+
+    assert errors == []

tests/test_validate_pr_body.py

@@ -0,0 +1,84 @@
+from __future__ import annotations
+
+import importlib.util
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+def _load_module():
+    script_path = Path(__file__).resolve().parents[1] / "scripts" / "validate_pr_body.py"
+    spec = importlib.util.spec_from_file_location("validate_pr_body", script_path)
+    assert spec is not None
+    assert spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_validate_pr_body_text_detects_escaped_newline() -> None:
+    module = _load_module()
+    errors = module.validate_pr_body_text("## Summary\\n- item")
+    assert any("escaped newline" in err for err in errors)
+
+
+def test_validate_pr_body_text_allows_literal_sequence_when_multiline() -> None:
+    module = _load_module()
+    text = "## Summary\n- escaped sequence example: \\\\n"
+    assert module.validate_pr_body_text(text) == []
+
+
+def test_validate_pr_body_text_detects_unbalanced_code_fence() -> None:
+    module = _load_module()
+    errors = module.validate_pr_body_text("## Summary\n- item\n```bash\necho hi\n")
+    assert any("unbalanced fenced code blocks" in err for err in errors)
+
+
+def test_validate_pr_body_text_detects_missing_structure() -> None:
+    module = _load_module()
+    errors = module.validate_pr_body_text("plain text only")
+    assert any("missing markdown section headers" in err for err in errors)
+    assert any("missing markdown list items" in err for err in errors)
+
+
+def test_validate_pr_body_text_passes_with_valid_markdown() -> None:
+    module = _load_module()
+    text = "\n".join(
+        [
+            "## Summary",
+            "- item",
+            "",
+            "## Validation",
+            "```bash",
+            "pytest -q",
+            "```",
+        ]
+    )
+    assert module.validate_pr_body_text(text) == []
+
+
+def test_fetch_pr_body_reads_body_from_tea_api(monkeypatch) -> None:
+    module = _load_module()
+
+    def fake_run(cmd, check, capture_output, text):  # noqa: ANN001
+        assert "tea" in cmd[0]
+        assert check is True
+        assert capture_output is True
+        assert text is True
+        return SimpleNamespace(stdout=json.dumps({"body": "## Summary\n- item"}))
+
+    monkeypatch.setattr(module.subprocess, "run", fake_run)
+    assert module.fetch_pr_body(391) == "## Summary\n- item"
+
+
+def test_fetch_pr_body_rejects_non_string_body(monkeypatch) -> None:
+    module = _load_module()
+
+    def fake_run(cmd, check, capture_output, text):  # noqa: ANN001
+        return SimpleNamespace(stdout=json.dumps({"body": 123}))
+
+    monkeypatch.setattr(module.subprocess, "run", fake_run)
+    with pytest.raises(RuntimeError):
+        module.fetch_pr_body(391)

workflow/session-handover.md

@@ -89,3 +89,51 @@
 - next_ticket: #316
 - process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
 - risks_or_notes: 모니터 판정을 liveness 중심에서 policy invariant(FORBIDDEN) 중심으로 전환
+
+### 2026-03-01 | session=codex-v3-stream-next-ticket
+- branch: feature/v3-session-policy-stream
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #368, #369, #370, #371, #374, #375, #376, #377, #381
+- next_ticket: #368
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: 비블로킹 소견은 합당성(정확성/안정성/유지보수성) 기준으로 반영하고, 미반영 시 근거를 코멘트로 남긴다.
+
+### 2026-03-01 | session=codex-issue368-start
+- branch: feature/issue-368-backtest-cost-execution
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #368
+- next_ticket: #368
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: TASK-V2-012 구현 갭 보완을 위해 cost guard + execution-adjusted fold metric + 회귀 테스트를 함께 반영한다.
+
+### 2026-03-02 | session=codex-v3-stream-next-ticket-369
+- branch: feature/v3-session-policy-stream
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #369, #370, #371, #374, #375, #376, #377, #381
+- next_ticket: #369
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: 구현 티켓은 코드/테스트/문서(요구사항 원장/구현감사/PR traceability) 동시 반영을 기본 원칙으로 진행한다.
+
+### 2026-03-02 | session=codex-issue369-start
+- branch: feature/issue-369-model-exit-signal-spec-sync
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #369
+- next_ticket: #369
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: v2 사양 기준으로 model_exit_signal을 직접 청산 트리거가 아닌 보조 트리거로 정합화하고 테스트/문서를 동기화한다.
+
+### 2026-03-02 | session=codex-v3-stream-next-ticket-377
+- branch: feature/v3-session-policy-stream
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #377, #370, #371, #375, #376, #381
+- next_ticket: #377
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: kill switch refresh 재시도 정책(횟수/간격/중단조건)을 코드/테스트/요구사항 원장/감사 문서에 동시 반영한다.
+
+### 2026-03-02 | session=codex-issue377-start
+- branch: feature/issue-377-kill-switch-refresh-retry
+- docs_checked: docs/workflow.md, docs/commands.md, docs/agent-constraints.md
+- open_issues_reviewed: #377
+- next_ticket: #377
+- process_gate_checked: process_ticket=#306,#308 merged_to_feature_branch=yes
+- risks_or_notes: refresh 단계를 최대 3회(초기+재시도2), 실패 시 지수 백오프로 재시도하고 성공 시 즉시 중단, 소진 시 오류를 기록한 뒤 다음 단계를 계속 수행한다.