governance: enforce runtime verification coverage gates (#301)

.gitea/ISSUE_TEMPLATE/runtime_verification.md

@@ -0,0 +1,41 @@
+---
+name: Runtime Verification Incident
+about: 실운영/스테이징 동작 검증 중 발견된 이상 징후 등록
+title: "[RUNTIME-VERIFY][SCN-XXX] "
+labels: runtime, verification
+---
+
+## Summary
+
+- 현상:
+- 최초 관측 시각(UTC):
+
+## Reproduction / Observation
+
+- 실행 모드(`live`/`paper`):
+- 세션(`NXT`, `US_PRE`, `US_DAY`, `US_AFTER`, ...):
+- 실행 커맨드:
+- 로그 경로:
+
+## Expected vs Actual
+
+- Expected:
+- Actual:
+
+## Requirement Mapping
+
+- REQ:
+- TASK:
+- TEST:
+
+## Temporary Mitigation
+
+- 즉시 완화책:
+
+## Close Criteria
+
+- [ ] Dev 수정 반영
+- [ ] Verifier 재검증 PASS
+- [ ] Runtime Verifier 재관측 PASS
+- [ ] `NOT_OBSERVED = 0`
+

.gitea/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,38 @@
+## Linked Issue
+
+- Closes #N
+
+## Scope
+
+- REQ: `REQ-...`
+- TASK: `TASK-...`
+- TEST: `TEST-...`
+
+## Main -> Verifier Directive Contract
+
+- Scope: 대상 요구사항/코드/로그 경로
+- Method: 실행 커맨드 + 관측 포인트
+- PASS criteria:
+- FAIL criteria:
+- NOT_OBSERVED criteria:
+- Evidence format: PR 코멘트 `Coverage Matrix`
+
+## Verifier Coverage Matrix (Required)
+
+| Item | Evidence | Status (PASS/FAIL/NOT_OBSERVED) |
+|---|---|---|
+| REQ-... | 링크/로그 | PASS |
+
+`NOT_OBSERVED`가 1개라도 있으면 승인/머지 금지.
+
+## Gitea Preflight
+
+- [ ] `docs/commands.md`와 `docs/workflow.md` 트러블슈팅 선확인
+- [ ] `tea` 사용 (`gh` 미사용)
+
+## Runtime Evidence
+
+- 시스템 실제 구동 커맨드:
+- 모니터링 로그 경로:
+- 이상 징후/이슈 링크:
+

.github/workflows/ci.yml

@@ -21,6 +21,12 @@ jobs:
       - name: Install dependencies
         run: pip install ".[dev]"
 
+      - name: Validate governance assets
+        run: python3 scripts/validate_governance_assets.py
+
+      - name: Validate Ouroboros docs
+        run: python3 scripts/validate_ouroboros_docs.py
+
       - name: Lint
         run: ruff check src/ tests/
 

docs/agent-constraints.md

@@ -12,6 +12,8 @@ It is distinct from `docs/requirements-log.md`, which records **project/product
 
 1. **Workflow enforcement**
    - Follow `docs/workflow.md` for all changes.
+   - Before any Gitea issue/PR/comment operation, read `docs/commands.md` and `docs/workflow.md` troubleshooting section.
+   - Use `tea` for Gitea operations; do not use GitHub CLI (`gh`) in this repository workflow.
    - Create a Gitea issue before any code or documentation change.
    - Work on a feature branch `feature/issue-{N}-{short-description}` and open a PR.
    - Never commit directly to `main`.
@@ -43,3 +45,8 @@ It is distinct from `docs/requirements-log.md`, which records **project/product
 - When work requires guidance, consult the relevant `docs/` policies first.
 - Any code change must be accompanied by relevant documentation updates.
 - Persist user constraints across sessions by recording them in this document.
+
+### 2026-02-27
+
+- All agents must pre-read `docs/commands.md` and `docs/workflow.md` troubleshooting before running Gitea issue/PR/comment commands.
+- `gh` CLI is prohibited for repository ticket/PR operations; use `tea` (or documented Gitea API fallback only).

docs/commands.md

@@ -4,6 +4,13 @@
 
 **Critical: Learn from failures. Never repeat the same failed command without modification.**
 
+## Repository VCS Rule (Mandatory)
+
+- 이 저장소의 티켓/PR/코멘트 작업은 Gitea 기준으로 수행한다.
+- `gh`(GitHub CLI) 명령 사용은 금지한다.
+- 기본 도구는 `tea`이며, `tea` 미지원 케이스만 Gitea API를 fallback으로 사용한다.
+- 실행 전 `docs/workflow.md`의 `Gitea CLI Formatting Troubleshooting`을 반드시 확인한다.
+
 ### tea CLI (Gitea Command Line Tool)
 
 #### ❌ TTY Error - Interactive Confirmation Fails
@@ -140,6 +147,12 @@ python -m src.main --mode=paper
 # Run with dashboard enabled
 python -m src.main --mode=paper --dashboard
 
+# Runtime verification monitor (NOT_OBSERVED detection)
+bash scripts/runtime_verify_monitor.sh
+
+# Follow runtime verification log
+tail -f data/overnight/runtime_verify_*.log
+
 # Docker
 docker compose up -d ouroboros          # Run agent
 docker compose --profile test up test   # Run tests in container

docs/ouroboros/50_tpm_control_protocol.md

@@ -34,6 +34,12 @@ Main Agent 아이디에이션 책임:
 - DCP-03 구현 착수: Phase 2 종료 전 Main Agent 승인 필수
 - DCP-04 배포 승인: Phase 4 종료 후 Main Agent 최종 승인 필수
 
+Main/Verifier 사고 재발 방지 규칙:
+- Main Agent는 검증 위임 시 `Directive Contract`를 충족하지 않으면 검증 착수 금지
+- Verifier Agent는 지시 누락/모호성 발견 시 즉시 `BLOCKED`를 선언하고 보완 요청
+- Verifier Agent는 `미관측(NOT_OBSERVED)` 항목을 PASS로 보고할 수 없다
+- Runtime 검증에서 요구 세션 증적이 없으면 "정상"이 아니라 `미검증 이상`으로 이슈화한다
+
 ## Phase Control Gates
 
 ### Phase 0: Scenario Intake and Scope Lock
@@ -112,6 +118,8 @@ Exit criteria:
 
 Control checks:
 - Verifier가 테스트 증적(로그/리포트/실행 커맨드) 첨부
+- Verifier가 `Coverage Matrix`(`REQ/TASK/TEST` x `PASS/FAIL/NOT_OBSERVED`) 첨부
+- `NOT_OBSERVED` 항목 수가 0인지 확인(0이 아니면 Gate 실패)
 - Runtime Verifier가 스테이징/실운영 모니터링 계획 승인
 - 산출물: 수용 승인 레코드
 
@@ -149,6 +157,9 @@ TPM 티켓 운영 규칙:
 - TPM은 합의된 변경을 이슈로 등록하고 우선순위(`P0/P1/P2`)를 지정한다.
 - PR 본문에는 TPM이 지정한 우선순위와 범위가 그대로 반영되어야 한다.
 - 우선순위 변경은 TPM 제안 + Main Agent 승인으로만 가능하다.
+- PM/TPM/Dev/Reviewer/Verifier/Runtime Verifier는 주요 의사결정 시점마다 PR 코멘트를 남겨 결정 근거를 추적 가능 상태로 유지한다.
+- PM/TPM/Dev/Reviewer/Verifier/Runtime Verifier는 이슈/PR/코멘트 조작 전에 `docs/commands.md`와 `docs/workflow.md`의 Gitea 트러블슈팅 섹션을 선참조해야 한다.
+- 저장소 협업에서 GitHub CLI(`gh`) 사용은 금지하며, Gitea 작업은 `tea`(필요 시 문서화된 API fallback)만 허용한다.
 
 브랜치 운영 규칙:
 - TPM은 각 티켓에 대해 `ticket temp branch -> program feature branch` PR 경로를 지정한다.
@@ -167,6 +178,8 @@ TPM 티켓 운영 규칙:
   - 시스템 실제 구동(스테이징/로컬 실운영 모드) 실행
   - 모니터링 체크리스트(핵심 경보/주문 경로/예외 로그) 수행
   - 결과를 티켓/PR 코멘트에 증적으로 첨부하지 않으면 완료로 간주하지 않음
+  - 세션별 필수 관측 포인트(`NXT`, `US_PRE`, `US_DAY`, `US_AFTER` 등) 중 미관측 항목은 `NOT_OBSERVED`로 기록
+  - `NOT_OBSERVED` 존재 시 승인 금지 + Runtime 이슈 발행
 
 ## Server Reflection Rule
 

docs/ouroboros/60_repo_enforcement_checklist.md

@@ -48,14 +48,20 @@ Updated: 2026-02-26
 병합 전 체크리스트:
 - 이슈 연결(`Closes #N`) 존재
 - PR 본문에 `REQ-*`, `TASK-*`, `TEST-*` 매핑 표 존재
+- Main -> Verifier Directive Contract(범위/방법/합격/실패/미관측/증적 형식) 기재
 - `src/core/risk_manager.py` 변경 없음
 - 주요 의사결정 체크포인트(DCP-01~04) 중 해당 단계 Main Agent 확인 기록 존재
+- 주요 의사결정(리뷰 지적/수정 합의/검증 승인)에 대한 에이전트 PR 코멘트 존재
 - 티켓 PR의 base가 `main`이 아닌 program feature branch인지 확인
 
 자동 점검:
 - 문서 검증 스크립트 통과
 - 테스트 통과
 - 개발 완료 시 시스템 구동/모니터링 증적 코멘트 존재
+- 이슈/PR 조작 전에 `docs/commands.md` 및 `docs/workflow.md` 트러블슈팅 확인 코멘트 존재
+- `gh` CLI 미사용, `tea` 사용 증적 존재
+- Verifier `Coverage Matrix` 첨부(PASS/FAIL/NOT_OBSERVED)
+- `NOT_OBSERVED` 항목 0 확인(0이 아니면 머지 금지)
 
 ## 5) 감사 추적
 

docs/workflow.md

@@ -16,12 +16,27 @@
 
 **Never commit directly to `main`.** This policy applies to all changes, no exceptions.
 
+## Agent Gitea Preflight (Mandatory)
+
+Gitea 이슈/PR/코멘트 작업 전에 모든 에이전트는 아래를 먼저 확인해야 한다.
+
+1. `docs/commands.md`의 `tea CLI` 실패 사례/해결 패턴 확인
+2. 본 문서의 `Gitea CLI Formatting Troubleshooting` 확인
+3. 명령 실행 전 `gh`(GitHub CLI) 사용 금지 확인
+
+강제 규칙:
+- 이 저장소 협업 명령은 `tea`를 기본으로 사용한다.
+- `gh issue`, `gh pr` 등 GitHub CLI 명령은 사용 금지다.
+- `tea` 실패 시 동일 명령 재시도 전에 원인/수정사항을 PR 코멘트에 남긴다.
+- 필요한 경우에만 Gitea API(`localhost:3000`)를 fallback으로 사용한다.
+
 ## Branch Strategy (Mandatory)
 
 - Team operation default branch is the **program feature branch**, not `main`.
 - Ticket-level development happens only on **ticket temp branches** cut from the program feature branch.
 - Ticket PR merges into program feature branch are allowed after verifier approval.
 - Until final user sign-off, `main` merge is prohibited.
+- 각 에이전트는 주요 의사결정(리뷰 지적, 수정 방향, 검증 승인)마다 PR 코멘트를 적극 작성해 의사결정 과정을 남긴다.
 
 ## Gitea CLI Formatting Troubleshooting
 
@@ -136,6 +151,22 @@ task_tool(
 
 Use `run_in_background=True` for independent tasks that don't block subsequent work.
 
+### Main -> Verifier Directive Contract (Mandatory)
+
+메인 에이전트가 검증 에이전트에 작업을 위임할 때, 아래 6개를 누락하면 지시가 무효다.
+
+1. 검증 대상 범위: `REQ-*`, `TASK-*`, 코드/로그 경로
+2. 검증 방법: 실행 커맨드와 관측 포인트(예: 세션별 로그 키워드)
+3. 합격 기준: PASS 조건을 수치/문구로 명시
+4. 실패 기준: FAIL 조건을 수치/문구로 명시
+5. 미관측 기준: `NOT_OBSERVED` 조건과 즉시 에스컬레이션 규칙
+6. 증적 형식: PR 코멘트에 `Coverage Matrix` 표로 제출
+
+`NOT_OBSERVED` 처리 규칙:
+- 요구사항 항목이 관측되지 않았으면 PASS로 간주 금지
+- `NOT_OBSERVED`는 운영상 `FAIL`과 동일하게 처리
+- `NOT_OBSERVED`가 하나라도 있으면 승인/머지 금지
+
 ## Code Review Checklist
 
 **CRITICAL: Every PR review MUST verify plan-implementation consistency.**
@@ -169,3 +200,7 @@ Before approving any PR, the reviewer (human or agent) must check ALL of the fol
 - [ ] PR references the Gitea issue number
 - [ ] Feature branch follows naming convention (`feature/issue-N-description`)
 - [ ] Commit messages are clear and descriptive
+- [ ] 이슈/PR 작업 전에 `docs/commands.md`와 본 문서 트러블슈팅 섹션을 확인했다
+- [ ] `gh` 명령을 사용하지 않고 `tea`(또는 허용된 Gitea API fallback)만 사용했다
+- [ ] Main -> Verifier 지시가 Directive Contract 6개 항목을 모두 포함한다
+- [ ] Verifier 결과에 `Coverage Matrix`(PASS/FAIL/NOT_OBSERVED)가 있고, `NOT_OBSERVED=0`이다

scripts/runtime_verify_monitor.sh

@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+# Runtime verification monitor with NOT_OBSERVED detection.
+
+set -euo pipefail
+
+ROOT_DIR="${ROOT_DIR:-/home/agentson/repos/The-Ouroboros}"
+LOG_DIR="${LOG_DIR:-$ROOT_DIR/data/overnight}"
+INTERVAL_SEC="${INTERVAL_SEC:-60}"
+MAX_HOURS="${MAX_HOURS:-24}"
+
+cd "$ROOT_DIR"
+
+OUT_LOG="$LOG_DIR/runtime_verify_$(date +%Y%m%d_%H%M%S).log"
+END_TS=$(( $(date +%s) + MAX_HOURS*3600 ))
+
+log() {
+  printf '%s %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$1" | tee -a "$OUT_LOG" >/dev/null
+}
+
+check_signal() {
+  local name="$1"
+  local pattern="$2"
+  local run_log="$3"
+
+  if rg -q "$pattern" "$run_log"; then
+    log "[COVERAGE] ${name}=PASS pattern=${pattern}"
+    return 0
+  fi
+  log "[COVERAGE] ${name}=NOT_OBSERVED pattern=${pattern}"
+  return 1
+}
+
+log "[INFO] runtime verify monitor started interval=${INTERVAL_SEC}s max_hours=${MAX_HOURS}"
+
+while true; do
+  now=$(date +%s)
+  if [ "$now" -ge "$END_TS" ]; then
+    log "[INFO] monitor completed (time window reached)"
+    exit 0
+  fi
+
+  latest_run="$(ls -t "$LOG_DIR"/run_*.log 2>/dev/null | head -n1 || true)"
+  if [ -z "$latest_run" ]; then
+    log "[ANOMALY] no run log found"
+    sleep "$INTERVAL_SEC"
+    continue
+  fi
+
+  # Basic liveness hints.
+  app_pid="$(cat "$LOG_DIR/app.pid" 2>/dev/null || true)"
+  wd_pid="$(cat "$LOG_DIR/watchdog.pid" 2>/dev/null || true)"
+  app_alive=0
+  wd_alive=0
+  port_alive=0
+  [ -n "$app_pid" ] && kill -0 "$app_pid" 2>/dev/null && app_alive=1
+  [ -n "$wd_pid" ] && kill -0 "$wd_pid" 2>/dev/null && wd_alive=1
+  ss -ltnp 2>/dev/null | rg -q ':8080' && port_alive=1
+  log "[HEARTBEAT] run_log=$latest_run app_alive=$app_alive watchdog_alive=$wd_alive port8080=$port_alive"
+
+  # Coverage matrix rows (session paths and policy gate evidence).
+  not_observed=0
+  check_signal "LIVE_MODE" "Mode: live" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "KR_LOOP" "Processing market: Korea Exchange" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "NXT_PATH" "NXT_PRE|NXT_AFTER|session=NXT_" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "US_PRE_PATH" "US_PRE|session=US_PRE" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "US_DAY_PATH" "US_DAY|session=US_DAY|Processing market: .*NASDAQ|Processing market: .*NYSE|Processing market: .*AMEX" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "US_AFTER_PATH" "US_AFTER|session=US_AFTER" "$latest_run" || not_observed=$((not_observed+1))
+  check_signal "ORDER_POLICY_SESSION" "Order policy rejected .*\\[session=" "$latest_run" || not_observed=$((not_observed+1))
+
+  if [ "$not_observed" -gt 0 ]; then
+    log "[ANOMALY] coverage_not_observed=$not_observed (treat as FAIL)"
+  else
+    log "[OK] coverage complete (NOT_OBSERVED=0)"
+  fi
+
+  sleep "$INTERVAL_SEC"
+done
+

scripts/validate_governance_assets.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""Validate persistent governance assets for agent workflow safety."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+
+def must_contain(path: Path, required: list[str], errors: list[str]) -> None:
+    if not path.exists():
+        errors.append(f"missing file: {path}")
+        return
+    text = path.read_text(encoding="utf-8")
+    for token in required:
+        if token not in text:
+            errors.append(f"{path}: missing required token -> {token}")
+
+
+def main() -> int:
+    errors: list[str] = []
+
+    pr_template = Path(".gitea/PULL_REQUEST_TEMPLATE.md")
+    issue_template = Path(".gitea/ISSUE_TEMPLATE/runtime_verification.md")
+
+    must_contain(
+        pr_template,
+        [
+            "Closes #N",
+            "Main -> Verifier Directive Contract",
+            "Coverage Matrix",
+            "NOT_OBSERVED",
+            "tea",
+            "gh",
+        ],
+        errors,
+    )
+    must_contain(
+        issue_template,
+        [
+            "[RUNTIME-VERIFY][SCN-XXX]",
+            "Requirement Mapping",
+            "Close Criteria",
+            "NOT_OBSERVED = 0",
+        ],
+        errors,
+    )
+
+    if errors:
+        print("[FAIL] governance asset validation failed")
+        for err in errors:
+            print(f"- {err}")
+        return 1
+
+    print("[OK] governance assets validated")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
+

src/analysis/backtest_cost_guard.py

@@ -0,0 +1,52 @@
+"""Backtest cost/slippage/failure validation guard."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import math
+
+
+@dataclass(frozen=True)
+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
+    unfavorable_fill_required: bool = True
+
+
+def validate_backtest_cost_model(
+    *,
+    model: BacktestCostModel,
+    required_sessions: list[str],
+) -> None:
+    """Raise ValueError when required cost assumptions are missing/invalid."""
+    if (
+        model.commission_bps is None
+        or not math.isfinite(model.commission_bps)
+        or model.commission_bps < 0
+    ):
+        raise ValueError("commission_bps must be provided and >= 0")
+    if not model.unfavorable_fill_required:
+        raise ValueError("unfavorable_fill_required must be True")
+
+    slippage = model.slippage_bps_by_session or {}
+    failure = model.failure_rate_by_session or {}
+
+    missing_slippage = [s for s in required_sessions if s not in slippage]
+    if missing_slippage:
+        raise ValueError(
+            f"missing slippage_bps_by_session for sessions: {', '.join(missing_slippage)}"
+        )
+
+    missing_failure = [s for s in required_sessions if s not in failure]
+    if missing_failure:
+        raise ValueError(
+            f"missing failure_rate_by_session for sessions: {', '.join(missing_failure)}"
+        )
+
+    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}")

src/analysis/backtest_execution_model.py

@@ -0,0 +1,103 @@
+"""Conservative backtest execution model."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import math
+from random import Random
+from typing import Literal
+
+
+OrderSide = Literal["BUY", "SELL"]
+
+
+@dataclass(frozen=True)
+class ExecutionRequest:
+    side: OrderSide
+    session_id: str
+    qty: int
+    reference_price: float
+
+
+@dataclass(frozen=True)
+class ExecutionAssumptions:
+    slippage_bps_by_session: dict[str, float]
+    failure_rate_by_session: dict[str, float]
+    partial_fill_rate_by_session: dict[str, float]
+    partial_fill_min_ratio: float = 0.3
+    partial_fill_max_ratio: float = 0.8
+    seed: int = 0
+
+
+@dataclass(frozen=True)
+class ExecutionResult:
+    status: Literal["FILLED", "PARTIAL", "REJECTED"]
+    filled_qty: int
+    avg_price: float
+    slippage_bps: float
+    reason: str
+
+
+class BacktestExecutionModel:
+    """Execution simulator with conservative unfavorable fill assumptions."""
+
+    def __init__(self, assumptions: ExecutionAssumptions) -> None:
+        self.assumptions = assumptions
+        self._rng = Random(assumptions.seed)
+        if assumptions.partial_fill_min_ratio <= 0 or assumptions.partial_fill_max_ratio > 1:
+            raise ValueError("partial fill ratios must be within (0,1]")
+        if assumptions.partial_fill_min_ratio > assumptions.partial_fill_max_ratio:
+            raise ValueError("partial_fill_min_ratio must be <= partial_fill_max_ratio")
+        for sess, bps in assumptions.slippage_bps_by_session.items():
+            if not math.isfinite(bps) or bps < 0:
+                raise ValueError(f"slippage_bps must be finite and >= 0 for session={sess}")
+        for sess, rate in assumptions.failure_rate_by_session.items():
+            if not math.isfinite(rate) or rate < 0 or rate > 1:
+                raise ValueError(f"failure_rate must be in [0,1] for session={sess}")
+        for sess, rate in assumptions.partial_fill_rate_by_session.items():
+            if not math.isfinite(rate) or rate < 0 or rate > 1:
+                raise ValueError(f"partial_fill_rate must be in [0,1] for session={sess}")
+
+    def simulate(self, request: ExecutionRequest) -> ExecutionResult:
+        if request.qty <= 0:
+            raise ValueError("qty must be positive")
+        if request.reference_price <= 0:
+            raise ValueError("reference_price must be positive")
+
+        slippage_bps = self.assumptions.slippage_bps_by_session.get(request.session_id, 0.0)
+        failure_rate = self.assumptions.failure_rate_by_session.get(request.session_id, 0.0)
+        partial_rate = self.assumptions.partial_fill_rate_by_session.get(request.session_id, 0.0)
+
+        if self._rng.random() < failure_rate:
+            return ExecutionResult(
+                status="REJECTED",
+                filled_qty=0,
+                avg_price=0.0,
+                slippage_bps=slippage_bps,
+                reason="execution_failure",
+            )
+
+        slip_mult = 1.0 + (slippage_bps / 10000.0 if request.side == "BUY" else -slippage_bps / 10000.0)
+        exec_price = request.reference_price * slip_mult
+
+        if self._rng.random() < partial_rate:
+            ratio = self._rng.uniform(
+                self.assumptions.partial_fill_min_ratio,
+                self.assumptions.partial_fill_max_ratio,
+            )
+            filled = max(1, min(request.qty - 1, int(request.qty * ratio)))
+            return ExecutionResult(
+                status="PARTIAL",
+                filled_qty=filled,
+                avg_price=exec_price,
+                slippage_bps=slippage_bps,
+                reason="partial_fill",
+            )
+
+        return ExecutionResult(
+            status="FILLED",
+            filled_qty=request.qty,
+            avg_price=exec_price,
+            slippage_bps=slippage_bps,
+            reason="filled",
+        )

src/analysis/walk_forward_split.py

@@ -0,0 +1,74 @@
+"""Walk-forward splitter with purge/embargo controls."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class WalkForwardFold:
+    train_indices: list[int]
+    test_indices: list[int]
+
+    @property
+    def train_size(self) -> int:
+        return len(self.train_indices)
+
+    @property
+    def test_size(self) -> int:
+        return len(self.test_indices)
+
+
+def generate_walk_forward_splits(
+    *,
+    n_samples: int,
+    train_size: int,
+    test_size: int,
+    step_size: int | None = None,
+    purge_size: int = 0,
+    embargo_size: int = 0,
+    min_train_size: int = 1,
+) -> list[WalkForwardFold]:
+    """Generate chronological folds with purge/embargo leakage controls."""
+    if n_samples <= 0:
+        raise ValueError("n_samples must be positive")
+    if train_size <= 0 or test_size <= 0:
+        raise ValueError("train_size and test_size must be positive")
+    if purge_size < 0 or embargo_size < 0:
+        raise ValueError("purge_size and embargo_size must be >= 0")
+    if min_train_size <= 0:
+        raise ValueError("min_train_size must be positive")
+
+    step = step_size if step_size is not None else test_size
+    if step <= 0:
+        raise ValueError("step_size must be positive")
+
+    folds: list[WalkForwardFold] = []
+    prev_test_end: int | None = None
+    test_start = train_size + purge_size
+
+    while test_start + test_size <= n_samples:
+        test_end = test_start + test_size - 1
+        train_end = test_start - purge_size - 1
+        if train_end < 0:
+            break
+
+        train_start = max(0, train_end - train_size + 1)
+        train_indices = list(range(train_start, train_end + 1))
+
+        if prev_test_end is not None and embargo_size > 0:
+            emb_from = prev_test_end + 1
+            emb_to = prev_test_end + embargo_size
+            train_indices = [i for i in train_indices if i < emb_from or i > emb_to]
+
+        if len(train_indices) >= min_train_size:
+            folds.append(
+                WalkForwardFold(
+                    train_indices=train_indices,
+                    test_indices=list(range(test_start, test_end + 1)),
+                )
+            )
+            prev_test_end = test_end
+        test_start += step
+
+    return folds

src/config.py

@@ -59,6 +59,8 @@ class Settings(BaseSettings):
     # KIS VTS overseas balance API returns errors for most accounts.
     # This value is used as a fallback when the balance API returns 0 in paper mode.
     PAPER_OVERSEAS_CASH: float = Field(default=50000.0, ge=0.0)
+    USD_BUFFER_MIN: float = Field(default=1000.0, ge=0.0)
+    OVERNIGHT_EXCEPTION_ENABLED: bool = True
 
     # Trading frequency mode (daily = batch API calls, realtime = per-stock calls)
     TRADE_MODE: str = Field(default="daily", pattern="^(daily|realtime)$")

src/db.py

@@ -31,8 +31,12 @@ def init_db(db_path: str) -> sqlite3.Connection:
             quantity INTEGER,
             price REAL,
             pnl REAL DEFAULT 0.0,
+            strategy_pnl REAL DEFAULT 0.0,
+            fx_pnl REAL DEFAULT 0.0,
             market TEXT DEFAULT 'KR',
             exchange_code TEXT DEFAULT 'KRX',
+            session_id TEXT DEFAULT 'UNKNOWN',
+            selection_context TEXT,
             decision_id TEXT,
             mode TEXT DEFAULT 'paper'
         )
@@ -53,6 +57,32 @@ def init_db(db_path: str) -> sqlite3.Connection:
         conn.execute("ALTER TABLE trades ADD COLUMN decision_id TEXT")
     if "mode" not in columns:
         conn.execute("ALTER TABLE trades ADD COLUMN mode TEXT DEFAULT 'paper'")
+    session_id_added = False
+    if "session_id" not in columns:
+        conn.execute("ALTER TABLE trades ADD COLUMN session_id TEXT DEFAULT 'UNKNOWN'")
+        session_id_added = True
+    if "strategy_pnl" not in columns:
+        conn.execute("ALTER TABLE trades ADD COLUMN strategy_pnl REAL DEFAULT 0.0")
+    if "fx_pnl" not in columns:
+        conn.execute("ALTER TABLE trades ADD COLUMN fx_pnl REAL DEFAULT 0.0")
+    # Backfill legacy rows where only pnl existed before split accounting columns.
+    conn.execute(
+        """
+        UPDATE trades
+        SET strategy_pnl = pnl, fx_pnl = 0.0
+        WHERE pnl != 0.0
+          AND strategy_pnl = 0.0
+          AND fx_pnl = 0.0
+        """
+    )
+    if session_id_added:
+        conn.execute(
+            """
+            UPDATE trades
+            SET session_id = 'UNKNOWN'
+            WHERE session_id IS NULL OR session_id = ''
+            """
+        )
 
     # Context tree tables for multi-layered memory management
     conn.execute(
@@ -171,8 +201,11 @@ def log_trade(
     quantity: int = 0,
     price: float = 0.0,
     pnl: float = 0.0,
+    strategy_pnl: float | None = None,
+    fx_pnl: float | None = None,
     market: str = "KR",
     exchange_code: str = "KRX",
+    session_id: str | None = None,
     selection_context: dict[str, any] | None = None,
     decision_id: str | None = None,
     mode: str = "paper",
@@ -187,24 +220,37 @@ def log_trade(
         rationale: AI decision rationale
         quantity: Number of shares
         price: Trade price
-        pnl: Profit/loss
+        pnl: Total profit/loss (backward compatibility)
+        strategy_pnl: Strategy PnL component
+        fx_pnl: FX PnL component
         market: Market code
         exchange_code: Exchange code
+        session_id: Session identifier (if omitted, auto-derived from market)
         selection_context: Scanner selection data (RSI, volume_ratio, signal, score)
         decision_id: Unique decision identifier for audit linking
         mode: Trading mode ('paper' or 'live') for data separation
     """
     # Serialize selection context to JSON
     context_json = json.dumps(selection_context) if selection_context else None
+    resolved_session_id = _resolve_session_id(market=market, session_id=session_id)
+    if strategy_pnl is None and fx_pnl is None:
+        strategy_pnl = pnl
+        fx_pnl = 0.0
+    elif strategy_pnl is None:
+        strategy_pnl = pnl - float(fx_pnl or 0.0) if pnl != 0.0 else 0.0
+    elif fx_pnl is None:
+        fx_pnl = pnl - float(strategy_pnl) if pnl != 0.0 else 0.0
+    if pnl == 0.0 and (strategy_pnl or fx_pnl):
+        pnl = float(strategy_pnl) + float(fx_pnl)
 
     conn.execute(
         """
         INSERT INTO trades (
             timestamp, stock_code, action, confidence, rationale,
-            quantity, price, pnl, market, exchange_code, selection_context, decision_id,
-            mode
+            quantity, price, pnl, strategy_pnl, fx_pnl,
+            market, exchange_code, session_id, selection_context, decision_id, mode
         )
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
         """,
         (
             datetime.now(UTC).isoformat(),
@@ -215,8 +261,11 @@ def log_trade(
             quantity,
             price,
             pnl,
+            strategy_pnl,
+            fx_pnl,
             market,
             exchange_code,
+            resolved_session_id,
             context_json,
             decision_id,
             mode,
@@ -225,6 +274,21 @@ def log_trade(
     conn.commit()
 
 
+def _resolve_session_id(*, market: str, session_id: str | None) -> str:
+    if session_id:
+        return session_id
+    try:
+        from src.core.order_policy import classify_session_id
+        from src.markets.schedule import MARKETS
+
+        market_info = MARKETS.get(market)
+        if market_info is not None:
+            return classify_session_id(market_info)
+    except Exception:
+        pass
+    return "UNKNOWN"
+
+
 def get_latest_buy_trade(
     conn: sqlite3.Connection, stock_code: str, market: str
 ) -> dict[str, Any] | None:

src/main.py

@@ -33,7 +33,11 @@ from src.core.blackout_manager import (
     parse_blackout_windows_kst,
 )
 from src.core.kill_switch import KillSwitchOrchestrator
-from src.core.order_policy import OrderPolicyRejected, validate_order_policy
+from src.core.order_policy import (
+    OrderPolicyRejected,
+    get_session_info,
+    validate_order_policy,
+)
 from src.core.priority_queue import PriorityTaskQueue
 from src.core.risk_manager import CircuitBreakerTripped, FatFingerRejected, RiskManager
 from src.db import (
@@ -63,6 +67,7 @@ BLACKOUT_ORDER_MANAGER = BlackoutOrderManager(
     windows=[],
     max_queue_size=500,
 )
+_SESSION_CLOSE_WINDOWS = {"NXT_AFTER", "US_AFTER"}
 
 
 def safe_float(value: str | float | None, default: float = 0.0) -> float:
@@ -429,6 +434,41 @@ def _determine_order_quantity(
     return quantity
 
 
+def _should_block_overseas_buy_for_fx_buffer(
+    *,
+    market: MarketInfo,
+    action: str,
+    total_cash: float,
+    order_amount: float,
+    settings: Settings | None,
+) -> tuple[bool, float, float]:
+    if (
+        market.is_domestic
+        or not market.code.startswith("US")
+        or action != "BUY"
+        or settings is None
+    ):
+        return False, total_cash - order_amount, 0.0
+    remaining = total_cash - order_amount
+    required = settings.USD_BUFFER_MIN
+    return remaining < required, remaining, required
+
+
+def _should_force_exit_for_overnight(
+    *,
+    market: MarketInfo,
+    settings: Settings | None,
+) -> bool:
+    session_id = get_session_info(market).session_id
+    if session_id not in _SESSION_CLOSE_WINDOWS:
+        return False
+    if KILL_SWITCH.new_orders_blocked:
+        return True
+    if settings is None:
+        return False
+    return not settings.OVERNIGHT_EXCEPTION_ENABLED
+
+
 async def build_overseas_symbol_universe(
     db_conn: Any,
     overseas_broker: OverseasBroker,
@@ -1194,6 +1234,23 @@ async def trading_cycle(
                         loss_pct,
                         take_profit_threshold,
                     )
+            if decision.action == "HOLD" and _should_force_exit_for_overnight(
+                market=market,
+                settings=settings,
+            ):
+                decision = TradeDecision(
+                    action="SELL",
+                    confidence=max(decision.confidence, 85),
+                    rationale=(
+                        "Forced exit by overnight policy"
+                        " (session close window / kill switch priority)"
+                    ),
+                )
+                logger.info(
+                    "Overnight policy override for %s (%s): HOLD -> SELL",
+                    stock_code,
+                    market.name,
+                )
     logger.info(
         "Decision for %s (%s): %s (confidence=%d)",
         stock_code,
@@ -1254,7 +1311,7 @@ async def trading_cycle(
     trade_price = current_price
     trade_pnl = 0.0
     if decision.action in ("BUY", "SELL"):
-        if KILL_SWITCH.new_orders_blocked:
+        if KILL_SWITCH.new_orders_blocked and decision.action == "BUY":
             logger.critical(
                 "KillSwitch block active: skip %s order for %s (%s)",
                 decision.action,
@@ -1292,6 +1349,24 @@ async def trading_cycle(
             )
             return
         order_amount = current_price * quantity
+        fx_blocked, remaining_cash, required_buffer = _should_block_overseas_buy_for_fx_buffer(
+            market=market,
+            action=decision.action,
+            total_cash=total_cash,
+            order_amount=order_amount,
+            settings=settings,
+        )
+        if fx_blocked:
+            logger.warning(
+                "Skip BUY %s (%s): FX buffer guard (remaining=%.2f, required=%.2f, cash=%.2f, order=%.2f)",
+                stock_code,
+                market.name,
+                remaining_cash,
+                required_buffer,
+                total_cash,
+                order_amount,
+            )
+            return
 
         # 4. Check BUY cooldown (set when a prior BUY failed due to insufficient balance)
         if decision.action == "BUY" and buy_cooldown is not None:
@@ -2285,6 +2360,25 @@ async def run_daily_session(
                         stock_code,
                         market.name,
                     )
+            if decision.action == "HOLD":
+                daily_open = get_open_position(db_conn, stock_code, market.code)
+                if daily_open and _should_force_exit_for_overnight(
+                    market=market,
+                    settings=settings,
+                ):
+                    decision = TradeDecision(
+                        action="SELL",
+                        confidence=max(decision.confidence, 85),
+                        rationale=(
+                            "Forced exit by overnight policy"
+                            " (session close window / kill switch priority)"
+                        ),
+                    )
+                    logger.info(
+                        "Daily overnight policy override for %s (%s): HOLD -> SELL",
+                        stock_code,
+                        market.name,
+                    )
 
             # Log decision
             context_snapshot = {
@@ -2325,7 +2419,7 @@ async def run_daily_session(
             trade_pnl = 0.0
             order_succeeded = True
             if decision.action in ("BUY", "SELL"):
-                if KILL_SWITCH.new_orders_blocked:
+                if KILL_SWITCH.new_orders_blocked and decision.action == "BUY":
                     logger.critical(
                         "KillSwitch block active: skip %s order for %s (%s)",
                         decision.action,
@@ -2360,6 +2454,24 @@ async def run_daily_session(
                     )
                     continue
                 order_amount = stock_data["current_price"] * quantity
+                fx_blocked, remaining_cash, required_buffer = _should_block_overseas_buy_for_fx_buffer(
+                    market=market,
+                    action=decision.action,
+                    total_cash=total_cash,
+                    order_amount=order_amount,
+                    settings=settings,
+                )
+                if fx_blocked:
+                    logger.warning(
+                        "Skip BUY %s (%s): FX buffer guard (remaining=%.2f, required=%.2f, cash=%.2f, order=%.2f)",
+                        stock_code,
+                        market.name,
+                        remaining_cash,
+                        required_buffer,
+                        total_cash,
+                        order_amount,
+                    )
+                    continue
 
                 # Check BUY cooldown (insufficient balance)
                 if decision.action == "BUY":

tests/test_backtest_cost_guard.py

@@ -0,0 +1,83 @@
+from __future__ import annotations
+
+import pytest
+
+from src.analysis.backtest_cost_guard import BacktestCostModel, validate_backtest_cost_model
+
+
+def test_valid_backtest_cost_model_passes() -> 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},
+        unfavorable_fill_required=True,
+    )
+    validate_backtest_cost_model(model=model, required_sessions=["KRX_REG", "US_PRE"])
+
+
+def test_missing_required_slippage_session_raises() -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": 10.0},
+        failure_rate_by_session={"KRX_REG": 0.01, "US_PRE": 0.08},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="missing slippage_bps_by_session.*US_PRE"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG", "US_PRE"])
+
+
+def test_missing_required_failure_rate_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},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="missing failure_rate_by_session.*US_PRE"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG", "US_PRE"])
+
+
+def test_invalid_failure_rate_range_raises() -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": 10.0},
+        failure_rate_by_session={"KRX_REG": 1.2},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="failure rate must be within"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])
+
+
+def test_unfavorable_fill_requirement_cannot_be_disabled() -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": 10.0},
+        failure_rate_by_session={"KRX_REG": 0.02},
+        unfavorable_fill_required=False,
+    )
+    with pytest.raises(ValueError, match="unfavorable_fill_required must be True"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])
+
+
+@pytest.mark.parametrize("bad_commission", [float("nan"), float("inf"), float("-inf")])
+def test_non_finite_commission_rejected(bad_commission: float) -> None:
+    model = BacktestCostModel(
+        commission_bps=bad_commission,
+        slippage_bps_by_session={"KRX_REG": 10.0},
+        failure_rate_by_session={"KRX_REG": 0.02},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="commission_bps"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])
+
+
+@pytest.mark.parametrize("bad_slippage", [float("nan"), float("inf"), float("-inf")])
+def test_non_finite_slippage_rejected(bad_slippage: float) -> None:
+    model = BacktestCostModel(
+        commission_bps=5.0,
+        slippage_bps_by_session={"KRX_REG": bad_slippage},
+        failure_rate_by_session={"KRX_REG": 0.02},
+        unfavorable_fill_required=True,
+    )
+    with pytest.raises(ValueError, match="slippage bps"):
+        validate_backtest_cost_model(model=model, required_sessions=["KRX_REG"])

tests/test_backtest_execution_model.py

@@ -0,0 +1,108 @@
+from __future__ import annotations
+
+import pytest
+
+from src.analysis.backtest_execution_model import (
+    BacktestExecutionModel,
+    ExecutionAssumptions,
+    ExecutionRequest,
+)
+
+
+def test_buy_uses_unfavorable_slippage_direction() -> None:
+    model = BacktestExecutionModel(
+        ExecutionAssumptions(
+            slippage_bps_by_session={"US_PRE": 50.0},
+            failure_rate_by_session={"US_PRE": 0.0},
+            partial_fill_rate_by_session={"US_PRE": 0.0},
+            seed=1,
+        )
+    )
+    out = model.simulate(
+        ExecutionRequest(side="BUY", session_id="US_PRE", qty=10, reference_price=100.0)
+    )
+    assert out.status == "FILLED"
+    assert out.avg_price == pytest.approx(100.5)
+
+
+def test_sell_uses_unfavorable_slippage_direction() -> None:
+    model = BacktestExecutionModel(
+        ExecutionAssumptions(
+            slippage_bps_by_session={"US_PRE": 50.0},
+            failure_rate_by_session={"US_PRE": 0.0},
+            partial_fill_rate_by_session={"US_PRE": 0.0},
+            seed=1,
+        )
+    )
+    out = model.simulate(
+        ExecutionRequest(side="SELL", session_id="US_PRE", qty=10, reference_price=100.0)
+    )
+    assert out.status == "FILLED"
+    assert out.avg_price == pytest.approx(99.5)
+
+
+def test_failure_rate_can_reject_order() -> None:
+    model = BacktestExecutionModel(
+        ExecutionAssumptions(
+            slippage_bps_by_session={"KRX_REG": 10.0},
+            failure_rate_by_session={"KRX_REG": 1.0},
+            partial_fill_rate_by_session={"KRX_REG": 0.0},
+            seed=42,
+        )
+    )
+    out = model.simulate(
+        ExecutionRequest(side="BUY", session_id="KRX_REG", qty=10, reference_price=100.0)
+    )
+    assert out.status == "REJECTED"
+    assert out.filled_qty == 0
+
+
+def test_partial_fill_applies_when_rate_is_one() -> None:
+    model = BacktestExecutionModel(
+        ExecutionAssumptions(
+            slippage_bps_by_session={"KRX_REG": 0.0},
+            failure_rate_by_session={"KRX_REG": 0.0},
+            partial_fill_rate_by_session={"KRX_REG": 1.0},
+            partial_fill_min_ratio=0.4,
+            partial_fill_max_ratio=0.4,
+            seed=0,
+        )
+    )
+    out = model.simulate(
+        ExecutionRequest(side="BUY", session_id="KRX_REG", qty=10, reference_price=100.0)
+    )
+    assert out.status == "PARTIAL"
+    assert out.filled_qty == 4
+    assert out.avg_price == 100.0
+
+
+@pytest.mark.parametrize("bad_slip", [-1.0, float("nan"), float("inf")])
+def test_invalid_slippage_is_rejected(bad_slip: float) -> None:
+    with pytest.raises(ValueError, match="slippage_bps"):
+        BacktestExecutionModel(
+            ExecutionAssumptions(
+                slippage_bps_by_session={"US_PRE": bad_slip},
+                failure_rate_by_session={"US_PRE": 0.0},
+                partial_fill_rate_by_session={"US_PRE": 0.0},
+            )
+        )
+
+
+@pytest.mark.parametrize("bad_rate", [-0.1, 1.1, float("nan")])
+def test_invalid_failure_or_partial_rates_are_rejected(bad_rate: float) -> None:
+    with pytest.raises(ValueError, match="failure_rate"):
+        BacktestExecutionModel(
+            ExecutionAssumptions(
+                slippage_bps_by_session={"US_PRE": 10.0},
+                failure_rate_by_session={"US_PRE": bad_rate},
+                partial_fill_rate_by_session={"US_PRE": 0.0},
+            )
+        )
+    with pytest.raises(ValueError, match="partial_fill_rate"):
+        BacktestExecutionModel(
+            ExecutionAssumptions(
+                slippage_bps_by_session={"US_PRE": 10.0},
+                failure_rate_by_session={"US_PRE": 0.0},
+                partial_fill_rate_by_session={"US_PRE": bad_rate},
+            )
+        )

tests/test_db.py

@@ -155,6 +155,9 @@ def test_mode_column_exists_in_schema() -> None:
     cursor = conn.execute("PRAGMA table_info(trades)")
     columns = {row[1] for row in cursor.fetchall()}
     assert "mode" in columns
+    assert "session_id" in columns
+    assert "strategy_pnl" in columns
+    assert "fx_pnl" in columns
 
 
 def test_mode_migration_adds_column_to_existing_db() -> None:
@@ -182,6 +185,13 @@ def test_mode_migration_adds_column_to_existing_db() -> None:
                 decision_id TEXT
             )"""
         )
+        old_conn.execute(
+            """
+            INSERT INTO trades (
+                timestamp, stock_code, action, confidence, rationale, quantity, price, pnl
+            ) VALUES ('2026-01-01T00:00:00+00:00', 'AAPL', 'SELL', 90, 'legacy', 1, 100.0, 123.45)
+            """
+        )
         old_conn.commit()
         old_conn.close()
 
@@ -190,6 +200,132 @@ def test_mode_migration_adds_column_to_existing_db() -> None:
         cursor = conn.execute("PRAGMA table_info(trades)")
         columns = {row[1] for row in cursor.fetchall()}
         assert "mode" in columns
+        assert "session_id" in columns
+        assert "strategy_pnl" in columns
+        assert "fx_pnl" in columns
+        migrated = conn.execute(
+            "SELECT pnl, strategy_pnl, fx_pnl, session_id FROM trades WHERE stock_code='AAPL' LIMIT 1"
+        ).fetchone()
+        assert migrated is not None
+        assert migrated[0] == 123.45
+        assert migrated[1] == 123.45
+        assert migrated[2] == 0.0
+        assert migrated[3] == "UNKNOWN"
         conn.close()
     finally:
         os.unlink(db_path)
+
+
+def test_log_trade_stores_strategy_and_fx_pnl_separately() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="AAPL",
+        action="SELL",
+        confidence=90,
+        rationale="fx split",
+        pnl=120.0,
+        strategy_pnl=100.0,
+        fx_pnl=20.0,
+        market="US_NASDAQ",
+        exchange_code="NASD",
+    )
+    row = conn.execute(
+        "SELECT pnl, strategy_pnl, fx_pnl FROM trades ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    assert row is not None
+    assert row[0] == 120.0
+    assert row[1] == 100.0
+    assert row[2] == 20.0
+
+
+def test_log_trade_backward_compat_sets_strategy_pnl_from_pnl() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="005930",
+        action="SELL",
+        confidence=80,
+        rationale="legacy",
+        pnl=50.0,
+        market="KR",
+        exchange_code="KRX",
+    )
+    row = conn.execute(
+        "SELECT pnl, strategy_pnl, fx_pnl FROM trades ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    assert row is not None
+    assert row[0] == 50.0
+    assert row[1] == 50.0
+    assert row[2] == 0.0
+
+
+def test_log_trade_partial_fx_input_does_not_infer_negative_strategy_pnl() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="AAPL",
+        action="SELL",
+        confidence=70,
+        rationale="fx only",
+        pnl=0.0,
+        fx_pnl=10.0,
+        market="US_NASDAQ",
+        exchange_code="NASD",
+    )
+    row = conn.execute(
+        "SELECT pnl, strategy_pnl, fx_pnl FROM trades ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    assert row is not None
+    assert row[0] == 10.0
+    assert row[1] == 0.0
+    assert row[2] == 10.0
+
+
+def test_log_trade_persists_explicit_session_id() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="AAPL",
+        action="BUY",
+        confidence=70,
+        rationale="session test",
+        market="US_NASDAQ",
+        exchange_code="NASD",
+        session_id="US_PRE",
+    )
+    row = conn.execute("SELECT session_id FROM trades ORDER BY id DESC LIMIT 1").fetchone()
+    assert row is not None
+    assert row[0] == "US_PRE"
+
+
+def test_log_trade_auto_derives_session_id_when_not_provided() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="005930",
+        action="BUY",
+        confidence=70,
+        rationale="auto session",
+        market="KR",
+        exchange_code="KRX",
+    )
+    row = conn.execute("SELECT session_id FROM trades ORDER BY id DESC LIMIT 1").fetchone()
+    assert row is not None
+    assert row[0] != "UNKNOWN"
+
+
+def test_log_trade_unknown_market_falls_back_to_unknown_session() -> None:
+    conn = init_db(":memory:")
+    log_trade(
+        conn=conn,
+        stock_code="X",
+        action="BUY",
+        confidence=70,
+        rationale="unknown market",
+        market="MARS",
+        exchange_code="MARS",
+    )
+    row = conn.execute("SELECT session_id FROM trades ORDER BY id DESC LIMIT 1").fetchone()
+    assert row is not None
+    assert row[0] == "UNKNOWN"

tests/test_main.py

@@ -15,6 +15,8 @@ from src.evolution.scorecard import DailyScorecard
 from src.logging.decision_logger import DecisionLogger
 from src.main import (
     KILL_SWITCH,
+    _should_force_exit_for_overnight,
+    _should_block_overseas_buy_for_fx_buffer,
     _trigger_emergency_kill_switch,
     _apply_dashboard_flag,
     _determine_order_quantity,
@@ -3690,6 +3692,81 @@ class TestOverseasBrokerIntegration:
         # DB도 브로커도 보유 없음 → BUY 주문이 실행되어야 함 (회귀 테스트)
         overseas_broker.send_overseas_order.assert_called_once()
 
+    @pytest.mark.asyncio
+    async def test_overseas_buy_blocked_by_usd_buffer_guard(self) -> None:
+        """Overseas BUY must be blocked when USD buffer would be breached."""
+        db_conn = init_db(":memory:")
+
+        overseas_broker = MagicMock()
+        overseas_broker.get_overseas_price = AsyncMock(
+            return_value={"output": {"last": "182.50"}}
+        )
+        overseas_broker.get_overseas_balance = AsyncMock(
+            return_value={
+                "output1": [],
+                "output2": [
+                    {
+                        "frcr_evlu_tota": "50000.00",
+                        "frcr_buy_amt_smtl": "0.00",
+                    }
+                ],
+            }
+        )
+        overseas_broker.get_overseas_buying_power = AsyncMock(
+            return_value={"output": {"ovrs_ord_psbl_amt": "50000.00"}}
+        )
+        overseas_broker.send_overseas_order = AsyncMock(return_value={"msg1": "주문접수"})
+
+        engine = MagicMock(spec=ScenarioEngine)
+        engine.evaluate = MagicMock(return_value=_make_buy_match("AAPL"))
+
+        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()
+
+        decision_logger = MagicMock()
+        decision_logger.log_decision = MagicMock(return_value="decision-id")
+
+        settings = MagicMock()
+        settings.POSITION_SIZING_ENABLED = False
+        settings.CONFIDENCE_THRESHOLD = 80
+        settings.USD_BUFFER_MIN = 49900.0
+        settings.MODE = "paper"
+        settings.PAPER_OVERSEAS_CASH = 50000.0
+
+        await trading_cycle(
+            broker=MagicMock(),
+            overseas_broker=overseas_broker,
+            scenario_engine=engine,
+            playbook=_make_playbook(market="US"),
+            risk=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,
+        )
+
+        overseas_broker.send_overseas_order.assert_not_called()
+
 
 # ---------------------------------------------------------------------------
 # _retry_connection — unit tests (issue #209)
@@ -3723,7 +3800,6 @@ class TestRetryConnection:
         with patch("src.main.asyncio.sleep") as mock_sleep:
             mock_sleep.return_value = None
             result = await _retry_connection(flaky, label="flaky")
-
         assert result == "ok"
         assert call_count == 2
         mock_sleep.assert_called_once()
@@ -3778,6 +3854,48 @@ class TestRetryConnection:
         assert call_count == 1  # No retry for non-ConnectionError
 
 
+def test_fx_buffer_guard_applies_only_to_us_and_respects_boundary() -> None:
+    settings = MagicMock()
+    settings.USD_BUFFER_MIN = 1000.0
+
+    us_market = MagicMock()
+    us_market.is_domestic = False
+    us_market.code = "US_NASDAQ"
+
+    blocked, remaining, required = _should_block_overseas_buy_for_fx_buffer(
+        market=us_market,
+        action="BUY",
+        total_cash=5000.0,
+        order_amount=4001.0,
+        settings=settings,
+    )
+    assert blocked
+    assert remaining == 999.0
+    assert required == 1000.0
+
+    blocked_eq, _, _ = _should_block_overseas_buy_for_fx_buffer(
+        market=us_market,
+        action="BUY",
+        total_cash=5000.0,
+        order_amount=4000.0,
+        settings=settings,
+    )
+    assert not blocked_eq
+
+    jp_market = MagicMock()
+    jp_market.is_domestic = False
+    jp_market.code = "JP"
+    blocked_jp, _, required_jp = _should_block_overseas_buy_for_fx_buffer(
+        market=jp_market,
+        action="BUY",
+        total_cash=5000.0,
+        order_amount=4500.0,
+        settings=settings,
+    )
+    assert not blocked_jp
+    assert required_jp == 0.0
+
+
 # run_daily_session — daily CB baseline (daily_start_eval) tests (issue #207)
 # ---------------------------------------------------------------------------
 
@@ -5193,6 +5311,88 @@ async def test_order_policy_rejection_skips_order_execution() -> None:
     broker.send_order.assert_not_called()
 
 
+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")):
+        settings = MagicMock()
+        settings.OVERNIGHT_EXCEPTION_ENABLED = True
+        try:
+            KILL_SWITCH.new_orders_blocked = True
+            assert _should_force_exit_for_overnight(market=market, settings=settings)
+        finally:
+            KILL_SWITCH.clear_block()
+
+
+@pytest.mark.asyncio
+async def test_kill_switch_block_does_not_block_sell_reduction() -> None:
+    """KillSwitch should block BUY entries, but allow SELL risk reduction orders."""
+    db_conn = init_db(":memory:")
+    decision_logger = DecisionLogger(db_conn)
+
+    broker = MagicMock()
+    broker.get_current_price = AsyncMock(return_value=(100.0, 0.5, 0.0))
+    broker.get_balance = AsyncMock(
+        return_value={
+            "output1": [{"pdno": "005930", "ord_psbl_qty": "3"}],
+            "output2": [
+                {
+                    "tot_evlu_amt": "100000",
+                    "dnca_tot_amt": "50000",
+                    "pchs_amt_smtl_amt": "50000",
+                }
+            ],
+        }
+    )
+    broker.send_order = AsyncMock(return_value={"msg1": "OK"})
+
+    market = MagicMock()
+    market.name = "Korea"
+    market.code = "KR"
+    market.exchange_code = "KRX"
+    market.is_domestic = True
+
+    telegram = MagicMock()
+    telegram.notify_trade_execution = AsyncMock()
+    telegram.notify_fat_finger = AsyncMock()
+    telegram.notify_circuit_breaker = AsyncMock()
+    telegram.notify_scenario_matched = AsyncMock()
+
+    settings = MagicMock()
+    settings.POSITION_SIZING_ENABLED = False
+    settings.CONFIDENCE_THRESHOLD = 80
+    settings.OVERNIGHT_EXCEPTION_ENABLED = True
+    settings.MODE = "paper"
+
+    try:
+        KILL_SWITCH.new_orders_blocked = True
+        await trading_cycle(
+            broker=broker,
+            overseas_broker=MagicMock(),
+            scenario_engine=MagicMock(evaluate=MagicMock(return_value=_make_sell_match())),
+            playbook=_make_playbook(),
+            risk=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="005930",
+            scan_candidates={},
+            settings=settings,
+        )
+    finally:
+        KILL_SWITCH.clear_block()
+
+    broker.send_order.assert_called_once()
+
+
 @pytest.mark.asyncio
 async def test_blackout_queues_order_and_skips_submission() -> None:
     """When blackout is active, order submission is replaced by queueing."""

tests/test_walk_forward_split.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+import pytest
+
+from src.analysis.walk_forward_split import generate_walk_forward_splits
+
+
+def test_generates_sequential_folds() -> None:
+    folds = generate_walk_forward_splits(
+        n_samples=30,
+        train_size=10,
+        test_size=5,
+    )
+    assert len(folds) == 4
+    assert folds[0].train_indices == list(range(0, 10))
+    assert folds[0].test_indices == list(range(10, 15))
+    assert folds[1].train_indices == list(range(5, 15))
+    assert folds[1].test_indices == list(range(15, 20))
+
+
+def test_purge_removes_boundary_samples_before_test() -> None:
+    folds = generate_walk_forward_splits(
+        n_samples=25,
+        train_size=8,
+        test_size=4,
+        purge_size=2,
+    )
+    first = folds[0]
+    # test starts at 10, purge=2 => train end must be 7
+    assert first.train_indices == list(range(0, 8))
+    assert first.test_indices == list(range(10, 14))
+
+
+def test_embargo_excludes_post_test_samples_from_next_train() -> None:
+    folds = generate_walk_forward_splits(
+        n_samples=45,
+        train_size=15,
+        test_size=5,
+        step_size=10,
+        embargo_size=3,
+    )
+    assert len(folds) >= 2
+    # Fold1 test: 15..19, next fold train window: 10..24.
+    # embargo_size=3 should remove 20,21,22 from fold2 train.
+    second_train = folds[1].train_indices
+    assert 20 not in second_train
+    assert 21 not in second_train
+    assert 22 not in second_train
+    assert 23 in second_train
+
+
+def test_respects_min_train_size_and_returns_empty_when_impossible() -> None:
+    folds = generate_walk_forward_splits(
+        n_samples=15,
+        train_size=5,
+        test_size=5,
+        min_train_size=6,
+    )
+    assert folds == []
+
+
+def test_embargo_uses_last_accepted_fold_when_intermediate_fold_skips() -> None:
+    folds = generate_walk_forward_splits(
+        n_samples=30,
+        train_size=5,
+        test_size=3,
+        step_size=5,
+        embargo_size=1,
+        min_train_size=5,
+    )
+    # 1st fold accepted, 2nd skipped by min_train_size, subsequent folds still generated.
+    assert len(folds) == 3
+    assert folds[0].test_indices == [5, 6, 7]
+    assert folds[1].test_indices == [15, 16, 17]
+    assert folds[2].test_indices == [25, 26, 27]
+
+
+@pytest.mark.parametrize(
+    ("n_samples", "train_size", "test_size"),
+    [
+        (0, 10, 2),
+        (10, 0, 2),
+        (10, 5, 0),
+    ],
+)
+def test_invalid_args_raise(n_samples: int, train_size: int, test_size: int) -> None:
+    with pytest.raises(ValueError):
+        generate_walk_forward_splits(
+            n_samples=n_samples,
+            train_size=train_size,
+            test_size=test_size,
+        )