fix(ops): live 신호 누락 방어 — bar 재정렬·datetime catch-up·N봉 재시도

lookback 롤링으로 bar_index가 밀려 매도/매수가 스킵되던 문제를 tick마다 재정렬하고,
last_processed_datetime 기반 catch-up과 OPS_CATCHUP_BARS(10) 2층 방어를 추가한다.

Co-authored-by: Cursor <cursoragent@cursor.com>

.env.example

@@ -90,6 +90,8 @@ OPS_SYNC_CANDLES=true
 # 비우면 DOWNLOAD_INTERVALS 전체 증분 sync
 # OPS_SYNC_INTERVALS=
 OPS_SIGNAL_TAIL_BARS=800
+# live 2층 방어: 최근 N봉(3분봉 기준) 미체결 신호 재시도 (10=30분)
+OPS_CATCHUP_BARS=10
 # OPS_PERSIST_SIGNAL_CACHE=false
 OPS_STATE_JSON=data/spot/operations/fractal_ops_state.json
 OPS_REPORT_JSON=docs/spot/3_operations/fractal_ops_report.json

src/bithumb/config.py

@@ -111,6 +111,7 @@ class Settings:
     ops_sync_intervals: list[int]
     ops_signal_tail_bars: int
     ops_persist_signal_cache: bool
+    ops_catchup_bars: int
     telegram_bot_token: str
     telegram_chat_id: str
     ops_telegram_enabled: bool
@@ -296,6 +297,7 @@ def load_settings(env_path: Path | None = None) -> Settings:
         ops_signal_tail_bars=int(os.getenv("OPS_SIGNAL_TAIL_BARS", "800")),
         ops_persist_signal_cache=os.getenv("OPS_PERSIST_SIGNAL_CACHE", "false").strip().lower()
         in ("1", "true", "yes", "on"),
+        ops_catchup_bars=int(os.getenv("OPS_CATCHUP_BARS", "10")),
         telegram_bot_token=os.getenv("COIN_TELEGRAM_BOT_TOKEN", "").strip(),
         telegram_chat_id=os.getenv("COIN_TELEGRAM_CHAT_ID", "").strip(),
         ops_telegram_enabled=_parse_ops_telegram_enabled(

src/bithumb/operations/runner.py

@@ -38,28 +38,150 @@ def sync_candles_if_enabled(settings: Settings) -> list[Any]:
     return sync_ops_candles(settings)
 
 
+def _parse_signal_dt(value: str) -> datetime:
+    """신호 datetime 파싱."""
+    return datetime.strptime(value, "%Y-%m-%d %H:%M:%S")
+
+
 def _signals_on_bar(signals: list[dict[str, Any]], bar_index: int) -> list[dict[str, Any]]:
     """특정 bar_index 신호만 반환."""
     return [s for s in signals if int(s.get("bar_index", -1)) == bar_index]
 
 
-def _pending_bar_indices(
+def _reconcile_processed_cursor(
+    state: dict[str, Any],
+    signals: list[dict[str, Any]],
+) -> None:
+    """tail/lookback 롤링 후 bar 커서가 신호보다 앞서 있으면 보정한다."""
+    last_dt = state.get("last_processed_datetime")
+    if not last_dt or not signals:
+        return
+    last_bar = int(state.get("last_processed_bar_index", -1))
+    max_bar_at_or_before = max(
+        (
+            int(sig.get("bar_index", -1))
+            for sig in signals
+            if str(sig.get("datetime", "")) <= str(last_dt)
+        ),
+        default=-1,
+    )
+    if last_bar > max_bar_at_or_before:
+        logger.warning(
+            "bar 커서 보정: %s → %s (last_dt=%s)",
+            last_bar,
+            max_bar_at_or_before,
+            last_dt,
+        )
+        state["last_processed_bar_index"] = max_bar_at_or_before
+
+
+def _pending_signals_for_ops(
     kept: list[dict[str, Any]],
+    *,
+    last_processed_datetime: str | None,
     last_bar_index: int,
     latest_bar_index: int,
-) -> list[int]:
-    """체결 대상 bar_index 목록 (최초 실행은 최신 봉만)."""
-    if last_bar_index < 0:
-        return [latest_bar_index] if any(
-            int(s.get("bar_index", -1)) == latest_bar_index for s in kept
-        ) else []
-    return sorted(
-        {
-            int(s.get("bar_index", -1))
-            for s in kept
-            if int(s.get("bar_index", -1)) > last_bar_index
-        }
+) -> list[dict[str, Any]]:
+    """아직 체결하지 않은 신호 목록 (datetime 우선, 시간순).
+
+    live tick은 최신 봉만이 아니라 마지막 처리 시각 이후 신호를 모두 처리한다.
+    """
+    last_dt = _parse_signal_dt(last_processed_datetime) if last_processed_datetime else None
+    pending: list[dict[str, Any]] = []
+    for sig in kept:
+        bar_idx = int(sig.get("bar_index", -1))
+        if bar_idx > latest_bar_index:
+            continue
+        sig_dt = _parse_signal_dt(str(sig["datetime"]))
+        if last_dt is not None:
+            if sig_dt <= last_dt:
+                continue
+        elif last_bar_index >= 0:
+            if bar_idx <= last_bar_index:
+                continue
+        elif bar_idx != latest_bar_index:
+            continue
+        pending.append(sig)
+    pending.sort(key=lambda s: (_parse_signal_dt(str(s["datetime"])), str(s.get("side", ""))))
+    return pending
+
+
+def _signal_key(sig: dict[str, Any]) -> tuple[str, str]:
+    """신호 고유 키 (datetime, side)."""
+    return (str(sig["datetime"]), str(sig["side"]))
+
+
+def _needs_catchup_execution(
+    sig: dict[str, Any],
+    trade_history: list[dict[str, Any]],
+) -> bool:
+    """최근 N봉 방어 대상인지 (성공·expected_skip 제외, 실패·미시도 포함)."""
+    dt, side = _signal_key(sig)
+    for record in trade_history:
+        if str(record.get("datetime")) != dt or str(record.get("side")) != side:
+            continue
+        trade = record.get("trade") or {}
+        if trade.get("executed"):
+            return False
+        if trade.get("expected_skip"):
+            return False
+        return True
+    return True
+
+
+def _catchup_signals_for_ops(
+    kept: list[dict[str, Any]],
+    *,
+    latest_bar_index: int,
+    catchup_bars: int,
+    trade_history: list[dict[str, Any]],
+) -> list[dict[str, Any]]:
+    """최근 N봉(신호 interval) 안에서 아직 성공 체결되지 않은 신호."""
+    if catchup_bars <= 0:
+        return []
+    min_bar = max(0, latest_bar_index - catchup_bars + 1)
+    catchup: list[dict[str, Any]] = []
+    for sig in kept:
+        bar_idx = int(sig.get("bar_index", -1))
+        if bar_idx < min_bar or bar_idx > latest_bar_index:
+            continue
+        if not _needs_catchup_execution(sig, trade_history):
+            continue
+        catchup.append(sig)
+    catchup.sort(
+        key=lambda s: (_parse_signal_dt(str(s["datetime"])), str(s.get("side", "")))
     )
+    return catchup
+
+
+def _merge_pending_signals(*groups: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """여러 pending 목록을 datetime·side 기준 병합 (중복 제거, 시간순)."""
+    merged: list[dict[str, Any]] = []
+    seen: set[tuple[str, str]] = set()
+    for group in groups:
+        for sig in group:
+            key = _signal_key(sig)
+            if key in seen:
+                continue
+            seen.add(key)
+            merged.append(sig)
+    merged.sort(
+        key=lambda s: (_parse_signal_dt(str(s["datetime"])), str(s.get("side", "")))
+    )
+    return merged
+
+
+def _pending_bar_indices(pending_signals: list[dict[str, Any]]) -> list[int]:
+    """pending 신호의 bar_index 목록 (시간순, 중복 제거)."""
+    seen: set[int] = set()
+    ordered: list[int] = []
+    for sig in pending_signals:
+        bar_idx = int(sig.get("bar_index", -1))
+        if bar_idx in seen:
+            continue
+        seen.add(bar_idx)
+        ordered.append(bar_idx)
+    return ordered
 
 
 def _cluster_pending(pending: list[dict[str, Any]]) -> list[tuple[str, list[dict[str, Any]]]]:
@@ -127,10 +249,11 @@ class OperationsRunner:
         latest_bar = int(len(df) - 1)
         gen = generate_raw_signals(self.settings, df=df, use_cache=True)
 
-        # tick: 최신 봉 후보만 MTF 평가 (전기간 MTF는 백테스트 전용)
+        filtered = filter_signals_for_ops(self.settings, gen["raw_signals"])
+        all_kept = filtered["kept"]
         bar_candidates = _signals_on_bar(gen["raw_signals"], latest_bar)
-        filtered = filter_signals_for_ops(self.settings, bar_candidates)
-        kept = filtered["kept"]
+
+        _reconcile_processed_cursor(self.state, gen["raw_signals"])
 
         reset_daily_trade_count(self.state)
         if self.settings.ops_mode == "live" and isinstance(self.executor, LiveExecutor):
@@ -146,13 +269,28 @@ class OperationsRunner:
                 self._notify_ops_error("portfolio_sync", exc)
 
         last_bar = int(self.state.get("last_processed_bar_index", -1))
-        target_bars = _pending_bar_indices(kept, last_bar, latest_bar)
+        last_dt = self.state.get("last_processed_datetime")
+        pending_signals = _pending_signals_for_ops(
+            all_kept,
+            last_processed_datetime=last_dt,
+            last_bar_index=last_bar,
+            latest_bar_index=latest_bar,
+        )
+        catchup_signals = _catchup_signals_for_ops(
+            all_kept,
+            latest_bar_index=latest_bar,
+            catchup_bars=self.settings.ops_catchup_bars,
+            trade_history=self.state.get("trade_history") or [],
+        )
+        catchup_keys = {_signal_key(s) for s in catchup_signals}
+        pending_signals = _merge_pending_signals(pending_signals, catchup_signals)
+        target_bars = _pending_bar_indices(pending_signals)
 
         executions: list[dict[str, Any]] = []
         max_daily = self.settings.ops_daily_max_trades
 
         for bar_idx in target_bars:
-            bar_signals = _signals_on_bar(kept, bar_idx)
+            bar_signals = _signals_on_bar(pending_signals, bar_idx)
             clusters = _cluster_pending(bar_signals)
             for side, cluster in clusters:
                 if self.state["trades_today_count"] >= max_daily:
@@ -161,7 +299,7 @@ class OperationsRunner:
                 cluster_size = len(cluster)
                 for sig in cluster:
                     full_sig = next(
-                        (k for k in kept if k["datetime"] == sig["datetime"]),
+                        (k for k in all_kept if k["datetime"] == sig["datetime"]),
                         sig,
                     )
                     try:
@@ -196,6 +334,7 @@ class OperationsRunner:
                         "signal_type": full_sig.get("signal_type"),
                         "price": full_sig["price"],
                         "bar_index": bar_idx,
+                        "catchup": _signal_key(full_sig) in catchup_keys,
                         "trade": trade.to_dict(),
                         "mtf_filter": full_sig.get("mtf_filter"),
                     }
@@ -231,9 +370,12 @@ class OperationsRunner:
         pipeline = {
             "technique_id": gen["technique_id"],
             "raw_count": len(bar_candidates),
-            "kept_count": len(kept),
+            "kept_count": len(all_kept),
             "rejected_count": len(filtered["rejected"]),
             "latest_bar_index": latest_bar,
+            "pending_signal_count": len(pending_signals),
+            "catchup_signal_count": len(catchup_signals),
+            "catchup_bars": self.settings.ops_catchup_bars,
         }
 
         now = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
@@ -260,6 +402,9 @@ class OperationsRunner:
             "raw_signals": pipeline["raw_count"],
             "filtered_signals": pipeline["kept_count"],
             "pending_bars": target_bars,
+            "pending_signal_count": pipeline["pending_signal_count"],
+            "catchup_signal_count": pipeline["catchup_signal_count"],
+            "catchup_bars": pipeline["catchup_bars"],
             "latest_bar_candidates": pipeline["raw_count"],
             "executions": executions,
             "portfolio": self.state["portfolio"],

src/bithumb/operations/signal_pipeline.py

@@ -67,6 +67,28 @@ def _merge_tail_signals(
     return merged
 
 
+def _reindex_signals_to_df(
+    signals: list[dict[str, Any]],
+    df: pd.DataFrame,
+) -> list[dict[str, Any]]:
+    """신호 bar_index를 현재 df 행 위치에 맞게 재부여한다.
+
+    lookback_days 롤링 윈도우로 앞쪽 봉이 빠지면 동일 datetime의 bar_index가
+    변하므로, 캐시 신호는 tick마다 datetime 기준으로 재정렬해야 한다.
+    """
+    if df.empty or not signals:
+        return signals
+    dt_to_idx = {str(row["datetime"]): int(i) for i, row in df.iterrows()}
+    reindexed: list[dict[str, Any]] = []
+    for signal in signals:
+        row = dict(signal)
+        idx = dt_to_idx.get(str(row.get("datetime", "")))
+        if idx is not None:
+            row["bar_index"] = idx
+        reindexed.append(row)
+    return reindexed
+
+
 def _load_technique_cached(cache_path: Path) -> TechniqueResult:
     """대용량 기법 JSON을 mtime 기준으로 메모리 캐시한다 (fractal_swing 등)."""
     path = Path(cache_path)
@@ -225,6 +247,7 @@ def generate_raw_signals(
                 len(cached.signals),
                 len(raw_signals),
             )
+        raw_signals = _reindex_signals_to_df(raw_signals, df)
         raw_signals = _apply_ops_min_score(
             cached.technique_id,
             raw_signals,
@@ -250,11 +273,12 @@ def generate_raw_signals(
     technique = get_technique(settings.ops_technique_id)
     params_obj = build_technique_params(settings)
     result = run_technique(technique, df, params_obj, gt_result=None)
+    signals = _reindex_signals_to_df(result.signals, df)
     return {
         "technique_id": result.technique_id,
         "technique_name": result.technique_name,
         "params": result.params,
-        "raw_signals": result.signals,
+        "raw_signals": signals,
         "data_end": data_end,
         "last_price": last_price,
         "from_cache": False,