fix(bot): multi-project signal freeze — cache-only _get_channel + per-tick scanner cap

Root cause: When 3+ projects generated pending simultaneously, Bot's
pending_approval_scanner made 20-40 Discord API calls in one tick
(sequential await), triggering Discord 429 rate limits which blocked
the entire scanner for 10-30s, freezing ALL signal delivery.

Two fixes:
1. _get_channel(): Replace guild.fetch_channels() (API call) with
   discord.utils.get(guild.channels) (in-memory cache). Eliminates
   redundant API calls + Lock contention when multiple projects arrive.
2. pending_approval_scanner: Per-tick caps (5 new + 5 status) prevent
   one tick from monopolizing Discord API quota. Excess items are
   naturally processed in subsequent 3-second ticks.

.agents/references/known-issues.md

@@ -514,6 +514,12 @@
 - **해결**: (1) writeChatSnapshot을 루프 바깥으로 이동(resolvedCount > 0일 때 1회), (2) `pd.conversation_id === activeSessionId` 조건 추가, (3) primaryCommand에서 'Deny'/'Allow' 텍스트 제외
 - **주의**: Bridge 파일 루프에서 외부 시스템(Discord)에 메시지를 보낼 때는 반드시 루프 바깥에서 집계 후 1회 발송
 
+### [2026-03-16] 멀티 프로젝트 동시 신호 정지 — Scanner O(N) Discord API + fetch_channels 병목
+- **증상**: 여러 프로젝트(3+)가 동시에 pending을 생성하면 모든 프로젝트의 신호 전달이 정지, AI 멈춤
+- **원인**: 2가지 복합: (1) `pending_approval_scanner`가 1 tick에 모든 pending(15건+)을 순차 처리 → 건당 `_get_channel()` + `channel.send/edit/fetch_message` Discord API 호출 → 20~40 API 호출이 Discord 429 rate limit 유발 → discord.py 내부 대기 → scanner tick 수십 초 블로킹 → 모든 프로젝트 신호 지연. (2) `_get_channel()`이 캐시 미스 시 `guild.fetch_channels()` (Discord API)를 asyncio Lock 내에서 호출 → Lock 대기 + API 호출 순차 발생
+- **해결**: (1) `_get_channel()`을 `discord.utils.get(guild.channels)` 캐시 기반으로 변경 — API 호출 제거, (2) scanner에 per-tick cap 추가 (Phase 1: 5건, Phase 2: 5건) — 나머지는 다음 tick으로 이월
+- **주의**: `guild.channels`는 discord.py 내부 캐시(Gateway 이벤트로 자동 갱신). 채널 삭제/생성은 캐시에 즉시 반영됨. per-tick cap은 최대 3초 지연을 유발하지만 전체 정지보다 훨씬 나음
+
 ### [2026-03-15] 이전 분석 오판(False Positive) — 교훈
 - **증상**: 시스템 감사 시 P0/P1으로 보고한 문제들이 실제로는 코드 방어 로직(멱등성, try-catch, 의도된 exact-match)으로 이미 방어되고 있었음
 - **원인**: 로컬 코드 스니펫만 보고 판단. 전체 데이터 생명주기를 끝까지 추적하지 않아 방어 로직을 놓침

bot.py

@@ -356,32 +356,34 @@ class GravityBot(commands.Bot):
         logger.info(f"Discovered {len(self.project_channels)} project channels")
 
     async def _get_channel(self, project_name: str) -> discord.TextChannel:
-        """Get or create a channel for a project. Lock-protected."""
+        """Get or create a channel for a project.
+
+        Uses guild.channels cache first (NO API call), only locks + creates
+        if channel truly doesn't exist. This prevents O(N) fetch_channels()
+        API calls when multiple projects arrive simultaneously.
+        """
         if project_name in self.project_channels:
             return self.project_channels[project_name]
 
+        channel_name = self._make_channel_name(project_name)
+
+        # 1. Check guild channel cache (NO API call — instant)
+        existing = discord.utils.get(
+            self.guild.channels, name=channel_name,
+            category_id=self.session_category.id,
+        )
+        if existing and isinstance(existing, discord.TextChannel):
+            self.project_channels[project_name] = existing
+            self.channel_to_project[existing.id] = project_name
+            logger.info(f"Found channel (cache): #{channel_name}")
+            return existing
+
+        # 2. Only lock + API call if truly creating new channel
         async with self._channel_lock:
-            # Double-check after lock
+            # Double-check after lock (another coroutine may have created it)
             if project_name in self.project_channels:
                 return self.project_channels[project_name]
 
-            channel_name = self._make_channel_name(project_name)
-
-            # Search existing channels FIRST (prevents duplicates)
-            try:
-                all_channels = await self.guild.fetch_channels()
-                for ch in all_channels:
-                    if (isinstance(ch, discord.TextChannel)
-                            and ch.name == channel_name
-                            and ch.category_id == self.session_category.id):
-                        self.project_channels[project_name] = ch
-                        self.channel_to_project[ch.id] = project_name
-                        logger.info(f"Found existing channel: #{channel_name}")
-                        return ch
-            except Exception as e:
-                logger.warning(f"fetch_channels failed: {e}")
-
-            # No existing channel — create new
             try:
                 ch = await self.guild.create_text_channel(
                     name=channel_name,
@@ -543,7 +545,11 @@ class GravityBot(commands.Bot):
 
     @tasks.loop(seconds=3)
     async def pending_approval_scanner(self):
-        """Scan bridge/pending/ for new approval requests + reload registrations."""
+        """Scan bridge/pending/ for new approval requests + reload registrations.
+
+        Per-tick caps prevent Discord API rate limit cascade when multiple
+        projects generate pending files simultaneously.
+        """
         try:
             # Reload conv→project registrations each cycle
             self._load_registrations()
@@ -551,8 +557,14 @@ class GravityBot(commands.Bot):
             # Channels are created on-demand when actual signals arrive
             # (via _get_channel in snapshot scanner / approval sender)
 
+            MAX_NEW_PER_TICK = 5     # Phase 1: max new pending to process per tick
+            MAX_STATUS_PER_TICK = 5  # Phase 2: max status changes to process per tick
+            phase1_processed = 0
+
             requests = self.bridge.get_pending_requests()
             for req in requests:
+                if phase1_processed >= MAX_NEW_PER_TICK:
+                    break
                 if req.request_id in self._sent_approval_ids:
                     continue
                 if req.discord_message_id != 0:
@@ -571,6 +583,7 @@ class GravityBot(commands.Bot):
                     if req.command.strip().lower() in reject_commands:
                         logger.warning(f"Auto-approve BLOCKED: command='{req.command}' is reject-word — skipping")
                         self._sent_approval_ids.add(req.request_id)
+                        phase1_processed += 1
                         continue
 
                     self._sent_approval_ids.add(req.request_id)
@@ -614,6 +627,7 @@ class GravityBot(commands.Bot):
                         embed.set_footer(text=f"auto-approve | {req.request_id[:12]}")
                         await channel.send(embed=embed)
                     logger.info(f"Auto-approved: {req.request_id[:12]} project={project} btn_idx={approve_btn_index}")
+                    phase1_processed += 1
                     continue
 
                 # Defer short-command pendings (e.g. "Run") by 4 cycles (~12s)
@@ -640,9 +654,13 @@ class GravityBot(commands.Bot):
                     self._sent_approval_ids.add(req.request_id)
                     self._sent_commands[req.request_id] = req.command
                     await self._send_approval_request(channel, req)
+                    phase1_processed += 1
 
             # ── Single-pass: handle auto_resolved, expired, and MERGE in one glob ──
+            phase2_processed = 0
             for f in self.bridge.pending_dir.glob("*.json"):
+                if phase2_processed >= MAX_STATUS_PER_TICK:
+                    break
                 try:
                     data = json.loads(f.read_text(encoding="utf-8-sig"))
                     status = data.get("status", "pending")
@@ -675,6 +693,7 @@ class GravityBot(commands.Bot):
                         self._sent_commands.pop(rid, None)
                         self._approval_messages.pop(rid, None)
                         self._sent_approval_ids.discard(rid)
+                        phase2_processed += 1
 
                     elif status == "expired":
                         msg_id = data.get("discord_message_id", 0)
@@ -697,6 +716,7 @@ class GravityBot(commands.Bot):
                         self._deferred_ids.pop(rid, None)
                         self._sent_commands.pop(rid, None)
                         self._sent_approval_ids.discard(rid)
+                        phase2_processed += 1
 
                     elif status == "pending":
                         # MERGE check: step_probe updated command in already-sent pending