fix: trend_windows now upserts instead of accumulating (7.5GB→4MB), add competitive signal retention cleanup

services/scheduler/app.py

@@ -475,6 +475,7 @@ async def main() -> None:
 
     logger.info("Scheduler started (tick=%ds)", SCHEDULER_TICK)
     recovery_counter = 0
+    cleanup_counter = 0
     try:
         while True:
             try:
@@ -486,6 +487,11 @@ async def main() -> None:
                         if recovery_counter >= 20:
                             recovery_counter = 0
                             await recover_stale_documents(pool, rds)
+                        # Run signal cleanup periodically (~25 minutes)
+                        cleanup_counter += 1
+                        if cleanup_counter >= CLEANUP_CYCLE_INTERVAL:
+                            cleanup_counter = 0
+                            await cleanup_old_signals(pool)
                     finally:
                         await release_lock(rds, "scheduler_cycle")
             except Exception:
@@ -542,5 +548,28 @@ async def recover_stale_documents(pool: asyncpg.Pool, rds: aioredis.Redis) -> in
     return enqueued
 
 
+# How often to run competitive signal cleanup (every ~100 cycles = ~25 minutes)
+CLEANUP_CYCLE_INTERVAL: int = 100
+# Keep competitive signals for this many days
+COMPETITIVE_SIGNAL_RETENTION_DAYS: int = 30
+
+
+async def cleanup_old_signals(pool: asyncpg.Pool) -> int:
+    """Delete competitive signal records older than the retention window.
+
+    Prevents the competitive_signal_records table from growing unbounded.
+    Returns the number of rows deleted.
+    """
+    result = await pool.execute(
+        "DELETE FROM competitive_signal_records WHERE computed_at < NOW() - INTERVAL '1 day' * $1",
+        COMPETITIVE_SIGNAL_RETENTION_DAYS,
+    )
+    # result is like "DELETE 1234"
+    count = int(result.split()[-1]) if result else 0
+    if count > 0:
+        logger.info("Cleaned up %d old competitive signal records", count)
+    return count
+
+
 if __name__ == "__main__":
     asyncio.run(main())