Nightly events retention: purge observability-spam events >14 days

The events table hit 3.5 GB / 1.8M rows, ~80% of it module_tick /
criteria_evaluated / signal_blocked / account_* snapshots that are only
useful for a few days of debugging. Trading-decision events are kept
forever. One-time purge (353k rows) + missing hot-path indexes
(raw_news_items payload_hash + fetched_at: 911k seq scans -> index-only)
and per-table autovacuum tuning applied directly to the DB.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

research/premarket_eval.py

@@ -89,6 +89,20 @@ def main() -> None:
                     f.write(line + "\n")
     except Exception as exc:  # noqa: BLE001 — never let the refetch break the eval
         print(f"SIP refetch skipped: {exc}")
+    # nightly DB housekeeping: observability-spam events (module ticks, per-symbol criteria
+    # evaluations, account snapshots) are only useful for a few days of debugging but were
+    # ~80% of the events table (3.5 GB before the 2026-07-04 purge). Trading-decision events
+    # are kept forever. Autovacuum (tuned to 2%) reclaims the space.
+    try:
+        spam = ("module_tick", "criteria_evaluated", "signal_blocked",
+                "account_summary_updated", "account_positions_updated", "account_orders_updated")
+        ph = ",".join(["?"] * len(spam))
+        cur = con.execute(
+            f"DELETE FROM events WHERE event_type IN ({ph}) "
+            "AND timestamp < now() - interval '14 days'", list(spam))
+        print(f"events retention: purged {getattr(cur, 'rowcount', '?')} spam rows >14d")
+    except Exception as exc:  # noqa: BLE001 — housekeeping must never break the eval
+        print(f"events retention skipped: {exc}")
     labeler.build_premarket(con, rebuild=True)          # rebuild from the grown DB
     m = _metrics(con)
     ts = datetime.now(timezone.utc).isoformat(timespec="seconds")