feat(manifold): add persistent cooldowns to reduce redundant evaluations

The trading loop re-evaluated the same ~22 politics/tech markets every ~60s,
flooding manifold_match_audit with ~76k rows (~3,500 attempts/market) of which
none carried new information, making the metrics uninterpretable.

Add per-market persistent cooldowns:
- New table manifold_eval_cooldown (poly_market_id PK, last_evaluated_at,
  last_status, retry_after, cooldown_reason) created via run_migrations.
- bayesian.evaluate() consults the cooldown BEFORE calling the matcher and skips
  the call entirely while now() < retry_after — no matcher call, no audit row,
  no signal (equivalent to no_results). After a real evaluation it upserts the
  cooldown with a verdict-dependent backoff: no_results/low_score/outcome_mismatch/
  ambiguous_inversion -> 24h, conditional_market -> 7d, accepted -> 1h.
- manifold.py stays a pure client/matcher with no DB access.
- New db methods get_manifold_cooldown / upsert_manifold_cooldown.
- /api/metrics/manifold-matches summary gains unique_markets
  {evaluated, accepted, coverage_rate} for per-market (not per-attempt) coverage.

Matching logic, MANIFOLD_MATCHER_VERSION, MANIFOLD_LOGODDS_WEIGHT, edge/exposure
thresholds and existing audit rows are unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

api/main.py

@@ -226,6 +226,10 @@ async def get_manifold_matches():
     summary.trades_dominated_by_mfld — non-excluded accepted-match trades where
       feat_mfld_lo is the largest signal (consistent with attribution/features,
       which also exclude excluded_from_metrics trades).
+    summary.unique_markets — distinct-market coverage (per-market, not per-attempt):
+      evaluated      — DISTINCT poly_market_id in manifold_match_audit (all versions)
+      accepted       — DISTINCT poly_market_id accepted by the current matcher
+      coverage_rate  — accepted / evaluated (null when evaluated=0)
 
     recent_matches: last 50 rows from manifold_match_audit, newest first, each
       tagged with matcher_version.

bot/data/db.py

@@ -553,6 +553,46 @@ class Database:
                 poly_outcome_type, mfld_outcome_type, matcher_version,
             )
 
+    async def get_manifold_cooldown(self, poly_market_id: str) -> Optional[dict]:
+        """Return the cooldown row for a market, or None if no cooldown is recorded.
+
+        The caller decides whether the cooldown is still active by comparing
+        now() against retry_after (see bayesian.evaluate()).
+        """
+        async with self._pool.acquire() as conn:
+            row = await conn.fetchrow(
+                "SELECT poly_market_id, last_evaluated_at, last_status, "
+                "retry_after, cooldown_reason FROM manifold_eval_cooldown "
+                "WHERE poly_market_id = $1",
+                poly_market_id,
+            )
+        return dict(row) if row else None
+
+    async def upsert_manifold_cooldown(
+        self,
+        poly_market_id: str,
+        last_status: str,
+        retry_after,
+        cooldown_reason: Optional[str] = None,
+    ) -> None:
+        """Insert or refresh the cooldown for a market after a real evaluation.
+
+        last_evaluated_at is stamped server-side with now(); retry_after is the
+        caller-computed earliest re-evaluation time.
+        """
+        async with self._pool.acquire() as conn:
+            await conn.execute("""
+                INSERT INTO manifold_eval_cooldown (
+                    poly_market_id, last_evaluated_at, last_status,
+                    retry_after, cooldown_reason
+                ) VALUES ($1, now(), $2, $3, $4)
+                ON CONFLICT (poly_market_id) DO UPDATE SET
+                    last_evaluated_at = now(),
+                    last_status       = EXCLUDED.last_status,
+                    retry_after       = EXCLUDED.retry_after,
+                    cooldown_reason   = EXCLUDED.cooldown_reason
+            """, poly_market_id, last_status, retry_after, cooldown_reason)
+
     async def mark_manifold_audit_used(self, audit_id: str) -> None:
         async with self._pool.acquire() as conn:
             await conn.execute(
@@ -592,6 +632,15 @@ class Database:
                 WHERE matcher_version = 'legacy_pre_outcome_guard'
                   AND match_status = 'accepted'
             """)
+            unique_markets = await conn.fetchrow("""
+                SELECT
+                    COUNT(DISTINCT poly_market_id) AS evaluated,
+                    COUNT(DISTINCT poly_market_id) FILTER (
+                        WHERE match_status = 'accepted'
+                          AND matcher_version = $1
+                    ) AS accepted
+                FROM manifold_match_audit
+            """, MANIFOLD_MATCHER_VERSION)
             mfld_dominated = await conn.fetchrow("""
                 SELECT COUNT(*) AS cnt FROM trades
                 WHERE (excluded_from_metrics IS NOT TRUE)
@@ -627,6 +676,14 @@ class Database:
                         int(legacy["accepted_without_outcome_type"] or 0),
                 },
                 "trades_dominated_by_mfld": int(mfld_dominated["cnt"] or 0),
+                "unique_markets": {
+                    "evaluated": int(unique_markets["evaluated"] or 0),
+                    "accepted":  int(unique_markets["accepted"] or 0),
+                    "coverage_rate": (
+                        float(unique_markets["accepted"]) / float(unique_markets["evaluated"])
+                        if unique_markets["evaluated"] else None
+                    ),
+                },
             },
             "recent_matches": [dict(r) for r in rows],
         }

bot/data/schema.sql

@@ -291,3 +291,29 @@ CREATE TABLE IF NOT EXISTS checkpoint_alerts (
     fired_at        TIMESTAMPTZ NOT NULL,
     last_fired_at   TIMESTAMPTZ
 );
+
+-- ─────────────────────────────────────────────────────────────────────────────
+-- Manifold evaluation cooldown — per-market backoff for the Manifold matcher
+--
+-- The trading loop re-evaluates the same ~stable set of politics/tech markets
+-- every cycle (~60s).  Most resolve to a stable terminal verdict (no Manifold
+-- coverage, low-score, outcome mismatch, conditional market) that will not change
+-- on the next cycle.  Re-querying them every minute floods manifold_match_audit
+-- with redundant rows and makes the metrics uninterpretable.
+--
+-- This table records, per poly_market_id, when the market was last evaluated and
+-- the earliest time it should be evaluated again (retry_after).  evaluate() in
+-- bot/strategy/bayesian.py consults it BEFORE calling the matcher and skips the
+-- call (and the audit write) entirely while now() < retry_after.
+--
+-- last_status / cooldown_reason are stored for observability only.
+-- ─────────────────────────────────────────────────────────────────────────────
+CREATE TABLE IF NOT EXISTS manifold_eval_cooldown (
+    poly_market_id    TEXT PRIMARY KEY,
+    last_evaluated_at TIMESTAMPTZ NOT NULL,
+    last_status       TEXT NOT NULL,
+    retry_after       TIMESTAMPTZ NOT NULL,
+    cooldown_reason   TEXT
+);
+
+CREATE INDEX IF NOT EXISTS idx_mfld_cooldown_retry ON manifold_eval_cooldown(retry_after);

bot/strategy/bayesian.py

@@ -14,7 +14,7 @@ import logging
 import math
 import uuid
 from dataclasses import dataclass, field
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 from typing import Optional, TYPE_CHECKING
 
 from bot.data.polymarket import Market, market_family_key
@@ -66,6 +66,34 @@ MANIFOLD_LOGODDS_WEIGHT = 0.6
 MAX_NEWS_QUERIES_PER_CYCLE = 5
 
 
+# ─────────────────────────────────────────────────────────────────────────────
+# Manifold evaluation cooldown
+#
+# Per-market backoff so the trading loop stops re-querying Manifold (and flooding
+# manifold_match_audit) for markets whose verdict is stable.  Computed from the
+# match result; longer for verdicts that essentially never change.
+#   no_results            → 24 h  (Manifold has no market on this topic)
+#   rejected/low_score    → 24 h  (best candidate below Jaccard threshold)
+#   rejected/outcome_mism. → 24 h  (outcome types differ)
+#   rejected/ambiguous    → 24 h  (party named but inversion unverifiable)
+#   rejected/conditional  →  7 d  (premise-gated market; structural, won't change)
+#   accepted              →  1 h  (signal is live; refresh probability hourly)
+# ─────────────────────────────────────────────────────────────────────────────
+def _cooldown_for(result: ManifoldMatchResult) -> tuple[timedelta, str]:
+    """Map a Manifold match result to (retry_delay, cooldown_reason)."""
+    if result.status == "accepted":
+        return timedelta(hours=1), "accepted"
+    if result.status == "no_results":
+        return timedelta(hours=24), "no_results"
+    # rejected — classify by the reason text the matcher produced
+    reason = result.match_reason or "rejected"
+    if "conditional_market" in reason:
+        return timedelta(days=7), reason
+    # outcome_mismatch, ambiguous_inversion, and low_score (jaccard<threshold)
+    # all settle in 24 h.
+    return timedelta(hours=24), reason
+
+
 # ─────────────────────────────────────────────────────────────────────────────
 # Phase 4 — Regime-based minimum edge (uses edge_NET, not edge_gross)
 # ─────────────────────────────────────────────────────────────────────────────
@@ -444,6 +472,29 @@ class BayesianStrategy:
         audit_id: Optional[str] = None
 
         if (is_politics or is_tech) and self._manifold is not None:
+            # ── Cooldown gate ────────────────────────────────────────────────
+            # Skip markets whose Manifold verdict was recently settled to a
+            # stable value.  A skip is equivalent to a no-signal: the matcher is
+            # NOT called and NO manifold_match_audit row is written, so only real
+            # evaluations are recorded.  See _cooldown_for() and the
+            # manifold_eval_cooldown table.
+            in_cooldown = False
+            if self._db is not None and market.id:
+                try:
+                    cd = await self._db.get_manifold_cooldown(market.id)
+                except Exception as exc:
+                    log.warning("Failed to read manifold cooldown: %s", exc)
+                    cd = None
+                if cd is not None and datetime.now(timezone.utc) < cd["retry_after"]:
+                    in_cooldown = True
+                    log.info(
+                        "MANIFOLD_COOLDOWN skip market=%s | last_status=%s "
+                        "retry_after=%s | %s",
+                        market.id, cd["last_status"],
+                        cd["retry_after"].isoformat(), market.question[:50],
+                    )
+
+            if not in_cooldown:
                 manifold_result = await self._manifold.get_match(market.question)
 
                 # Persist audit record for ALL outcomes (accepted / rejected / no_results)
@@ -478,6 +529,21 @@ class BayesianStrategy:
                             log.warning("Failed to save manifold audit: %s", exc)
                             audit_id = None
 
+                    # Record the cooldown so this market is not re-queried every
+                    # cycle.  Written even if the audit save above failed — we
+                    # still performed a real evaluation.
+                    if market.id:
+                        delay, cd_reason = _cooldown_for(manifold_result)
+                        try:
+                            await self._db.upsert_manifold_cooldown(
+                                poly_market_id=market.id,
+                                last_status=manifold_result.status,
+                                retry_after=datetime.now(timezone.utc) + delay,
+                                cooldown_reason=cd_reason,
+                            )
+                        except Exception as exc:
+                            log.warning("Failed to save manifold cooldown: %s", exc)
+
                 # Structured log — both forms for compatibility
                 log.info(
                     "MANIFOLD_MATCH poly='%s' mfld='%s' score=%s raw=%s final=%s"