[codex] Cache PM recovery index

publisher_verify.py

@@ -856,6 +856,31 @@ def _reactivate(conn):
            pm_closed.status_code == 503
            and pm_closed.json().get("detail") == "per_miner_seed_secret_missing")
         os.environ["CATHEDRAL_PERMINER_SEED_SECRET"] = "publisher-verify-stable-seed"
+        _pm._instance_index.cache_clear()
+        os.environ["CATHEDRAL_PERMINER_ALLOTMENT_T1"] = "3"
+        cache_epoch = _pm.current_epoch()
+        cache_id = _pm.instance_id("cache-hotkey", cache_epoch, 1, 2)
+        cache_first = _pm.recover_tier_seq_for("cache-hotkey", cache_epoch, cache_id)
+        cache_info_first = _pm._instance_index.cache_info()
+        cache_second = _pm.recover_tier_seq_for("cache-hotkey", cache_epoch, cache_id)
+        cache_info_second = _pm._instance_index.cache_info()
+        cache_foreign = _pm.recover_tier_seq_for("cache-other-hotkey", cache_epoch, cache_id)
+        os.environ["CATHEDRAL_PERMINER_ALLOTMENT_T1"] = "2"
+        cache_reduced_allotment = _pm.recover_tier_seq_for("cache-hotkey", cache_epoch, cache_id)
+        os.environ["CATHEDRAL_PERMINER_ALLOTMENT_T1"] = "3"
+        os.environ["CATHEDRAL_PERMINER_SEED_SECRET"] = "publisher-verify-cache-rotated"
+        cache_rotated = _pm.recover_tier_seq_for("cache-hotkey", cache_epoch, cache_id)
+        ck("per-miner recovery index is cached and seed-scoped",
+           cache_first == (1, 2)
+           and cache_second == (1, 2)
+           and cache_info_first.misses >= 1
+           and cache_info_second.hits > cache_info_first.hits
+           and cache_foreign is None
+           and cache_reduced_allotment is None
+           and cache_rotated is None)
+        os.environ["CATHEDRAL_PERMINER_SEED_SECRET"] = "publisher-verify-stable-seed"
+        os.environ["CATHEDRAL_PERMINER_ALLOTMENT_T1"] = "1"
+        _pm._instance_index.cache_clear()
         def _map_pm_coldkey(conn):
             conn.execute(
                 "INSERT OR REPLACE INTO coldkey_map(hotkey, coldkey, updated_at_iso) "

scaffold/publisher/per_miner.py

@@ -125,6 +125,12 @@ def allotment_for(tier: int) -> int:
                             {1: 10_000, 2: 10_000}.get(tier, 10_000)))
 
 
+# Worst-case entries are roughly cache_size * allotment_for(tier).
+_RECOVER_INDEX_CACHE_SIZE = max(
+    1, _env_int("CATHEDRAL_PERMINER_RECOVER_INDEX_CACHE", 64)
+)
+
+
 def assignment_page_limit_max() -> int:
     return max(1, min(500, _env_int("CATHEDRAL_PERMINER_MAX_PAGE_LIMIT", 50)))
 
@@ -237,6 +243,10 @@ def _seed_secret_bytes() -> bytes:
     return _EPHEMERAL_SEED_SECRET
 
 
+def _seed_secret_fingerprint() -> str:
+    return hashlib.sha256(_seed_secret_bytes()).hexdigest()
+
+
 # --------------------------------------------------------------------------
 # Per-miner instance set generation
 # --------------------------------------------------------------------------
@@ -369,19 +379,33 @@ def verify_miner_submission_for(
 
 
 def recover_tier_seq_for(hotkey: str, epoch: int, challenge_id: str) -> tuple[int, int] | None:
-    """Find (tier, seq) for a challenge_id by scanning the miner's allotment.
+    """Find (tier, seq) for a challenge_id in the miner's allotment.
     Returns None if the challenge_id was not generated for this hotkey+epoch.
     """
     parsed = parse_challenge_id(challenge_id)
-    candidate_tiers = [parsed["tier"]] if parsed and parsed["tier"] in TIERS else TIERS
-    for tier in candidate_tiers:
-        for seq in range(allotment_for(tier)):
-            cid = instance_id(hotkey, epoch, tier, seq)
-            if cid == challenge_id:
-                return tier, seq
+    if not parsed or parsed["epoch"] != int(epoch) or parsed["tier"] not in TIERS:
+        return None
+    tier = parsed["tier"]
+    seq = _instance_index(_recover_index_key(hotkey, epoch, tier)).get(challenge_id)
+    if seq is not None:
+        return tier, seq
     return None
 
 
+def _recover_index_key(hotkey: str, epoch: int, tier: int) -> tuple[str, int, int, int, str]:
+    return hotkey, int(epoch), int(tier), allotment_for(tier), _seed_secret_fingerprint()
+
+
+@lru_cache(maxsize=_RECOVER_INDEX_CACHE_SIZE)
+def _instance_index(key: tuple[str, int, int, int, str]) -> dict[str, int]:
+    """Shared read-only cid->seq cache; do not mutate the returned dict."""
+    hotkey, epoch, tier, allotment, _secret_fingerprint = key
+    return {
+        instance_id(hotkey, epoch, tier, seq): seq
+        for seq in range(allotment)
+    }
+
+
 def recover_seq_for(hotkey: str, epoch: int, challenge_id: str) -> int | None:
     """Find the seq number for a challenge_id. Kept for backward compatibility."""
     result = recover_tier_seq_for(hotkey, epoch, challenge_id)