fix(clustering): raise coverage floor 0.5 -> 0.6 to gate degenerate stablecoin fits

services/clustering/backend/app/config.py

@@ -93,12 +93,15 @@ class Settings(BaseSettings):
     # construction and a re-cluster reproduces the same DBSCAN-noise. Below the
     # floor, drift is treated as STRUCTURAL (not staleness): recluster_recommended
     # returns False and the scheduler stops the ~60s auto-refit loop (see
-    # model_unclusterable). Observed coverage: ~0.16 uniform/tight, ~0.43
-    # heavy-tail, ~0.88 a healthy dominant-mode fit, so 0.5 separates the
-    # pathological fits from the healthy one. Raising it only treats MORE fits as
-    # un-clusterable (fewer re-fits, more honest UI): recall-safe, since classify +
-    # anomaly scoring run every tick regardless.
-    min_cluster_coverage: float = Field(default=0.5, ge=0.0, le=1.0, alias="MIN_CLUSTER_COVERAGE")
+    # model_unclusterable). Calibrated on mainnet 2026-07-23: the degenerate
+    # stablecoin-validator fits cluster at 0.52 (Strike) and 0.58 (Djed at a 2000
+    # window), both with drift pinned at ~1.0 and a re-fit reproducing the same
+    # noise; a genuinely healthy fit (Djed at a ~1100 window) sits at 0.97. 0.6
+    # falls in the wide empty gap between those two regimes, so it gates the
+    # pathological fits while leaving healthy ones clusterable. Raising it only
+    # treats MORE fits as un-clusterable (fewer re-fits, more honest UI): recall-
+    # safe, since classify + anomaly scoring run every tick regardless.
+    min_cluster_coverage: float = Field(default=0.6, ge=0.0, le=1.0, alias="MIN_CLUSTER_COVERAGE")
 
     # Number of transactions whose (tx, utxos) pairs are fetched concurrently
     # within a page during ingest. Admission is still serialized by the token

services/clustering/backend/tests/test_scheduler.py

@@ -12,9 +12,15 @@
 
 from typing import Any
 
-from app.config import get_settings
+from app.config import Settings, get_settings
 from app.service.scheduler import _decide
 
+# Real mainnet fit coverages observed 2026-07-23 (see config.min_cluster_coverage):
+# the two degenerate stablecoin-validator fits vs a genuinely healthy one.
+_COV_STRIKE_DEGENERATE = 0.522
+_COV_DJED_DEGENERATE = 0.578
+_COV_DJED_HEALTHY = 0.97
+
 _NOW = 1_000_000_000  # fixed clock; tests offset last_fit_at relative to it
 
 
@@ -99,3 +105,24 @@ def test_legacy_unknown_coverage_behaves_as_before() -> None:
     deploy changes nothing until the first fit records a real coverage (self-heal)."""
     c = _contract(drift_score=_high_drift(), fit_coverage=-1.0, last_fit_at=0)
     assert _decide(c, get_settings(), now=_NOW) == ("onboard", 1)
+
+
+def test_default_floor_gates_degenerate_fits_but_not_healthy_ones() -> None:
+    """Calibration lock: the shipped default (0.6) must classify the observed
+    degenerate stablecoin fits (0.52, 0.58) as un-clusterable while leaving a
+    healthy fit (0.97) clusterable. Guards against silently drifting the floor back
+    below the degenerate regime (which would re-enable the futile re-fit loop) or
+    above the healthy regime (which would suppress legitimate re-clusters)."""
+    s = Settings()
+    assert s.model_unclusterable(_COV_STRIKE_DEGENERATE)
+    assert s.model_unclusterable(_COV_DJED_DEGENERATE)
+    assert not s.model_unclusterable(_COV_DJED_HEALTHY)
+
+
+def test_default_floor_stops_strike_style_tight_loop() -> None:
+    """End-to-end at the real Strike numbers: drift pinned at 1.0 with coverage 0.522
+    and a recent fit must fall through to classify, not the ~60s re-fit loop it did
+    before the floor was raised. (It still re-baselines once per slow interval, per
+    test_unclusterable_rebaselines_on_slow_cadence.)"""
+    c = _contract(drift_score=1.0, fit_coverage=_COV_STRIKE_DEGENERATE, last_fit_at=_recent())
+    assert _decide(c, Settings(), now=_NOW) == ("classify", 0)