Cut XGBoost/LSTM heat predictor from pipeline and demo

Parallels the MI MOS cut. HeatWavePredictor (XGBoost classifier +
ensembled LSTM) was trained on 2 years of pure-synthetic AR(1) noise
using CITY_CLIMATE seasonal curves (heat_forecast.py:331, lstm auto-
retrain at :530). It wrote trigger_probability_7d / prediction_confidence
/ xgb_probability / lstm_probability into the predictions table, but
nothing downstream used those numbers for a decision -- GraphCast drives
trigger firing, burn analysis drives pricing. Classic orphan layer.

Changes:
- Drop HeatWavePredictor import + self._predictor in pipeline.py
- Drop the predictor.predict() call inside _step_predict; the step now
only does GraphCast/observed trigger logic and burn-analysis pricing
- predictions row: NULL for trigger_probability_7d, prediction_confidence,
xgb_probability, lstm_probability. model_tier and ensemble_method
retained (they describe the active forecast source, not the dropped
ensemble). annual_cost_per_worker / payout_factor / learned_frequency
keep coming from burn analysis
- api.py _generate_demo_data: drop HeatWavePredictor, replace with a
composite/100 heuristic for the demo trigger_probability_7d shape

Kept intact:
- UHI XGBoost (uhi_model.py) -- thin wrapper over published UHI
literature formula, values match Dar UHI measurements, low cost
- heat_forecast.py / lstm_model.py files on disk -- not wired from
runtime anymore but left for the WIP Chronos-Bolt retrain work the
user has in models/heat_lstm.pt + scripts/retrain_trigger_heads.py

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/api.py

@@ -27,7 +27,6 @@ from fastapi.responses import HTMLResponse
 from config import ZONES, ZONE_MAP, CITIES, HEAT_THRESHOLDS, PAYOUT_PER_EVENT_USD
 from src.indexing.heat_index import calculate_wbgt, calculate_heat_index, count_consecutive_days, count_trigger_days
 from src.downscaling.uhi_model import UHICorrector
-from src.prediction.heat_forecast import HeatWavePredictor
 from src.pricing.burn_analysis import BurnAnalysisPricer
 from src.pricing.budget_optimizer import BudgetOptimizer
 from src.database.crud import init_db, upsert_zone
@@ -123,7 +122,6 @@ def _generate_demo_data():
 
     # Initialize ML models
     uhi_corrector = UHICorrector()
-    predictor = HeatWavePredictor()
 
     # City base temperatures (ERA5-Land grid-level — before UHI correction)
     city_climate = {
@@ -173,11 +171,6 @@ def _generate_demo_data():
             daily_hi.append(hi)
             daily_uhi_deltas.append(round(uhi_delta, 1))
 
-        # ML heat wave prediction (pass last 30 days for proper anomaly features)
-        pred_prob, pred_conf, pred_tier = predictor.predict(
-            z, daily_temps[-30:], daily_humidity[-30:], daily_wbgt[-30:],
-        )
-
         max_temp = max(daily_temps)
         max_wbgt = max(daily_wbgt)
         recent_temps = daily_temps[-7:]
@@ -208,6 +201,11 @@ def _generate_demo_data():
 
         enrolled = int(z.worker_population_est * rng.uniform(0.15, 0.45))
 
+        # Composite-driven trigger probability for demo shape. Not a model output.
+        pred_prob = round(min(1.0, composite / 100), 2)
+        pred_conf = 0.5
+        pred_tier = "composite_heuristic"
+
         zone_data = {
             "zone_id": z.zone_id,
             "name": z.name,

src/pipeline.py

@@ -27,7 +27,6 @@ from src.healing.healer import HealingAgent, RuleBasedFallback, HealedData
 from src.indexing.heat_risk import compute_heat_risk, HeatTriggerEvent
 from src.indexing.heat_index import calculate_wbgt, calculate_heat_index
 from src.downscaling.uhi_model import UHICorrector
-from src.prediction.heat_forecast import HeatWavePredictor
 from src.calibration.basis_risk import assess_all_zones
 from src.explanation.explainer import TriggerExplainer, TemplateExplainer
 
@@ -102,7 +101,6 @@ class HeatRiskPipeline:
 
         # ML models
         self._uhi_corrector = UHICorrector()
-        self._predictor = HeatWavePredictor()
         self._burn_pricer = None  # initialized lazily in _step_predict
 
         # Load ERA5-Land historical data for padding short climate histories
@@ -587,17 +585,7 @@ class HeatRiskPipeline:
                 heat = self._heat_data.get(zone_id, {})
                 corrected = heat.get("corrected_temps", temps)
 
-                if len(corrected) >= 7:
-                    prob, conf, tier = self._predictor.predict(
-                        zone, corrected[-30:], humidities[-30:], wbgts[-30:]
-                    )
-                else:
-                    prob, conf, tier = 0.1, 0.3, "climatology"
-
                 self._heat_data.setdefault(zone_id, {}).update({
-                    "trigger_probability": prob,
-                    "prediction_confidence": conf,
-                    "model_tier": tier,
                     "temps": corrected,
                     "humidities": humidities,
                     "wbgts": wbgts,
@@ -679,11 +667,13 @@ class HeatRiskPipeline:
                 self._db_write(insert_prediction, self.db, {
                     "zone_id": zone_id,
                     "date": today,
-                    "trigger_probability_7d": heat.get("trigger_probability", 0),
-                    "prediction_confidence": heat.get("prediction_confidence"),
+                    # trigger probability columns stay NULL: the synthetic-trained
+                    # XGBoost/LSTM predictor was cut. GraphCast drives triggers.
+                    "trigger_probability_7d": None,
+                    "prediction_confidence": None,
                     "model_tier": forecast_source,
-                    "xgb_probability": heat.get("trigger_probability"),
-                    "lstm_probability": heat.get("neural_trigger_prob"),
+                    "xgb_probability": None,
+                    "lstm_probability": None,
                     "ensemble_method": "graphcast_forecast" if gc_wbgt else "observed_fallback",
                     "annual_cost_per_worker": heat.get("annual_cost_per_worker"),
                     "payout_factor": heat.get("payout_factor"),