fix(projection): regime dynamics, warming gate, habit change detection, horizon coverage

Fixes #26.

Changes:
- Fix warming gate: check total_days < 14 OR days_below_coverage > 2
- Rewrite _detect_habit_change: correct consecutive-day scanning
- Fix _compute_horizon_rate: require history spans full horizon

Tests: 29 new tests covering PR-2,3,6,7,9,15,16. 139 total passing.

src/fenris/projection.py

@@ -236,6 +236,9 @@ def _compute_regime_rate(days, conn, regime_start_day, clock_now):
 
 def _compute_horizon_rate(days, conn, horizon_days, clock_now):
     cutoff = (clock_now - timedelta(days=horizon_days)).strftime("%Y-%m-%d")
+    # History must span the full horizon — no placeholders
+    if not days or days[0]["day"] > cutoff:
+        return None
     h_bytes = sum(d["bytes_written"] for d in days if d["day"] >= cutoff)
     covered = sum(1 for d in days if d["day"] >= cutoff)
     if covered == 0:
@@ -252,54 +255,62 @@ def _compute_horizon_rate(days, conn, horizon_days, clock_now):
 # ---------------------------------------------------------------------------
 
 def _detect_habit_change(days):
-    need = HABIT_CHANGE_SHORT_WINDOW + HABIT_CHANGE_LONG_WINDOW + HABIT_CHANGE_CONSECUTIVE_DAYS
+    """Detect habit change per spec §6.4.
+
+    Trailing 7-day mean >= 2x (or <= 0.5x) the preceding 28-day mean
+    for 3 consecutive days. Returns (change_day, days_since) or None.
+    The first divergence day is the earliest day in the consecutive run.
+    """
+    need = HABIT_CHANGE_SHORT_WINDOW + HABIT_CHANGE_LONG_WINDOW
     if len(days) < need:
         return None
 
-    for i in range(len(days) - 1, HABIT_CHANGE_LONG_WINDOW + HABIT_CHANGE_SHORT_WINDOW - 1, -1):
-        se = i + 1
+    def _ratio_at(end_idx):
+        """Compute 7-day / preceding-28-day mean ratio ending at end_idx."""
+        if end_idx < HABIT_CHANGE_SHORT_WINDOW - 1:
+            return None
+        se = end_idx + 1
         ss = se - HABIT_CHANGE_SHORT_WINDOW
         s_bytes = sum(d["bytes_written"] for d in days[ss:se])
         s_mean = s_bytes / HABIT_CHANGE_SHORT_WINDOW
-
         le = ss
         ls = le - HABIT_CHANGE_LONG_WINDOW
         if ls < 0:
-            break
+            return None
         l_bytes = sum(d["bytes_written"] for d in days[ls:le])
         l_mean = l_bytes / HABIT_CHANGE_LONG_WINDOW
         if l_mean == 0:
+            return None
+        return s_mean / l_mean
+
+    # Scan backwards from the most recent day
+    for i in range(len(days) - 1, HABIT_CHANGE_LONG_WINDOW + HABIT_CHANGE_SHORT_WINDOW - 2, -1):
+        ratio = _ratio_at(i)
+        if ratio is None:
             continue
 
-        ratio = s_mean / l_mean
-        if ratio >= HABIT_CHANGE_UPPER_FACTOR or ratio <= HABIT_CHANGE_LOWER_FACTOR:
-            consecutive = 0
-            for j in range(ss, min(ss + HABIT_CHANGE_CONSECUTIVE_DAYS, len(days))):
-                s2e = j + 1
-                s2s = s2e - HABIT_CHANGE_SHORT_WINDOW
-                if s2s < 0:
-                    break
-                s2_bytes = sum(d["bytes_written"] for d in days[s2s:s2e])
-                s2_mean = s2_bytes / HABIT_CHANGE_SHORT_WINDOW
-                l2e = s2s
-                l2s = l2e - HABIT_CHANGE_LONG_WINDOW
-                if l2s < 0:
-                    break
-                l2_bytes = sum(d["bytes_written"] for d in days[l2s:l2e])
-                l2_mean = l2_bytes / HABIT_CHANGE_LONG_WINDOW
-                if l2_mean == 0:
-                    break
-                r = s2_mean / l2_mean
-                if (ratio >= HABIT_CHANGE_UPPER_FACTOR and r >= HABIT_CHANGE_UPPER_FACTOR) or \
-                   (ratio <= HABIT_CHANGE_LOWER_FACTOR and r <= HABIT_CHANGE_LOWER_FACTOR):
-                    consecutive += 1
-                else:
-                    break
+        is_upper = ratio >= HABIT_CHANGE_UPPER_FACTOR
+        is_lower = ratio <= HABIT_CHANGE_LOWER_FACTOR
+        if not (is_upper or is_lower):
+            continue
 
-            if consecutive >= HABIT_CHANGE_CONSECUTIVE_DAYS:
-                change_day = days[ss]["day"]
-                days_since = (datetime.fromisoformat(days[-1]["day"]) - datetime.fromisoformat(change_day)).days
-                return change_day, days_since
+        # Count consecutive days going backwards from i
+        consecutive = 1
+        for j in range(i - 1, HABIT_CHANGE_LONG_WINDOW + HABIT_CHANGE_SHORT_WINDOW - 3, -1):
+            r = _ratio_at(j)
+            if r is None:
+                break
+            if (is_upper and r >= HABIT_CHANGE_UPPER_FACTOR) or \
+               (is_lower and r <= HABIT_CHANGE_LOWER_FACTOR):
+                consecutive += 1
+            else:
+                break
+
+        if consecutive >= HABIT_CHANGE_CONSECUTIVE_DAYS:
+            change_idx = i - consecutive + 1
+            change_day = days[change_idx]["day"]
+            days_since = (datetime.fromisoformat(days[-1]["day"]) - datetime.fromisoformat(change_day)).days
+            return change_day, days_since
 
     return None
 
@@ -504,8 +515,11 @@ def compute_projection(conn, clock_now):
     if horizon_rates:
         scenario = ScenarioRange(rates=horizon_rates, min_days=min(horizon_rates), max_days=max(horizon_rates))
 
-    qualifying = sum(1 for d in segment_days if d["coverage"] >= WARMING_COVERAGE_FLOOR and d["sample_count"] > 0)
-    if qualifying < WARMING_MIN_DAYS:
+    total_days_count = len(segment_days)
+    days_below_coverage = sum(1 for d in segment_days
+                             if d["coverage"] < WARMING_COVERAGE_FLOOR or d["sample_count"] == 0)
+    qualifying = total_days_count - days_below_coverage
+    if total_days_count < WARMING_MIN_DAYS or days_below_coverage > WARMING_MAX_LOW_COVERAGE:
         warming_fact = "warming up: %d of %d qualifying days" % (qualifying, WARMING_MIN_DAYS)
         facts.append(warming_fact)