feat(hour/day derivation): hour classification, monitoring periods, day aggregates, pruning

Hour classification (PR-4):
- Powered-off: POH delta < 90% of wall-clock span
- Active: DUW delta >= 256 MiB
- Idle: powered on + sampled + below active threshold
- Unknown: unsampled without POH evidence
- Four splits sum to exactly wall_clock_seconds
- Disabled time is never an hour state

Monitoring periods (FL-8):
- ensure_period_open: opens period at run moment if none exists
- close_period: closes with end cause
- is_inside_period: checks timestamp against period bounds
- Never backdated; wall-clock outside periods excluded from denominator

Day aggregates (ST-4, PR-5):
- Derived monotonically from hour rows
- UTC-bounded; no 23/25-hour days
- Coverage: known seconds / period wall-clock
- Gap hours inside periods contribute unknown seconds
- Hours outside periods excluded entirely
- No absent hour interpolated/estimated/fabricated (FL-3)

Raw sample pruning (ST-5):
- Prunes samples older than 14 days
- Hour observations and day aggregates retained indefinitely

Closes #22

src/fenris/pruning.py

@@ -0,0 +1,31 @@
+"""Raw sample pruning per spec §3.4, ST-5.
+
+Raw samples are pruned opportunistically to 14 days.
+Hour observations and day aggregates are retained indefinitely.
+"""
+import sqlite3
+from datetime import datetime, timedelta, timezone
+
+# Spec §3.4: Raw-sample retention
+RAW_SAMPLE_RETENTION_DAYS = 14
+
+
+def prune_old_samples(
+    conn: sqlite3.Connection,
+    now: datetime,
+    retention_days: int = RAW_SAMPLE_RETENTION_DAYS,
+) -> int:
+    """Remove raw samples older than retention_days.
+
+    Args:
+        conn: Connection to the observation store.
+        now: Current UTC time.
+        retention_days: Number of days to retain (default 14).
+
+    Returns:
+        Number of samples removed.
+    """
+    cutoff = (now - timedelta(days=retention_days)).isoformat()
+    cursor = conn.execute("DELETE FROM samples WHERE ts < ?", (cutoff,))
+    conn.commit()
+    return cursor.rowcount