fix: guard local-day pruning across monitoring gaps (#100)

src/fenris/collect.py

@@ -122,6 +122,11 @@ def main() -> None:
 
         if result["ok"]:
             print(f"Collection successful: {result['sample_count']} sample(s)")
+            if result.get("retention_error"):
+                print(
+                    f"Retention deferred: {result['retention_error']}",
+                    file=sys.stderr,
+                )
             sys.exit(0)
         else:
             print(f"Collection failed: {result['error']}", file=sys.stderr)

src/fenris/collector.py

@@ -477,9 +477,21 @@ def run_collection(
             break
 
         published_count += _recover_pending_for_collection(conn)
+        retention_error = None
+        try:
+            if _pending_count(conn) == 0:
+                from .pruning import prune_old_samples
+
+                prune_old_samples(conn, clock.utcnow())
+        except Exception as exc:  # noqa: BLE001 - publication already committed.
+            # Publication already committed. Keep collection successful and
+            # retry atomic retention on the next normal collection.
+            conn.rollback()
+            retention_error = str(exc)
         return {
             "ok": True,
             "sample_count": published_count,
+            "retention_error": retention_error,
             "store_path": str(store_path),
         }
 

src/fenris/pruning.py

@@ -1,131 +1,332 @@
-"""Raw sample pruning per spec §3.4, ST-5.
+"""Expire detail only after UTC and local_days replacement evidence is durable.
 
-Raw samples are pruned opportunistically to 14 days.
-Hour observations and day aggregates are retained indefinitely.
-Boundary anchors required for successor evidence are retained.
-
-Local-day summaries (local_days table) are never touched by pruning.
-They are persisted at collection time from hour observations and survive
-raw-sample pruning because they depend on hour observations, not on raw
-samples.  This is the mechanism that keeps local-day history trustworthy
-after detail expires (issue #93).
+The 14-day cutoff never overrides local-day preservation, shared boundary
+evidence, publication recovery, or the newest sample's successor-anchor role.
 """
 import sqlite3
-from datetime import datetime, timedelta, timezone
+from datetime import date, datetime, timedelta, timezone
 
-# Spec §3.4: Raw-sample retention
 RAW_SAMPLE_RETENTION_DAYS = 14
 
 
+def _utc_datetime(value: datetime) -> datetime:
+    """Normalize a clock value to aware UTC."""
+    if value.tzinfo is None:
+        value = value.replace(tzinfo=timezone.utc)
+    return value.astimezone(timezone.utc)
+
+
+def _parse_sample_time(value: str) -> datetime | None:
+    """Parse old timestamps conservatively; malformed legacy values stay."""
+    try:
+        parsed = datetime.fromisoformat(value)
+    except (TypeError, ValueError):
+        return None
+    return _utc_datetime(parsed)
+
+
+def _sample_rows(conn: sqlite3.Connection) -> list[tuple]:
+    return conn.execute(
+        "SELECT id, ts, segment_id, bytes_written, bytes_read, local_tz "
+        "FROM samples ORDER BY id"
+    ).fetchall()
+
+
+def _required_hour_window(
+    start: datetime,
+    end: datetime,
+) -> tuple[datetime, datetime, int] | None:
+    """Return first hour, exclusive end and count for [start, end)."""
+    if end <= start:
+        return None
+    first_hour = start.replace(minute=0, second=0, microsecond=0)
+    last_hour = end.replace(minute=0, second=0, microsecond=0)
+    try:
+        exclusive_end = last_hour if end == last_hour else last_hour + timedelta(hours=1)
+    except OverflowError:
+        return None
+    count = int((exclusive_end - first_hour).total_seconds() // 3600)
+    return first_hour, exclusive_end, count
+
+
+def _has_utc_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    window = _required_hour_window(start, end)
+    if window is None:
+        return False
+    first_hour, exclusive_end, expected_hour_count = window
+
+    actual_hour_count = conn.execute(
+        "SELECT COUNT(*) FROM hour_observations WHERE hour >= ? AND hour < ?",
+        (
+            first_hour.strftime("%Y-%m-%dT%H:00:00+00:00"),
+            exclusive_end.strftime("%Y-%m-%dT%H:00:00+00:00"),
+        ),
+    ).fetchone()[0]
+    if actual_hour_count < expected_hour_count:
+        return False
+
+    start_written, end_written = previous[3], current[3]
+    start_read, end_read = previous[4], current[4]
+    if None in (start_written, end_written, start_read, end_read):
+        return False
+    if end_written < start_written or end_read < start_read:
+        return False
+    bytes_written = end_written - start_written
+    bytes_read = end_read - start_read
+    if expected_hour_count == 1:
+        totals = conn.execute(
+            "SELECT bytes_written_delta, bytes_read_delta FROM hour_observations "
+            "WHERE hour = ?",
+            (first_hour.strftime("%Y-%m-%dT%H:00:00+00:00"),),
+        ).fetchone()
+    else:
+        totals = conn.execute(
+            "SELECT unattributed_bytes_written, unattributed_bytes_read "
+            "FROM day_aggregates WHERE day = ?",
+            (start.date().isoformat(),),
+        ).fetchone()
+    if totals is None or totals[0] < bytes_written or totals[1] < bytes_read:
+        return False
+
+    last_day = (exclusive_end - timedelta(microseconds=1)).date()
+    expected_day_count = (last_day - first_hour.date()).days + 1
+    actual_day_count = conn.execute(
+        "SELECT COUNT(*) FROM day_aggregates WHERE day >= ? AND day <= ?",
+        (first_hour.date().isoformat(), last_day.isoformat()),
+    ).fetchone()[0]
+    return actual_day_count >= expected_day_count
+
+
+def _local_day_row_exists(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+) -> bool:
+    return conn.execute(
+        "SELECT 1 FROM local_days WHERE local_date = ? AND tz_name = ?",
+        (local_date, tz_name),
+    ).fetchone() is not None
+
+
+def _local_day_bounds(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+) -> tuple[datetime, datetime] | None:
+    row = conn.execute(
+        "SELECT utc_start, utc_end FROM local_days "
+        "WHERE local_date = ? AND tz_name = ?",
+        (local_date, tz_name),
+    ).fetchone()
+    if row is None:
+        return None
+    start = _parse_sample_time(row[0])
+    end = _parse_sample_time(row[1])
+    if start is None or end is None or end <= start:
+        return None
+    return start, end
+
+
+def _contains_instant(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+    instant: datetime,
+) -> bool:
+    bounds = _local_day_bounds(conn, local_date, tz_name)
+    return bounds is not None and bounds[0] <= instant < bounds[1]
+
+
+def _monitoring_period_covers(
+    conn: sqlite3.Connection,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    """Require continuous monitoring before a local-day total replaces detail."""
+    for started_at, ended_at in conn.execute(
+        "SELECT started_at, ended_at FROM monitoring_periods"
+    ):
+        period_start = _parse_sample_time(started_at)
+        period_end = _parse_sample_time(ended_at) if ended_at is not None else None
+        if period_start is None or period_start > start:
+            continue
+        if ended_at is None or (period_end is not None and end <= period_end):
+            return True
+    return False
+
+
+def _has_local_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    start_tz = previous[5]
+    end_tz = current[5]
+    if not end_tz:
+        return False
+
+    start_id, end_id = previous[0], current[0]
+    segment_id = current[2]
+    if (
+        start_tz
+        and start_tz == end_tz
+        and segment_id is not None
+        and _monitoring_period_covers(conn, start, end)
+    ):
+        known_days = conn.execute(
+            "SELECT local_days.utc_start, local_days.utc_end "
+            "FROM local_days JOIN local_day_segment_totals "
+            "  ON local_day_segment_totals.local_day_id = local_days.id "
+            "WHERE local_days.tz_name = ? "
+            "  AND local_days.activity_precision IN ('measured', 'coarse') "
+            "  AND local_days.last_sample_id >= ? "
+            "  AND local_days.activity_intervals > 0 "
+            "  AND local_day_segment_totals.segment_id = ? "
+            "  AND local_day_segment_totals.activity_intervals > 0",
+            (end_tz, end_id, segment_id),
+        ).fetchall()
+        for row in known_days:
+            bounds_start = _parse_sample_time(row[0])
+            bounds_end = _parse_sample_time(row[1])
+            if (
+                bounds_start is not None
+                and bounds_end is not None
+                and bounds_start <= start
+                and end < bounds_end
+            ):
+                return True
+
+    evidence = conn.execute(
+        "SELECT bytes_written, bytes_read, reason, start_local_date, end_local_date, "
+        "       start_tz_name, end_tz_name, started_at, ended_at "
+        "FROM local_day_unallocated_evidence "
+        "WHERE start_sample_id = ? AND end_sample_id = ?",
+        (start_id, end_id),
+    ).fetchone()
+    if evidence is None:
+        return False
+
+    start_written, end_written = previous[3], current[3]
+    start_read, end_read = previous[4], current[4]
+    if None in (start_written, end_written, start_read, end_read):
+        return False
+    if end_written < start_written or end_read < start_read:
+        return False
+    # The source IDs uniquely identify this interval. Confirm its durable
+    # deltas and retain every recorded local-day boundary it touches.
+    if (end_written - start_written, end_read - start_read) != evidence[:2]:
+        return False
+    if evidence[5] != start_tz or evidence[6] != end_tz:
+        return False
+    if evidence[2] not in {
+        "counter_discontinuity",
+        "legacy_timezone_unknown",
+        "timezone_change",
+        "monitoring_period",
+        "local_midnight",
+        "segment_unknown",
+    }:
+        return False
+    if _parse_sample_time(evidence[7]) != start or _parse_sample_time(evidence[8]) != end:
+        return False
+
+    affected_days: set[tuple[str, str]] = set()
+    if start_tz:
+        affected_days.add((evidence[3], start_tz))
+    affected_days.add((evidence[4], end_tz))
+    if start_tz and start_tz == end_tz and evidence[3] < evidence[4]:
+        day = date.fromisoformat(evidence[3]) + timedelta(days=1)
+        last_day = date.fromisoformat(evidence[4])
+        while day < last_day:
+            affected_days.add((day.isoformat(), end_tz))
+            day += timedelta(days=1)
+    if not all(
+        _local_day_row_exists(conn, local_date, zone)
+        for local_date, zone in affected_days
+    ):
+        return False
+    if start_tz and not _contains_instant(conn, evidence[3], start_tz, start):
+        return False
+    return _contains_instant(conn, evidence[4], end_tz, end)
+
+
+def _interval_has_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+) -> bool:
+    start = _parse_sample_time(previous[1])
+    end = _parse_sample_time(current[1])
+    if start is None or end is None or end <= start:
+        return False
+    try:
+        return (
+            _has_utc_replacement(conn, previous, current, start, end)
+            and _has_local_replacement(conn, previous, current, start, end)
+        )
+    except (sqlite3.Error, ValueError, OverflowError, TypeError):
+        # Missing or unreadable derived evidence never authorizes deletion.
+        return False
+
+
+def _sample_needs_preservation(
+    conn: sqlite3.Connection,
+    rows: list[tuple],
+    index: int,
+    newest_id: int,
+) -> bool:
+    sample = rows[index]
+    segment_id = sample[2]
+    if segment_id is None or conn.execute(
+        "SELECT 1 FROM controller_segments WHERE id = ?",
+        (segment_id,),
+    ).fetchone() is None:
+        return True
+    if sample[0] == newest_id:
+        # Keep newest sample as the source anchor for the next collection.
+        return True
+
+    pairs = []
+    if index > 0 and rows[index - 1][2] == segment_id:
+        pairs.append((rows[index - 1], sample))
+    if index + 1 < len(rows) and rows[index + 1][2] == segment_id:
+        pairs.append((sample, rows[index + 1]))
+    return any(
+        not _interval_has_replacement(conn, previous, current)
+        for previous, current in pairs
+    )
+
+
 def needs_boundary_anchor(
     conn: sqlite3.Connection,
     sample_ts: str,
     now: datetime,
 ) -> bool:
-    """Check if a sample is needed as a boundary anchor for derivation.
-
-    A sample is a boundary anchor if:
-    1. It's older than retention_days (strictly before cutoff)
-    2. It has a next sample that forms an interval spanning the retention boundary
-    3. The interval hasn't been derived yet
-
-    The interval spans the boundary if:
-    - The sample is before the cutoff, AND
-    - The next sample is strictly after the cutoff (or within retention)
-    """
-    from .derive import _parse_ts
-    
-    sample_dt = _parse_ts(sample_ts)
-    retention_cutoff = now - timedelta(days=RAW_SAMPLE_RETENTION_DAYS)
-    
-    # If sample is within retention (strictly after cutoff), not an anchor
-    if sample_dt > retention_cutoff:
+    """Return whether an old sample still carries unreplaced evidence."""
+    sample_time = _parse_sample_time(sample_ts)
+    cutoff = _utc_datetime(now) - timedelta(days=RAW_SAMPLE_RETENTION_DAYS)
+    if sample_time is None:
+        return True
+    if sample_time >= cutoff:
         return False
-    
-    # Check if this sample has a next sample
-    cursor = conn.execute(
-        """SELECT ts, segment_id FROM samples WHERE ts > ? ORDER BY ts LIMIT 1""",
-        (sample_ts,),
+
+    rows = _sample_rows(conn)
+    matching = [index for index, row in enumerate(rows) if row[1] == sample_ts]
+    if not matching:
+        return False
+    newest_id = rows[-1][0] if rows else -1
+    return any(
+        _sample_needs_preservation(conn, rows, index, newest_id)
+        for index in matching
     )
-    next_row = cursor.fetchone()
-    
-    if next_row is None:
-        # No next sample - this is the last sample
-        # It's not needed for derivation (no interval to derive)
-        return False
-    
-    next_ts_str = next_row[0]
-    next_segment_id = next_row[1]
-    next_dt = _parse_ts(next_ts_str)
-    
-    # Check if the next sample is strictly after the cutoff (i.e., interval spans boundary)
-    if next_dt > retention_cutoff:
-        # The interval spans the retention boundary
-        # Check if the interval needs derivation
-        # Get current sample's segment_id
-        cursor = conn.execute(
-            "SELECT segment_id FROM samples WHERE ts = ?",
-            (sample_ts,),
-        )
-        current_segment_row = cursor.fetchone()
-        current_segment_id = current_segment_row[0] if current_segment_row else None
-        
-        # If different segments, no interval to derive
-        if current_segment_id != next_segment_id:
-            return False
-        
-        # Check if the interval [sample_ts, next_ts] needs derivation
-        # It needs derivation if any hour in the span lacks an observation
-        current_hour = sample_dt.replace(minute=0, second=0, microsecond=0)
-        end_hour = next_dt.replace(minute=0, second=0, microsecond=0)
-        
-        while current_hour <= end_hour:
-            cursor = conn.execute(
-                "SELECT id FROM hour_observations WHERE hour = ?",
-                (current_hour.isoformat(),),
-            )
-            if cursor.fetchone() is None:
-                # This hour lacks an observation - interval needs derivation
-                return True
-            current_hour += timedelta(hours=1)
-        
-        # All hours in the span have observations - interval is derived
-        return False
-    else:
-        # The interval doesn't span the boundary (both samples are old)
-        # Check if the interval needs derivation
-        # Get current sample's segment_id
-        cursor = conn.execute(
-            "SELECT segment_id FROM samples WHERE ts = ?",
-            (sample_ts,),
-        )
-        current_segment_row = cursor.fetchone()
-        current_segment_id = current_segment_row[0] if current_segment_row else None
-        
-        # If different segments, no interval to derive
-        if current_segment_id != next_segment_id:
-            return False
-        
-        # Check if the interval [sample_ts, next_ts] needs derivation
-        current_hour = sample_dt.replace(minute=0, second=0, microsecond=0)
-        end_hour = next_dt.replace(minute=0, second=0, microsecond=0)
-        
-        while current_hour <= end_hour:
-            cursor = conn.execute(
-                "SELECT id FROM hour_observations WHERE hour = ?",
-                (current_hour.isoformat(),),
-            )
-            if cursor.fetchone() is None:
-                # This hour lacks an observation - interval needs derivation
-                # But only keep if the interval is significant (spans multiple hours)
-                # or if the next sample is the last sample before a gap
-                gap = (next_dt - sample_dt).total_seconds()
-                if gap > 24 * 3600:  # Significant gap (> 24 hours)
-                    return True
-            current_hour += timedelta(hours=1)
-        
-        # All hours in the span have observations or gap is not significant
-        return False
 
 
 def prune_old_samples(
@@ -133,36 +334,50 @@ def prune_old_samples(
     now: datetime,
     retention_days: int = RAW_SAMPLE_RETENTION_DAYS,
 ) -> int:
-    """Remove raw samples older than retention_days.
+    """Delete expired samples only after every dependent fact is durable.
 
-    Retains boundary anchors required for successor evidence.
-
-    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.
+    Pending publications are stored separately and never selected here. A
+    sample stays when it is the newest successor anchor, has an unpublishable
+    neighbour interval, lacks UTC or local-day replacement evidence, or has a
+    timestamp that cannot be safely interpreted.
     """
-    cutoff = (now - timedelta(days=retention_days)).isoformat()
-    
-    # Get all samples older than cutoff
-    cursor = conn.execute(
-        "SELECT id, ts FROM samples WHERE ts < ? ORDER BY ts",
-        (cutoff,),
-    )
-    old_samples = cursor.fetchall()
-    
-    removed = 0
-    for sample_id, sample_ts in old_samples:
-        # Check if this sample is a boundary anchor
-        if needs_boundary_anchor(conn, sample_ts, now):
-            continue  # Skip - it's a boundary anchor
-        
-        # Remove the sample
-        conn.execute("DELETE FROM samples WHERE id = ?", (sample_id,))
-        removed += 1
-    
-    conn.commit()
-    return removed
+    cutoff = _utc_datetime(now) - timedelta(days=retention_days)
+    owns_transaction = not conn.in_transaction
+    savepoint = "fenris_sample_retention"
+    if owns_transaction:
+        conn.execute("BEGIN IMMEDIATE")
+    else:
+        conn.execute(f"SAVEPOINT {savepoint}")
+
+    try:
+        rows = _sample_rows(conn)
+        if not rows:
+            if owns_transaction:
+                conn.commit()
+            else:
+                conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+            return 0
+
+        newest_id = rows[-1][0]
+        expired = []
+        for index, row in enumerate(rows):
+            sample_time = _parse_sample_time(row[1])
+            if sample_time is None or sample_time >= cutoff:
+                continue
+            if _sample_needs_preservation(conn, rows, index, newest_id):
+                continue
+            expired.append(row[0])
+
+        conn.executemany("DELETE FROM samples WHERE id = ?", ((sample_id,) for sample_id in expired))
+        if owns_transaction:
+            conn.commit()
+        else:
+            conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+        return len(expired)
+    except Exception:
+        if owns_transaction:
+            conn.rollback()
+        else:
+            conn.execute(f"ROLLBACK TO SAVEPOINT {savepoint}")
+            conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+        raise

tests/test_issue_100.py

@@ -0,0 +1,341 @@
+"""Normal-collection retention acceptance tests for issue #100."""
+import sqlite3
+import sys
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris.collector import run_collection
+from fenris.local_day import query_local_day_summary
+from fenris.pruning import prune_old_samples
+from fenris.status import read_status
+from fenris.store import init_store
+
+
+@pytest.fixture
+def sysfs_controller(tmp_path):
+    controller = tmp_path / "sys" / "class" / "nvme" / "nvme0"
+    controller.mkdir(parents=True)
+    (controller / "subsysnqn").write_text("nqn.example:drive-1\n")
+    (controller / "model").write_text("Fenris Test Drive\n")
+    (controller / "serial").write_text("drive-1\n")
+    (controller / "firmware_rev").write_text("1.0\n")
+    transport = controller / "transport"
+    transport.mkdir()
+    (transport / "address").write_text("0000:00:01.0\n")
+    (transport / "trstring").write_text("pcie\n")
+    return controller
+
+
+class FakeClock:
+    def __init__(self, now):
+        self.now = now
+
+    def utcnow(self):
+        return self.now
+
+
+def _smartctl(written, read):
+    return {
+        "nvme_smart_health_information_log": {
+            "critical_warning": 0,
+            "temperature": 35,
+            "available_spare": 100,
+            "percentage_used": 1,
+            "data_units_written": written,
+            "data_units_read": read,
+            "power_on_hours": 100,
+            "power_cycles": 1,
+            "unsafe_shutdowns": 0,
+            "media_errors": 0,
+        },
+        "user_capacity": {"bytes": 1_000_000_000_000},
+        "model_name": "Fenris Test Drive",
+        "serial_number": "drive-1",
+        "firmware_version": "1.0",
+    }
+
+
+def _collect(config, controller, now, written, read):
+    return run_collection(
+        smartctl_data=_smartctl(written, read),
+        sysfs_path=controller,
+        config=config,
+        clock=FakeClock(now),
+    )
+
+
+def test_normal_collection_expires_detail_and_keeps_local_day_evidence(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+
+    first = _collect(config, sysfs_controller, first_time, 1000, 2000)
+    assert first["ok"] is True, first
+    second = _collect(config, sysfs_controller, second_time, 1001, 2003)
+    assert second["ok"] is True, second
+
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (second_time.isoformat(),)
+        ]
+        assert conn.execute(
+            "SELECT bytes_written, bytes_read, reason "
+            "FROM local_day_unallocated_evidence"
+        ).fetchall() == [(512_000, 1_536_000, "local_midnight")]
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM controller_segments").fetchone()[0] == 1
+        assert conn.execute("SELECT COUNT(*) FROM day_aggregates").fetchone()[0] == 15
+
+    monkeypatch.setenv("TZ", "Asia/Kolkata")
+    with read_status(store_path, second_time, query_services=False) as (reader, state):
+        assert reader is not None
+        assert state.store_fault is None
+        old_day = query_local_day_summary(
+            reader, "2026-09-01", "UTC", second_time,
+        )
+        assert old_day["utc_start"] == "2026-09-01T00:00:00+00:00"
+        assert old_day["utc_end"] == "2026-09-02T00:00:00+00:00"
+        assert old_day["bytes_written"] is None
+        assert old_day["shared_bytes_written"] == 512_000
+        assert old_day["shared_bytes_read"] == 1_536_000
+        assert old_day["shared_evidence_count"] == 1
+
+
+def test_pruning_keeps_samples_when_replacement_evidence_is_missing(tmp_path):
+    store_path = tmp_path / "incomplete.db"
+    conn = init_store(store_path)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    old_time = now - timedelta(days=20)
+    next_time = now - timedelta(days=1)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 100, 100, 1, 'UTC')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 200, 200, 1, 'UTC')",
+        (next_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 2
+    conn.close()
+
+
+def test_later_same_day_activity_does_not_replace_gap_interval(tmp_path):
+    store_path = tmp_path / "unrelated-local-activity.db"
+    conn = init_store(store_path)
+    gap_start = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    gap_end = gap_start + timedelta(minutes=10)
+    later_end = gap_start + timedelta(minutes=20)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (gap_start.isoformat(),),
+    )
+    for sample_time, written, read in (
+        (gap_start, 1000, 100),
+        (gap_end, 2000, 200),
+        (later_end, 2100, 210),
+    ):
+        conn.execute(
+            "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+            "VALUES (?, '/dev/nvme0', ?, ?, 1, 'UTC')",
+            (sample_time.isoformat(), written, read),
+        )
+
+    # UTC aggregates can replace the old counter pair, but the local-day
+    # summary contains only the later B->C interval, after monitoring resumed.
+    conn.execute(
+        "INSERT INTO hour_observations (hour, bytes_written_delta, bytes_read_delta) "
+        "VALUES ('2026-09-01T00:00:00+00:00', 1000, 100)"
+    )
+    conn.execute(
+        "INSERT INTO day_aggregates (day) VALUES ('2026-09-01')"
+    )
+    conn.execute(
+        "INSERT INTO monitoring_periods (started_at, ended_at) VALUES (?, ?)",
+        (gap_end.isoformat(), later_end.isoformat()),
+    )
+    last_sample_id = conn.execute("SELECT MAX(id) FROM samples").fetchone()[0]
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, activity_intervals, activity_precision, last_sample_id) "
+        "VALUES ('2026-09-01', 'UTC', '+00:00', ?, ?, 100, 10, 1, 'measured', ?)",
+        (
+            gap_start.isoformat(),
+            (gap_start + timedelta(days=1)).isoformat(),
+            last_sample_id,
+        ),
+    )
+    local_day_id = conn.execute("SELECT MAX(id) FROM local_days").fetchone()[0]
+    conn.execute(
+        "INSERT INTO local_day_segment_totals "
+        "(local_day_id, segment_id, bytes_written, bytes_read, activity_intervals) "
+        "VALUES (?, 1, 100, 10, 1)",
+        (local_day_id,),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 3
+    conn.close()
+
+
+def test_pruning_keeps_oldest_unpaired_sample_as_successor_anchor(tmp_path):
+    store_path = tmp_path / "anchor.db"
+    conn = init_store(store_path)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    old_time = now - timedelta(days=20)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 100, 100, 1, 'UTC')",
+        (old_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT ts FROM samples").fetchone()[0] == old_time.isoformat()
+    conn.close()
+
+
+def test_pruning_never_expires_pending_publications(tmp_path):
+    store_path = tmp_path / "pending.db"
+    conn = init_store(store_path)
+    old_time = datetime(2026, 9, 1, tzinfo=timezone.utc)
+    conn.execute(
+        "INSERT INTO pending_publications (sample_ts, payload) VALUES (?, '{}')",
+        (old_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(
+        conn, datetime(2026, 9, 30, tzinfo=timezone.utc),
+    ) == 0
+    assert conn.execute(
+        "SELECT sample_ts FROM pending_publications"
+    ).fetchone()[0] == old_time.isoformat()
+    conn.close()
+
+
+def test_pruning_preserves_unparseable_legacy_timestamp(tmp_path):
+    store_path = tmp_path / "legacy.db"
+    conn = init_store(store_path)
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read) "
+        "VALUES ('legacy timestamp', '/dev/nvme0', 100, 100)"
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, datetime(2026, 9, 30, tzinfo=timezone.utc)) == 0
+    assert conn.execute("SELECT ts FROM samples").fetchone()[0] == "legacy timestamp"
+    conn.close()
+
+
+def test_collection_defers_interrupted_pruning_and_retries_next_run(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "interrupted.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+    assert _collect(config, sysfs_controller, first_time, 1000, 2000)["ok"] is True
+
+    writer = sqlite3.connect(store_path)
+    writer.execute(
+        "CREATE TRIGGER stop_retention BEFORE DELETE ON samples "
+        "BEGIN SELECT RAISE(ABORT, 'simulated retention interruption'); END"
+    )
+    writer.commit()
+    writer.close()
+
+    second = _collect(config, sysfs_controller, second_time, 1001, 2001)
+    assert second["ok"] is True, second
+    assert "simulated retention interruption" in second["retention_error"]
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 2
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM local_day_unallocated_evidence").fetchone()[0] == 1
+
+    writer = sqlite3.connect(store_path)
+    writer.execute("DROP TRIGGER stop_retention")
+    writer.commit()
+    writer.close()
+
+    third_time = second_time + timedelta(days=15)
+    third = _collect(config, sysfs_controller, third_time, 1002, 2002)
+    assert third["ok"] is True, third
+    assert third["retention_error"] is None
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (third_time.isoformat(),)
+        ]
+
+
+def test_collection_recovers_old_pending_work_before_pruning(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "pending-recovery.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+    assert _collect(config, sysfs_controller, first_time, 1000, 2000)["ok"] is True
+
+    writer = sqlite3.connect(store_path)
+    writer.execute(
+        "CREATE TRIGGER interrupt_publication "
+        "BEFORE INSERT ON local_day_unallocated_evidence "
+        "BEGIN SELECT RAISE(ABORT, 'simulated publication interruption'); END"
+    )
+    writer.commit()
+    writer.close()
+
+    failed = _collect(config, sysfs_controller, second_time, 1001, 2001)
+    assert failed["ok"] is False
+    assert "simulated publication interruption" in failed["error"]
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 1
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 1
+
+    writer = sqlite3.connect(store_path)
+    writer.execute("DROP TRIGGER interrupt_publication")
+    writer.commit()
+    writer.close()
+
+    third_time = second_time + timedelta(minutes=5)
+    recovered = _collect(config, sysfs_controller, third_time, 1002, 2002)
+    assert recovered["ok"] is True, recovered
+    assert recovered["retention_error"] is None
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (second_time.isoformat(),),
+            (third_time.isoformat(),),
+        ]
+        assert conn.execute(
+            "SELECT COUNT(*) FROM local_day_unallocated_evidence"
+        ).fetchone()[0] == 1

tests/test_issue_93.py

@@ -151,8 +151,8 @@ class TestDetailExpiresSummariesRemain:
         # Run pruning
         pruned = prune_old_samples(conn, now, retention_days=14)
 
-        # Old samples should be pruned, recent ones retained
-        assert pruned >= 1
+        # The handcrafted totals lack replacement provenance, so retain samples.
+        assert pruned == 0
 
         # Local-day summaries must still be queryable
         old_result = query_local_day_summary(conn, "2026-09-16")
@@ -434,10 +434,10 @@ class TestExistingEntryPoints:
             _insert_sample(conn, ts, bw=days_ago * 100)
 
         pruned = prune_old_samples(conn, now, retention_days=14)
-        assert pruned == 15
+        assert pruned == 0
 
         cursor = conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 14
+        assert cursor.fetchone()[0] == 29
         conn.close()
 
     def test_repair_with_real_temp_store(self, tmp_path):

tests/test_pruning.py

@@ -27,9 +27,14 @@ def store_conn(tmp_path: Path):
 
 
 def _insert_sample(conn, ts_iso, device="/dev/nvme0"):
+    conn.execute(
+        "INSERT OR IGNORE INTO controller_segments "
+        "(id, opened_at, identity_key) VALUES (1, '2026-01-01T00:00:00+00:00', 'test')"
+    )
     conn.execute(
         "INSERT INTO samples (ts, device, data_units_written, data_units_read, "
-        " bytes_written, bytes_read, percentage_used) VALUES (?, ?, 0, 0, 0, 0, 0)",
+        " bytes_written, bytes_read, percentage_used, segment_id, local_tz) "
+        "VALUES (?, ?, 0, 0, 0, 0, 0, 1, 'UTC')",
         (ts_iso, device),
     )
     conn.commit()
@@ -50,33 +55,30 @@ class TestPruneOldSamples:
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
         assert cursor.fetchone()[0] == 1
 
-    def test_removes_old_samples(self, store_conn):
+    def test_keeps_old_samples_without_replacement_evidence(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
-        # Insert samples at 10, 14, and 15 days ago
-        # All three are before the cutoff (2026-09-01T12:00:00)
-        # The 15-day-old sample is not a boundary anchor because
-        # the next sample (14 days ago) is also before the cutoff
+        # Old samples stay until UTC and local-day evidence replace them.
         for days_ago in [10, 14, 15]:
             ts = (now - timedelta(days=days_ago)).isoformat()
             _insert_sample(store_conn, ts)
 
         pruned = prune_old_samples(store_conn, now, retention_days=14)
-        assert pruned == 1  # Only the 15-day-old sample removed
+        assert pruned == 0
 
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 2
+        assert cursor.fetchone()[0] == 3
 
-    def test_removes_many_old_samples(self, store_conn):
+    def test_keeps_many_old_samples_without_replacement_evidence(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
         for days_ago in range(1, 30):
             ts = (now - timedelta(days=days_ago)).isoformat()
             _insert_sample(store_conn, ts)
 
         pruned = prune_old_samples(store_conn, now, retention_days=14)
-        assert pruned == 15  # Days 15-29 removed
+        assert pruned == 0
 
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 14  # Days 1-14 kept
+        assert cursor.fetchone()[0] == 29
 
     def test_empty_store_no_error(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
@@ -130,8 +132,8 @@ class TestPruneOldSamples:
         )
         assert cursor.fetchone()[0] == 1
 
-    def test_old_sample_with_derived_interval_removed(self, store_conn):
-        """Old samples with fully derived intervals are removed."""
+    def test_hour_rows_alone_do_not_replace_local_day_evidence(self, store_conn):
+        """An hour row alone cannot authorize source-sample deletion."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Old sample with derived interval
@@ -148,8 +150,8 @@ class TestPruneOldSamples:
         # Run pruning
         pruned = prune_old_samples(store_conn, now, retention_days=14)
         
-        # Old sample should be removed (interval is derived)
+        # Keep the source samples until local-day replacement evidence exists.
         cursor = store_conn.execute(
             "SELECT COUNT(*) FROM samples WHERE ts = '2026-09-10T10:00:00+00:00'"
         )
-        assert cursor.fetchone()[0] == 0
+        assert cursor.fetchone()[0] == 1

tests/test_repair.py

@@ -235,8 +235,8 @@ class TestBoundaryAnchorRetention:
         # the interval spans the retention boundary
         assert needs_boundary_anchor(store_conn, "2026-09-15T23:55:00+00:00", now)
 
-    def test_not_boundary_anchor_if_fully_derived(self, store_conn):
-        """Sample that's fully derived is not a boundary anchor."""
+    def test_hour_only_derivation_does_not_replace_local_day_evidence(self, store_conn):
+        """UTC hour evidence alone does not make a sample safe to prune."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Insert sample and fully derive its interval
@@ -247,8 +247,7 @@ class TestBoundaryAnchorRetention:
         _insert_hour(store_conn, "2026-09-14T10:00:00+00:00", 
                      bytes_written_delta=1000000)
         
-        # Not a boundary anchor
-        assert not needs_boundary_anchor(store_conn, "2026-09-14T10:00:00+00:00", now)
+        assert needs_boundary_anchor(store_conn, "2026-09-14T10:00:00+00:00", now)
 
     def test_pruning_retains_boundary_anchors(self, store_conn):
         """Pruning keeps samples needed as boundary anchors."""
@@ -274,8 +273,8 @@ class TestBoundaryAnchorRetention:
         )
         assert cursor.fetchone()[0] == 1
 
-    def test_pruning_removes_old_sample_with_derived_interval(self, store_conn):
-        """Pruning removes old samples when interval is fully derived."""
+    def test_pruning_keeps_samples_without_local_day_replacement(self, store_conn):
+        """Pruning keeps samples when only the UTC-hour row exists."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Old sample with derived interval
@@ -289,11 +288,11 @@ class TestBoundaryAnchorRetention:
         # Run pruning
         pruned = prune_old_samples(store_conn, now, retention_days=14)
         
-        # Old sample should be removed (interval is derived)
+        # The source remains until local-day evidence is also durable.
         cursor = store_conn.execute(
             "SELECT COUNT(*) FROM samples WHERE ts = '2026-09-10T10:00:00+00:00'"
         )
-        assert cursor.fetchone()[0] == 0
+        assert cursor.fetchone()[0] == 1
 
 
 # ---------------------------------------------------------------------------