navi-offroute: numpy-pack TrailheadIndex (memory opt #2) (#32)

Apply the proven OSMParkingIndex (PR #31) memory-pack pattern to TrailheadIndex,
which the 2026-05-26 memory audit flagged at ~300-450 MB/worker using the old
list[dict]+list[Point] storage.

- mvum_transitions.py: store coords as packed float64 numpy arrays (_lats/_lons)
  and attributes as interned lists (_names/_road_classes); build candidate record
  dicts lazily via _record(i) in query_trailheads_near_line instead of holding
  740k dicts + 740k shapely Point objects. Points are built only to construct the
  STRtree, then released. Adds a records property (lazy, for tests/introspection)
  and tracks build_time_seconds + memory_estimate_mb (psutil RSS delta) like
  OSMParkingIndex. Query logic (coarse STRtree bbox + precise degree-distance
  check) unchanged.
- admin.py: GET /api/admin/trailhead/info -> {count, build_time_seconds,
  memory_estimate_mb}, mirroring /api/admin/osm-parking/info.

Tests: existing test updated (len(records)==count; the removed _points assertion)
plus a numpy-backing test (_lats/_lons dtype float64, len==count). Full offroute
suite: 83 passed.

Real-DB sanity (not deployed): index loads 740,430 entry points in ~4.8 s using
~285 MB RSS (down from the audit's inferred ~300-450 MB; same ~40% pack ratio as
parking), query returns 317 trailheads on a Redfish Lake corridor.

Co-authored-by: Matt <mj@k7zvx.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/services/navi_offroute/admin.py

@@ -163,3 +163,17 @@ def osm_parking_info():
         'build_time_seconds': round(idx.build_time_seconds, 3),
         'memory_estimate_mb': round(idx.memory_estimate_mb, 1),
     })
+
+
+@bp.route('/api/admin/trailhead/info')
+@require_auth
+def trailhead_info():
+    """Read-only stats for the in-memory trailhead index (Layer 3a)."""
+    idx = current_app.config.get('MVUM_TRAILHEAD_INDEX')
+    if idx is None:
+        return jsonify({'status': 'error', 'message': 'trailhead index not loaded'}), 503
+    return jsonify({
+        'count': idx.count,
+        'build_time_seconds': round(idx.build_time_seconds, 3),
+        'memory_estimate_mb': round(idx.memory_estimate_mb, 1),
+    })

backend/services/navi_offroute/mvum_transitions.py

@@ -9,12 +9,20 @@ is pure spatial lookup — no routing logic — mirroring the MVUMSpatialIndex
 The router (``_route_auto``) uses these points as drive->offroad transition
 candidates: a hybrid "drive to a trailhead, switch vehicles, continue offroad"
 plan is considered when it beats the single-mode winner by a comfortable margin.
+
+Coordinates are stored in packed numpy arrays and the attribute columns as plain
+(interned) lists; the shapely Point objects exist only long enough to build the
+STRtree and are then released. Record dicts are reconstructed lazily in
+query_trailheads_near_line — same memory-pack pattern as OSMParkingIndex.
 """
 import logging
 import sqlite3
+import sys
 import time as _time
 from pathlib import Path
 
+import numpy as np
+import psutil
 from shapely.geometry import Point, LineString
 from shapely.strtree import STRtree
 
@@ -26,17 +34,21 @@ logger = logging.getLogger("navi_offroute.mvum_transitions")
 class TrailheadIndex:
     """In-memory STRtree over ``trail_entry_points`` (trailhead/road access points).
 
-    Keeps the STRtree of point geometries plus a parallel ``records`` list of
-    ``{lat, lon, name, road_class}`` dicts aligned with the tree's geometries.
-    (The DB column is ``highway_class``; it is surfaced here as ``road_class`` for
+    Storage is columnar: ``_lats``/``_lons`` (float64 numpy arrays) plus
+    ``_names``/``_road_classes`` (lists, aligned by index). query_trailheads_near_line()
+    builds the ``{lat, lon, name, road_class}`` record dicts lazily from these columns.
+    (The DB column is ``highway_class``; it is surfaced as ``road_class`` for
     consistency with the entry-point records the router already emits.)
     """
 
     def __init__(self, db_path=None):
         t0 = _time.perf_counter()
+        proc = psutil.Process()
+        rss_before = proc.memory_info().rss
+
         self.db_path = Path(db_path) if db_path else navi_db_path()
-        self.records = []          # aligned with self._points
-        self._points = []
+        lats, lons = [], []
+        self._names, self._road_classes = [], []
 
         conn = sqlite3.connect(f"file:{self.db_path}?mode=ro", uri=True)
         conn.row_factory = sqlite3.Row
@@ -46,26 +58,48 @@ class TrailheadIndex:
                 "WHERE lat IS NOT NULL AND lon IS NOT NULL"
             )
             for row in cur:
-                lat = float(row["lat"])
-                lon = float(row["lon"])
-                self.records.append({
-                    "lat": lat,
-                    "lon": lon,
-                    "name": row["name"] or "",
-                    "road_class": row["highway_class"] or "",
-                })
-                self._points.append(Point(lon, lat))
+                lats.append(float(row["lat"]))
+                lons.append(float(row["lon"]))
+                self._names.append(row["name"] or "")
+                # intern the small-cardinality road-class strings so duplicate values
+                # share one object instead of ~740k separate ones.
+                rc = row["highway_class"] or ""
+                self._road_classes.append(sys.intern(rc))
         finally:
             conn.close()
 
-        self._tree = STRtree(self._points) if self._points else None
-        self.count = len(self.records)
+        self._lats = np.asarray(lats, dtype=np.float64)
+        self._lons = np.asarray(lons, dtype=np.float64)
+
+        # Build the STRtree from transient Point objects, then release them; the tree
+        # internalizes its own geometry storage and we reconstruct points on demand.
+        points = [Point(lon, lat) for lon, lat in zip(lons, lats)]
+        self._tree = STRtree(points) if points else None
+        del points
+
+        self.count = len(self._lats)
         self.build_time_seconds = _time.perf_counter() - t0
+        self.memory_estimate_mb = max(
+            0.0, (proc.memory_info().rss - rss_before) / (1024 * 1024))
         logger.info(
             "Trailhead index loaded: %d entry points in %.2f seconds",
             self.count, self.build_time_seconds,
         )
 
+    def _record(self, i):
+        """Construct a trailhead record dict for column index ``i``."""
+        return {
+            "lat": float(self._lats[i]),
+            "lon": float(self._lons[i]),
+            "name": self._names[i],
+            "road_class": self._road_classes[i],
+        }
+
+    @property
+    def records(self):
+        """All records, built lazily (used by tests / introspection — not the hot path)."""
+        return [self._record(i) for i in range(self.count)]
+
     def query_trailheads_near_line(self, coords, buffer_m=2000):
         """Trailhead records within ~``buffer_m`` of a (lat, lon) polyline.
 
@@ -81,8 +115,8 @@ class TrailheadIndex:
         buffer_deg = _buffer_degrees_for_meters(buffer_m, avg_lat)
         out = []
         for i in self._tree.query(geom.buffer(buffer_deg)):
-            if geom.distance(self._points[i]) <= buffer_deg:
-                out.append(self.records[i])
+            if geom.distance(Point(self._lons[i], self._lats[i])) <= buffer_deg:
+                out.append(self._record(i))
         return out
 
 

backend/services/navi_offroute/tests/test_mvum_transitions.py

@@ -6,6 +6,8 @@ stubbed self.route, so no Valhalla/DEM dependencies are exercised.
 """
 import sqlite3
 
+import numpy as np
+
 import pytest
 
 from services.navi_offroute.mvum_transitions import TrailheadIndex
@@ -38,13 +40,26 @@ def test_trailhead_index_loads(tmp_path):
     ])
     idx = TrailheadIndex(db_path=db)
     assert idx.count == 2
-    assert len(idx.records) == len(idx._points) == 2
+    assert len(idx.records) == idx.count == 2
     rec = idx.records[0]
     assert rec["name"] == "Trailhead A"
     assert rec["road_class"] == "track"   # highway_class surfaced as road_class
     assert rec["lat"] == 44.00 and rec["lon"] == -114.00
 
 
+def test_trailhead_index_numpy_backing(tmp_path):
+    db = _trailhead_db(tmp_path, [
+        (44.00, -114.00, "track", "A"),
+        (44.01, -114.02, "residential", "B"),
+        (44.02, -114.03, "path", "C"),
+    ])
+    idx = TrailheadIndex(db_path=db)
+    assert idx._lats.dtype == np.float64
+    assert idx._lons.dtype == np.float64
+    assert len(idx._lats) == len(idx._lons) == idx.count == 3
+    assert idx._lats[1] == 44.01 and idx._lons[1] == -114.02
+
+
 def test_query_trailheads_near_line_returns_close_only(tmp_path):
     # One point sits right on the line; one is ~30 km away (well outside 2 km).
     db = _trailhead_db(tmp_path, [