docs: add BUG-23 — Pi/server trip counts diverge (57 vs 10)

Root-caused the discrepancy by replicating the server's trip builder over
live data: 58 raw groups -> 51 after the server's duration/distance floor
(7 short trips dropped that the Pi keeps) -> 10 after skipping 41 deleted
trip markers. Documents the min-trip filter mismatch and the row-set
mismatch as the actual bugs.

BUGS.md

@@ -26,6 +26,34 @@ When CLOSED, move the entire entry to Resolved — don't duplicate.
 
 ---
 
+### BUG-23 — Pi and server trip counts diverge (server drops short trips the Pi keeps)
+- **Status**: `OPEN` 2026-08-15
+- **Phase**: 29
+- **Symptom**: local Pi history shows **57 trips**; the portal shows **10**. Investigated
+  2026-08-15: Pi `tracking` = 8562 rows / 57 trips; server `tracking_points` = 8841,
+  `trips` = 10, `deleted_trip_markers` = 41.
+- **Cause** (replicated `server/modules/db.py::compute_trips_for_device()` over live data —
+  58 raw trip groups → 10 final; three compounding factors):
+  1. **Deleted markers (not a bug, but most of the gap)**: 41 trips were manually deleted
+     on the portal (Phase 26). Recompute intentionally skips their `start_time`, so
+     58 − 41 = 17.
+  2. **Min-trip filter mismatch (the actual bug)**: the server drops any trip with
+     `duration_s < 30 OR distance < 0.05` km. The Pi has no equivalent — its only floors
+     are "≥3 GPS-motion samples **and ≥10 s**" for GPS-only trips, and OBD trips have no
+     duration/distance floor at all. **7 real short trips** (30–80 s, 26–47 m — driveway /
+     parking moves) are kept by the Pi but dropped by the server. 17 − 7 = 10.
+  3. **Row-set mismatch (minor)**: the server segments only GPS-bearing rows
+     (`gps_lat/lon IS NOT NULL`), while the Pi normalizes all rows including speed-only
+     NULL-GPS rows (FEAT-12). The server also holds 8841 points vs the Pi's 8562 (the Pi's
+     local DB was recreated at some point and lost ~279 old rows). Net effect ≈ 1 extra
+     trip on the server (58 raw vs 57 on the Pi).
+- **Fix**: reconcile the two trip builders — drop the server's `duration/distance` floor to
+  match the Pi's "10 s for GPS-only" rule (or add the same floor to the Pi), align the
+  NULL-GPS row handling, and decide how deletions are mirrored between Pi and server (the
+  41 markers currently exist only server-side).
+
+---
+
 ## Resolved
 
 ### BUG-22 — Trip max speed ignores GPS-only stretch (shows OBD max only)