fix: refine hysteresis recalculation with MA pre-smoothing and lower thresholds

- dem.py: pre-smooth elevation with 30s moving average before hysteresis
  in recalculate_elevation_hysteresis(); thresholds drop from 5m/10m to
  1m (barometric) / 3m (GPS) — accurate after noise is smoothed out
- dem.py: widen DEM median-filter window 45s → 60s
- dem.py: rename response key source → altitude_source for consistency
- writer.py: write altitude_source into detail JSON at extract time
- tests/test_dem.py: 21 unit tests for pure functions and file-level hysteresis
- tests/test_edit_server.py: 11 TestClient API tests for both recalculate endpoints
- add httpx as dev dependency (required by FastAPI TestClient)

bincio/extract/dem.py

@@ -35,7 +35,32 @@ _DEM_HYSTERESIS_M = 10.0
 # Median filter window (seconds / samples at 1 Hz) applied to DEM-interpolated
 # series before hysteresis.  45 s smooths SRTM tile steps while keeping real
 # climbs (typical cycling ramp > 100 m over > 2 min).
-_MEDIAN_WINDOW_S = 45
+_MEDIAN_WINDOW_S = 60
+
+# Moving-average window (seconds) applied to the 1 Hz elevation series before
+# hysteresis in the on-demand recalculation.  Pre-smoothing lets us use a
+# much lower dead-band (capturing real small climbs) while still suppressing
+# GPS jitter and barometric quantization noise.
+_MA_WINDOW_S = 30
+
+
+def _moving_average(values: list[float], window: int) -> list[float]:
+    """Apply a centred sliding-window moving average to *values*.
+
+    Edge handling: window shrinks symmetrically at both ends (same effective
+    behaviour as scipy's 'nearest' / numpy's 'reflect' mode).
+    """
+    half = window // 2
+    n = len(values)
+    out: list[float] = []
+    cumsum = [0.0] * (n + 1)
+    for i, v in enumerate(values):
+        cumsum[i + 1] = cumsum[i] + v
+    for i in range(n):
+        lo = max(0, i - half)
+        hi = min(n, i + half + 1)
+        out.append((cumsum[hi] - cumsum[lo]) / (hi - lo))
+    return out
 
 
 def _median_filter(values: list[float], window: int) -> list[float]:
@@ -275,18 +300,32 @@ def recalculate_elevation(
 def recalculate_elevation_hysteresis(user_dir: Path, activity_id: str) -> dict:
     """Recompute elevation gain/loss from the original recorded elevation data.
 
-    Uses the same source-aware hysteresis thresholds as the extract pipeline:
+    Algorithm
+    ---------
+    1. Read ``elevation_m_original`` (backup from a prior DEM run) if present,
+       otherwise read ``elevation_m`` from the timeseries.
+    2. Apply a :data:`_MA_WINDOW_S` (30 s) moving average to smooth out
+       barometric quantization steps and GPS jitter.
+    3. Apply a low dead-band threshold to the smoothed series:
+       - **1 m** for barometric altimeters (FIT files with ``enhanced_altitude``)
+       - **3 m** for GPS-derived altitude (GPX, TCX, FIT without enhanced_altitude)
 
-    - 5 m for barometric altimeters (FIT files with ``enhanced_altitude``)
-    - 10 m for GPS-derived altitude (GPX, TCX, FIT without barometric)
+    The 30 s pre-smoothing makes the low thresholds safe: after averaging,
+    0.2 m barometric quantization noise and short-period GPS jitter are
+    suppressed below the threshold, while real terrain changes (which persist
+    across the window) are preserved.
 
-    The elevation array in the timeseries is **not** modified.  If a DEM
-    correction was previously applied, the backup in ``elevation_m_original``
-    is used as the source so the original sensor data is recovered.
+    The elevation array in the timeseries is **not** modified — only the
+    summary stats in the detail JSON and ``index.json`` are patched.
+
+    ``altitude_source`` is read from the detail JSON (written by the extractor
+    for activities recorded after this field was added).  For older activities
+    it falls back to ``"unknown"`` → 3 m GPS threshold.
 
     Returns
     -------
-    dict with keys ``elevation_gain_m``, ``elevation_loss_m``.
+    dict with keys ``elevation_gain_m``, ``elevation_loss_m``,
+    ``threshold_m``, ``altitude_source``.
     """
     acts_dir = user_dir / "activities"
     json_path = acts_dir / f"{activity_id}.json"
@@ -299,7 +338,7 @@ def recalculate_elevation_hysteresis(user_dir: Path, activity_id: str) -> dict:
 
     ts = json.loads(ts_path.read_text(encoding="utf-8"))
 
-    # Use original elevation if a DEM backup exists, otherwise use current
+    # Prefer the pre-DEM backup; fall back to the current elevation array
     ele_arr: list[Optional[float]] = (
         ts.get("elevation_m_original") or ts.get("elevation_m") or []
     )
@@ -307,12 +346,14 @@ def recalculate_elevation_hysteresis(user_dir: Path, activity_id: str) -> dict:
     if len(elevations) < 2:
         raise ValueError("Not enough elevation data to compute gain/loss")
 
-    # Determine threshold from altitude_source stored in detail JSON
+    # Determine source-aware threshold
     detail = json.loads(json_path.read_text(encoding="utf-8"))
     altitude_source = detail.get("altitude_source", "unknown")
-    threshold = 5.0 if altitude_source == "barometric" else 10.0
+    threshold = 1.0 if altitude_source == "barometric" else 3.0
 
-    gain, loss = _hysteresis_gain_loss(elevations, threshold)
+    # Pre-smooth to suppress noise, then accumulate with low dead-band
+    smoothed = _moving_average(elevations, _MA_WINDOW_S)
+    gain, loss = _hysteresis_gain_loss(smoothed, threshold)
     gain_r = round(gain, 1)
     loss_r = round(loss, 1)
 
@@ -337,5 +378,5 @@ def recalculate_elevation_hysteresis(user_dir: Path, activity_id: str) -> dict:
         "elevation_gain_m": gain_r,
         "elevation_loss_m": loss_r,
         "threshold_m": threshold,
-        "source": altitude_source,
+        "altitude_source": altitude_source,
     }