Previous thresholds (10 m GPS, 5 m barometric, 30 s MA) were calibrated for
raw noisy GPS. Strava-exported FIT files carry elevation already pre-processed
by Strava (smooth 1 m quantisation, no steps > 5 m), so the aggressive
filtering suppressed real climbing — avg −17 % error across 37 reference
activities.
New strategy, keyed on source + altitude_source:
strava_export → MA 5 s, threshold 1.0 m
fit_file / barometric → no MA, threshold 1.5 m
fit_file / gps → MA 5 s, threshold 2.0 m
unknown non-strava → MA 5 s, threshold 1.5 m
Result on 37 cross-referenced activities: avg −2.8 %, std 4.6 %,
37/37 within ±15 % (was 0/37).
Both paths — initial import (metrics._elevation) and bulk recalculate
(dem.recalculate_elevation_hysteresis) — now use the same elevation_params()
function from metrics.py.
- Exclude per-activity VAM contributions where climbing_time_s < 10 min; short
punchy efforts don't represent aerobic fitness and were skewing monthly averages
- Store climbing_time_s alongside climbing_vam_mh in metrics, detail JSON, and
summary JSON so the frontend has the data to reason about confidence
- Accumulate total climbing time per period; opacity scales from 0.25 (10 min,
minimum threshold) to 1.0 (≥ 1 h) so thin-evidence months read as faint dots
- Render VAM as dots only (no lines) since each period is an independent average,
not a cumulative — lines implied continuity that isn't there
- Tooltip now shows "1060 m/h · 38 min climbing"
Remove the per-duration VAM curve everywhere (metrics, summaries, detail
JSON, athlete.json, VamChart.svelte, AthleteView VAM tab). Keep only
climbing_vam_mh per activity. Add it to activity summaries so NerdCorner
can plot average climbing VAM per week/month year-over-year alongside
distance/elevation/time. Add --backfill-vam-summary flag to copy the
field from existing detail JSONs into index.json without re-extracting.
Refactors core VAM logic into _vam_from_ele_1hz() and _build_ele_1hz()
so both the DataPoint-based extract path and the timeseries-based backfill
path share the same implementation.
render --recompute-vam reads stored *.timeseries.json files and updates
climbing_vam_mh + vam_curve in activities/*.json and index.json in-place,
without re-parsing the original FIT/GPX files.
Extract pipeline now computes two VAM metrics per activity (cycling,
running, hiking, walking):
- climbing_vam_mh: VAM on ascending segments only, using 30 s forward
lookahead to classify climbing vs. flat/descent (stored in detail JSON)
- vam_curve: [[duration_s, vam_mh], ...] best VAM per standard duration
(60 s – 1 h), sliding window on 30 s smoothed elevation, only windows
with ≥ 10 m net gain count (stored in summary + detail)
Athlete JSON aggregates vam_curve across all activities (all_time,
last_365d, last_90d), same structure as power_curve.
Frontend:
- ActivityDetail shows "Climbing VAM" stat (grouped with elevation)
- AthleteView adds a "VAM Curve" tab that appears only when the athlete
has climbing data; renders VamChart (new component, mirrors MmpChart)
vam_curve stripped from combined global feed; kept in user year shards
for season-based on-the-fly aggregation in VamChart.
Requires bincio reextract to backfill existing activities.
Devices (Apple Watch, some GPS units) record 0.0 when they lose barometric/GPS
lock mid-activity. The old accumulation committed these as real sea-level points,
inflating both gain and loss by the current elevation (e.g. 792m dropout on the
Cosmo Walk added ~1584m of phantom gain+loss).
Fix: skip any elevation value < 1.0m when the current committed elevation is
significantly above zero (> threshold). Gradual legitimate descents to sea level
are unaffected because intermediate values are committed along the way.
Add --recompute-elevation flag to bincio render to backfill existing activities.
- _rebuild_athlete_json now applies sidecar edits (sub_sport, sport, etc.)
in-memory before passing summaries to write_athlete_json, so activities
marked indoor via sidecar are correctly excluded from records.
- _best_climb now runs Kadane's over cumulative distance (not 1Hz dense
time) so recording pauses don't create None gaps that falsely reset the
climbing window. Grappa: 811m→1603m; Nivolet: 311m→2009m.
- Add bincio render --recompute-climbs to backfill existing activities
from their stored timeseries.
FIT files from older devices (and GPX/TCX files) often omit the speed field.
The sliding-window best-effort algorithm was treating all such points as speed=0,
so no records were ever produced for these activities.
Fix: when p.speed_kmh is None but consecutive lat/lon are available, compute
haversine segment speed and spread it evenly across the 1Hz interval slots.
This mirrors what _gps_stats already does for avg/max speed computation.
test_writer: _dummy_metrics() and test_build_summary_required_fields were
missing np_power_w=None after the field was added to ComputedMetrics.
test_metrics: the leading-zero elevation heuristic fired on a single 0.0
start value, incorrectly skipping the first legitimate elevation step.
Guard now requires at least 2 consecutive near-zero leading values before
activating the Apple Watch lock-acquisition workaround.
Apple Watch and similar devices record exactly 0.0 for elevation while
waiting for barometric/GPS lock, then jump to the real altitude. The
hysteresis accumulator was seeding from 0.0, counting the full jump as
ascent. Fix: detect a leading near-zero run followed by a large jump
and seed the accumulator from the first real value instead.
Applied in both _elevation() (fresh extractions) and
recalculate_elevation_hysteresis() (recompute path). Added a bulk
admin endpoint POST /api/admin/users/{handle}/recompute-elevation and
corresponding button to fix existing stored activities.
GPS jitter and barometric quantization noise caused systematic overestimation
of elevation gain — in extreme cases 100% of reported gain was sub-1m noise.
Implements source-aware hysteresis: elevation is only committed when it
deviates from the last committed value by ≥5m (barometric) or ≥10m (GPS/GPX/TCX).
- ParsedActivity gains `altitude_source` field ("barometric"/"gps"/"unknown")
- FIT parser sets "barometric" when enhanced_altitude is present, else "gps"
- GPX and TCX parsers always set "gps"
- metrics._elevation() uses the threshold matching the source
- 5 new parametric tests covering flat GPS noise, threshold differences, and real climbs
Strava originals with absolute Unix timestamps stored as elapsed-second
offsets produce a t_max of ~1.6 billion. compute_mmp and compute_best_efforts
both create dense 1Hz arrays via range(t_min, t_max+1), which for a 1.6B
span allocates 44+ GB and OOM-kills the process. Add a >1-week sanity
check and return None early for corrupt streams.
Root cause: old Strava activities (seen from 1970-epoch start_date)
where the time stream contains absolute Unix timestamps instead of
elapsed seconds.
Davide Scaini