Diagnosed live over the WebView DevTools protocol: a 0 km track had exactly one
usable point. The map block runs (points.length is truthy), but with a single
coordinate fitBounds is skipped, so the Leaflet map never gets a view and never
fires 'load'. With a saved route active, the overlay code then does
`await whenMapReady(map)`, which waits on that 'load' event forever — hanging the
rest of showSessionDetail, so the Delete/Recalc/Export handlers past it never
bind. No exception, just a silent stall (hence "nothing happens on Delete").
Fix: when there are < 2 coords, setView to the single point so the map loads and
whenMapReady resolves. Verified end-to-end via DevTools: the button now binds and
the 0 km track deletes.
(The prior commit moved the handlers out of the `points.length` guard, which
covers a truly 0-point track; this covers the 1-point-with-route hang.)
The Delete / Recalculate / Refine / Compare handlers were bound inside
`if (mapEl && points.length)` in showSessionDetail, so a track with no points
(started recording, no GPS fix, stopped → 0 km) skipped the whole block and its
"Удалить" button did nothing. Moved the session-action button handlers outside
the map/points guard so they bind regardless. (Also covers a track whose points
all get dropped by the new accuracy filter.)
DB.deleteGpsSession already handled zero points fine (DELETE of 0 rows is a
no-op); the bug was purely the unbound click handler.
Exclude real GPS fixes coarser than the "Макс. погрешность GPS" setting
(minAccuracyM) from a track's stats and its drawn line. Non-destructive: the
points stay in the DB, so raising the threshold brings them back.
- GPS.filterByAccuracy(points, threshold?): drops points with a measured
accuracy above the threshold; always keeps points without accuracy (imported
GPX/FTB, older recordings) and estimated dead-reckoning fills (sentinel >=999).
- Wired into GPS.computeStats (top) so every recompute path funnels through it —
the "Пересчитать" button, "Пересчитать всё", and the post-de-jitter recompute.
Also into GPS.filterNoise so the detail-screen line matches the stats.
- Removes mild jitter detours that the spike filter misses (verified: 342 -> 334m
when a poor-accuracy nudge is dropped). Lower the threshold + Пересчитать to
apply; raise it to undo.
- Route editor (showRouteEditor): create ("Новый маршрут") or edit an existing
route's geometry in-app — tap to add, drag a point to move, tap a point to
delete. Save: new prompts a name (addRoute), edit rewrites points_json
(updateRoutePoints; drops stale elevations, re-fetch from the preview).
- Slope colouring in discrete 1% bands (green→yellow→red→brown→black, ≥10% black)
instead of degree buckets; colour-strip legend (slopeLegendHtml).
- Diagnostic logs compiled-in but silenced in release: get_app_info now returns
debug = cfg!(debug_assertions); JS installs a console.log filter that drops
lines starting with a diagnostic tag ([poll]/[recover]/[stop]/[GPS]/[HR]/[SD])
in release only. No more hand-adding/removing logs.
- Fix: saved-route overlay was invisible on the recording (rotating) map — it
flashed on pinch-zoom then vanished. leaflet-rotate keeps tile/overlay panes
inside _rotatePane, but the commented-out createPane override meant our custom
routePane attached to mapPane (outside rotation). Now parent routePane to
map._rotatePane when present; non-rotating maps unchanged.
See how hilly a hand-drawn route is BEFORE riding it.
- parseRouteGPX reads <ele>; GPS.routeElevStats computes cumulative distance,
grade-to-next, ascent/descent (1m hysteresis) and max grade over smoothed
elevations (filterAndSmoothAlt kills SRTM vertical noise).
- Route preview screen (tap a route name): map with a wide colored casing +
tappable altitude/grade points, elevation profile chart (steep >8% in red),
length/ascent/descent/max-grade, fullscreen expand/collapse.
- Casing colored by slope (deg: lime <2, yellow 2-4, red 4-6, black >6) or by
elevation (green->yellow->brown), toggled with an on-map ∡/⛰ button. Casing
drawn in a dedicated pane below the overlay pane, so a recorded track stays on
top; the same toggle is on the track-detail map too.
- fetchElevations tries, in order: GPXZ (Copernicus GLO-30, key from GPS
settings — stored in localStorage, never in git) -> open-elevation (SRTM) ->
open-meteo -> OpenTopoData (keyless). Import auto-fetches elevations.
- db.js: updateRoutePoints persists fetched elevations into points_json;
getRoutes reports hasEle. Route points are now {lat, lon, ele?}.
Native prompt at route import replaced with in-app appPrompt.
Fixes to the smoothing/routes UI reported on device:
- Blank detail after save/revert: showSessionDetail was called from inside the
open detail modal, stacking a second overlay; querySelector("#sd-map") then
hit the stale, already-initialized Leaflet map and threw, aborting the rest
of the render (map + speed chart). New reopenDetail() closes the current
detail (and its fullscreen-map child) and reopens fresh, like recalc/refine.
- Native confirm()/prompt() replaced with in-app appConfirm/appPrompt dialogs
(winStack + history integrated, hardware back cancels). Used for smoothed
save, route delete, route rename.
- Slow smoothed save: one UPDATE per point (4000+ IPC round-trips) replaced
with a chunked CASE update — 18 statements for 4366 points, params kept
under SQLite's 999 limit. Verified value-for-value on a copy DB.
- Closing the fullscreen map now also closes the smoothing panel (it is a
fixed element that otherwise lingered over the collapsed map).
- "≈" button active state now reflects panel-open only, not session.smoothed
(it looked permanently pressed after a save).
Docs: README + PROJECT_MAP updated for saved routes, de-jitter, accuracy_m,
routes table and new gps_points/gps_sessions columns. Version 0.1.393.
Routes:
- routes table + CRUD helpers (db.js), GPX rtept/trkpt/wpt parser (gps.js)
- "Saved routes" screen (import/rename/recolor/delete, one active at a time)
- Route overlay on both the recording map and the past-track detail map,
toggled by an "M" button; default ON at Start when a route is selected
- Draw after real map-load via async whenMapReady(), no setTimeout
Road colors:
- per-activity configurable colors (walk/bike/run); fix walk roads (atype "w")
- get_road_points maps run -> "r" (lib.rs)
De-jitter (Savitzky-Golay):
- GPS.smoothTrackSG (quadratic LS per window, iterated, endpoints anchored)
- reversible storage: lat_smoothed/lon_smoothed + session.smoothed flag
- detail-screen tuning panel (window/iters sliders, live preview,
save smoothed / revert original, kcal recomputed on new points)
- accuracy_m column per point for later accuracy-gated recompute
Active map-button state restyled to inverted (dark glyph on translucent
white) for OLED; detail map safe-area handled in fullscreen.
On stretches where the fix rate drops (dt > gapThresholdSec, incl. dead-
reckoning) the straight chord mis-measures the path, so computeStats now
estimates gap distance by avg(speed)*dt (capped at max speed; a tiny chord
over a long gap means the rider was stopped and the reported speed is stale
-> trust the position). A fast jump WITHOUT a time gap is a teleport glitch
and is dropped. Distance now depends only on ts + speed, not on the
`estimated` flag (which live points carry but DB rows don't), so the live
counter, Stop and recompute finally agree. Verified on a real ride: at
gapThreshold=30 -> 35.7 km / 1024 kcal (map-measured ~35 km); clean tracks
without gaps unchanged.
Settings: read GPS config on input+change (reliable on mobile), add a
"Save GPS settings" button and a "saved" toast so it's clear the value
persisted; changing gapThreshold + recompute now visibly applies.
The old stop path re-added every native point "not already in the DB",
which raw-inserted points the live poll had intentionally dropped
(stationary / auto-pause / wrong state) with scrambled timestamps. The
detail view draws by seq order, so those late-written points with early
timestamps rendered as spikes across the map (distance stayed correct —
computeStats sorts by ts). Now stop imports only points past
_lastProcessedTs (the poll's processed marker): the genuinely un-polled
tail, including a screen-off stretch, without re-adding filtered points.
- native_root(): pick app root by presence of gps_live/gps_points files.
app_data_dir() is the root directly on this Tauri; the old .parent().parent()
read the wrong dir, so background (screen-off) points were never imported.
- recoverFromFiles: match file->session by id/date, plus session id stamped
into the native file name (gps_points_<ms>_s<id>.jsonl) for exact binding
(gps_set_session_id -> Kotlin bindSessionIfPending). Proximity dedup +-2.5s,
recalc with noDelete (never runs the destructive glitch/>24h filter that
once wiped real rides), delete already-imported files. Progress bar UI.
- bike kcal is NET (daily balance already books BMR*activityLevel): pedaling =
max(physics power/efficiency, MET-net by speed), coasting (power<=0, the bike
would roll unpowered) = basal only. Elevation via cumulative hysteresis.
recomputeAllStats + "recompute elevation/kcal" menu refresh stored tracks
without touching points.
- live recording: the native file is the DB source of truth; watchPosition
only drives the live marker, so a fix is not stored twice with two timestamps.
- track list pagination, SAF local-file picker, tile regions from router.
- applyBearing: use h, not h - rotOffset. The .rec-screen (map + panel) is
CSS-rotated by rotOffset in landscape and setBearing adds another rotate on
the map pane; subtracting rotOffset cancelled the screen rotation for the
map only, gluing course to the portrait top edge while the panel turned.
Using h lets the map turn with the screen -> course points to physical top.
- Poll: guard the requestAnimationFrame yield with !document.hidden. rAF never
fires while the screen is off, so awaiting it there hung the import loop and
froze point import mid-walk (regression from the progressive-draw change).
- recoverSession: import missing native points by the SET of existing
timestamps instead of ts > maxDbTs. Sparse screen-gated points can have a
late max ts, so the old check imported nothing and left the track full of
gaps; also pass sensorData so bearing/accel/gyro are stored.
- stopRecording: replace the dead _lastNativeCount backfill (the variable
is never assigned → guard is always false → missed points were never
saved) with a full timestamp-based re-import from the native points
file, then recompute stats over the complete set. Fixes losing ~90% of a
track when recording with the screen off — the JS import poll is frozen
by the WebView while the screen is off, so only the native file had all
points.
- Poll: draw each imported point via a lightweight onImportPoint and yield
a frame every ~20 points so the track visibly grows as points transfer
on screen-wake catch-up (restores pre-batch behaviour, kept cheap).
- Native: hard-reject fixes coarser than the Min-accuracy setting (drops
coarse cell ~800m that stuck the track); remove low-accuracy and adaptive
relaxation branches so only accurate fixes are recorded, honest gaps otherwise
- Store bearing_deg per point (gps_points migration + addGpsPoint)
- Default Min-accuracy 20 -> 50m (one-time migration, respects user override)
- Fix: http_request was synchronous with no timeout; an unreachable host
(LAN sync on startup) blocked the whole UI ~30s. Now async + spawn_blocking
+ connect(4s)/global(12s) timeouts, fails in background without freezing
- Bundled prior work: course-up map orientation, compass calibration,
online elevation refine (open-elevation), distance gap-gate fix
showSessionDetail rendered its map/charts/buttons inside setTimeout blocks
guarded by renderToken. On WebKitGTK the deferred render was voided, so the
track detail showed no map, charts, or fullscreen button (while the
synchronously-rendered roads view worked).
- Remove all setTimeout/requestAnimationFrame/IIFE from showSessionDetail;
render map, charts and wire buttons synchronously right after the modal
markup exists (like showRoadsMap). Early `return`s become `if` guards;
goals/food-cost keep real `await`. Force layout (offsetHeight) + call
invalidateSize so Leaflet reads the real container size in the modal.
- Fix large i64 session-id corruption from tauri-plugin-sql 2.4.0 (ids come
back off by 1 as JS numbers). Read/pass session ids as text via
CAST(id AS TEXT) AS id_str everywhere they are reused: Trek list, Today
click, compare modal/view, goal-marker prev sessions, map default-center,
and importGpxSession (whose points were orphaned by the corrupted id).
SQLite coerces the string to the INTEGER column. Point ids (max ~135k)
are small and unaffected.
Regression from per-session points file change (1feb2ef): the service
returns START_STICKY, so when the OS kills it during screen-off, it
restarts with a null intent and onStartCommand minted a NEW points file,
repointing gps_live_points.jsonl and orphaning accumulated background
points. Evidence: many tiny orphaned gps_points_*.jsonl fragments.
Fix: onStartCommand creates a new session file only on a genuine start
from JS (intent with extras). A START_STICKY restart (null intent) keeps
writing to the existing session file from the pointer — one file per
session until Stop.
Verified on device: am kill (= START_STICKY restart) no longer creates
a new file, pointer unchanged, trail resumes on return. Background
writing confirmed: 49 points in 70s with screen off (Dozing) at 1Hz.
Compass:
- Gimbal lock compensation: when phone tilts past horizontal (|beta|>90
in landscape, |gamma|>90 in portrait), heading flips 180 — compensated
- EMA smoothing for heading (factor 0.3, ~300ms settle) — reduces jitter
from road bumps while keeping responsive turns
- 3 orient modes: X(off) A(auto) L(lock) — lock freezes layout, compass
still works with gimbal compensation
- Debug overlay on map shows alpha/beta/gamma/heading/mode
Build:
- ./build.sh install — install last APK without building
- ./build.sh -d install — install debug APK
- install_apk respects -d flag for debug/release selection
Products:
- health-diet.ru scraper (Tampermonkey) + import script
- 11469 products with 37 nutrients, Dodo 258, McDonalds 619, KFC 361, Ashan 2000
- Rust import_seed_products: dynamic column list, auto-create missing columns
- Seed version bumped to 10
GPS session safety:
- Separate gps_points_{ts}.jsonl per session, deleted only after DB save
- Toast on DB errors, NaN guard in finishGpsSession
Other:
- build.sh: router/VPS upload only with -u flag
- Orientation buttons bottom-right, weather via Yandex Pogoda
- flex-direction:row in landscape — panel 200px on the right, stats vertical
- CSS rotation compensates phone orientation, user sees element coords directly
- Bottom nav hidden in landscape, restored on exit
- GPS placeholder overlay: separate div from map container to avoid Leaflet conflict
- display:contents removed, clean column layout for stats in landscape
- leaflet-rotate plugin: map rotates by compass heading via deviceorientation
- Landscape detection: CSS viewport rotation when phone held sideways (gamma > 50)
- Bottom nav hidden in landscape mode
- GPS placeholder: pixel art overlay shown until GPS signal arrives
- invalidateSize() after map init to prevent black screen
GPS altitude: accel-based glitch filter (accel=0 + speed = GPS drift), altitude
jump >100m filter, timestamp sanity (>24h = remove). Per-point Crr from accel
vibration (smooth/mixed/gravel). Gap interpolation on map with dashed lines
when accel confirms movement. Walk/run separate speed limits (configurable).
Adaptive GPS accuracy (auto x2 when signal bad). Dead reckoning points when
GPS silent but accel shows movement. Locale.US fix for sensor JSON.
Sensors: accelerometer + gyroscope in GpsRecorderService (RMS, max, pitch, yaw).
Columns in gps_points. Dynamic addGpsPoint INSERT (safe if columns missing).
Manual routes: draw on map by tapping points, elevation from past tracks +
open-meteo API, realistic timestamps by activity speed. Manual sessions
skip glitch filtering on recalculate. Speed from GPS.config not hardcoded.
Water tracking: water_pct field in product add/edit forms, searchProducts
returns water_pct, "ml" label for liquids, recipe water from ingredients.
BMR: moving average weight (7 days), fixed charts using dailyBase not raw bmr.
BLE: ble_clear_hr on stop, recovery toast, serde alias for ts/timestamp.
Products: 217 Gridnev products (bakery, desserts, bread, ready meals).
Magnit scraper fix (NUXT_DATA name/value pairs, not indices). Supplements
addSupplement with all 41 columns. Nutrient fallback RDI includes omega/EPA/DHA.
Export/import: custom products + recipes + supplements in single JSON bundle.
Export all or single product. Dedup on import by name_lower + source.
OTA: router upload (LAN, always), VPS upload (-u flag). Kotlin checks router
first (2s timeout), then VPS. Update URL from update_info.json.
Session detail: renderToken prevents stale setTimeout callbacks after recalculate.
WAL disabled (PRAGMA journal_mode=DELETE).
FOREIGN KEY constraint failed because product_units table has FK to
products(id) and wasn't cleaned before DELETE. Now: delete orphaned
product_units rows first, then delete dupe products. No PRAGMA needed.
Verified on desktop: 16675 → 8330 with FK enforcement ON.
FOREIGN KEY constraint failed on desktop (rusqlite enables FK by default).
Android was silently failing too. Fix: disable FK during dedup, re-enable after.
Tested: desktop 16675 → 8330 products, idx_products_dedup created.
Also: alert() instead of boot-label for seed errors (visible to user).
Root cause: app_data_dir() = files/dev.baldav.app/ but tauri-plugin-sql
stores baldav.db in app root /data/user/0/pkg/baldav.db (two levels up).
Rust created empty DB at wrong path, dedup ran on empty DB = no-op.
Fix: added p_root = app_data_dir/../../../baldav.db check.
Also: show seed error on boot label instead of swallowing silently.
Dedup fix:
- Dedup SQL now runs BEFORE seed version check (was inside "if new version" block,
never executed when version matched). 16675→8330 products.
- UNIQUE INDEX on (name_lower, source, COALESCE(store_id,'')) prevents future dupes
- Errors propagated via map_err (no more silent failures)
RDI standards from DB (not hardcoded):
- rdi_standards table: RU, EU, US
- rdi_values table: nutrient_key + value per standard
- 26 nutrients seeded for all 3 standards
- BMR.loadRDI(code) reads from DB, replaces hardcoded RDI object
- Settings: "Стандарт РСП" dropdown (РФ/ЕС/США)
- Profile stores selected standard, loaded at init
- Fallback: hardcoded RU values until DB ready
execute_batch() silently failed on multi-statement dedup SQL.
Split into 4 separate execute() calls with map_err — errors
now propagate to JS and show to user instead of being swallowed.
Dedup products in Rust import (not JS — JS was too slow/hung):
- Key: (name_lower, source, COALESCE(store_id,''))
- Same product from same store = duplicate, deleted
- Different sources (magnit vs ion vs manual) = separate products
- Reassigns food_log + recipe_ingredients refs before deleting
- UNIQUE INDEX prevents future dupes (INSERT OR IGNORE now works)
- Tested: 16675 -> 8330 in 0.15s on device DB copy
Removed JS dedup migration (replaced by Rust).
Order in import_seed_products:
1. Index name_lower
2. Dedup (reassign refs + delete)
3. UNIQUE INDEX
4. INSERT OR IGNORE from seed
5. UPDATE store products
6. seed_version + COMMIT
- "Уд." button (red) next to "Ред." on product card
- deleteProduct: nullifies food_log refs, zeros recipe_ingredients refs, deletes product
- List refreshes after deletion
Was: for loop with 3 IPC calls per duplicate group = thousands of
round-trips = 10+ minutes on phone. Now: 3 bulk UPDATE/DELETE
statements that run in SQLite directly. One-time migration,
subsequent launches skip (COUNT=0).
1. Deduplicate products: 13482 -> 8325 in seed DB. JS migration
deduplicates app DB on launch: reassigns food_log and
recipe_ingredients to kept product before deleting dupes.
2. MICRO_FIELDS: full chemical names + alternatives for all nutrients.
"B3 / PP / ниацин", "B9 / фолиевая", "Кальций / Ca" etc.
3. Track comparison: newer track always compared against older.
Green = newer faster (progress), Red = newer slower (regression).
Removed separate cyan dashed track — only colored segments.
Order of selection doesn't matter — date determines which is newer.
4. Source badges in search: source column was already in SELECT,
badges render via sourceBadge(p.source) in prod card.
5. Gap threshold already cosmetic-only (confirmed, no changes needed).
Seed DB v9, VACUUM'd, indexed.
ATTACH UPDATE with correlated subqueries on name_lower was doing
full table scans on 13000 rows. 56 seconds on desktop, 2-5 min
on phone. User saw "загрузка..." for minutes on every launch.
Fix: CREATE INDEX on name_lower in both main and seed DBs before
the UPDATE. Also pre-built index in seed DB file.
Benchmarked: 56.5s -> 0.09s on desktop. Phone should be <1s.
- sourceBadge() helper renders colored badge by source field
- searchProducts SELECT now includes source column
- getFoodForDate SELECT includes p.source as product_source
- Badge shown on product cards in search and food log entries
- CSS: .src-badge with per-source colors (pastel, dark-theme friendly)
Root cause: addProduct INSERT didn't include name_lower column.
Search uses WHERE name_lower LIKE — products with NULL name_lower
never found. User entered 20 min of strawberry data for nothing.
Fixes:
- addProduct: includes name_lower = name.toLowerCase()
- Added phosphorus_mg and carotene_mcg to addProduct INSERT
- PRODUCT_MICRO_FIELDS: subset of MICRO_FIELDS for product forms
(excludes supplement-only fields: omega, EPA/DHA, fish oil etc.)
- showAddProductModal and showEditProductModal use PRODUCT_MICRO_FIELDS
- Rust: removed stale inserted+=1 line
- Seed DB v8: added "Клубника свежая" (alias for "Земляника садовая")
Was: 13000 individual INSERT OR IGNORE in a loop = 2+ minutes.
Now: one INSERT ... SELECT FROM seed.products = seconds.
Flow:
1. ATTACH seed DB
2. INSERT OR IGNORE ... SELECT (bulk, new products)
3. UPDATE store products with KBJU + micronutrients (user_edited=0 only)
4. UPDATE nutrients_guessed flag
5. INSERT OR REPLACE seed_version into profile
6. COMMIT (all or nothing — crash-safe)
7. DETACH
seed_version written INSIDE transaction — if app crashes before COMMIT,
next launch re-does the import. No partial state.
New column: user_edited INTEGER DEFAULT 0 on products table.
- JS migration adds column
- showEditProductModal sets user_edited=1 on save
- Rust seed import: UPDATE store products (source IN magnit/5ka/kdv/ion)
WHERE user_edited IS NULL OR user_edited = 0
- Now updates KBJU + all micronutrients from seed (not just INSERT OR IGNORE)
- User edits preserved: user_edited=1 products never touched by seed import
- Manual/recipe_import products also never touched (source check)
Flow: seed update fixes wrong KBJU (e.g. bread 268→243) automatically,
but if user already fixed it manually — their fix stays.
Product editing:
- "Ред." button on expanded product card
- showEditProductModal: edit name, KBJU, all micronutrients
- Updates product in DB directly, refreshes search
Bread fix:
- "Хлеб Тостовый пшеничный Аютинский 280г" was 268 in Magnit data
- Package says 243 kcal/100g — corrected in seed DB
- Seed version 7
Users can now fix any wrong KBJU directly from the product card
without re-importing or editing the seed DB manually.
