# Port progress log

Running record of what actually happened, task by task — including what went wrong.
The v2 equivalent of this file caught real bugs by making risks explicit before they
were walked into, so the practice carries over.

Plan: [PLAN.md](PLAN.md) · Source of truth for *why*: the native repo's
`docs/ARCHITECTURE.md`, `docs/TESTING.md`, `docs/v2/PROGRESS.md`.

---

## T00 — Disk space + toolchain · **complete**

**Outcome:** `flutter doctor` reports no issues in any category. Flutter 3.47.0 (Dart
3.13.0), Xcode 26.0.1, CocoaPods 1.17.0, Android SDK 36.0.0, iOS 26.0.1 simulator runtime.

### The disk panic was largely a false alarm — but measure twice

Planning measured **16 GiB free at 92%**, which drove a whole cleanup strategy. A second
measurement minutes later, before deleting anything, showed **36 GiB free at 81%**. Most
likely APFS local snapshots aging out.

**Nothing was deleted.** Gradle caches, AVDs, and `~/.cargo` were all left intact.
Post-install the machine sits at ~22 GiB free.

**Lesson:** re-measure immediately before acting on a disk-space number. Had the plan been
followed literally, ~9 GB of still-useful caches would have been destroyed for no reason.

### Things that actually needed fixing

| Problem | Fix |
|---|---|
| `cmdline-tools component is missing` | `sdkmanager --sdk_root=$ANDROID_HOME "cmdline-tools;latest"` — the exact trap already documented in the native repo's `DEVELOPMENT.md`: brew's `sdkmanager` resolves its own SDK root and does not see `~/Library/Android/sdk` unless `--sdk_root` is passed explicitly |
| Android licenses unaccepted | `yes \| sdkmanager --sdk_root=$ANDROID_HOME --licenses` |
| Default JDK is 25, which AGP rejects | `flutter config --jdk-dir=<temurin-21>` — same constraint as the native build, now pinned in Flutter's config rather than relying on an exported `JAVA_HOME` |
| No iOS simulator runtime installed | `xcodebuild -downloadPlatform iOS` (8.05 GB). Note simulator *devices* already existed for a runtime that did not — `simctl list devices` looked populated while `list runtimes` was empty |
| CocoaPods absent, system Ruby 2.6.10 | `brew install cocoapods`, deliberately not `gem install` |

---

## T01 — Repo scaffold · **complete**

`~/dojo/rippr-flutter`, `flutter create --org com.rippr --project-name rippr`.

### Two deliberate deviations from the generated defaults

**Bundle id is `com.rippr.port`, not `com.rippr`.** `--org com.rippr` + name `rippr`
produces `com.rippr.rippr`, which is wrong either way. The choice of `com.rippr.port` is
deliberate and temporary: the plan requires the native app to stay installable as a
reference and fallback, and **two apps cannot share an applicationId**. Keeping them
distinct means both can sit on the same phone — which also enables the strongest possible
validation in T25: record the same ride on both simultaneously and compare the numbers.

> **T27 must switch this to `com.rippr`** at cutover. Recorded here because it is exactly
> the kind of temporary decision that silently becomes permanent.

The Android `namespace` stays `com.rippr` and the Kotlin source was moved from
`kotlin/com/rippr/rippr/` to `kotlin/com/rippr/` to match.

### Builds verified on both platforms

`✓ build/ios/iphonesimulator/Runner.app` and `✓ build/app/outputs/flutter-apk/app-debug.apk`.

Android needed three changes to the generated `build.gradle.kts`:

```kotlin
compileSdk = 37      // a dependency demands it; the build fails outright on 36
minSdk     = 26      // parity with the native app (Android 8.0)
targetSdk  = 36
```

`sdkmanager` cannot fetch `platforms;android-37` from the stable channel — it reports
"Failed to find package". **AGP installed it automatically** during the build (as
`android-37.0`), along with CMake 3.22.1 and a 2.8 GB NDK, because the licences had
already been accepted. Convenient, but see the disk note below.

### ⚠ `flutter_foreground_task` is on two deprecation paths

Both warnings name the same package — the one chosen for Android background liveness in
T12:

- **iOS:** does not support Swift Package Manager. *"This will become an error in a future
  version of Flutter."*
- **Android:** applies the Kotlin Gradle Plugin. *"Future versions of Flutter will fail to
  build if your app uses plugins that apply KGP."*

Neither breaks today's build. Both should be re-checked at T12, and they strengthen the
case for the `LocationSource` seam in T10 — the background layer needs to stay swappable.

### The disk problem was real, just not where the plan predicted

The plan braced for SDK *installs* filling the disk. The actual consumption was **builds**:
free space fell from 22 GiB to **3.9 GiB** during the first Android build — Gradle caches
grew 3.7 → 8.0 GB, the NDK added 2.8 GB, and `build/` alone reached 2.7 GB.

Recovery, in order of how safe each step was:

| Action | Reclaimed |
|---|---|
| Delete `build/` + `flutter clean` + the native app's `app/build` | ~3.8 GiB |
| Prune Gradle caches for versions **no project uses** (9.5.0, 9.7.0), `build-cache-1`, and the native project's 8.11.1 distribution | ~3 GiB |

Ended at **11 GiB free (94% used)**. `modules-2` (1.9 GB) was deliberately **kept** —
deleting it forces a re-download of every dependency, which is a real time cost for space
we do not currently need. Only two Gradle versions are actually in use: 9.3.1 (this
project) and 8.11.1 (the native app, whose distribution cache re-downloads on demand).

**Standing risk for T24:** the Android emulator needs a fixed 7.4 GiB free to boot. At 11
GiB that works, but one more full build cycle could eat the margin. Run `flutter clean`
before booting the emulator.

### Dependencies

`flutter_riverpod` · `go_router` · `drift` + `drift_flutter` · `path_provider` ·
`geolocator` · `flutter_foreground_task` · `permission_handler` · `flutter_map` +
`latlong2` · `share_plus` · `shared_preferences` · `http` · `synchronized`.
Dev: `drift_dev`, `build_runner`, `mocktail`, `integration_test`.

**`sqlite3_flutter_libs` resolves to an `+eol`-tagged release (0.6.0+eol).** It was added
explicitly at first, then removed — `drift_flutter` depends on it transitively regardless,
so the pin belongs to drift, not to us. Worth watching when drift next majors, but not
actionable now.

---

## T02 — Geo utilities · **complete**

`lib/src/geo/geo.dart` + `test/geo_test.dart`. **23/23 passing.**

Ported structurally faithfully from `com.rippr.geo.Geo`: haversine (with the
`asin(sqrt(a))` conditioning note), iterative Douglas–Peucker with an explicit stack,
equirectangular perpendicular distance with clamped projection, null-on-empty bounds,
path length. Every test case and tolerance carried over unchanged, including the
Calgary–Edmonton 280.9 km figure that was corrected during v2 after the *test* proved
wrong rather than the code.

**Deliberate API divergence:** Kotlin's `object Geo` namespace became top-level functions,
which is idiomatic Dart. `LatLon` is our own type rather than `latlong2`'s `LatLng` — the
pure layer must not depend on the map package; conversion happens at the render boundary.

### Two things went wrong

**`library;` after the import.** Dart requires the library directive before all other
directives. Caught immediately by the compiler — noted only because a file-level doc
comment is otherwise easy to attach wrongly.

**A hang that was not a hang.** `flutter test` and `flutter build ios` were run
concurrently and both sat at 0% CPU for minutes. The suspicion was Rosetta, because
`flutter_tester` lives under `artifacts/engine/darwin-x64/` — **that was wrong**: the
binary there is arm64 and the directory name is legacy. The real cause was
`ibtool`/`actool` spawning `IBAgent-iOS` and `AssetCatalogSimulatorAgent`, which deadlock
against a booted simulator. Run alone with simulators shut down, the same suite finishes
in under a second.

Two lessons, both echoing v2's harness troubles:
- **Do not run an iOS build against a booted simulator** if anything else needs it.
- **A killed background job still reports exit code 0.** Both jobs "completed
  successfully" *because they were killed*. Never read a success code from a process you
  terminated — re-run it cleanly.

---

## T03 — Telemetry, formatting, ephemeral state · **complete**

`telemetry.dart` (msToKmh, sanitizeSpeedKmh, isUsableFix, formatDuration, encodeBatch),
`ui/format.dart`, `telemetry/live_telemetry.dart`. **8 tests.**

Also added `domain/models.dart` — `Trip`, `Segment`, `TrackPoint`, `TripState`,
`RideStats` as plain Dart with **no persistence dependency**. Drift will map *to* these
in T08 rather than the domain depending on the database. This is the Dart equivalent of
the discipline that made the Kotlin logic testable on the JVM.

**Kotlin `Float` becomes Dart `double`.** Dart has no float32. Widening is the right
call — a shim would be friction for sub-millimetre precision on GPS-derived values — but
it means speed-derived values cannot be compared bit-for-bit. See T07 for the measured
consequence.

**`Format` builds its `DateFormat` per call**, unlike the Kotlin original which captured
`Locale.getDefault()` once at class-init. Doing the same in Dart would freeze the format
for the process lifetime and ignore a locale change.

---

## T04 — Ride statistics · **complete**

`stats/ride_statistics.dart` including `ElevationAccumulator`. **21 tests.**

Ported structurally faithfully — moving average, reversal hysteresis,
`gainIncludingPending()`, and the `finish()` reconciliation against `lastRaw`. Nothing
was "improved" during translation, per the plan.

### The failure that proved the port correct

The ported elevation test failed: **50.86 m against Kotlin's 35 m bound.** That looks
exactly like a porting bug in the hardest code in the project.

It was not. The two languages' `Random(42)` are different streams. Rather than tune the
bound blind — which `docs/v3/BACKLOG.md` explicitly warns against — the question was
settled by building the T07 harness early and driving **both implementations from one
shared LCG**:

```
noisy_gain = 38.959594555022136   ← Kotlin
noisy_gain = 38.959594555022136   ← Dart
```

Bit-identical. The port is exact.

**This also found something about the native app.** On the shared fixture the algorithm
yields ~39 m, which would **fail Kotlin's own 35 m bound**. The native test passes on
seed luck, not on a property of the algorithm. A sweep of 25 Dart seeds spanned
24.7–46.7 m (median 36). The Dart test now uses the shared LCG, asserts bit-equality
with Kotlin, and sets its bound from measured behaviour with headroom.

> Worth carrying back to the native repo if it is ever revived: that guard is weaker
> than it looks.

---

## T05 — Live accumulator · **complete**

`recording/ride_accumulator.dart`. **12 tests, green on the first run.**

The cross-batch anchor is intact — `distance across many small batches matches one big
batch` is the guard, and its absence under-reports distance by a few percent invisibly.

---

## T06 — Export writers · **complete**

`export/ride_export.dart`. **15 tests, green on the first run.** Parsed with a real XML
parser and `jsonDecode`, not substring matching, exactly as the Kotlin suite did.

---

## T07 — Cross-language parity harness · **complete**

`tool/parity/` — `run.sh`, `main.kt` (Kotlin oracle), `probe.dart`. Run it with
`JAVA_HOME` set to a 17–21 JDK; needs `kotlinc` (`brew install kotlin`, ~95 MB).

It copies `Geo.kt`, `RideStatistics.kt` and `RideExport.kt` **verbatim** from the native
repo and compiles them against minimal stand-ins for the Room-annotated holders and one
`Telemetry` constant. The files under test are never reimplemented.

### Result

Every key byte-identical across 20 fixtures, including:

| | |
|---|---|
| `noisy_gain` | `38.959594555022136` — the elevation accumulator, to the last digit |
| `simplify_count` | 218 of 21,600 points, identical Douglas–Peucker decisions |
| `gpx_len` / `gpx_fnv` | GPX output byte-identical, including the escaped hostile name |
| `geojson_len` / `geojson_fnv` | GeoJSON byte-identical |

**The single accepted difference** is `run_avg_speed`: `40.030228` (Kotlin `Float`) vs
`40.03022888407912` (Dart `double`) — precisely the divergence predicted in T03. The
harness prints this explanation on failure so a future reader is not left guessing.

### Two traps hit while building it

**`String.hashCode` is not comparable across languages.** The first version compared
Java's and Dart's hashes of the GPX output — which would have "failed" forever for no
reason. Replaced with an FNV-1a implemented identically in both.

**A missing jar silently passed as success.** When `RideExport.kt` was not yet copied,
compilation failed, `java -jar` errored, and the pipeline continued. `run.sh` now checks
for the jar and exits non-zero. This is the same class of bug as the v2 sweep that
returned four identical results because a compile error hid behind `/dev/null` — the
comment in `run.sh` says so explicitly.

---

## Phase 1 complete

**79 tests passing, `flutter analyze` clean.** ~890 lines of Kotlin logic ported, with
its ~965 lines of tests, and proven equivalent rather than assumed equivalent.

Next: **T08 (Drift schema)**, the first task that touches persistence.

---

## T08 — Drift schema · **complete**

`lib/src/data/database.dart` (+ generated `database.g.dart`). **16 tests.**

Mirrors `app/schemas/com.rippr.data.AppDatabase/2.json`: three tables, CASCADE foreign
keys, indices on `tripId` / `segmentId` / `synced`, WAL with `synchronous = NORMAL`.
Starts at Dart schema version 1 with a real `MigrationStrategy` — the destructive
fallback never comes back.

### Foreign keys are OFF by default in SQLite

Room switched them on for us. **Drift does not.** Without `PRAGMA foreign_keys = ON` in
`beforeOpen`, every `CASCADE` in the schema is decorative and deleting a trip silently
orphans all of its points. There is now a test that reads the pragma back and asserts it
is `1`, because this is invisible until data is already wrong.

### Two collisions worth recording

**Drift generates row classes named after the table.** `Trips` → `Trip`, colliding with
the domain model of the same name and producing 21 confusing analyzer errors of the form
*"Trip can't be assigned to Trip"*. Fixed with `@DataClassName('TripRow')` etc. The
mapping functions `_toTrip` / `_toSegment` / `_toPoint` convert row → domain, so the
domain layer stays unaware Drift exists.

**Drift snake_cases column names.** `speedKmh` became `speed_kmh`, which broke the one
raw-SQL query (the live stats aggregate) with `no such column: speedKmh`. Rather than 30
`.named()` annotations, `build.yaml` sets `case_from_dart_to_sql: preserve`. That keeps
the schema column-for-column identical to Room's, lets the raw SQL stay byte-identical to
the Kotlin DAO query it was ported from, and leaves a Room-file importer possible later.

### The instrumented-to-unit win, realised

`SchemaTest` needed a device and an emulator. The Drift equivalent runs on the Dart VM in
well under a second with nothing booted. `TripRepositoryTest` and `MergeTest` (441 more
lines) should convert the same way in T09.

**95 tests passing, analyze clean.**

---

## T09 — Trip repository · **complete**

`lib/src/data/trip_repository.dart`. **26 tests, green on the first run.**

Every lifecycle transition ported: `startTrip` (adopting, never duplicating), `pauseTrip`,
`resumeTrip`, `completeTrip`, `discardTrip`, `renameTrip`, `deleteTrip`, `mergeTrips`,
`recomputeAggregates`. Each runs inside `_db.transaction` and each is idempotent, because
the platform can restart the recorder from any state.

`Room.withTransaction` maps onto Drift's `transaction()` almost exactly, so this was the
most mechanical port so far.

### What the tests protect

- **Adoption over rejection.** `startTrip` on an already-active trip returns the *same*
  handle rather than opening a second trip — the behaviour that makes a process kill
  survivable.
- **Merge never joins segments.** The boundary between two merged rides stays a segment
  boundary, exactly like a pause. The regression test puts the two rides a degree of
  latitude apart and asserts the ~111 km gap never reaches `distanceM`.
- **Aggregates are recomputed, not summed** — because distance is not additive across
  that gap.
- **Rename collapses empty and whitespace-only input to null**, so a stored `""` can
  never diverge from the UI's date-label branch.
- **Merge rejects** self-merge, an active trip, and a missing id, and leaves no orphans.

### One piece of speculative code removed

A `mergeTableUpdates` helper was written to nudge Drift's stream queries after
re-parenting, then deleted before commit: Drift's own `update()` already notifies
dependent streams, so it earned nothing and would have been misleading scaffolding.

### The instrumented-to-unit win, totalled

`TripRepositoryTest` + `MergeTest` + `SchemaTest` were **666 lines of instrumented tests
requiring a booted emulator**. All three are now plain unit tests finishing in about two
seconds with nothing running.

---

## Phase 2 complete

**121 tests passing, `flutter analyze` clean.** The data layer is done and the domain,
statistics and export layers above it are proven equivalent to the Kotlin original.

Next: **Phase 3, the recording engine** — the risky phase. T10's `LocationSource` seam
first, then the pipeline, then the two platform liveness stories. The
`flutter_foreground_task` deprecation warnings recorded under T01 become relevant at T12.

---

## T10 — Location source seam · **complete**

`lib/src/recording/location_source.dart`: `LocationFix`, `LocationSource`,
`LocationException`, and `FakeLocationSource`.

`LocationFix` is deliberately **not** `TrackPoint` — a fix has no trip or segment
identity. Those ids are stamped on by the engine at creation, which is the whole basis of
the pause guarantee.

### `geolocator` can replace `flutter_foreground_task` entirely

`geolocator_android` ships `ForegroundNotificationConfig`, which raises a foreground
service with `foregroundServiceType=location` for as long as the position stream is
active, and exposes `enableWakeLock` and `setOngoing`. That is **everything**
`TrackingService` used a foreground service and a `PARTIAL_WAKE_LOCK` for.

Dropping `flutter_foreground_task` would remove both deprecation paths recorded under
T01 (no Swift Package Manager on iOS; applies KGP on Android) at no cost to the design.

**One parity casualty:** geolocator's notification config has no support for *actions*,
so the notification would be display-only — the native app's Pause/Resume buttons in the
shade would be lost. Decision deferred to T12, where it actually bites. The seam means
neither choice touches the engine.

---

## T11 — Recording pipeline · **complete**

`lib/src/recording/recording_engine.dart`. **24 tests.**

The Kotlin `Channel(UNLIMITED)` + blocking-`receive` writer coroutine becomes a plain
`List` buffer plus a periodic `Timer`. Dart has no blocking receive and its single
threaded event loop makes one unnecessary — the guarantee is unchanged: the fix callback
only appends and returns, so disk latency can never stall GPS. `Mutex` becomes
`synchronized`'s `Lock`, serialising the periodic flush against explicit drains at pause,
stop and discard.

All five actions ported: start (adopting), pause, resume (via start), stop, discard, plus
`restoreAfterProcessDeath`.

---

## 🐞 A real bug found in the native app

**`TrackingService.restoreAfterProcessDeath` does not do what its comment says.**

```kotlin
// Resume into a *new* segment: the time the process was dead is a real
// gap in the recording and should render as one.
trips.resumeTrip(System.currentTimeMillis())?.let { handle ->
```

`resumeTrip` calls `adoptOrOpenSegment`, which returns the **existing open segment** if
there is one. After a *pause* that is correct, because pausing closes the segment first.
After a *crash* nothing closed it — so the adopt branch wins and the stated intent is
silently not met.

### The consequence is data corruption, not cosmetics

Points either side of the dead time land in one segment. `RideStatistics.compute` groups
by segment, and it is the **authoritative** pass that overwrites the live estimate when
the trip completes. So the gap gets measured as if it had been ridden.

Measured, by reverting the fix and running the guard:

```
the dead time leaked into distance: 111217.31924957958 m
```

**111 km of phantom distance** added to a ride because the process died and the rider
relaunched somewhere else. The map would also draw a straight line across roads never
ridden — exactly the artefact segments exist to prevent.

The live accumulator gets this right (it is re-seeded with a null anchor). The
authoritative recomputation then overwrites the correct figure with the wrong one.

### The fix

`TripRepository.resumeIntoNewSegment` closes the stale segment and opens a fresh one.
The stale segment is closed **at its last recorded point**, not at `now` — recording
genuinely stopped when the process died, and `computeSummary` sums closed segment spans
for elapsed time, so closing at `now` would bill the dead time as ride time.

This is a deliberate, documented **departure from parity**. The plan says port bugs
faithfully, and that holds for the elevation drift where a "fix" would make differential
testing ambiguous. It does not hold here: the code contradicts its own stated intent and
the result is silently wrong data.

> Worth carrying back to the native app if it is ever revived. Second finding of its kind
> after the seed-lucky elevation bound.

### And a near-miss worth recording

The first version of the guard **passed with the bug still present**. The fixture called
`pause()` to flush points to disk — but pausing *closes* the segment, so the crash state
was never reproduced. It was only caught by deliberately reverting the fix and checking
the test failed.

The rewritten fixture writes points through the repository directly, leaving the segment
open exactly as a crash does. It now fails at 111 km with the native behaviour and passes
with the fix.

**A regression test nobody has watched fail is not yet a regression test.**

**145 tests passing, analyze clean.**