background_geo_tracker

• Breaking: GeoUploadConfig takes one url, not baseUrl and path. The uploader never had a use for the two halves apart — it joined them back together and posted to the result. All the split bought was a way to be wrong about who owns the slash between them, re-answered by every caller, with the caller that got it wrong finding out as a 404 inside a background uploader nobody watches. Callers pass the finished endpoint:

  GeoUploadConfig.standard(url: 'https://api.example.com/v1/points', …)
  

  The method-channel key changes with it — base_url and path become url — and both platforms read only the new one. Nothing reads the old pair and nothing migrates it: an install carrying a stored config from an earlier build reads an empty URL and posts nowhere until something calls configure again. Reinstall, or make the host re-configure on launch.

• The status says why the queue is not draining. GeoTrackingStatus gains uploadUrl — the endpoint the native uploader holds, which is not always what the app believes it configured — and lastUpload, how the last drain ended: ok, no url — never configured, http 500, network: …, halted: credentials refused. The uploader gives up at five separate guards and did so in silence at every one, which made a growing queue on an otherwise healthy-looking collector impossible to explain from a log. Both fields decode with fallbacks, so a platform that has not implemented them still produces a usable status.

• Android: an unusable endpoint no longer kills the worker. Request.url throws IllegalArgumentException, which the worker's IOException catch never covered, so an empty URL took the drain down with nothing to show for it. Parsed up front now, and recorded as lastUpload.

• iOS: a session comes back after a reboot again. The relaunch path ran entirely through the plugin's own application delegate, and a plugin is registered only once the implicit FlutterViewController exists. In a scene-based app that view controller is instantiated when a UI scene connects — and iOS waking the app in the background for a significant location change connects none. So on the one launch the whole path was built for, nothing ran: no CLLocationManager, no drain, no points. A phone switched off and back on stayed silent, and the position the backend served for it was the one from the moment it powered down, until its owner opened the app by hand.
• One line for hosts: call AttractorGeoLaunch.resumeIfTracking() from your AppDelegate's application(_:didFinishLaunchingWithOptions:). That method is what a background launch does call. See the iOS setup section of the README, which until now said no AppDelegate changes were needed.
• The collector's "a point arrived" hook onto the uploader moved next to the resume, so both ways of bringing the native stack up wire it identically.

First release on pub.dev. Nothing about how tracking works changed; everything below is naming, packaging and documentation.

• Renamed from attractor_geo to background_geo_tracker, and AttractorGeoController to BackgroundGeoTracker with it. The old name said who wrote the package rather than what it does, which is no use to anyone finding it on pub.dev. Update the dependency, the import and the one class name; nothing else about the API moved.
• Storage identifiers deliberately kept as attractor_geo* — the SQLite filenames, the Keychain service, the UserDefaults prefix, the SharedPreferences files, the notification channel and the WorkManager job names. Renaming them would orphan the queue and the stored credentials of every install that upgrades, and leave an already-scheduled Android job that the new code no longer knows how to cancel. An upgrade from 0.2.0 keeps its undelivered points.
• Documentation rewritten for use outside the repository it was extracted from. The install snippet is a version, not a path, and the passages that pointed at wrapper classes living in that app are gone — reset() on sign-out and re-configure() on start() are now stated as the host's job, because outside that app nothing else does them for you.
• Dropped the App Transport Security and Android cleartext-HTTP sections. Both documented how to weaken a host app's network security to reach a plaintext development backend, which is a local concern and not something a package should be teaching.
• Every public member now carries documentation, and the package ships with a .pubignore, MIT licence and pub.dev metadata.

• Added reset() — ends the session and forgets what belonged to the account that was signed in: the stored upload credentials and every queued point. stop() deliberately keeps both. Without this the credentials and the queue outlived a sign-out, and the next account on the device would upload the previous one's track under its own name.
• Android now reports status changes. It previously published a status only when the backend refused a credential, so a UI watching statusChanges never learned that a session had started, that a permission dialog had been answered, or that permission had been revoked mid-session — while iOS reported all of them. Answered permission dialogs are picked up through a RequestPermissionsResultListener, which the plugin did not register at all.
• Android asks for POST_NOTIFICATIONS once at start, on API 33+. The foreground service's notification is required to be visible for the whole session and did not appear without the runtime grant. Refusing it hides the notification and does not stop the session.
• Stopping cancels the one-shot upload job too. Only the periodic one was cancelled, so a drain enqueued moments earlier could still wake up and upload using credentials the app had already discarded.
• Android honours uploadIntervalSeconds. The interval was served by WorkManager, whose floor for periodic work is 15 minutes, so a phone sitting still under batchSize held its points for a quarter of an hour while iOS shipped them in a minute — one setting meaning two different things. The foreground service, alive for the whole session anyway, now drives the drain itself and only enqueues when the queue is non-empty. The periodic worker stays as the fallback for a collector the OS has killed.
• iOS honours its own backoff. The periodic drain ran on its own schedule regardless of a backoff in progress, so a backend that was down was retried every upload interval and the computed backoff only ever delayed whichever attempt lost the race.
• One OkHttpClient across worker runs on Android, instead of one per run throwing away the connection pool.