# Architecture

Everything lives in [`app.py`](../app.py). There is no package structure.

## Components

| Class / function | Responsibility |
|------------------|----------------|
| `DataStore` | Owns `orte.csv`. Loads it into a pandas `DataFrame`, exposes read helpers (`get_map_data`, `get_autocomplete_strings`, `total_count`, `find_duplicates`) and write helpers (`append_rows`, `update_row`, `delete_rows`). All writes go through `_save_csv` → rotating backup + atomic write. All columns are read as strings. Falls back to an empty frame (or the newest backup) if `orte.csv` is missing/empty/corrupt. |
| `geocode_city()` | Pure function: `(city, plz) -> (lat, lon) | None`. Tries several Nominatim queries with a configurable regional bias (`GEOCODE_REGIONS`). Raises `GeocodingUnavailable` if *every* query failed on a service/network error (vs. simply not finding the place). |
| `GeocoderWorker` | A `threading.Thread` (daemon). Holds a FIFO list of `(row_id, city, plz, callback)` items guarded by a lock, woken by an `Event`. Processes one item per second, calls `geocode_city`, then invokes `callback` **from the worker thread**. Each item is wrapped in `try/except` so one failure never kills the thread; after `SERVICE_ERROR_THRESHOLD` consecutive service errors it fires the optional `on_service_error` callback. |
| `parse_date()` / `parse_coord()` / helpers | Module-level pure functions for normalizing user input (date → ISO, coordinate strings → float, `"City (PLZ)"` splitting). Shared by the entry tab and `EditDialog`. |
| `generate_map()` | Builds `karte.html` from `DataStore.get_map_data()` with Folium. |
| `EditDialog` | `tk.Toplevel` modal dialog for editing one row. Returns its result via `self.result`. |
| `App` | `ttkbootstrap.Window`. Builds the UI, owns the `DataStore`, the `GeocoderWorker`, and the in-memory `_queue` list for Tab 1. |

## Data flow

### Adding entries
```
user types -> _on_add_row()
  -> append dict to self._queue
  -> insert row into Tab-1 Treeview (coords = "⏳ wird gesucht…")
  -> geocoder.enqueue(row_id, city, plz, self._geocode_done)

GeocoderWorker thread (1/sec):
  coords = geocode_city(...)
  -> self._geocode_done(row_id, coords)        [worker thread]
       -> self.after(0, self._apply_geocode_result, ...)   [hop to UI thread]
            -> mutate self._queue entry + update Treeview cell

user clicks "Alle speichern" -> _on_save_all()
  -> DataStore.find_duplicates() (status-bar hint only)
  -> DataStore.append_rows()  ->  CSV append + in-memory concat
  -> clear queue, refresh Tab 2, update autocomplete
```

### Editing / deleting
`EditDialog` -> `DataStore.update_row(df_idx, values)` or
`DataStore.delete_rows(indices)` -> full `df.to_csv()` rewrite -> refresh.

## Threading model

- Tkinter is single-threaded; all widget access must happen on the main thread.
- The geocoder runs off-thread so the UI never blocks for the 1-second-per-item
  rate limit.
- The **callback is invoked on the worker thread**, and every callback in
  `App` immediately does `self.after(0, ...)` to marshal back onto the UI
  thread. This is the load-bearing convention — any new callback must follow it.
  (`on_service_error` does the same.)
- The worker catches every exception per item, so a bug in a callback or a
  network error logs a traceback but never stops the queue.
- On window close, `_on_close()` calls `geocoder.stop()` (sets a flag + wakes
  the event) and then `destroy()`. The thread is a daemon, so a missed stop
  won't hang the process.

## Persistence model

- `orte.csv` is the single source of truth. It is read once at startup and then
  kept in sync in memory.
- **Every** write (append, update, delete) rebuilds the full `DataFrame` and
  goes through `DataStore._save_csv`:
  1. `_make_backup()` – copy the current file to
     `data/backups/orte-YYYYMMDD-HHMMSS.csv` (skipped if byte-identical to the
     newest backup); prune to `MAX_BACKUPS` (20).
  2. `_write_atomic()` – write to `orte.csv.tmp`, `fsync`, then `os.replace()`
     onto `orte.csv` (atomic on POSIX). A crash mid-write leaves the old file
     intact.
- A backup is also taken once at startup.
- Still **no file locking**: if the CSV is edited in another program while the
  app is open, the next in-app save overwrites those changes (but the pre-write
  backup captures them).

## Logging

`_setup_logging()` (called from `__main__`) attaches a `RotatingFileHandler` to
`data/app.log` (512 KB × 3). Logger name `arbeitsorte`, module-level `log`.
Records saves, map generation, geocoding failures, and any uncaught exception.
This is the file to ask for when diagnosing a problem on the user's laptop.

## External dependencies at runtime

| Dependency | Needed for | Offline behaviour |
|------------|-----------|-------------------|
| Nominatim (OSM) | geocoding new/edited entries | entries save with blank coords, shown as `⚠ nicht gefunden` |
| CDN (jsdelivr, jquery, cloudflare) | rendering `karte.html` | map page loads blank / unstyled |
| CartoDB tiles | map background | grey tiles |

Existing entries that already have coordinates do **not** need the network.