garmin-img-format-parsing/stage-2-parse-stage-1/readme.md

Found the two real problems.

First, your crash:

* `minidom.parseString(...)` was reparsing the entire generated XML in memory
* one of the landmark names or tag values contains an XML-illegal control character
* on a huge file, that blows up exactly like your `ExpatError: not well-formed (invalid token)`

Second, the sampling:

* the previous script was taking the first few examples in a group
* now it samples **spread out across the group**, so the examples are taken from different positions in the dataset instead of only the front

Updated script:
[landmarks_csv_to_osmand_v3.py](sandbox:/mnt/data/landmarks_csv_to_osmand_v3.py)

What changed:

* removed the fragile `minidom.parseString()` round-trip
* writes XML directly with `ElementTree`
* sanitizes text and tag values to strip XML-invalid control characters
* supports `.gz` outputs correctly
* group examples are now **spread-out samples**
* interactive group renaming still works, but now the shown examples are better distributed
* loader still accepts both raw landmark CSVs and slimmer derived CSVs

Use it instead of the old one:

```bash
python landmarks_csv_to_osmand_v3.py *.csv --osm bgmountains_poi.osm.gz
```

If you want to inspect grouping before writing:

```bash
python landmarks_csv_to_osmand_v3.py *.csv --show-groups --group-examples 5
```

If you want interactive naming with spread-out examples:

```bash
python landmarks_csv_to_osmand_v3.py *.csv --show-groups --interactive-group-names --group-examples 5 --gpx landmarks.gpx
```

If you want POI-only OSM for OsmAndMapCreator:

```bash
python landmarks_csv_to_osmand_v3.py *.csv --osm landmarks.osm.gz
```

Then:

```bash
java.exe -Xms256M -Xmx4096M -cp "OsmAndMapCreator.jar;lib/*" net.osmand.MainUtilities generate-poi C:\path\to\landmarks.osm.gz
```

The main fix is this: the script now treats dirty text safely instead of trusting all names/tags to already be valid XML.

If you want, the next refinement should be a `--debug-bad-rows` mode that writes out the original rows whose text had to be sanitized, so you can see exactly which source landmarks were malformed.