Job search in Denmark: aggregate jobnet.dk and jobindex.dk, rank matches with Laya

# Job classifier search
</div>


Collects job ads from jobnet.dk and jobindex.dk, filters them against a
candidate profile, ranks them with the Laya decision model, prints two tables
(IT, warehouse/logistics) and writes the same result to `jobs.md`.


## Example of work

| Ranked IT jobs | Warehouse / logistics |
| --- | --- |
| <img src="img/1.png" alt="Terminal table with ranked IT jobs" width="480"> | <img src="img/2.png" alt="Continuation of the IT table and the warehouse/logistics table" width="480"> |

## Run

Python 3.13 and `uv`. Commands are run from the project root: `profile.yaml`,
`.cache`, `labels.csv` and `jobs.md` are resolved relative to the working
directory.

```
uv sync
uv run python src/jobfit.py
```

The first run downloads the Laya multilingual checkpoint and fetches every job
page, about four minutes. Later runs use the cache and take seconds.

## Pipeline

1. Sources. `sources/jobnet.py` calls the Jobnet search API, `sources/jobindex.py`
   calls the Jobindex search API. Pages are fetched in parallel and cached.
2. Enrichment. `sources/enrich.py` opens each Jobindex ad page: internal ads give
   the full text, external ones only a teaser. Jobnet returns full descriptions.
3. Normalization. Both sources map to one dict: title, company, location, url,
   description, deadline, address, coordinates, source. Duplicates are dropped
   by title and company.
4. Filters. A keyword gate keeps ads that mention a profile skill or role.
   `blacklist_companies`, `blacklist_title_keywords` and
   `blacklist_description_patterns` from `profile.yaml` remove the rest.
5. Scoring. Laya answers nine typed questions per ad: job family, seniority,
   skill overlap, role fit, English/Danish, Copenhagen, spam, sponsorship. Raw
   answers are cached, so percents can be