You are an agent, your current working directory is /app.

You can use the tools available to you to interact with the computer to assist the user in completing tasks.

# Repair uneven-row joins in petl

An ETL pipeline built on the frozen petl tree in `/app/vendor/petl` receives
iterable tables whose data-row widths do not always match their headers. The
current relational join implementations can raise `IndexError`, shift a value
under the wrong output field, or emit a row shorter or longer than the output
header.

Repair the existing implementation in both:

- `/app/vendor/petl/petl/transform/joins.py`
- `/app/vendor/petl/petl/transform/hashjoins.py`

The public sort-based and hash-based join families must share these semantics:

1. Before relational processing, each data row behaves as a row of exactly its
   source header width. Missing positions use the operation's `missing` value
   when that API has one, otherwise `None`. Values beyond the source header do
   not leak into the joined schema.
2. Key selection by field name, index, natural key, distinct `lkey`/`rkey`, or
   compound key remains correct even when a key component is absent from a
   short row.
3. Inner, left, right, outer, lookup, anti, and cross joins preserve documented
   match cardinality and duplicate-key behavior. The equivalent hash joins
   must agree semantically while retaining their normal input-order behavior.
4. Prefix handling, custom missing sentinels, lazy iteration, cached and
   uncached execution, and repeated materialization continue to work. Do not
   mutate caller-owned input rows.

Do not modify the visible workflow, existing tests, or other library modules.
The verifier protects those assets and exercises additional private tables.

After repairing the source, run:

```sh
cd /app
python3 -m py_compile vendor/petl/petl/transform/joins.py vendor/petl/petl/transform/hashjoins.py
python3 workflow.py /app/output.json
```

`/app/output.json` must use schema `petl-uneven-joins.v1` with the keys
`case_order`, `cases`, `row_counts`, and `all_rectangular`; it must be produced
by the supplied workflow from the repaired source tree.