Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/test_baseline.py`:
- Around line 127-131: Replace the shallow registration check with an end-to-end
CLI invocation: use click.testing.CliRunner to invoke the Click group `main`'s
"baselines" command (or add a new test) and pass a pytest `tmp_path` as the
output directory, then assert the runner.exit_code is 0 and that the expected
output files or artifacts are actually created under `tmp_path`; target the
`main` Click group and the `baselines` command so you exercise the command path
rather than only checking `main.commands`.

In `@vera_bench/baseline_runner.py`:
- Around line 127-168: The ProblemResult returns for the timeout, non-zero-exit,
and bad-JSON branches currently omit the real test count and default to
tests_total=0; update each of those ProblemResult constructors (the one returned
on timeout, the one when result.returncode != 0, and the one in the
json.JSONDecodeError except block) to include tests_total=tests_total so failed
runs retain the actual test count, keeping other fields (problem_id,
model="baseline", language="python", attempt, check_pass, run_correct,
error_message, wall_time_s, timestamp=_now()) unchanged.
- Around line 194-215: The temp directory created by tempfile.mkdtemp() stored
in work_dir leaks on errors; change to use tempfile.TemporaryDirectory() as a
context manager and set work_dir = Path(tmpdir.name) (or Path(tmpdir) depending
on API) so the directory is automatically removed on exit; wrap the block that
uses work_dir (the with Progress(...) loop that calls run_python_baseline)
inside the TemporaryDirectory() context and import tempfile.TemporaryDirectory
if necessary to ensure cleanup on both success and failure.

In `@vera_bench/cli.py`:
- Around line 223-238: The output JSONL file at output_path must be truncated
before appending new results so stale rows don't persist; before calling
run_all_baselines (which opens the file in append mode), clear or recreate
output_path (e.g., open output_path for write/truncate or unlink+touch) using
the existing output_path variable (and keep solutions_dir and problems as-is),
so subsequent compute_metrics()/report reads only the fresh run's records;
ensure this truncation happens after output_dir.mkdir(...) and before the
run_all_baselines(...) call.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@vera_bench/baseline_runner.py`:
- Around line 188-193: run_all_baselines currently accepts a language parameter
but always calls run_python_baseline; either remove the unused language
parameter or implement dispatching: inside run_all_baselines check the language
value and call run_python_baseline for "python" and raise NotImplementedError
(or add a TODO) for unsupported languages like "typescript"; update the function
signature and any callers if you remove the parameter, or add the dispatch logic
and a clear error for unsupported languages to avoid silently ignoring the
parameter.
- Around line 21-30: The _find_baseline_file function currently collapses both
zero and multiple glob matches into None; change it to return None only when no
matches are found and raise a clear ValueError when multiple matches exist to
surface naming conflicts. Specifically, in _find_baseline_file inspect the
matches list: if len(matches) == 0 return None; if len(matches) == 1 return
matches[0]; if len(matches) > 1 raise a ValueError that includes the prefix,
lang_dir and the list of matching paths (the variables matches, prefix, and
lang_dir) so callers and logs can debug the conflict.