[Feature]: Add vakra small training data and make it the default --m3-data source

[haroldship]

Feature Request

Bundle the vakra small M3 training data with the repository and make it the default data source when --m3-data is passed without a path argument.

Motivation / Problem

Today, ./benchmarks/m3/eval.sh --m3-data requires an explicit path to a zip or directory:

if [[ -z "${2:-}" || "$2" == --* ]]; then
    echo "Error: --m3-data requires a path (zip file or directory)" >&2
    exit 2
fi

There is no out-of-the-box default, so every user must know the location of the M3 data before they can run --m3-data mode. The comment in m3_registry_m3_data.yaml already anticipates a "default zip" invocation (./benchmarks/m3/eval.sh --m3-data # default zip), but that path does not exist yet.

Use Case

• Developers and CI jobs want to run ./benchmarks/m3/eval.sh --m3-data against a well-known, representative dataset without specifying an external path every time.
• The vakra small dataset is compact enough to ship alongside the repo and covers the capability domains already listed in m3_registry_m3_data.yaml.
• Having a canonical default lowers the barrier to entry for new contributors running M3 evals locally.

Proposed Solution

1. Add the vakra small data — commit (or reference via a just download-m3-data task) the vakra small zip/directory under benchmarks/m3/data/vakra_small/ (or as a tracked zip artifact).

2. Wire it as the default — in eval.sh, change the --m3-data argument parsing so that omitting a path falls back to the bundled vakra small data:

   --m3-data)
       M3_DATA=true
       if [[ -z "${2:-}" || "$2" == --* ]]; then
           # No path supplied — use bundled vakra small data
           M3_DATA_PATH="$SCRIPT_DIR/data/vakra_small"
       else
           M3_DATA_PATH="$2"
           shift
       fi
       shift
       ;;

3. Update help text — document that --m3-data (with no path) runs against the bundled vakra small dataset.

4. Update m3_registry_m3_data.yaml — align the domains lists with whatever capabilities and domains vakra small actually contains.

Alternatives Considered

• Keeping the required-path behavior and adding a separate --m3-data-default flag. Rejected: adds flag surface with no benefit; the implicit default is cleaner.
• Downloading the data at eval time via setup_m3.sh. This works but requires network access and makes the no-arg invocation slower and less reproducible.

Priority

Medium

Additional Context

• Related files: benchmarks/m3/eval.sh, benchmarks/m3/config/m3_registry_m3_data.yaml, benchmarks/m3/m3_data_loader.py
• The vakra scoring pipeline (benchmarks/m3/m3_vakra_score.py) already consumes the data shape that M3DataLoader produces, so no scoring-side changes are expected.