feat(adapters): add the task-source adapter foundation and clawbench-sources CLI

[vaibhavdabas16]

Advances #72. This is step 1 of the incremental plan in that issue — _base.py + schema.py + native.py + __init__.py + the sources CLI — with no behaviour change for bundled tasks. The benchmark adapters themselves (claw-eval, wild-claw-bench, webarena, mind2web, webvoyager) follow in later PRs, so this does not close the issue.

What's here

src/clawbench/adapters/:

• schema.py — ClawBenchTask, the shared target type. A superset of test-cases/task.schema.json plus provenance (source, source_id), with ScoringLayer, AdapterWarning, and ExtraInfo.
• _base.py — AdapterBase, the @register registry, source-spec parsing (name or name:/path), cache-dir resolution, and the CLAWBENCH_OFFLINE gate.
• native.py — identity loader for the bundled corpora; the reference implementation for the field-mapping docstring convention.
• cli.py — clawbench-sources.

Decisions worth reviewing

Two deliberate deviations from the issue text:

1. CLI shape. The issue sketches clawbench sources, but clawbench is the TUI and every non-interactive command here is its own clawbench-<verb> entrypoint. This adds clawbench-sources with list / show / cases rather than growing a subcommand tree on the TUI. sources refresh is deliberately absent — it needs the _pins.yaml fetch machinery, which is step 5.

2. Dataclasses, not Pydantic. The issue specifies a Pydantic ClawBenchTask. ClawBench has no Pydantic dependency, and its other registry (run_support/harness_registry.py) validates frozen dataclasses with hand-rolled checks and explicit error messages. This follows that rather than adding a runtime dependency for one module. Happy to switch if you'd rather take the dependency.

The scoring-layer contract, from §4 of the issue, is the part most worth a look. An adapter declares which layers its tasks can be judged by; a layer a source cannot support is meant to score null rather than 0, so leaderboard aggregation never conflates "the agent failed" with "this task was never scored on that axis". ClawBenchTask.to_task_json() enforces the corresponding invariant on the write side: it refuses to render a native task.json for a task with no interception contract instead of inventing an eval_schema. Wiring null through the recording itself belongs with the first adapter that actually needs it.

Time limits. ClawBenchTask.time_limit is in minutes, matching task.json and the container watchdog. Most upstream schemas use seconds, so each adapter converts — the docstring says so explicitly to keep that from being silently wrong.

Native default_path() is the bundled test-cases/ root, and load() accepts either a suite directory (*/task.json) or a corpus root (*/*/task.json), so clawbench-sources cases clawbench-native --path test-cases/v2 and the bare form both work.

Testing

tests/test_task_source_adapters.py — 26 cases covering the registry and its error message, source-spec parsing (including that a bare Windows path isn't mistaken for name:path), cache-dir precedence, the offline flag, schema validation and task.json rendering checked against the committed task.schema.json, warning formatting, the native loader on a suite dir, a corpus root, and the real bundled V2 corpus, plus each CLI subcommand. clawbench.adapters.cli is added to the test_cli_entrypoints help-module list.

Full suite passes locally (301 passed, 10 skipped); ruff check, ruff format --check, and pyright are clean on the new files.