feat(benchmarks): add 25-defect dataset, manifest loader, and Pass@k evaluation harness

[Abdurrafay19]

What this adds

Implements the benchmark evaluation harness described in the project spec: a
25-defect synthetic dataset with oracle tests, a typed manifest loader, and a
harness that runs the full LangGraph agent against each defect and grades the
result against a hidden oracle test — never against the SLM's own
self-written tests, since a model that writes weak tests would otherwise
score as "fixed" even when it isn't.

Files added

• benchmarks/manifest.py — Pydantic schema (DefectRecord, DefectManifest)
  and loaders for the dataset manifest, source files, and oracle tests. Fails
  fast (raises) on missing files or malformed JSON, unlike the tools in
  patchwork/tools/, since this loads a fixed local dataset rather than
  untrusted SLM output or flaky subprocesses. Validates defect IDs are unique
  at load time via a model_validator, not just in tests.
• benchmarks/dataset/ — 25 buggy Python files across 5 categories (mutable
  default arguments, unhandled NoneType, off-by-one slicing, float precision,
  resource leaks), each with a corresponding oracle test in
  benchmarks/dataset/oracle_tests/.
• benchmarks/evaluate.py — evaluate_defect() runs one defect through
  build_patchwork_graph() and grades the final patched code against the
  oracle; evaluate_all() runs the full 25-defect set and computes Pass@1 /
  Pass@overall rates plus VRAM/duration telemetry via profile_call.
  Crash-isolated per defect: a transport-level failure
  (RuntimeError/ConnectionError/TimeoutError/OSError) on one defect
  is caught and recorded, not allowed to kill the rest of the batch.
• tests/test_manifest.py, tests/test_evaluate.py — real-dataset,
  real-subprocess integration tests. structured_llm is mocked everywhere
  (no live Ollama call in CI); grading, sandbox execution, and file I/O are
  real.

Config changes

• pyproject.toml — added [tool.pytest.ini_options] (testpaths),
  [tool.ruff] (target-version, extend-exclude for
  benchmarks/dataset/, which is intentionally-broken specimen code not
  meant to pass lint/type checks), [tool.mypy] (strict = true made
  permanent, exclude for the same dataset path), and a pynvml
  stub-suppression override consumed by patchwork/telemetry/profiler.py.

Testing

128 tests passing. Coverage: manifest.py 100%, evaluate.py 74% (remaining
gap is main()/save_results()/_print_summary() — CLI glue meant to run
once against a real model, not unit-tested against a mock). Total project
coverage 90%, above the ≥80% target in the project spec.

ruff check, ruff format --check, and mypy --strict all clean across
patchwork/, tests/, benchmarks/.