feat: add debug-pytest subcommand

[mvanhorn]

Summary

vibe-debug debug-pytest <test_id> runs a pytest test under debugpy and stops at the failing assertion frame, returning locals + traceback in the same shape as debug-python. The path most agent debugging actually takes - "this test is failing, why?" - now works without needing the agent to know the file/line of the assertion ahead of time.

Why this matters

Today debug-python covers python script.py, but pytest's collection / fixture / conftest machinery doesn't fit through that surface. Two adjacent agent-debugger projects ship a pytest path with different shapes:

• samefarrar/mcp-pdb ships start_debug(file_path, use_pytest=True) as a parameter on its main MCP tool
• markomanninen/mcp-debugpy ships pytest JSON reports for machine-readable failures

Microsoft's debug-gym paper (arXiv 2503.21557), which mcp-pdb cites as inspiration, used pytest as the harness for measuring agent debugger gains: Claude 3.7 Sonnet went from 37.2% on SWE-bench Lite without a debugger, to 48.4% with one, to 52.1% with debug(5). SWE-bench is mechanically "make this failing pytest pass." Without a pytest mode, vibe-debug isn't runnable against that benchmark.

Demo

The included examples/test_buggy_discount.py asserts apply_discount(120.0, "gold") == 102.0 against the existing buggy examples/buggy_discount.py. debug-pytest runs it, stops at the AssertionError frame, and dumps:

Pytest: test_buggy_discount.py::test_gold
Outcome: failed
Exception: AssertionError: assert 119.85 == 102.0
Stopped: test_buggy_discount.py:8 in test_gold
Locals:
  actual = 119.85
  loyalty_level = 'gold'
  price = 120.0

The agent gets the full state at the failure point in one round trip - no need to read the traceback, infer where to break, and re-run.

Changes

• cli.py: new debug-pytest subcommand parallel to debug-python with --pytest-arg, --break, --break-on-failure/--no-break-on-failure (default ON), --eval, --cwd, --python, --timeout, --locals-limit, --json. New _debug_pytest_payload mirrors _debug_python_payload's shape and adds a pytest block (rootdir, test_id, pytest_args, outcome, duration_seconds). New _continue_pytest_execution skips exception stops in _pytest, pluggy, site-packages, and anything outside the test's working tree so the debugger lands at the user's failing assertion, not pytest's internal exception bookkeeping.
• session.py: new DebugSession.launch_pytest spawns python -m debugpy --listen --wait-for-client -m pytest <test_id> and attaches via DAPClient. New set_exception_breakpoints and exception_info wrap the corresponding DAP requests. The launcher uses the user's --python (or default) so pytest must be installed in the target's environment - vibe-debug does NOT add pytest to its own dependencies.
• mcp_server.py: new debug_pytest MCP tool with the same input shape. debug_launch and debug_attach schemas gain optional exception_filters for symmetry with debug_pytest's breakpoint-on-failure behavior.
• examples/test_buggy_discount.py: a 7-line failing pytest against the existing buggy_discount.py. Used by runtime_proof.py to exercise the new path.
• tools/runtime_proof.py: extended to call debug_pytest and assert the failed AssertionError stop. The proved[] list grows from 14 to 15 entries.
• tests/test_cli.py: new tests for debug-pytest happy path (failing test → AssertionError stop), parser validation, and --no-break-on-failure mode.

Testing

18 tests pass (python -m unittest discover -s tests -v). New tests cover:

• tests/test_cli.py::test_debug_pytest_stops_at_failing_assertion - end-to-end run on the included example
• tests/test_cli.py::test_debug_pytest_no_break_on_failure_runs_to_completion - opt-out behavior
• tests/test_cli.py::test_debug_pytest_passes_through_pytest_args - --pytest-arg "-k pattern" shlex parsing

vibe-debug doctor passes. python tools/runtime_proof.py proves debug_pytest end-to-end.

Notes

This is the second cold PR in two days (the first is #1, --log for non-pausing logpoints). Both touch independent seams (logpoints adds to set_breakpoints; this adds a new subcommand and launch_pytest), so they can land in either order or independently.

The --break-on-failure machinery (set_exception_breakpoints + exception_info in session.py) is the same building block plan 019 (general --break-on-exception flag for debug-python) would use - this PR adds the infrastructure but only exposes it via pytest's failure mode. Generalizing it to debug-python --break-on-exception {raised,uncaught,all} would be a small follow-up if you want it.

Happy to convert this to a Discussion or close if the surface (subcommand vs flag, default break mode, pytest-not-as-dep) doesn't match how you'd want it shaped.