M2: run_tests tool and the edit-test-repair loop

The fourth v0 tool: run pytest against a fixed path and return the result. Deliberately not a shell — narrow enough to be safe, and the only execution feedback this task class needs.

The interesting part is what comes back. Raw pytest output is long, and most of it is noise that will crowd a small context window. Truncate and summarise: the failing test name, the assertion, and the relevant traceback frame. How aggressively to trim is an empirical question the transcripts will answer.

Done when, given a spec and a failing test file, LIZA writes a module, runs the tests, reads the failure and revises — at least one unaided cycle against a local model. This is the project's early definition of done.

Begin by running pytest against the specified failing test file and inspect the current v0 tool entry point. Implement the narrow test runner and output trimming so the failing test name, assertion, and relevant traceback frame are returned; done when LIZA completes at least one unaided edit-test-repair cycle with a local model.