evals: prevent silent overwrites in results writer artifacts

[renaudcepre]

What

Closes #119. Two silent data-loss gaps in the per-case markdown artifacts (results_writer.py).

1. Case-name collisions

Names were sanitized with re.sub(r"[^\w\-]", "_", name), so a/b and a:b both became a_b.md — the second silently overwriting the first. The writer now tracks the stems used in each run dir and appends a numeric suffix on collision (a_b.md, a_b-2.md), so every case keeps its own file. (Also falls back to case when a name sanitizes to empty.)

2. Same-second / racing run directories

Run dirs are timestamped to the second with mkdir(exist_ok=True), so two runs of the same suite within one second — or two racing processes — merged their case files into one directory. _make_run_dir now creates with exist_ok=False and bumps a numeric suffix until mkdir wins. mkdir is atomic, so this is also safe across concurrent processes.

Tests

New tests/evals/test_results_writer.py (kept separate to avoid growing the already-large test_e2e.py, cf. #98):

• colliding sanitized names produce two distinct files, each with its own content;
• empty name falls back to case.md;
• successive run dirs are distinct and all exist;
• a pre-existing timestamp dir forces the suffix branch.

Full suite: 1246 passed, ruff clean, mypy --strict clean.

---

Generated by Claude Code