[babaliauskas]

Bumps the version to 1.6.2 and dates the CHANGELOG section (#69, #70).

Highlights:

• max_cost_increase / max_latency_increase are judged per example: latency on a 20%-trimmed geometric mean of per-example ratios, cost on the paired total. Values are signed and the denominator counts example pairs.
• semantic and llm_judge no longer score narration beside a tool call.
• The report shows "—" for a skipped test's effect and range, and shows effects signed.

After merge: tag the merge commit v1.6.2 to publish (CONTRIBUTING.md › Releasing).

🤖 Generated with Claude Code