Treat the negate key as a strict boolean in the skill-evals runner

[justinmclean]

Summary

tools/skill-evals/src/skill_evals/runner.py reads the assertion negate key
with a truthiness check, so any non-empty value inverts the assertion.

Background

negate: true inverts a predicate so a case can assert the absence of
something, which is how several suites test that injected content was not
obeyed. The current check accepts any truthy value, so "negate": "false",
"negate": 0.1 or "negate": [] all silently do something other than what the
fixture author meant. A typo flips the assertion instead of erroring, and an
inverted security assertion passes for the wrong reason.

Where to look

• runner.py line 719 - if spec.get("negate") and holds is not None:.
  Lines 39 and 712 document the key as "negate": true.
• tools/skill-evals/tests/ - where a regression test belongs.

Acceptance criteria

• negate is honoured only when it is the boolean true.
• A non-boolean value is rejected with a clear error rather than silently inverting.
• A test covers boolean true, boolean false, and one non-boolean value.
• uv run --directory tools/skill-evals --group dev pytest passes.

Estimated effort

~1 hour for someone new to the codebase.

[surajthedev]

Hi @justinmclean, I'd like to work on this issue!

I'll update the check in runner.py:719 to only honor negate when it's a strict boolean true, and raise a clear error for any non-boolean value instead of silently inverting the assertion. I'll also add a regression test in tools/skill-evals/tests/ covering boolean true, boolean false, and a non-boolean value case.

[justinmclean]

Thanks @surajthedev that sounds like the right approach.