bench: re-run with-superpowers vs without-superpowers after grader fix

[christso]

Objective

Re-run the with-superpowers vs without-superpowers skill benchmarks now that the LLM grader bug (#982, PR #983) is fixed.

Context

The initial benchmark runs in agentv-bench-skills produced misleading results because the with-superpowers eval used prompt: in assertions, which triggered bug #982 (bare prompt replacing entire grader template). The scores showed with-superpowers performing worse, but this was a grader artifact, not a real skill degradation.

Previous results (tainted by bug)

Experiment	Target	Pass Rate	Avg Score
without-superpowers	gemini	100%	1.000
without-superpowers	azure	50%	0.735
with-superpowers	gemini	50%	0.750
with-superpowers	azure	0%	0.250

Steps

1. Merge PR fix(core): extract system messages in prompt builder for LLM grader #983 (grader fix)
2. Install superpowers: cd agentv-bench-skills && ./scripts/setup-superpowers.sh
3. Re-run both experiments against at least 2 providers:

   agentv eval evals/reasoning/logic-puzzle.EVAL.yaml --target gemini --experiment without-superpowers
   agentv eval evals/reasoning/logic-puzzle.EVAL.yaml --target gemini --experiment with-superpowers
   agentv eval evals/reasoning/logic-puzzle.EVAL.yaml --target azure --experiment without-superpowers
   agentv eval evals/reasoning/logic-puzzle.EVAL.yaml --target azure --experiment with-superpowers

4. Verify the comparison view in studio shows accurate results
5. Document findings

Blocked by

• bug: LLM grader reports 'no response provided' when system prompt is present in input #982 / PR fix(core): extract system messages in prompt builder for LLM grader #983 (grader fix must be merged first)

Acceptance criteria

• Clean benchmark data with grader fix applied
• Comparison matrix in studio shows valid with/without-superpowers comparison
• Results committed to agentv-bench-skills repo

[christso]

Completed in #995. Clean benchmark data after grader fix (#983) committed to agentv-bench-skills. Both experiments score ~99-100% on azure and gemini — previous 0% for with-superpowers/azure was entirely a grader artifact.