bench: record transfer-eval-8fold skill-transfer pilot (run 20260902001)

[bcarmeli]

Adds the missing dashboard record for the zero-shot skill-transfer pilot (transfer_eval_v1), whose raw results were already merged via #421 into skillsbench-history (results/transfer-eval-8fold/) but never surfaced on the benchmarks page.

• 8 folds, 5 tasks, 2 domains. Each tasks[] row is one train→test fold (named <train>→<test>), not a SkillsBench task.
• reward_baseline/reward_opt intentionally do not reproduce cap-evolve's own per-fold test_delta (always 0.0 by construction under --max-iterations 0 — the frozen donor skill is the only artifact evaluated, so it's compared against itself). Instead they show the comparison the source summary actually makes: the target task's own native seed vs. the transferred skill's score on it.
• experiment: "transfer-eval-8fold" (using the tier/experiment split from bench: record skillsbench task-by-task-87 sweep #456/site(benchmarks): derive Benchmark/Type/Experiment filters, add Experiment column #457), tier: "full", iterations: 0, optimizer_model: "none" — no optimizer loop, pure evaluation.
• benchmarks.json/meta.json regenerated via ci/benchmarks/lib/record.py aggregate.

Net effect across the 8 folds is a regression (0.2125 → 0.1625): transfer gives a small lift only where the native seed was near zero, and hurts where the target already had a strong seed. Directional only (small sample) — full detail in results/transfer-eval-8fold/summary.md on skillsbench-history.