experiment 21: enable surface scan second pass — 15 matches (18 PRs, 12 scored)

Enabled the dormant _surface_scan method as a second review pass after the main
agentic loop. Surface scan runs a separate 5-round conversation targeting concrete
value-level bugs (typos, wrong variables, inconsistent names, missing definitions).

Raw matches nearly doubled (15 vs 8). Gained matches on 4 previously-zero PRs.
Mean PR F1=0.385 (lower than 0.410 due to averaging artifact from partial matches).
KEEP: +5 net matches on shared PRs, surface scan precision 35% > main review 18%.