Quantify every metric-definition variant before debating which to pick

When a spec leaves a metric definition open (set vs multiset vs ordered, adjacent-only vs anywhere, raw-string vs canonicalised key), compute every variant across the full corpus before arguing about it — most variants turn out numerically indistinguishable and the real deliverable is a defensible written rationale, not a different number. In one agentic-trace benchmark all four redundancy rules landed on 51 or 52 and flagged the identical trace set, and an ordered precision metric was provably dead weight because every trace whose called multiset matched gold also matched in order. Pick the variant that avoids a latent bug rather than the one that shifts the score: keying redundancy on parsed args collapses every unparseable call to the same None key, so two genuinely different malformed calls register as duplicates. Also test adjacency rules against a real interrupted run — a strict back-to-back rule silently undercounts a retry loop when one malformed call is wedged mid-streak.