v6_t2 on the 67-task feedback split

Task-by-task optimization, one bundle per task. Each row is a feedback task: its Seed reward (v6_g2_e1's 10 Seed trials) and, where v6_t2 ran on it, the reward of that task's own champion bundle. These are fit numbers on the task each bundle was optimized for, not a merged or held-out result.

34optimized
34reached 1.0
8unfixable (routing)
25not run (≥ 0.95)
0.830 → 0.937feedback mean, Seed → per-task best
$319optimizer + eval
0.0 0.5 1.0 unfixable with skills: routed to a specialist missing a required tool
taskdomainclass SeedT2 bestΔ itersstepswhy not optimized

Classes: optimized — a v6_t2 project ran; at-ceiling — Seed 1.0 on all 10 trials; above-threshold — Seed in [0.95, 1.0), left out by the 0.95 cutoff; unfixable-routing — parsec's regex fast path (classify_fast) sends the request to a specialist that lacks a tool the verifier requires, so no skill edit can reach it. Steps: A = accepted candidate, r = rejected, x = the optimizer failed to run (the CLI never ran, so no edit was made and the unchanged bundle was re-measured). Suite v6_t2_e1.