v3 Slice 6: Pareto frontier + over-time tracking + saturation + classical item stats

[mark-allwyn]

Parent

#13

What to build

The fairness and over-time analysis on top of the stats core - turning a roster of scored results into the honest comparison the benchmark exists to provide.

• Pareto frontier: accuracy against thinking-tokens, latency, and cost, so a cheap fast model and an expensive reasoning model are each judged on their own frontier.
• Model tiering (reasoning / non-reasoning, size class, open / closed) so like is compared with like.
• Over-time tracking: results tagged with benchmark version and model launch date; a frontier-over-time view.
• Saturation detector: signal when the top-N models' confidence intervals overlap (the cue to re-harden to a new version).
• Classical item statistics: per-item p-value and item-total discrimination, flagging dead items (everyone right) and broken items (everyone wrong). IRT explicitly deferred.

Acceptance criteria

• Pareto frontier computed and exported from real roster results; dominated points correctly excluded
• Saturation detector fires when top-N CIs overlap on synthetic data and stays quiet when they do not
• Item stats flag a planted dead item and a planted broken item in a fixture set
• Over-time view orders models by launch date and tags benchmark version
• All numbers come from the shared metric helpers (no second computation path)
• pytest green

Blocked by

• v3 Slice 3: Stats core - bootstrapped CIs, marginals, single-source metrics, export #16
• v3 Slice 5: Provider breadth + per-vendor native-config policy + parallel runner #18