v3 Slice 3: Stats core - bootstrapped CIs, marginals, single-source metrics, export

[mark-allwyn]

Parent

#13

What to build

The statistics layer, decoupled from collection - it reads persisted results only and never calls a provider. Establishes the single source of truth for every reported number.

• Shared metric helpers (domain accuracy, composite) that every downstream consumer must use - no second computation path may exist.
• Bootstrapped 95% confidence intervals on the headline conjunctive accuracy.
• Per-part marginal accuracies and per-skill / per-bundle breakdowns (which causal sub-skills fail).
• Refusal and invalid reported as separate columns, never folded into accuracy.
• stats (console summary) and export (JSON) CLI commands.

Acceptance criteria

• Metrics run over synthetic persisted results with no provider calls
• CI brackets the point estimate; each per-part marginal is >= the conjunctive accuracy (sanity invariant)
• Refusal/invalid column counts reconcile exactly with raw stop_reason counts
• Only one accuracy computation path exists (consumers call the shared helper)
• export output is deterministic from the results files
• pytest green

Blocked by

• v3 Slice 2: Tracer bullet - one causal item end-to-end (generate, run, score, leaderboard) #15