calibration: seed provider track records via per-provider counterfactual replay campaigns

[JSONbored]

Problem

#8229 stage 1 (the would-have-routed report) waits on live reviewer_vote rows accumulating. The counterfactual harness (#8221) can generate the same evidence offline TODAY: a replay campaign per provider/model over the fixture corpus yields per-fixture verdicts that map 1:1 onto ProviderReviewSignal, feeding computeProviderTrackRecords (#8228) with clearly-labeled replay-derived data while live votes mature.

Requirements

⚠️ Required pattern. Replay-derived signals must be UNMISTAKABLY segregated from live votes — separate provenance field end-to-end, never mixed into the live reviewer_vote event type. The #8221 artifacts discipline applies (cached raw outputs, deterministic seeds, budget caps); BYOK providers only behind explicit flags.

• A thin adapter: replay --out artifacts → ProviderReviewSignal[] (provider = the variant modelSpec, vote from the verdict, abstentions excluded) → computeProviderTrackRecords → a rendered per-provider report.
• Run the campaign for the local smoke model plus each production reviewer model (BYOK-flagged), same seed across providers so the sample is identical.
• Post the first per-provider track-record table on calibration: evidence-weighted reviewer routing (flag-gated, report-first) #8229 as stage 1's replay-derived baseline.

Deliverables

• Adapter + renderer with unit coverage
• First same-seed multi-provider campaign recorded on calibration: evidence-weighted reviewer routing (flag-gated, report-first) #8229

Links & Resources

#8229 (stages 1-2), #8221 (harness), #8228 (computeProviderTrackRecords), scripts/counterfactual-replay.ts

Boundaries

Report-only; no routing change; replay-derived data never masquerades as live votes.

maintainer-only — model spend.

[linear-code]

JSO-5725