feat(0.50.2): actionability fixes from real-data dogfood

[tangletools]

Summary

Surfaced by running analyzeRuns() against three real consumer datasets (legal-agent canonical, agent-builder canonical campaigns, gtm-agent golden run). Where the report was silent when it should have been loud, fix the silence.

Changes (all additive):

• ScalarDistribution.tailRuns?: Array<{runId, score}> populated for the composite distribution — the report now names the 5 worst runs to inspect first.
• InsightReport.costQuality.degraded?: {cost?, pareto?} — explicit reasons when costUsd is all zero or only a single candidate exists. Replaces the silently-misleading single-point Pareto.
• buildRecommendations fires on poor composite distribution: critical/investigate below 0.3, high/investigate below 0.5, with worst-N runIds named in the detail.
• buildRecommendations flags missing-judges with a medium/expand-corpus pointing at outcome.judgeScores enrichment.

Dogfood result. Legal-agent canonical (n=36, composite mean=0.002) previously returned recommendations: []. Now returns:

critical/investigate — Composite mean 0.002 is below the 0.3 floor — the agent is broken on this corpus
  Worst 5 runs: restaurant-formation=0, crypto-exchange-licensing=0, nuclear-startup-nrc=0,
  cannabis-dispensary=0, existing-business-audit=0
medium/expand-corpus — No judge scores recorded — per-dimension + calibration insights unavailable

Test plan

• pnpm test tests/contract-analyze-runs.test.ts — 15/15 pass (4 new tests for the new branches)
• pnpm test — 1431/1431 overall
• pnpm typecheck — clean
• Re-dogfooded on real legal-agent + agent-builder data, verified the recommendations + tailRuns + degraded fields populate correctly

[drewstone]

✅ Auto-approved tangletools PR — 54a98a85

This PR was opened by the trusted tangletools automation account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: tangletools_author · 2026-05-27T20:18:49Z}