3.14.0: Canonical Data Lineage

Highlights

• Canonical Evidence Infrastructure: Introduced a typed, fail-closed contract (evidence_recorder, claim_policy, score_lineage) ensuring all material claims retain source provenance and positive score credits require verified fact lineage before publication.
• BUY Rationale Discipline: Required BUY recommendations to cite at least one eligible supporting fact rather than relying solely on passing scorecard metrics.
• Earnings Durability & Management Guidance: Added deterministic management-guidance retrieval and an earnings durability classifier (TEMPORARILY_BOOSTED) to suppress moat/capital-efficiency bonuses earned during temporary cyclical peaks.
• Live-First FX Rate Cache: Implemented FxRateCache with a 1-hour TTL for live-first currency conversion with fallback, updating stale spot rate defaults across IBKR portfolio valuation and reconciliation.
• Enhanced Cost Accounting: Added cost rollups by provider/model/tier, retry cost re-attribution, a cost_report.py analysis script, and custom OPENAI_API_BASE routing for OpenAI-compatible providers (e.g., Kimi/Moonshot).

Key Improvements & Fixes

• Evidence Gate Refinements: Fixed false rejections triggered by rounded-precision citation matching and raw N/A guidance reads.
• Scorecard Accuracy: Resolved false N/A collapses on GROWTH_ADJ and HEALTH_ADJ scores caused by unbacked rubric criteria vetoes.
• Observability: Added deterministic logging for forensic-auditor rejected hosts and increased byte/timeout caps for primary disclosures.
• Model & Chart Fixes: Updated Gemini model health checks and pricing, bumped Value Trap Detector thinking level, and corrected article-relative chart path resolution.