fix: per-item drop of invalid brief facts; relax LLM-produced singleton fields

fix: tolerate malformed LLM-copied EvidenceRef fields

fix: allow metrics/analyst sources in CategorizedRisk and GuidancePoint

feat: multi-provider LLM support, prominent chat, design pass, and new analytics

fix: coerce tax_rate_and_effective_leverage to consensus_beat_mix + harden prompts

test: strengthen uppercase canonical test + add _detect_trend boundary cases

test: add SubtextRead + _detect_trend tests

fix: coerce compound LLM source strings to prevent BriefOutput validation failures

deploy: Primer app for HF Spaces — clean orphan history