Research question across independent angles and cross-check claims

Research questions needing ≥3 sources cross-checked for consistency (SDK/API behavior, vendor pricing, version facts); replaces serial autoresearch loop.

• · question (string, required): research question about SDK/API/vendor behavior
• · maxAngles (optional, 1-8, clamped; default 5)
• · maxClaims (optional, 1-100, clamped; default 24)
• · all passed as args or args.question

• · per-angle findings (discrete claims with source URLs)
• · cross-checked voted result: {status: verified|contradicted|unverified, reason, bestSource (authoritative URL or 'none')}
• · cited report with UNVERIFIED claims flagged
• · sources list with URLs

• · agent fan-out (1 agent per research angle, parallel)
• · WebSearch + WebFetch (each angle does its own, lean question)
• · cross-check agents (vote per claim against independent sources)

• · question must be specific enough to research directly (not underspecified like 'what car to buy')
• · question is about SDK/API behavior, vendor pricing, version facts (not subjective)
• · prefer official/primary sources over secondary
• · no pre-fetched page dumps injected into sibling agents (each angle searches independently)

• · question too vague → no crisp claims to verify
• · claim cannot be web-verified → reported UNVERIFIED instead of dropped
• · sibling angle agents share context (violates #385 guardrail)
• · fan-out spawns >ANGLE_CAP agents (clamped to 8)
• · cross-check spawns >CLAIM_CAP agents (clamped to 100)
• · source quality mixed (unofficial + official both cited, not rank-separated)

• · 3-phase structure: Sweep → CrossCheck → Synthesize
• · ANGLE_CAP bounded [1,8]; CLAIM_CAP bounded [1,100] (fan-out guardrails, see #385)
• · 5 predefined research angles: official docs, package registry, release notes, third-party corroboration, known issues
• · VOTE_SCHEMA: status enum (verified|contradicted|unverified) + reason + bestSource (authoritative URL field)
• · UNVERIFIED explicitly flagged in report (not dropped)
• · confidence enum per claim: high|medium|low