[G2] Agentic audit v2 — implement the per-stage AI agents

[anmolg1997]

Workstream: agentic-audit | Estimate: 12d normal → 6d AI-assisted | Phase: 3

Why

BLOCKED on Akhilesh's spec (committed on the 2026-06-12 call): page-by-page input, working memory across page-spanning steps, selective context (filter 10-20 checklists via headers/segmentation), tools orchestrated in a workflow.

Scope

• Implement spec on the G1 architecture
• Checkpointed runs; HITL interrupts; telemetry + C3 version stamping
• Migrate existing agentic_audit rules

Acceptance criteria

• Spec acceptance scenarios pass on reference docs
• Run survives worker restart mid-document

---

From the enterprise-readiness backlog (codebase audit + pharma/GxP research + tooling survey, 2026-06-12).

[anmolg1997]

Redefined to concrete scope. 'Agentic audit v2' is now the 8 named per-stage agents (Onboarding, Planning, Evidence, Analysis, Review Support, Finding, CAPA, Learning — VISION.md §9). The spec we were waiting on has arrived: the agentic_audit authoring guide landed in-repo via #155 (.claude/skills/audit-rule-author/references/agentic_audit_guide.md). MVP needs the Evidence/Analysis/Review/Finding agents (Stages 2–5); Onboarding/Planning/CAPA/Learning are Phase 2.

See planning/RECONCILIATION.md (PR #156) for the full product↔engineering mapping.