Cleanup Artifacts — Orchestrator

Single-agent — orchestrator IS the runner; critic in-procedure. Audit + groom docs/forsvn/artifacts/: classify every file, surface refs + risk, archive non-KEEP on --apply. Never deletes. Metadata: routing.yaml. Methodology / when-not-to-use: references/playbook.md.

Core question: load-bearing, or cruft from weeks ago?

Critical Gates — load first

1. MOVE, never delete. mv into docs/forsvn/artifacts/.archive/[YYYY-MM-DD]/. --purge-archive (v1 out of scope) is the only deletion path.
2. HARD-NEVER refused. Canonical (brand/, research/, architecture/); VCS (.git/, .gitmodules, submodules); infra (.forsvn/index/manifest.json); append-only (docs/forsvn/experience/); anchors (docs/forsvn/artifacts/meta/{roadmap,tasks}.md).
3. --dry-run default. No move without --apply + per-category confirm.
4. Critic gate non-negotiable. Grep 5 random STALE/ORPHAN refs across docs/forsvn/artifacts/, brand/, research/, architecture/. Live ref → escalate (no prompt), refs surfaced.
5. Manifest-sync. Re-run bun scripts/manifest-sync.ts after any move.

Before Starting + Pre-Dispatch

Apply references/_shared/{before-starting-check, pre-dispatch-protocol, mode-resolver}.md. budget: standard; --fast bypasses prompts — Critical Gates supersede --fast. Read .forsvn/index/{manifest.json,artifact-index.md} + docs/forsvn/experience/technical.md (prior excludes). Manifest missing/stale (>1d) → NEEDS_CONTEXT. Dimensions: scope, mode, threshold (90d), excludes. Session execution profile (single-vs-multi): inherit per references/_shared/execution-policy.md.