ArtemisKit 0.6.3 — Faults and durable recovery

Controlled faults and private workflow recovery

ArtemisKit adds explicit CLI/SDK pause and resume with the same workflow identity, original deadline,
consumed budgets, queued tool calls and saved state. The CLI uses workflow resume and private
--checkpoint-dir storage; a confirmed safe pause exits 10. The SDK exposes matching options and
session.pause().

• Six declared fault kinds test unavailable tools, timeouts, stale/incomplete data, malformed read
  results and conflicting instructions. Only declared pre-effect failures retry automatically.
• Bounded retries consume normal action/tool budgets and preserve failed-attempt evidence.
• Public V3 records retain bounded fault, retry, state-change and recovery metadata. Sensitive
  conversations, tool data and files stay in explicitly selected private checkpoints.
• Invalid identities, corrupt checkpoints, exhausted deadlines, ambiguous pending operations and
  terminal restores refuse replay. Safe simulated and Docker restart share the same contract.
• Default V2 runs and historical V1/V2 reading retain their prior meaning.

Milestone 0.6.3 maps to core 0.6.3 and CLI/SDK 0.6.2; exact dependency versions are in
docs/releases/0.6.3.json. Local qualification passed 1,683 tests, all 13 builds/typechecks and fresh
Node/Bun/TypeScript/real Docker consumers. Read docs/releases/0.6.3.md for publication status and
operational limits.

Local checkpoints are sensitive same-user data, not distributed locks or external exactly-once
guarantees. Live model quality was not measured. Workflow reports are released separately in 0.6.4.