Add prompt provenance, tracing, evals, and failure taxonomy

[Konan69]

Current state

Garden imported the core prompt posture and has document citations through findInDocument, issue/automation run traces, and bundled document skills. It does not have a product-wide provenance, prompt snapshot, tracing, or eval layer.

Remaining gaps

• General citation/provenance system across tools and claims.
• Unified clarification contract across chat, issues, automations, and external messengers.
• File-delete and destructive-operation gating.
• Observability/eval suites.
• Prompt/version tracking and per-run snapshot storage.
• Consolidated chat tool trace UI/export.
• Failure taxonomy.
• Confirm whether the planned schedule skill is installable or remove it from planning docs.

Acceptance criteria

• Each run can be inspected with prompt version/snapshot, selected tool traces, and model/provider metadata.
• Claims that cite documents or external sources have structured provenance where the tool supports it.
• Destructive operations share a clear approval policy.
• Eval coverage exists for prompt/tool safety regressions.

Source

docs/known-gaps/prompt-and-system-spec.md