Feat v0.9.7 workflow starlark write (#81)

Co-authored-by: Jeffrey Smith <jasafpro@gmail.com>
Co-committed-by: Jeffrey Smith <jasafpro@gmail.com>

CHANGELOG.md

@@ -2,6 +2,37 @@
 
 All notable changes to Armature are documented here.
 
+## v0.9.7 — Full Read/Write Workflow Starlark Module
+
+Extensions with `workflow.access` permission can now start, advance,
+cancel, and signoff workflow instances directly from Starlark scripts.
+
+**WorkflowEngine interface**
+
+- Defined in `sandbox` package to break the `workflow ↔ sandbox` circular
+  import (follows `NotificationSender` / `ConnectionResolver` pattern)
+- 4 methods: `Start`, `Advance`, `Cancel`, `SubmitSignoff`
+- `workflow.Engine` satisfies the interface implicitly — zero changes to
+  the engine package
+
+**New Starlark builtins**
+
+- `workflow.start(workflow_id, data={})` → instance dict
+- `workflow.advance(instance_id, data={})` → instance dict
+- `workflow.cancel(instance_id)` → None
+- `workflow.submit_signoff(instance_id, decision, comment="")` → signoff dict
+- All write ops require authenticated user context (`RunContext.UserID`)
+- Graceful errors when engine or user context unavailable
+
+**Refactors**
+
+- `instanceToDict` helper extracted — shared by `get_instance`, `start`,
+  `advance`
+- `signoffToDict` helper for signoff return values
+- `dictToJSON` helper for Starlark dict → `json.RawMessage` conversion
+
+**Tests:** 6 new unit tests (mock engine, happy path + error guards)
+
 ## v0.9.6 — Deprecate stage_type, Collapse stage_mode
 
 Simplifies the workflow stage classification model by removing