delivery-looplisted

# Delivery loop Execute one approved Markdown task packet without reopening settled product or architecture decisions. Read [references/delivery-checklist.md](references/delivery-checklist.md) at the start of the run and keep its concise evidence block current. ## Entry contract - Identify the exact task Markdown named by the request. Stop if more than one file could be the authority. - Treat the task as an execution contract, not proof that its repository claims are current. - The current request must authorize implementation. A packet produced by another skill does not grant permission by itself. - Research and decomposition are complete before this loop. For an explicit Qwen or DGX target, route missing evidence to `spark-task-research` and unresolved planning to `spark-task-planning`. For another execution context, route missing repository evidence to `task-research` and return unresolved product or architecture decisions to the task owner. Never solve either silently during implementation. - Repository instructions and project workflow skills outrank the packet when they conflict. ## Four phases 1. **Validate task Markdown** - Read the packet and applicable repository instructions completely. - Record the full `HEAD`, branch, upstream ref, and worktree status. Verify the packet's committed evidence with its pinned commit and keep dirty or ignored evidence separate. - Run pinned evidence commands against the named object. If a mutable ref matters, assert it