chore: audit orchestrators for skill-forward conformance

## Summary

Rescope this audit from finding dual-mode agent candidates to assessing current orchestrators against HVE Builder's skill-forward architecture.

The audit determines whether reusable capability is located in skills, whether user-facing agents remain appropriately thin, whether durable conventions are correctly scoped to instructions, and whether isolated workers are background-only. It does not convert artifacts or assume that file size alone indicates nonconformance.

## Audit Scope

Inventory every current collection-distributed user-facing agent that coordinates one or more skills, instruction-defined workflows, or subagents. Use that inventory to select systems requiring detailed assessment.

The detailed assessment must include:

* RPI Agent
* Code Review
* Security Reviewer
* DT Coach
* ADO Backlog Manager
* GitHub Backlog Manager
* Jira Backlog Manager
* RAI Planner
* SSSC Planner

Include additional current orchestrators when the inventory shows the same responsibility-allocation questions.

## Acceptance Criteria

* [ ] Produce an inventory with agent path, user-facing purpose, reusable capability owner, delegated workers, and relevant instruction surfaces.
* [ ] Assess capability duplication, wrapper responsibilities, worker isolation, worker invocability, handoff contracts, and load timing.
* [ ] Classify each detailed target as conformant/no change, guidance-only correction, skill extraction candidate, wrapper-thinning candidate, background-worker exposure correction, or justified exception.
* [ ] Cite current workspace path and line evidence for every classification.
* [ ] Record why apparently large multi-turn agents remain conformant when their state, gates, or conversation responsibilities legitimately belong in the agent.
* [ ] Treat `user-invocable: false` as the default for orchestrator-only workers; require demonstrated independent user value for exceptions.
* [ ] Do not perform architecture conversions in this audit.
* [ ] Search for existing issues before recommending remediation and create or propose only deduplicated follow-up work.
* [ ] Produce a summary matrix of recommended changes and explicit no-change decisions.

## Related

* Parent: #2531
* Guidance task: #1311
* Historical parent: #1310
* Reference architectures: #2100 and #2474

Begin by inventorying the current collection-distributed user-facing agents named in the scope, then inspect their reusable capabilities, delegated workers, instruction surfaces, and handoff behavior. Check related issues #2531, #1311, #1310, #2100, and #2474 before assessing targets. Done means a path-and-line-evidence inventory, classifications, deduplicated follow-up recommendations, and a summary matrix including explicit no-change decisions.

Written by the indexing model from the issue text.