research(agent): validate direction-aware model handoffs

[ss-o]

Objective

Validate whether direction-aware model handoff interfaces improve the existing
generator-verifier workflow before changing organization guidance.

The pilot replays 12 historical tasks across z-shell/.github, z-shell/zi,
and z-shell/zsh-lint: six low-capability-to-high-capability escalations and
six high-capability-to-low-capability downshifts. It compares raw trajectory
transfer, sender-authored compaction, and trajectory removal while preserving
the task working tree.

Reference: The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

Acceptance criteria

• Run no more than 84 Codex prompts using GPT-5.6 Luna and Sol with medium
  reasoning and standard speed.
• Use included Codex allowance only, with no API-key billing, credit purchase,
  or automatic overage.
• Keep every replay in a temporary isolated repository copy and preserve all
  current workspace worktrees.
• Record deterministic validation, task outcome, model direction, interface,
  and duplicated work for every condition.
• Treat the directional result as locally supported only when the preferred
  interface wins at least four of six tasks in each direction without a
  critical regression.
• Change canonical guidance only if the threshold is met.

Instruction impact review

1. Classification: scoped guidance for the generator-verifier workflow,
   informed by a bounded runtime evaluation. It is not shared baseline policy
   or deterministic enforcement.
2. Consumers: Codex, Claude Code, Copilot, Gemini CLI, and humans performing
   generator-verifier work in organization repositories.
3. Canonical owner:
   .github/instructions/generator-verifier-workflow.instructions.md remains
   the correct owner.
4. Duplication and conflicts: .github/AGENT_MEMORY.md remains the durable
   GitHub handoff protocol and must not duplicate live trajectory-transfer
   guidance. Context-engineering guidance remains complementary.
5. Manifest impact: no new surface or route is expected because the public
   and private manifests already route generator-verifier tasks to the
   canonical file. Recheck after drafting the change.
6. Runtime delivery: the existing manifest-declared scoped instruction is
   available to all supported runtimes without relying on an optional skill or
   hook.
7. Generated output and limits: the target is not part of the generated
   root AGENTS.md composite. Run public policy validation and the applicable
   private checks to confirm no routing or generation drift.

Execution boundaries

• No commit, push, pull request, merge, release, or deployment is authorized.
• No API key or paid credit purchase is authorized.
• Stop if included Codex usage is exhausted or a replay cannot remain isolated.