fix(core): skip skill context when locating invoking turns

[K-Mistele]

Summary

• preserve AgentLayer's existing fork policy: remove the parent turn that invoked the fork, then append the delegated child task
• when locating that parent turn, skip injected <skill ...>...</skill> user-role messages and continue backward to the actual human message
• use the same distinction for counting fork_turns: N
• add a regression test matching the production skill-injection message shape

Design Note

This does not copy native Codex's full-history fork algorithm. Codex full-history forks do not remove the invoking turn; they filter the complete rollout by item type. This change only adopts Codex's useful distinction between real human messages and injected skill context inside AgentLayer's existing turn-removal behavior.

Verification

• bun check

[github-actions]

Docs Agent Review

Agent finished with reason: error

---

To apply these recommendations, comment: @docs-agent apply