docs(skill): canonical analyst-loop wiring pattern

[drewstone]

Summary

Adds a 'Closing the loop — analyst-loop wiring pattern' section to the agent-eval SKILL between 'Capture integrity' and 'Pitfalls'. Codifies the shape every vertical agent (tax / legal / gtm / creative) copies into `tests/eval/lib/analyst-loop.ts` — only the per-vertical `ImprovementAdapter` differs.

Covers:

• The minimal orchestrator wiring (`AnalystRegistry` + `FindingsStore` + `OtlpFileTraceStore` + `runAnalystLoop`)
• Subject-grammar routing table (`agent-knowledge:` → `KnowledgeAdapter`; `system-prompt|tool-doc|rubric|persona:` → vertical `ImprovementAdapter`)
• Five directives covering ledger location, subject discipline, auto-apply policy, event forwarding, and fail-loud semantics for missing surfaces
• Reference implementation pointer at `tangle-network/tax-agent#67`

This is the source of truth the per-agent PRs reference; future agents copy from here.

Test plan

• Markdown renders correctly (CommonMark)
• No code changes — docs only

[drewstone]

Moving to dotfiles per CLAUDE.md convention; skills live in ~/dotfiles/claude/skills/, not in product repos. Will re-add as a generic section in agent-eval-adoption.