docs(analysis): define agent-side NL review boundary

[rizumita]

Summary

• Documents that natural-language interpretation over fslc analyze output is an AI-agent responsibility, not an fslc feature.
• Adds agent-facing rules to cite exact source text and TSG node ids, keep formal_status:"not_a_violation", avoid CI/verifier failures from NL suggestions, and require explicit opt-in before external model calls.
• Mirrors the rule in skills/fsl/SKILL.md and skills/fsl/reference.md so agents get the instruction at point of use.

Why

• Goal: Resolve Design an external natural-language review layer for fslc analyze outputs #117 without adding an in-repo NL reviewer command or hard-coded language heuristics.
• Reason: The AI agent already owns natural-language reading and privacy decisions; fslc should remain deterministic and language-agnostic.
• Alternative: Add a prototype fslc-nl-review command. Rejected for now because it would move LLM/privacy policy into the core repo without a stronger product need.

Invariants

• fslc analyze remains structural observation, not verifier semantics.
• Agent-side NL suggestions remain non-authoritative review candidates.
• No source text is sent to an external service without explicit opt-in.

Evidence

• git diff --check

Rollback

• Revert the docs/skill commit; no runtime state or schema migration is involved.

Related

Closes #117