docs(prompt-optimizer): fill in three guidance gaps (#120, #121, #122)

[dcramer]

Close three open enhancement issues on the prompt-optimizer skill by filling
gaps where the skill was either silent or actively nudging authors toward
weaker patterns. All edits land in the existing skill structure — no new
workflow steps, no new reference files.

What changed

#120 — tool policy vs. native tool disclosure

• Renamed the <tools> marker to <tool_policy> in the portable skeleton,
  Example 1, and the canonical "good section names" list.
• Added a note in core-patterns.md and one line per provider section in
  model-family-notes.md clarifying that tool schemas are disclosed via the
  provider-native tools parameter (Anthropic, OpenAI, Gemini) — the prompt
  should carry policy, not schema restatements.
• Tightened the "Layer the prompt correctly" long-prompt bullet to match.

#121 — layered prompts with multiple owners

• New ### Layered prompts with multiple owners subsection in
  core-patterns.md covering the platform-layer vs. deployer-persona-layer
  split (SOUL.md, CLAUDE.md, AGENTS.md).
• The invariant: the deployer layer is voice-only; platform behavior rules
  must still fire if the deployer layer is empty or sparse.
• Cross-reference added from SKILL.md Step 3.

#122 — directives in descriptive markers

• New ### Where rules live subsection in core-patterns.md noting that
  descriptive markers (<context>, <state>, <turn-state>,
  <environment>) read as data and state-conditional directives buried
  in them can underperform the same directive in a canonical rules section
  (<behavior>, <constraints>, <tool_policy>, <workflow>).
• New bullet in "High-value prompt moves".
• New ## Example 4 in transformed-examples.md with the resume-notice
  before/after (0.5 → ≥0.75 on the relevant eval with no other change).

Why

Each issue came with a concrete field finding. The skill was either silent
on the topic (#121 layered prompts, #122 rule placement) or actively nudging
authors toward the weaker pattern (#120 enumerating tools in the prompt).
The edits keep the skill's "empirical finding → rule" voice and stay inside
the existing file structure.

Prior art consulted

Rather than rely purely on the field observations in the issues, I checked
authoritative sources for each claim:

• Anthropic tool-use docs:
  "When you call the Claude API with the tools parameter, the API constructs
  a special system prompt from the tool definitions, tool configuration, and
  any user-specified system prompt" — schemas are injected automatically.
• pi-agent-core packages/agent/src/agent-loop.ts:
  LLM is invoked with llmContext = { systemPrompt, messages, tools } — tools
  are a distinct native parameter alongside prompt text.
• OpenAI Codex CLI system prompt:
  references tools contextually as policy ("prefer rg for searching") and
  does not restate schemas.
• Hermes Agent SOUL.md
  and soul-md.xyz: direct industry prior art for
  the platform + persona layering model (SOUL.md for voice, AGENTS.md for
  project, platform defaults code-level). Hermes explicitly falls back to a
  built-in default identity when SOUL.md is empty — the pattern the skill
  now documents.

Notably, Codex and Claude Code system prompts both do embed directives in
descriptive sections in places, so the #122 guidance is scoped to
state-conditional rules where the author's eval showed a measurable gap,
not an absolute rule against context-section prose.

Fixes #120
Fixes #121
Fixes #122