fix(prompts): disambiguate skills from tools to prevent direct-call failures

[sridhar-3009]

Summary

Closes #292

Weaker models (Qwen72B and similar) read the skills listing in the system prompt and call the skill name directly as a tool — e.g. tavily-search(query="...") — instead of the correct skill(name="tavily-search"). The call fails with "unknown tool: tavily_search" because skills are not registered as tools; they are accessed exclusively through the skill built-in tool.

Root cause (confirmed in issue comment): The previous instruction said:

"invoke it with skill(name="<skill_name>")"

The abstract placeholder <skill_name> is easy for a model to treat as a formatting hint rather than a literal pattern. Combined with the listing format - **tavily-search**: Search the web..., the model mistakes the skill name for a registered tool name.

Fix: Three changes to _build_skills_section():

1. Lead with an explicit prohibition — IMPORTANT: Skills are NOT tools. Do NOT call them as direct function calls.
2. Show a concrete invocation example using the first real skill name instead of <skill_name>, so the model sees skill(name="tavily-search") verbatim.
3. Explain the failure mode — "Calling a skill name directly as a tool will fail with 'unknown tool'" — giving the model a clear signal to remember.

Before:

# Available Skills

The following skills are available via the `skill` tool. When a user's request matches a skill, invoke it with skill(name="<skill_name>") to load detailed instructions…

- **tavily-search**: Search the web with Tavily

After:

# Available Skills

IMPORTANT: Skills are NOT tools. Do NOT call them as direct function calls.
Skills must be accessed exclusively through the `skill` tool:
  skill(name="tavily-search")

When a user's request matches a skill below, call `skill(name="<skill_name>")` to load the full instructions…
Calling a skill name directly as a tool will fail with 'unknown tool'.

Available skills (access each via `skill(name="<name>")`):

- **tavily-search**: Search the web with Tavily

Test plan

• With a weaker model (Qwen72B or similar), install a tavily-search skill and ask a question that matches it — verify the model calls skill(name="tavily-search") instead of tavily-search(query="...").
• With a stronger model (claude-sonnet, gpt-4o), verify existing skill invocation still works correctly.
• Verify that when no skills are installed, the section is still omitted from the system prompt.