feat(skills): support SystemContentBlock[] in system prompt injection

[agent-of-mkmeral]

Problem

The AgentSkills plugin's _on_before_invocation hook only handles string system prompts. When the system prompt is set as list[SystemContentBlock] (e.g. with cache points for prompt caching), the plugin:

1. Reads agent.system_prompt → gets concatenated string (cache points invisible)
2. Does string manipulation to append skills XML
3. Sets agent.system_prompt = new_string → destroys content block structure

This silently drops cache points and other non-text blocks on every invocation.

Fix

Two changes, no new helper methods:

1. Agent.system_prompt_content property (agent.py)

Public getter for _system_prompt_content so plugins can read the content block representation:

@property
def system_prompt_content(self) -> list[SystemContentBlock] | None:
    return self._system_prompt_content

2. Content block branch in _on_before_invocation (agent_skills.py)

One method, two clear codepaths:

content_blocks = getattr(agent, "system_prompt_content", None)
has_structured_blocks = content_blocks is not None and any("text" not in block for block in content_blocks)

if has_structured_blocks:
    # Filter out old skills text block, append new one, preserve cache points
    ...
    agent.system_prompt = filtered  # Sets list back, preserving structure
else:
    # Existing string manipulation behavior
    ...

Parity with TypeScript SDK

This matches the TypeScript implementation which already handles SystemContentBlock[] via a two-branch design in _injectSkillsXml.

Testing

• 3 new agent tests: system_prompt_content property for string, list, and None inputs
• 8 new skills tests: cache point preservation, idempotency, skills swap with cache points, string regression, all-text-blocks fallback, warning on missing XML, None fallback
• 133/133 skills tests passing, 0 regressions
• 12/12 system_prompt agent tests passing, 0 regressions

@mkmeral