fix(tool/skill): cap description preview at 500 chars and load full S…

[duguwanglong]

The skill tool exposes two surfaces, both of which were silently dropping critical content:

1. The tool's meta-description (shipped in every system prompt) listed each skill's full frontmatter description (up to 1024 chars). Long descriptions bloated the prompt and the model often acted on a partial preview without loading the full file.

2. When the model called skill(name=...), ToolRegistry's auto-truncate path silently cropped the returned SKILL.md at 10 KB / 200 lines (head-only), dropping the workflow steps, references, and constraints that authors typically place at the end of the file. Users reported "skill.md 后面的 感觉就完全丢失了".

This change implements progressive disclosure (mirroring hermes-agent's skills_list / skill_view split):

• build_description() now caps each skill's preview at MAX_SKILL_DESCRIPTION_PREVIEW_CHARS (500) using head + tail truncation, preserving both the opening (scope/triggers) and the closing (hard constraints like "必须先加载本 skill"). Adds an explicit instruction telling the model the previews are summaries and it MUST call skill(name=...) to load the full SKILL.md before acting.

• skill_tool_impl() sets truncated=True on its ToolResult to opt out of the registry's auto-truncate pass, so the full SKILL.md (verified end- to-end with 18.8 KB / 500 lines agent-browser) reaches the model intact. Also emits a skill.load.full_content log event for auditability.

Empirical impact on the 12 existing skills:

• System prompt index: 4.6 KB total (only tool-builder 614→500 is trimmed)
• Load on demand: previously 8/12 skills lost their tail; now all 12 are delivered in full

Tests: 15 new in tests/tool/test_skill_tool_description.py covering truncation algorithm, build_description output, and end-to-end no-truncation through ToolRegistry.execute().