Page-level chunks (__PAGE__) store frontmatter in their context for
semantic search quality during embedding. However, when formatting
these chunks for LLM prompts, the frontmatter was duplicated:
- Once in the <properties> section (parsed from page outline)
- Again in the <block> content (from stored RAG chunk context)

This wasted ~50-200 tokens per page depending on property count.

Solution: Reuse existing _clean_context_for_llm() function from
page_indexer.py to strip frontmatter from page-level chunks during
prompt formatting. This elegant approach reuses tested code instead
of duplicating logic.

Changes:
- Import _clean_context_for_llm() in llm_helpers.py
- Detect page-level chunks (::__PAGE__) in format_chunks_for_llm()
- Apply _clean_context_for_llm() to page-level chunks only
- Regular blocks unchanged (already cleaned during indexing)
- Frontmatter remains in <properties> section for LLM context

Impact:
- Before: "tags:: foo, bar" appears twice in prompt (properties + block)
- After: "tags:: foo, bar" appears once (properties only)
- Token savings: ~50-200 per page with properties
- No impact on semantic search quality (frontmatter still embedded)

Tests:
- Added test_llm_helpers.py with 3 integration tests
- Tests cover page-level chunk stripping, regular block preservation
- All existing llm_wrappers tests pass (no regressions)

Assisted-by: Claude Code