fix(skill-validator): preserve blank lines inside block scalars in parse_frontmatter

[andreahlert]

What

Fixes a latent correctness bug in skill-validator's parse_frontmatter, surfaced by a code review.

The bug

parse_frontmatter treated any blank line as a terminator for the current key. A YAML block scalar (| or >) with a paragraph break (a blank line between two indented paragraphs) silently lost everything after the first paragraph.

Two consequences:

1. validate_frontmatter measures len(fm["description"]) + len(fm["when_to_use"]) against MAX_METADATA_CHARS (the Claude Code truncation budget). With the dropped paragraphs the measurement was too small, so frontmatter that actually exceeded the budget could pass the check.
2. validate_principle_compliance and validate_trigger_preservation only saw the truncated string, missing forbidden patterns or trigger phrasing that lived in the dropped paragraphs.

Why latent

No existing SKILL.md frontmatter currently uses a paragraph break in a block scalar, so nothing in-tree exercised the bug. But init_skill.py scaffolding emits multi-line description and when_to_use blocks, and any author writing a two-paragraph description would have been silently mis-validated.

The fix

In real YAML, a blank line inside a block scalar is part of the value, not a terminator. Only a new top-level key finalises the current value. The parser now appends blank lines to the value lines and lets .strip() at finalisation discard leading/trailing blanks so single-line values are unaffected.

Changes

• src/skill_validator/__init__.py — parse_frontmatter: blank line is content, not terminator
• tests/test_validator.py — regression test for a | block scalar with an internal blank line

Validation

• pytest: 77 passed
• ruff check / ruff format / mypy: clean
• prek including skill-validate (.claude/skills/**): all hooks pass — every existing SKILL.md still validates after the parser change