Fix three robustness defects in skill-and-tool-validator

[justinmclean]

Summary

Three small, independent defects in the validator and its test suite: a
docstring unit error, a test that aborts instead of skipping, and an unguarded
directory read.

Background

All three were raised in review, agreed as non-blocking, and merged without
being actioned. None changes behaviour on a healthy checkout; each degrades
badly in a specific environment.

Where to look

• .../skill_and_tool_validator/__init__.py line 2488 - docstring says the
  content is shorter than _MIN_LICENSE_FILE_SIZE characters, but line 2498
  compares path.stat().st_size, which is bytes. Equivalent for ASCII; a
  file with a BOM or multi-byte content measures differently.
• tools/skill-and-tool-validator/tests/test_validator.py line 1686 -
  test_real_repo_tool_python_files_all_have_headers calls find_repo_root()
  with no guard. In a shallow or out-of-tree run it aborts the session rather
  than skipping. Add a skipif.
• .../__init__.py line 3285 - validate_eval_coverage calls .iterdir() with
  no try/except. A restrictive CI runner gets an unhandled PermissionError
  instead of a graceful violation. collect_tool_python_files shows the pattern.

Acceptance criteria

• The docstring states bytes.
• The repo-wide header test skips rather than errors, with a test proving it.
• validate_eval_coverage handles OSError as collect_tool_python_files does.
• uv run --directory tools/skill-and-tool-validator --group dev pytest passes.

Estimated effort

~2 hours for someone new to the codebase.

[Shaurya2k06]

I'll take this up!
cc @justinmclean