Covers the tag validator (normalize, resolve, morph stemming, fuzzy,
salvage extraction, strict/non-strict, max_tags, min_post_count,
exclude_categories, dedupe), the node against a mocked server (payload
shape, think switch, Ollama reasoning fallback, template handling,
error surfacing), and the dataset builder (stable hash split, stale
system-prompt filtering, chat-format output). Includes regression tests
for the max_tags salvage overflow and unclosed <think> leak.
Also corrects the extract() docstring examples, which the tests caught:
"black" alone is not a Danbooru tag and "dark" is.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
