feat(take-notes): add a manual tag vocabulary and gallery tag chips

[davertor]

Adds a manually curated tag vocabulary so the archive can answer "show me my AI notes" — today the gallery offers only a free-text filter over title/byline/meta, so browsing by topic means remembering a word that happens to appear in the text.

The vocabulary is closed: it lives in ~/take-notes/config.json, the skill picks from it and never invents a tag, so the taxonomy stays the user's rather than whatever the model felt like that day. Default state is a single tag, Unknown, which is also the silent fallback when nothing fits.

{ "language": "es", "tags": ["Unknown", "AI", "Investing", "Engineering"] }

How it looks in the gallery

→ docs/gallery-tags.png — added in this PR, and rendered inline in the README's new Tags section.

_{(The repo is private, so GitHub can't proxy the image into this body. The Files changed tab shows it too.)}

Two things are new on this page:

• A chip per tag in use, under the filter bar. Click one to narrow the grid, click it again to clear. The chip goes pen-coloured while active and the hit counter reads 2/4. Chip and text filter combine with AND — the existing apply() grew one piece of state rather than a second filtering engine.
• The primary tag on each card, beside the kind in the foot (VIDEO AI, ARTICLE ENGINEERING). Extras don't show there but are still filterable: every tag goes into data-tags, folded and pipe-delimited at both ends so a whole tag matches without ai also hitting air gap.

The four notes are the real committed examples; the tags on them are illustrative, since the vocabulary is yours to define. Re-rendering docs/examples/ with genuine tags — and rebuilding the Pages gallery at docs/index.html, which predates chips — is deliberately left out of this PR.

Notes written before tagging existed parse as untagged and group under the Unknown chip, so nothing falls out of the grid.

What's in it

scripts/notes.py	becomes the single config reader (read_config, configured_tags) and parses the rail's tag row back out; Note gains tag + tags
scripts/tags.py (new)	uv run tags.py [--add X] [--remove X] — read-modify-write, so language survives; refuses to remove Unknown
scripts/render.py	repeatable --tag, first occurrence is primary; stays a dumb renderer and validates nothing
scripts/gallery.py	the chips, the card tag, data-tags
scripts/export.py	tags reach Obsidian's frontmatter tags: list and the Anki tags column
SKILL.md	Step 0 for tag management, the selection rule in Step 2, --tag in Step 5
both note templates	{{TAGS}} after the .watch link, plus .tags/.tag/.tag.is-primary styling

Managing the vocabulary works from either side — /take-notes --tags | --add-tag "AI" | --remove-tag "AI", or tags.py directly. Both are worth having: the skill is disable-model-invocation: true, so saying "add a tag" in ordinary conversation never reaches it.

Verified

for s in render notes gallery export transcript tags; do uv run skills/take-notes/scripts/$s.py --selftest; done — all six pass, and tags.py is now in the CI loop.

• No regression on untagged notes — the four committed examples still render, grouped under one Unknown chip.
• CLI round-trip — --add AI --add Investing, then --remove Investing; "language": "es" intact throughout, and Unknown refused removal.
• Render → parse — --tag AI --tag Engineering reads back as tag == "AI", tags == ("AI", "Engineering").
• Chips — the template's own filter IIFE driven under node against a DOM stub: narrow, clear-on-second-click, non-primary tags filterable, chip+text ANDing, single-select.
• Exports — tags: ["take-notes", "AI", "Engineering"] in the frontmatter; take-notes key-point AI Engineering in the Anki column, whitespace hyphenated since Anki splits that field on spaces.

Still open: /take-notes --add-tag and a full end-to-end run against a real source, both of which need the skill symlink pointing at this branch.

🤖 Generated with Claude Code