[mgiovani]

Summary

Adds a style layer so rules can read CSS, plus the first three rules of the new design group. Closes #86 and #87. Part of #85 and epic #98.

• Lang::Css covers .css and .scss, parsed with tree-sitter-css. It joins neither CODE_LANGS nor PROSE_LANGS, so no existing rule starts running on stylesheets.
• StyleDoc (src/style.rs) is built once per file and exposed as ctx.style. It holds CSS declarations with absolute byte spans and Tailwind class tokens grouped per class attribute. Sources:
  • CSS/SCSS: declarations from the tree.
  • HTML: <style> bodies, re-parsed through tree-sitter set_included_ranges so offsets stay absolute, plus style= and class= attributes.
  • TSX: className/class strings (including cn(...) arguments) and style={{...}} objects, with camelCase keys mapped to kebab-case.
• src/color.rs: pure color parsing (hex, rgb/rgba, hsl/hsla, named colors) and HSL conversion.
• Gate: StyleDoc is built only when a design-group rule is enabled for the file's lang. Built unconditionally, it tripled HTML lint time. With the gate, the default run is unchanged: 231 ms before vs 229 ms after, on 300 Design2Code pages.

Rules (design group; SLOP052, SLOP053, SLOP055 and SLOP056 Tier B on, SLOP054 Tier C off)

Code	Detects
SLOP052	text filled with a gradient (background-clip: text plus a gradient, or bg-clip-text plus a gradient utility)
SLOP053	purple-to-blue gradients: CSS stops in the violet/indigo hue band, Tailwind from-purple/violet/indigo with a matching via/to, bracketed stops
SLOP054	hand-copied Tailwind violet hexes outside gradients and outside custom-property definitions
SLOP055	violet Tailwind accent utilities (text, bg, border, ring, ... with purple, violet or indigo 400-700, variants and opacity allowed) in class lists; flags the look, not the author
SLOP056	shadcn violet theme preset: --primary, --ring, --accent, --sidebar-primary, --chart-1 set to a violet HSL triplet, oklch(), hex or rgb in a block that also defines at least two shadcn marker tokens (--card, --popover, --primary-foreground, ...); a lone brand --primary and zinc, slate or blue presets stay quiet

The five rules own disjoint spans (clip, gradient, bare hex, accent utility, theme token), and none overlaps SLOP009, SLOP018, SLOP021 or SLOP044.

First measurement (superseded)

WebSight's 2023-era AI pages contain none of these tells (0 hits for every rule), so they couldn't show separation. The app-builder cells in Tier decision below replace them as the AI split. The WebSight rows stay in bench/candidates.toml.

Dependency

tree-sitter-css = "0.25" (MIT). Regex cannot track nested blocks, strings and comments in CSS, and tree-sitter is the parser family the crate already uses. cargo tree -i tree-sitter shows a single core (0.26.10). Cargo.lock gains only tree-sitter-css; cc and tree-sitter-language were already locked.

Tier decision

Scored on pinned public repos from four AI app builders (64 repos, 1,455 files; tsx, css, html under src/, shadcn components/ui/ and *.min.* excluded) against verified human files (Ant Design TSX, 11 named-maintainer Tailwind React apps pinned to late 2022 with their CSS, three pinned CSS repos; 1,275 files). Builder provenance markers: Lovable gpt-engineer-app[bot] commits (28 repos, 2026-06 to 2026-10), bolt.new .bolt/config.json (12, 2024-10 to 2026-03), Same.new .same/todos.md (12, 2025-05 to 2026-09), Replit Agent Replit-Commit-Author: Agent trailer (12, 2025-06 to 2026-10). No pre-2020 Tailwind baseline exists.

Bar: AI rate at least 3x the human rate, human rate at most 1% (2% AI floor as guidance only). Re-scored after the second review pass: SLOP053 now requires stops on both the blue side and the purple core (alpha at least 0.5) and pairs stops only within a variant; the SLOP056 marker and violet-floor changes moved no corpus count.

Rule	bolt.new	Lovable	Same.new	Replit Agent	AI pooled	Human verified	Tailwind 2022 alone	Design2Code
SLOP052	13/526	6/558	4/183	3/188	26/1455 (1.79%)	1/1275 (0.08%)	1/589	3/484
SLOP053	31/526	1/558	19/183	14/188	65/1455 (4.47%)	1/1275 (0.08%)	0/589	0/484
SLOP054	1/526	2/558	0/183	0/188	3/1455 (0.21%)	0/1275	0/589	1/484
SLOP055 (tsx, html)	100/509	13/513	23/171	16/165	152/1358 (11.19%)	6/1000 (0.60%)	6/500 (1.20%)	0/484
SLOP056 (css, html)	1/27	5/58	0/15	1/29	7/129 (5.43%)	0/275	0/89	0/484

• SLOP052: Tier B on. Under the 2% guidance floor (1.79%) but 22x the human rate.
• SLOP053: Tier B on.
• SLOP055: Tier B on. A product call: it flags the look, not the author; humans on Tailwind reach for these shades at 1.2% in the 2022 baseline.
• SLOP056: Tier B on. Requires two shadcn marker tokens (--card, --primary-foreground, --radius, ...) in the block, so a lone --primary: #9146ff stays quiet. Small denominators: most builder output keeps its theme in one index.css.
• SLOP054: AI rate 0.21%. Stays Tier C. A violet hue-band widening was measured and rejected: syntax-highlighting themes (codemirror 38%, highlight.js 33%) and Tailwind apps push the human rate far above 1%.

Tailwind v3/v4. SLOP052, SLOP053 and SLOP055 read bg-gradient-to-*, bg-linear-*, bg-radial, bg-conic, the ! important marker before (v3) or after (v4) a utility, the v4 tw: prefix, bracketed stops including oklch() (underscores as spaces, color: hints, /opacity), and ignore the v4 bg-(--brand) shorthand. color.rs parses oklch()/oklab() (hue in deg, grad, rad, turn or none) and shadcn bare HSL triplets. Ceiling: a v3 tw- config prefix is not read.

Rows and per-rule notes are in bench/candidates.toml.

Notes

• tests/integration.rs: a data-only "css" entry in LANG_DIRS, the same precedent as "html".
• LintContext.style adds style: None to every test literal. This is one mechanical commit.
• .css and .scss files now enter the walk, so existing custom rules can match them. CHANGELOG recommends excluding *.min.css.
• Known ceilings, documented: SCSS $vars and @include can hide a neighboring declaration, and /* ai-slop-ignore */ inside an HTML <style> block is not honored (use an HTML comment around the element). // suppression works in .scss.

Verification

• just check passes. The lib has 1066 tests, plus the integration, natlang and image harnesses.
• cargo install --path . followed by stopslop .: zero findings (the repo config selects ALL).
• With --select design on the fixtures, only the slop_design.* files fire.
• Reviewed with an 8-lane review (correctness, performance, style, tests, errors, simplicity, architecture, deps). No Critical or Major findings survived verification. The Minor and Nit fixes are in the last four commits.