fix(gate-45): a reduced-motion block that exists is not a fallback — judge every motion selector under the cascade

[WilcoLouwerse]

The defect

Gate-45 (prefers-reduced-motion) asked one question per file: is there a @media (prefers-reduced-motion …) block in it? thematiq's css/systems/nldesign/theme.css had one, so the gate said PASS at 6dcbbaf9 — where c495c99c had added :not(.action-button) to four !important motion selectors while the reset block kept the bare .button-vue, button, .button. Equal specificity before, so the reset won on source order; strictly less specific after, so it lost the cascade on every button in the app — with the block that was supposed to stop the motion sitting right there in the file, under a comment that says "Keep this block in step with the motion declarations above."

The gate said PASS again at 67caeb85, after the selectors were mirrored back.

Measured 2026-09-21 with the package runner at full scope, one variable:

theme.css at	reset block	gate-45 before	gate-45 after this PR
6dcbbaf9 (reset no longer names the motion selectors)	exists	PASS	FAIL — 3 finding(s), naming exactly .button-vue:not(.action-button), .button-vue:not(.action-button) .button-vue__text, button:not(.action-button)
67caeb85 (reset names them again)	exists	PASS	PASS

Same verdict on the broken and the fixed file: the gate was not observing the property it is named after. It surfaced as a 🔴 in a human review three rounds in (thematiq#604), which is the wrong place for a mechanically decidable fact.

The change

The checker now flattens the stylesheet into rules (resolving SCSS / CSS-Nesting & against the parent, skipping @keyframes bodies, keeping the comma inside :is(.a, .b)), and requires every selector carrying motion to be overridden by some rule inside a reduced-motion block under the cascade's actual rules:

• equal importance → the more specific selector wins, then source order (the guard is later, so equal specificity is enough);
• unequal importance → !important wins regardless of specificity.

motion M vs guard G	covered when
G == M	unless M is !important and G is not
G extends M (.btn:not(.x) over .btn, html .btn over .btn)	same condition — G is more specific
M extends G (.btn:hover over .btn)	only when G is !important and M is not
G is *	only when G is !important and M is not — * has specificity 0

That last row is the one that changes the meaning of the repo-wide universal reset the pre-pass detects: it is now a * guard fed into the same check instead of an early exit. It still silences every plain motion declaration in the repo (ARM 6 of the scope suite is unchanged) and silences no !important one — which is exactly what thematiq's own comment above its reset says, and why that file repeats its selectors verbatim instead of writing *.

Recognised as correct without a matching selector: @media (prefers-reduced-motion: no-preference) { …motion… } (the motion itself is conditional), and a duration token — transition: opacity var(--dur) with --dur: 0ms redefined inside the reduced-motion block (the atom-design convention). Brace-less .sass falls back to the presence question.

The finding names the selector and, for !important motion, says what the fallback needs:

css/systems/nldesign/theme.css: stylesheet rule=motion-selector-not-overridden-by-reduced-motion-fallback selector=.button-vue:not(.action-button) (!important — the fallback needs equal or higher specificity, also !important)

Anti-widening — measured, not asserted

A per-selector question is exactly the kind of change that turns a gate into a noise generator, so the controls are pinned first and the fleet was run second.

scripts/lib/test_gate_45_selector_parity.sh (new, discovered by run-helper-suites.sh automatically) — 24 assertions: the thematiq shape as the failing arm (the finding must name the three narrowed selectors and must not name .button, which the reset still covers), and a control for every correct idiom: mirrored selectors, an !important guard over a plain motion (and the plain-over-plain case that genuinely loses), * over plain vs !important motion, no-preference, a zeroed duration token, &-nesting (resolved selector is what gets reported), :is(.a, .b), an html-prefixed guard, and the repo-wide reset in another file (covers plain motion there, not !important motion — and the finding lands on the motion file, not the reset).

Existing suites stay green: test_gate_45_stylesheet_scope.sh 13/13, test_gate_a11y_markup_scope.sh, test_gate_45_to_55_acceptance.sh ALL GREEN (with ajv resolvable), tests/test-hydra-gates-bin.sh 70/70.

Fleet dry run — the full runner at default (full) scope on every local ConductionNL app clone with a css/:

repo @ sha	gate-45	of which the new per-selector rule
thematiq @ 7466a5a (development, pre-c495c99c)	PASS	0
openregister @ cabd4106	PASS	0
portaliq @ aa6dc05	PASS	0
opencatalogi @ 291273fe	PASS	0
pipelinq @ 045e9048	FAIL — 14	0 — all 14 are motion-without-reduced-motion-fallback, and the pre-PR checker reports the identical 14
procest @ fa02ce95	FAIL — 25	0 — presence rule only
openconnector @ c7bd7b3c	FAIL — 5	0 — presence rule only
softwarecatalog @ c224f89	FAIL — 10	0 — presence rule only

Across eight apps the new rule fires on nothing but the fixture it was built from. The reds on pipelinq / procest / openconnector / softwarecatalog are <style> blocks with motion and no reduced-motion block at all — the question the gate already asked, on stale local clones (May 2026 for pipelinq and procest).

Also in this PR

• scripts/test-fixtures/gate-acceptance/COVERED-ELSEWHERE.md — the gate-45 row names the new suite and its planted/clean arms (the ledger that enumerates a gate's tests has to move with the tests, which is the same rule this gate now enforces on stylesheets).
• The _fail 45 message now says "N finding(s): motion with no reduced-motion fallback, or a motion selector the fallback does not override under the cascade" — it used to count "stylesheet(s)", which is no longer what a line in the log is.

Not run: shellcheck (not installed on the machine that produced this; the bash side of the diff is one message string, the rest is inside the PYRM heredoc).

The hydra-side docs (hydra-gate-prefers-reduced-motion/SKILL.md, which still said the gate reads .vue only, and the gate table in hydra-gates/SKILL.md) are updated in ConductionNL/hydra#688 to match.

🤖 Generated with Claude Code

[WilcoLouwerse]

Inline findings for the Thorough review of #788 — verdict follows in a separate submission.

[WilcoLouwerse]

Verdict: APPROVE (Thorough) — self-review posted as COMMENT (GitHub blocks self-APPROVE)

No 🔴. Six non-blocking findings posted inline: three 🟡 on the cascade model, three 🟢.

What I verified rather than took on trust — every claim below came from running the runner at a8c8ba9c and at a8c8ba9c^ against the same fixture, one variable at a time:

• The new suite is 24/24 green locally and PASS — test_gate_45_selector_parity.sh appears in the hydra-gates package CI log, with ALL discovered helper suites PASSED. run-helper-suites.sh picks it up through find "${LIB}" -maxdepth 1 -type f -name 'test_*', so "discovered automatically" holds in CI, not only on the authoring machine.
• test_gate_45_stylesheet_scope.sh 13/13 and test_gate_a11y_markup_scope.sh all green — no regression from the rewrite.
• The body's "Not run: shellcheck" caveat is closed: the ShellCheck workflow ran on this PR and passed.
• The one red at this SHA — quality.yml on the push event, 0 jobs — is a repo-wide startup failure that fails identically on main and on every recent branch. Not this PR's, and not a reason to hold it.
• The change earns more than the fixture it was built from. A reduced-motion block that names the selector but declares no motion property (@media(reduce){ .btn { color: red } }) now FAILs where the pre-PR checker passed — a true positive the PR body doesn't claim.

Why the 🟡s don't block. All three are misses or noise on idioms the fleet does not currently write, and the fleet dry run's zero-new-findings result reproduces. The one I'd most like addressed is the first — an ancestor-scoped guard like .app-wrapper .btn over .btn goes PASS → FAIL across this PR while winning the cascade outright, and :root .btn on the identical fixture passes. Since gate 45 has no exclude hatch, an app that writes that idiom has no appeal. It is defensible conservatism (html/body/:root always match, an arbitrary class ancestor may not) — it just isn't written down as a choice.

The second 🟡 is a one-line fix in new code (gi is derived per rule, not per declaration). The third is the comment at :7741 asserting a source-order property that _covered never checks — the same "assert instead of observe" shape this PR exists to remove, one level down.

None of that outweighs the change. The gate now observes the property it is named after, the controls are pinned before the behaviour widened rather than after, and the anti-widening measurement is real. 👍

[WilcoLouwerse]

Verdict: APPROVE (Quick, re-review) — self-review posted as COMMENT (GitHub blocks self-APPROVE)

All six findings from the first round are addressed in 0c1c449a and their threads are resolved. No new findings in the delta.

Each fix is verified twice, because a gate fix that only makes its own new test pass is the failure mode this gate exists to catch:

1. the fixture that misbehaved now matches the cascade — measured through the runner, not reasoned about; and
2. a new suite arm discriminates: run the extended suite against a8c8ba9c and exactly P12, P13, P14, P15, P16 and P17 go red while every pre-existing arm stays green. The controls that should not discriminate — P12b, P14b, P16b — are green in both directions.

finding	fix	control
ancestor-scoped guard reported although it wins	_descends, one combinator-boundary test replacing four hardcoded prefixes	P12 / P12b
!important read per rule	per-declaration importance and property-family matching	P13 / P17
source order assumed	brace offset carried through _rules, applied where specificity ties	P14 / P14b
literal duration beside a zeroed token	_nonzero_time on the motion value	P15
token redefinition treated as zeroing	_zeroish, 0.01ms still counting	P16 / P16b
file-local guards undocumented	stated in the comment block; behaviour unchanged	—

One of these needed more than the thread asked for. Finding 2's suggested one-line fix — read !important per declaration — still passed its fixture, because the !important sat on animation while the motion was a transition and nothing matched the two families. It closes only with family matching as well. Recording that here rather than quietly widening the fix.

Green: the new suite is 24 → 35 assertions; test_gate_45_stylesheet_scope.sh 13/13 and test_gate_a11y_markup_scope.sh unchanged; the full local run-helper-suites.sh reports 124 passed, 0 failed (1 pre-existing documented quarantine); CI on 0c1c449a is green on all four pull_request workflows including hydra-gates package and ShellCheck. The quality.yml push run is the repo-wide startup failure that fails identically on main — not this PR's.

Ready to merge as far as I'm concerned; I won't merge my own PR.