[robobun]

Problem

• bun build --minify prints input:-moz-read-only::placeholder{color:red} plus an extra input:read-only::placeholder{color:red} that is not in the source and matches in Chrome and Safari (lightningcss does the same).
• Cause: StyleRule::to_css prints a rule once per vendor prefix pass. A prefixed component printed unprefixed in a pass it had no variant for (src/css/selectors/selector.rs), get_prefix gave author-mixed rules an unprefixed pass, and merge_style_rules merged the passes of equivalent rules but not their components' prefixes.
• The pseudo class path also used | for &: .a:autofill with chrome 103 printed .a:autofill{} twice. The first commit fixes only that and stands alone.

Fix

• A prefixed component prints the pass prefix only when it has that variant, else as written. An :is() that needs :-webkit-any() becomes Component::Any with those prefixes.
• A rule's passes are the extra prefixes its components got from the targets or from merged rules, plus one pass as written. Merging equivalent rules merges the prefix set of each component pair.
• Correct because a prefix prints only for a component written with it, needing it for the targets, or merged with an equivalent rule that has it.
• Verified: test/js/bun/css/css.test.ts (84 new cases, 61 fail on 1.4.2), nested-vendor-prefix-duplication.test.ts, all of test/js/bun/css/ and test/bundler/css/. Self-reviewed: 5 concerns raised, 4 addressed. Not taken: a separate PR for the & fix. It is the first commit instead.

Background

• src/css ports lightningcss. A prefixable pseudo class or element stores a VendorPrefix bit set: its written prefix, or NONE plus what the targets need. StyleRule::to_css serializes the rule once per pass: .a::placeholder becomes .a::-webkit-input-placeholder{…}.a::placeholder{…} for old Safari.
• The minifier merges adjacent rules with equal declarations and selectors equal up to prefixes into one rule with several passes.

Notes

• Upstream lightningcss 1.33.0 prints the same output as bun 1.4.3 for every case here except the | typo, which is bun-only. The upstream mixing rule in get_prefix is also order dependent: .b:autofill .a:placeholder-shown (chrome 103) round-trips through a second minify unchanged, .a:placeholder-shown .b:autofill grows a duplicate rule. Both are stable now.
• More symptoms of the same cause, all fixed: .p::placeholder,.p::-moz-placeholder{} gained a .p::placeholder,.p::placeholder{} rule; .a:read-only{c}.a:-moz-read-only{c} printed .a:read-only{c} twice and lost the -moz- rule; .a::-webkit-input-placeholder{c}.a::-moz-placeholder{c} printed the second rule as ::placeholder (bulma.css); .a:not(:-moz-read-only) printed as .a:not(:read-only); .a:is(.x,.y):-webkit-autofill printed as :-webkit-any(.x,.y):-webkit-autofill plus :is(.x,.y):autofill; a nested .x:-moz-read-only { .a::placeholder {} } printed the parent as .x:read-only.
• :is() in a -webkit- pass: .x:is(.a .b,.c):autofill with chrome 103 printed .x:-webkit-any(.a .b,.c):-webkit-autofill, which Chrome rejects (-webkit-any() takes compound selectors only), so the -webkit-autofill styling was lost. It now prints .x:is(.a .b,.c):-webkit-autofill. :is() still becomes :-webkit-any() / :-moz-any() when the targets lack :is(): downlevel_selectors converts it (and the :is() it builds for :not(a, b)) to Component::Any with the needed prefixes, which prints the same strings as before (test/bundler/css/is-selector-21169.test.ts and test/regression/issue/25794.test.ts are unchanged).
• update_prefix runs again on every style rule of an @media rule when an adjacent @media with the same query is merged into it. Every new test case also runs inside such a pair, so the converted :is(), the merged prefix sets and the passes have to come out the same the second time. An earlier revision of this branch printed .a:is(.b,.c) only (no -webkit-any/-moz-any copies) in that position.
• Merging: selector::merge_prefixes ORs the prefix set of each component pair and first checks, from the sets alone, that the merged rule prints exactly the variants the two rules print. .a:-moz-read-only::placeholder{c} + .a:read-only::-moz-placeholder{c} stays two rules: one merged rule would print -moz-read-only::-moz-placeholder and read-only::placeholder. With targets, a rule with no explicit prefixes absorbs any equivalent rule regardless, and a component with an unprefixed variant keeps exactly the prefixes the targets need (Targets::prefixes). That is lightningcss's removal of unneeded prefixed copies, now in either rule order: .a:-moz-read-only{c} next to .a:read-only{c} is .a:read-only{c} for firefox 100 whichever comes first (before: only when the unprefixed rule came second). A kept :is() becomes Component::Any when an explicit :-webkit-any() rule merges into it.
• Nested rules with nesting compiled away: a nested rule with prefixes of its own is printed once, in its own passes, with the parent selectors substituted for & (css: don't duplicate vendor-prefixed nested rules per ancestor prefix pass #31270). Those passes now also include the extra prefixes of the parents' components (selector::nested_prefix_passes, used by StyleRule::to_css when a parent context is set). :fullscreen { .x::selection {} } with safari 8 printed only :fullscreen .x::selection on main and now also prints :-webkit-full-screen .x::selection; :fullscreen { .a::-webkit-input-placeholder {} } prints the :-webkit-full-screen and :fullscreen variants, both with ::-webkit-input-placeholder. A rule that mixes two explicit vendors (.a:-webkit-autofill:-moz-read-only) still gets no passes of its own and prints in each parent pass.
• The :is() that :not(a, b) lowers to now gets the same combinator check as a written :is(): .x:not(.a .b,.c) with safari 8 prints .x:not(:is(.a .b,.c)) instead of an extra :not(:-webkit-any(.a .b,.c)) copy that browsers reject.
• Cross-vendor passes are kept: input:-moz-read-only::placeholder with firefox 30 and safari 8 targets prints a ::-webkit-input-placeholder pass that keeps :-moz-read-only. No browser matches it, but it is valid, it never applies where the source does not, and dropping it would put a vendor heuristic on top of the prefix tables.
• The tests from css: bound vendor-prefix fan-out when serializing nested rules #31642 in nested-vendor-prefix-duplication.test.ts reached the 64 MB prefix-expansion bound through author-mixed selectors (.a:placeholder-shown .x, .b:-webkit-autofill .y nested 20 deep fanned out 2^20 times). Those inputs now print once per level, so the tests assert the exact linear output instead of the error. The bound stays but has no test now: reaching it needs a multi-pass rule with nested rules printed with nesting preserved, and with the current prefix tables every browser version that needs a selector prefix predates nesting support, while merged rules never have nested rules. It still guards a minify whose targets differ from the printer's.
• The 20k-rule flat fan-out test in the same file spends 3.6 s in minify in a debug build, before and after this change, and timed out at the 5 s default here, so it now uses 10k rules.
• Merge order side effect without targets: ::-ms-input-placeholder{c} ::placeholder{c} .x{c} (foundation.css) now merges the first two rules, so .x no longer joins ::placeholder's selector list: one more rule, same meaning (+19 bytes on foundation.css with --target=bun, -62 bytes with browser targets).
• Corpus: test/js/bun/css/files/*.css with --minify for --target=browser and --target=bun: 4 of 30 outputs change, all in placeholder rules. foundation keeps ::-moz-placeholder and bulma keeps ::-webkit-input-placeholder; both were printed as ::placeholder before.
• Overlap: css: keep a selector list whole when a member is vendor prefixed #42100 (keep selector lists with a vendor-prefixed member whole) adds a guard to update_prefix that prints a rule mixing NONE with a vendor prefix once as written, with no downlevel of its unprefixed members; this change covers that case (with the downlevel) and the guard can go when both land. css: stop emitting :-webkit-any() copies of :is() that outrank the original rule #42029 changes the same :is() lowering arm and serializer branch (it gates -webkit-any() copies on specificity); its gate would decide whether downlevel_selectors converts an :is() here. css: keep incompatible selectors that fail the same way in one rule #40464 shares update_prefix. Whichever lands second needs a rebase.

---

[human-review] gate passed · iteration 0 · 8 files touched

fails on main (without fix)

ASAN without fix: 66 failed, 67 skipped
$ BUN_DEBUG_QUIET_LOGS=1 bun scripts/build.ts --profile=debug --quiet test "--reporter=junit" "--reporter-outfile=/tmp/pr_gate.xml" test/js/bun/css/css.test.ts test/js/bun/css/nested-vendor-prefix-duplication.test.ts
bun test v1.4.3 (f42e98025)

test/js/bun/css/css.test.ts:
Output .flexrow {
  flex-direction: row;
}

.flexcol {
  flex-direction: column;
}

.hello {
  flex-wrap: wrap;
}

.world {
  flex-wrap: nowrap;
}

(pass) css tests > .flexrow {
	flex-direction: row;
}

.flexcol {
	flex-direction: column;
}

.hello {
	flex-wrap: wrap;
}

.world {
	flex-wrap: nowrap;
} [3.01ms]
Output .foo {
  content: "+";
}

(pass) css tests > .foo {
  content: "\2b";
} [1.08ms]
Output div {
  --foo: 1 1 1 #0101011a, 2 2 2 #02020233;
  --bar: 1 1 1 #01010116, 2 2 2 #02020233;
}

(pass) css tests > custom property cases > div {
        --foo: 1 1 1 rgba(1, 1, 1, 0.1), 2 2 2 rgba(2, 2, 2, 0.2);
        --bar: 1 1 1 #01010116, 2 2 2 #02020233;
      } [1.01ms]
Output :root {
  --my-color: red;
  --my-bg: white;
  --my-font-size: 16px;
}

.element {
  color: var(--my-color);
  background-color: var(--my-bg);
  font-size: var(--my-font-size);
}

(pass) css tests > custom property cas
... (truncated)

release without fix: 12 failed, 67 skipped
bun test v1.4.3-canary.1 (c8b90a7ee)

test/js/bun/css/css.test.ts:
Output .flexrow {
  flex-direction: row;
}

.flexcol {
  flex-direction: column;
}

.hello {
  flex-wrap: wrap;
}

.world {
  flex-wrap: nowrap;
}

(pass) css tests > .flexrow {
	flex-direction: row;
}

.flexcol {
	flex-direction: column;
}

.hello {
	flex-wrap: wrap;
}

.world {
	flex-wrap: nowrap;
} [0.11ms]
Output .foo {
  content: "+";
}

(pass) css tests > .foo {
  content: "\2b";
} [0.02ms]
Output div {
  --foo: 1 1 1 #0101011a, 2 2 2 #02020233;
  --bar: 1 1 1 #01010116, 2 2 2 #02020233;
}

(pass) css tests > custom property cases > div {
        --foo: 1 1 1 rgba(1, 1, 1, 0.1), 2 2 2 rgba(2, 2, 2, 0.2);
        --bar: 1 1 1 #01010116, 2 2 2 #02020233;
      } [0.03ms]
Output :root {
  --my-color: red;
  --my-bg: white;
  --my-font-size: 16px;
}

.element {
  color: var(--my-color);
  background-color: var(--my-bg);
  font-size: var(--my-font-size);
}

(pass) css tests > custom property cases > :root {
        --my-color: red;
        --my-bg: white;
        --my-font-size: 16px;
      }
      .element {
        color: var(--my-color);
        background-color: var(--my-bg);
        font-size: 
... (truncated)

passes on PR (with fix)

ASAN with fix: 67 skipped
$ BUN_DEBUG_QUIET_LOGS=1 bun scripts/build.ts --profile=debug --quiet test "--reporter=junit" "--reporter-outfile=/tmp/pr_gate.xml" test/js/bun/css/css.test.ts test/js/bun/css/nested-vendor-prefix-duplication.test.ts
bun test v1.4.3 (f42e98025)

test/js/bun/css/css.test.ts:
Output .flexrow {
  flex-direction: row;
}

.flexcol {
  flex-direction: column;
}

.hello {
  flex-wrap: wrap;
}

.world {
  flex-wrap: nowrap;
}

(pass) css tests > .flexrow {
	flex-direction: row;
}

.flexcol {
	flex-direction: column;
}

.hello {
	flex-wrap: wrap;
}

.world {
	flex-wrap: nowrap;
} [2.89ms]
Output .foo {
  content: "+";
}

(pass) css tests > .foo {
  content: "\2b";
} [1.01ms]
Output div {
  --foo: 1 1 1 #0101011a, 2 2 2 #02020233;
  --bar: 1 1 1 #01010116, 2 2 2 #02020233;
}

(pass) css tests > custom property cases > div {
        --foo: 1 1 1 rgba(1, 1, 1, 0.1), 2 2 2 rgba(2, 2, 2, 0.2);
        --bar: 1 1 1 #01010116, 2 2 2 #02020233;
      } [1.11ms]
Output :root {
  --my-color: red;
  --my-bg: white;
  --my-font-size: 16px;
}

.element {
  color: var(--my-color);
  background-color: var(--my-bg);
  font-size: var(--my-font-size);
}

(pass) css tests > custom property cas
... (truncated)

release with fix: 67 skipped
$ bun scripts/build.ts --profile=release
[configured] bun-profile → bun (stripped) in 618ms (unchanged)
ninja: Entering directory `/workspace/bun/build/release'
[1/138] gen generated_host_exports.rs
generated_host_exports.rs: 122 exports (host=5, lazy=10, generic=107, rust=0); 242 extern-C blocks audited
[2/138] gen ZigGeneratedClasses.{cpp,h,rs}
Found 2 classes from /workspace/bun/src/jsc/resolve_message.classes.ts
  - ResolveMessage (15 fields)
  - BuildMessage (10 fields)
Found 1 classes from /workspace/bun/src/runtime/api/Archive.classes.ts
  - Archive (4 fields, 1 class fields)
Found 2 classes from /workspace/bun/src/runtime/api/BunObject.classes.ts
  - ResourceUsage (8 fields)
  - Subprocess (20 fields)
Found 1 classes from /workspace/bun/src/runtime/api/cron.classes.ts
  - CronJob (5 fields)
Found 3 classes from /workspace/bun/src/runtime/api/filesystem_router.classes.ts
  - FileSystemRouter (5 fields)
  - FrameworkFileSystemRouter (2 fields)
  - MatchedRoute (8 fields)
Found 1 classes from /workspace/bun/src/runtime/api/Glob.classes.ts
  - Glob (5 fields)
Found 1 classes from /workspace/bun/src/runtime/api/h2.classes.ts
  - H2FrameParser (32 fields)
Found 
... (truncated)

diff hotspot

src/css/lib.rs                                     |  10 +
 src/css/rules/mod.rs                               |  25 +-
 src/css/rules/style.rs                             |  18 +-
 src/css/selectors/parser.rs                        |  40 ++-
 src/css/selectors/selector.rs                      | 372 ++++++++++++++++-----
 test/js/bun/css/css.test.ts                        | 230 ++++++++++++-
 .../css/nested-vendor-prefix-duplication.test.ts   | 139 ++++----
 test/js/bun/css/util.ts                            |   6 +
 8 files changed, 654 insertions(+), 186 deletions(-)

gate history · 1 passed · 0 rejected · iteration 0

evidence per changed file

file                                                      reads  edits  tests
src/css/lib.rs                                                1      1     45
src/css/rules/mod.rs                                          3      4     45
src/css/rules/style.rs                                        2      3     45
src/css/selectors/parser.rs                                   5      2     45
src/css/selectors/selector.rs                                17     23     45
test/js/bun/css/css.test.ts                                   5      7     38
test/js/bun/css/nested-vendor-prefix-duplication.test.ts      4      3     18
test/js/bun/css/util.ts                                       1      1     47

[robobun]

Status: closed, not taken. A maintainer declined this change when #47025 consolidated the open CSS pull requests (see the comment below). No further work is planned here.

The & fix for pseudo class prefix passes from the first commit is on main through #47025.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Essentials

Run ID: 2351f3d0-4396-463f-99e5-41bbcd6aeae3

📥 Commits

Reviewing files that changed from the base of the PR and between e1ed95f and e97b1dd.

📒 Files selected for processing (3)

• src/css/rules/style.rs
• src/css/selectors/selector.rs
• test/js/bun/css/css.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 0 remain after this review.

---

Walkthrough

Changes

CSS selector vendor-prefix handling now calculates prefix passes, preserves explicit variants, merges equivalent selectors, and serializes prefixes according to active passes. Style-rule merging and nested CSS regression tests were updated.

CSS vendor-prefix processing

Layer / File(s)	Summary
Selector prefix processing src/css/lib.rs, src/css/selectors/parser.rs, src/css/selectors/selector.rs	Centralized prefix mappings, added canonical prefix and prefix-pass handling, updated selector downleveling, and changed serialization to use active prefix passes.
Style-rule prefix merging src/css/rules/style.rs, src/css/rules/mod.rs	Style rules now derive prefixes from selector passes. Rule merging uses merge_prefixes, updates cached prefix state, and unions source prefixes.
Prefix regression coverage test/js/bun/css/css.test.ts, test/js/bun/css/nested-vendor-prefix-duplication.test.ts, test/js/bun/css/util.ts	Tests cover explicit and target-dependent prefixes, selector merging, nested rules, prefix removal, and linear output expansion.

Suggested reviewers: jarred-sumner

Priority: ➖ Normal

Merge Risk: ⚪ Minimal · up to e97b1

This change corrects CSS selector vendor-prefix serialization and adds regression coverage for prefixed, nested, and merged rules. No concrete unresolved merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and concisely describes the main change: preserving explicitly prefixed selectors across vendor-prefix serialization passes.
Description check	✅ Passed	The description explains the problem, fix, background, and verification results. It does not use the exact template headings, but it provides the required information and detailed test evidence.

---

_{Comment @coderabbitai help to get the list of available commands.}

[robobun]

The automated review above has no actionable items, and there are no open review threads. Of the checks so far, lint, format, clippy and source lints pass; claude-find-issues is red because the action itself errored at startup (833 ms, is_error: true, no findings), not because of the diff. Buildkite #113422 is still running.

[robobun]

On the comment-length bot notes: the flagged comments are doc comments on the new functions (prefix_passes, merge_prefixes, pass_prefix, VendorPrefix::canonical, prefix_mut) that state each one's contract, plus two pre-existing comments that only moved. None of them justifies a workaround. e1ed95f shortens the longest ones (prefix_passes and merge_prefixes from 8 and 10 lines to 5 and 6, pass_prefix and the :is() note to 3 and 1). I resolved those threads; say so if any of them should go further.

[robobun]

e97b1dd addresses both review notes above: a nested rule printed with nesting compiled away now takes its prefix passes from its own components and the parent selectors substituted for &, so an explicit prefix in the child no longer hides a parent variant (and :fullscreen { .x::selection {} } with safari 8 gains the :-webkit-full-screen variant that main also missed), and the :is() that :not(a, b) lowers to gets the same combinator check as a written :is(). Four test cases added, PR description updated. The comment-length bot notes on this push are two- and three-line doc comments; resolved.

[robobun]

The automated incremental review of e97b1dd has no actionable items either, and all review threads are resolved. The PR description now covers the nested-pass and :not() changes from that commit. Buildkite #113448 is running for it.

[claude]

Both earlier findings are addressed in e97b1dd — passes_for now always includes the as-written pass and nested_prefix_passes folds in parent selector sets so :fullscreen{.a:read-only:-webkit-any-link{…}} prints the :-webkit-full-screen variant again, and the :not(a, b) lowering now goes through is_prefixes so .x:not(.a .b,.c) no longer emits an invalid :-webkit-any(.a .b,.c) copy; both have regression tests. I re-checked the simplified passes_for against the single-explicit-prefix case and the merge/printed_variants guard and found nothing further. Given this reworks the vendor-prefix pass computation and rule merging across selector.rs/style.rs and intentionally diverges from upstream lightningcss, a human look is still worthwhile.

Extended reasoning...

Overview

This PR reworks how Bun's CSS minifier decides which vendor-prefix passes to print a style rule in and how it merges adjacent equivalent rules. The core changes are in src/css/selectors/selector.rs (~370 lines: new prefix_sets/passes_for/printed_variants/merge_prefixes/is_prefixes/any_prefixes/nested_prefix_passes helpers, Component::Is → Component::Any conversion in downlevel_component, and a pass_prefix serializer helper replacing the old per-pass prefix override), plus src/css/rules/style.rs (to_css now folds parent-selector passes in when nesting is compiled away), src/css/rules/mod.rs (merge_style_rules rewired to the new merge helper), src/css/lib.rs (VendorPrefix::canonical()), and src/css/selectors/parser.rs (prefix_mut() on pseudo-classes/elements). Tests add ~230 lines to css.test.ts and rewrite nested-vendor-prefix-duplication.test.ts to assert exact linear output instead of the previous byte-limit error.

Security risks

None identified. This is pure CSS AST transformation and serialization operating on already-parsed selector components; no untrusted length arithmetic, no filesystem/network I/O, no FFI or JSC interaction. Allocations are short-lived Vec<VendorPrefix> scratch buffers.

Level of scrutiny

High. The change replaces the pass-computation and rule-merge logic that governs minified output for every rule containing a vendor-prefixable pseudo, and it deliberately diverges from lightningcss upstream on several inputs. Two earlier rounds of this review each found a real behavioral regression; both were fixed in e97b1dd with the exact test cases suggested. On this pass I traced the new simplified passes_for (always NONE ∪ (each set − canonical)) through the single-explicit-prefix case ([WEBKIT] → NONE pass, component prints as written via pass_prefix), the mixed case ([MOZ|NONE, WEBKIT] with parent [WEBKIT|MOZ|NONE] → WEBKIT|MOZ|NONE), and confirmed merge_prefixes' printed_variants union/subset check still refuses the differ-two-ways case. No new issues found.

Other factors

Test coverage is substantial (84 new cases per the description, including the two regression cases from the prior review, @ media-wrapped re-runs of update_prefix, and the :not() combinator guard). The nested-vendor-prefix-duplication.test.ts rewrite drops the fan-out from 20k to 10k rules to stay under debug-build timing. Given the scope of the pass-computation rewrite and the intentional upstream divergence, a maintainer familiar with the lightningcss port should sign off rather than this being auto-approved.

[robobun]

Thanks, that matches my understanding: both earlier findings are fixed in e97b1dd with regression cases, and no review threads are open. Agreed that a maintainer who knows the lightningcss port should give the final look, mainly at the three places this now differs from upstream on purpose: a prefixed component prints as written in a pass it has no variant for (pass_prefix), an :is() that needs :-webkit-any() becomes Component::Any at downlevel time, and equivalent-rule merging ORs per-component prefix sets behind the printed_variants check. CI on e97b1dd is green apart from the two unrelated lanes noted in the status comment above.

[robobun]

Heads-up for review order: #42131 is the standalone version of this PR's first commit (the | to & change in write_prefixed), plus the matching contains() to == changes in the :fullscreen, ::placeholder and ::file-selector-button arms and tests for both. If #42131 lands first, commit 7efa52f here becomes empty on rebase and the pass_prefix rewrite applies on top unchanged. #42131 adds no test for author-written mixed-prefix lists, so it does not pin the behaviour this PR changes.

[robobun]

Note for review: this PR also fixes a second bug that has the same cause in get_prefix. The bug came in as a separate report. I checked it with a build of main merged with this branch.

Symptom on main and on 1.4.2. Put a logical property on a rule whose last compound has :fullscreen. With the default targets, bun prints an invalid selector and leaves :dir() as written. Browsers drop both rules, so the declaration is lost. There is no warning.

$ printf '.c0:fullscreen { inset-inline-start:4px }\n' > in.css && bun build in.css --minify
.c0:full-screen:dir(ltr){left:4px}.c0:full-screen:dir(rtl){right:4px}

Cause. The first update_prefix widens Fullscreen(NONE) to Fullscreen(WEBKIT | NONE). get_additional_rules copies those selectors and appends :dir(). The copies go through update_prefix again. On main, get_prefix then sees WEBKIT | NONE followed by NONE, compares WEBKIT == NONE, and returns empty. The rule gets no passes and no downlevel, and the :fullscreen arm prints the legacy name for the multi-bit set. .c0:fullscreen .c1 { ... } works on main only because :dir() comes first in that selector. The get_prefix in this PR does not depend on the order and returns WEBKIT | NONE. The output is then:

.c0:-webkit-full-screen:not(:-webkit-any(:lang(ae),...)){left:4px}
.c0:fullscreen:not(:-moz-any(:lang(ae),...)){left:4px}
.c0:fullscreen:not(:is(:lang(ae),...)){left:4px}
(and the same three selectors without :not() for right:4px)

Other inputs with the same cause. All of them print correctly with this PR.

• .a:autofill { inset-inline-start:4px } prints .a:autofill:dir(ltr) on main. There is no :-webkit-autofill pass, and :dir() stays as written.
• .a:fullscreen::file-selector-button { inset-inline-start:4px } prints .a:full-screen:dir(ltr)::file-upload-button on main.
• A logical property is not necessary. When an adjacent @media with the same query is merged in, the rule is minified a second time. .a:fullscreen:not(.b), .a:fullscreen .b:fullscreen and .a:fullscreen::file-selector-button then print as .a:full-screen:not(.b), .a:full-screen .b:full-screen and .a:full-screen::file-upload-button.

#42131 alone does not fix this. I built main with its diff: the spelling becomes :fullscreen, but the rule still gets no -webkit- pass and :dir() stays as written. lightningcss 1.33.0 prints that same output.

Tests. 0d9a56e adds 14 cases for these inputs to css.test.ts. They use the minify_prefix_test_again helper from this PR. It is one test-only commit on top of e97b1dd, on the branch robobun/47308683/css-fullscreen-dir-tests. 9 of the 14 cases fail on 1.4.2. With main, this PR and that commit together, all of css.test.ts passes (1279 pass, 0 fail). The commit is not part of this PR yet. Cherry-pick it to pin this behaviour here.

[robobun]

Cross-reference from #42100 (keep a selector list whole when a member is vendor prefixed).

#42100 no longer touches update_prefix or the selector printer, so it merges cleanly with this branch (checked with git merge-tree at a6dc48f against e97b1dd). It only changes the split gate in minify_style_arm, and it is deliberately conservative about this PR's territory: selector::prints_as_written() keeps a vendor-prefixed list whole only when main's printer writes every member as written. A list that mixes a prefixed pseudo with an unprefixed prefixable one (.p::placeholder, .p::-moz-placeholder, .q:-webkit-autofill, .q:autofill) still takes the split path, because a prefix pass for another prefix would print the prefixed member unprefixed.

With the printer from this PR that restriction is not needed. Whichever of the two lands second can drop the prints_as_written() check in StyleRule::should_compile_selector_list so that every Compatibility::VendorPrefixed list stays whole. One more input for this PR's merge logic: once lists stay whole, bulma.css line 5474 (.select select::-moz-placeholder, .input::-moz-placeholder, … followed by the same list with ::-webkit-input-placeholder and the same body) merges through is_equivalent, and main's printer then writes the second list as ::placeholder.

[robobun]

Correction to my note above about #42100.

"Merges cleanly" holds only for the text. A self-review of #42100 built the two branches together and found that the combination is wrong as it stands: this PR ORs the prefixes of merged twins into the selectors in place, and #42100's prints_as_written() then reads a merged MOZ|WEBKIT set as two author-written prefixes, so a list that was kept whole is split again when a same-query @media block is minified a second time.

The plan for #42100 changed because of that and of other regressions it has on main's printer:

• css: keep a selector list whole when a member is vendor prefixed #42100 stays a draft and lands last, after this PR, css: keep nested rules in source order across vendor prefix passes #43407 and css: keep target-added vendor prefix passes on cloned and re-minified rules, and never print a bare :full-screen #43339.
• It is then rebased to the gate only: prints_as_written() and its scan are deleted, and a VendorPrefixed list stays whole unless the rule has nested rules.
• Nothing in this PR needs to change for it.

One input from that review that belongs here: with whole lists, x:-moz-full-screen, .b {color:red} followed by x:-webkit-full-screen, .b {color:red} reaches the is_equivalent merge, and main's printer writes the second as x:fullscreen,.b, which Safari 14 to 16.3 do not parse. The review reports that this PR's merge prints both lists as written.

[Jarred-Sumner]

Not taken in #47025 (7942008), which consolidated the open CSS parser pull requests. It is right or harmless, but not worth its code, the bigger output or the behavior it changes.

[robobun]

Understood. I am leaving this closed and doing no further work on it.

For later readers, two parts of this are on main at 7942008 through #47025:

• the & fix for pseudo class prefix passes from the first commit here (src/css/selectors/selector.rs:1165)
• StyleRule::prints_selectors_as_written, which gives a selector list with an author-written vendor prefix no prefix passes

The per-component prefix passes and the rule-merge changes from this PR are not on main.