Plan step 41 (ii): the design guard needs the set of files an application's
CSS entry pulls in, transitively, to check whether a package tag's
stylesheet is among them — reusing bundle()'s own import resolution and
cache rather than re-implementing @import parsing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 37 review. The Vite test counted each of four files' plain
root rule, which the minifier folds into one even when a build repeats
the whole stylesheet: it passed against the Workbench's Tailwind entry,
which repeated 231 rules. It now asserts that no innermost rule repeats
under the same at-rules, which that build fails and a deduplicating one
passes, and it builds a second entry shaped like an application's —
outside the package, the foundation, then four component stylesheets
that each import button.css, and button.css again — where button.css
must also keep its first position. The fixture config refuses to run
without DEDUP_OUT_DIR, so it never writes into the package tree.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 37 review. bundle() left an @import it did not recognise
exactly as written, in the middle of the output, where a browser ignores
it: a layer(), supports() or media condition was silently dropped, an
unquoted url(./x.css) import was left in place with its path mangled by
the url() rewrite, and a bare specifier or absolute URL vanished the same
way. An @import after a rule or inside a @layer block was inlined anyway,
nesting a whole file's layers inside another. Each now throws, naming the
file; unquoted url() imports and @IMPORT inline like the other forms.
The cache keyed only the top-level files' mtimes, so a changed
button.css left a cached all.css bundle standing in a long-lived worker;
a cached bundle is now served only while every file it inlined keeps its
mtime. A null and an empty $base no longer share a cache key, a leading
@charset or byte-order mark is dropped from each inlined file, and a
test pins the import graph free of cycles, the one case where a
depth-first bundle would place a stylesheet before a file it imports.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 37: the showcase and the error page's fallback serve CSS
without the application's Vite build, so nothing deduplicates their
@imports for them the way Vite's bundled postcss-import does — a
browser's native @import fetches every occurrence, it does not skip a
file it already loaded. Stylesheets::bundle(array $files, ?string
$base = null): string does in PHP what that build step does: it
inlines every @import depth-first, each file once, first occurrence
kept; leaves a bare specifier or an absolute URL untouched; rewrites a
relative url() against $base (or, without one, against the directory
of $files[0]); breaks a cycle instead of looping; throws naming the
importer when an import is missing; and caches per resolved file list
and mtime, with resetCache() for tests.
tests/Feature/StylesheetsBundleTest.php covers bundle()'s own
behaviour (dedup, url() rewriting, cycles, the missing-import
exception, the cache) and, at the end, pins the Vite deduplication
bundle() is modelled on against a real build — of all.css alone
(tests/Fixtures/dedup.vite.config.mjs), whose docblock explains why:
the Workbench's own entry still shares one file with @tailwindcss/vite
until plan step 39 removes it, and that plugin bundles its whole
reachable module graph itself, without the same dedup guarantee.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
