perf(css): join an at-rule with a later one stating the same condition - #22174

…the changeset

Reported by CodeRabbit on #22174.

The changeset line ran to 86 characters where the guide caps it at 80.

`@keyframes phi` was asserted by a substring neither the merged nor the
separated form contains, so it could not fail; it now states the whole
two-block sequence the case is about. An anonymous `@layer` was not
covered at all — `style.css` had only the named one — and it is the case
worth pinning, since two anonymous layers are two layers rather than one
condition written twice.

`_gatherDistantAtRule` now names the layer rule itself rather than
leaving it to `_joinRuleEntries`, which already refuses both spellings
(`ANONYMOUS_LAYER_RE`, and its `mergeRules` bail). No output moves — the
new case passes on the parent commit too — but the streaming path now
reads like the nested one, and a layer no longer occupies a held piece
that could never be folded.

…the changeset

Reported by CodeRabbit on #22174.

The changeset line ran to 86 characters where the guide caps it at 80.

`@keyframes phi` was asserted by a substring neither the merged nor the
separated form contains, so it could not fail; it now states the whole
two-block sequence the case is about. An anonymous `@layer` was not
covered at all — `style.css` had only the named one — and it is the case
worth pinning, since two anonymous layers are two layers rather than one
condition written twice.

`_gatherDistantAtRule` now names the layer rule itself rather than
leaving it to `_joinRuleEntries`, which already refuses both spellings
(`ANONYMOUS_LAYER_RE`, and its `mergeRules` bail). No output moves — the
new case passes on the parent commit too — but the streaming path now
reads like the nested one, and a layer no longer occupies a held piece
that could never be folded.

…the changeset

Reported by CodeRabbit on #22174.

The changeset line ran to 86 characters where the guide caps it at 80.

`@keyframes phi` was asserted by a substring neither the merged nor the
separated form contains, so it could not fail; it now states the whole
two-block sequence the case is about. An anonymous `@layer` was not
covered at all — `style.css` had only the named one — and it is the case
worth pinning, since two anonymous layers are two layers rather than one
condition written twice.

`_gatherDistantAtRule` now names the layer rule itself rather than
leaving it to `_joinRuleEntries`, which already refuses both spellings
(`ANONYMOUS_LAYER_RE`, and its `mergeRules` bail). No output moves — the
new case passes on the parent commit too — but the streaming path now
reads like the nested one, and a layer no longer occupies a held piece
that could never be folded.

…the changeset

Reported by CodeRabbit on #22174.

The changeset line ran to 86 characters where the guide caps it at 80.

`@keyframes phi` was asserted by a substring neither the merged nor the
separated form contains, so it could not fail; it now states the whole
two-block sequence the case is about. An anonymous `@layer` was not
covered at all — `style.css` had only the named one — and it is the case
worth pinning, since two anonymous layers are two layers rather than one
condition written twice.

`_gatherDistantAtRule` now names the layer rule itself rather than
leaving it to `_joinRuleEntries`, which already refuses both spellings
(`ANONYMOUS_LAYER_RE`, and its `mergeRules` bail). No output moves — the
new case passes on the parent commit too — but the streaming path now
reads like the nested one, and a layer no longer occupies a held piece
that could never be folded.

…the changeset

Reported by CodeRabbit on #22174.

The changeset line ran to 86 characters where the guide caps it at 80.

`@keyframes phi` was asserted by a substring neither the merged nor the
separated form contains, so it could not fail; it now states the whole
two-block sequence the case is about. An anonymous `@layer` was not
covered at all — `style.css` had only the named one — and it is the case
worth pinning, since two anonymous layers are two layers rather than one
condition written twice.

`_gatherDistantAtRule` now names the layer rule itself rather than
leaving it to `_joinRuleEntries`, which already refuses both spellings
(`ANONYMOUS_LAYER_RE`, and its `mergeRules` bail). No output moves — the
new case passes on the parent commit too — but the streaming path now
reads like the nested one, and a layer no longer occupies a held piece
that could never be folded.

Reported by CodeRabbit on #22174, which read `_mergeDistantAtRule` as
merging an escaped `@keyframes` because the name check would miss it.

It does not, and the reason is one level up: entries are collected only
for a name `MERGEABLE_AT_RULES` holds, read off the parser, so an escaped
spelling carries no children and both merge paths stop at
`entry.children === null`. The control is `@media` written the same way —
the one at-rule a plain spelling does merge — which stays as written too.

Pinned as unit tests rather than in the config case: an escaped at-rule
anywhere in a sheet stops an earlier distant rule merge, on `main` as
much as here, so putting one in that fixture would quietly retire its own
first assertion rather than add to it.

`mergeDistantRules` landed with two expectations written before #22201 made
an at-keyword read as the name its escapes spell. They assert that
`@\6d edia` is left as it stands and that an escaped `@keyframes` keeps
`from` — both of which were only true while the tables were keyed by the raw
spelling, so they have failed on `main` since #22174 merged.

The behaviour is the one #22201 established and is what the plain spelling
already gets: the two escaped conditions join, the sheet echoes the spelling
back as written, and the keyframe selector shortens. Added alongside is the
refusal under an escaped name, since reading the name is what reaches it.

…t of lib/ root (#22213)

* refactor: group the module graph's nodes and edges into lib/graph/

`lib/graph/` already held the two graphs a compilation carries; it now holds
what they are graphs of. The edges join it — `Dependency`, the
`DependenciesBlock` holding them and the `AsyncDependenciesBlock` for a block
loaded on demand — and so does the chunk side, `Chunk`, `ChunkGroup`,
`Entrypoint` and `HotUpdateChunk`. `lib/dependencies/` keeps the concrete
subclasses and `lib/template/` the `DependencyTemplate` they print through;
these are the base classes every build has. Root goes from 49 files to 45.

Three re-exports are owed and kept. `ignore-webpack-plugin` requires
`lib/Dependency` at the top level and peers webpack >=5;
`external-svg-sprite-loader` and its pixolith fork require `lib/Chunk` the
same way; `@vue/preload-webpack-plugin` requires `lib/Entrypoint` at module
scope and peers webpack 5. `lib/Chunk`'s and `lib/Entrypoint`'s `tsType` in
the schema point at the real path, since a type cannot be read through the
re-export.

`lib/AsyncDependenciesBlock` is recorded as unreachable instead: its only
importer is hard-source-webpack-plugin, which also imports
`lib/DependenciesBlockVariable` and `lib/JavascriptGenerator`, both removed
in webpack 5. The two serializable classes keep their old request restorable
through `registerLegacyRequest`, so a cache pack written before this loads.

`AsyncDependenciesBlock`'s removed-property tombstone gains the JSDoc its
`Dependency` counterpart already carried — without it the getter infers an
implicit `any` that `tsc` reports once the file moves.

* refactor: group the define plugins into lib/define/

`DefinePlugin` replaces a free identifier with a constant at parse time, and
`EnvironmentPlugin` and `DotenvPlugin` are each that same plugin fed from
somewhere else — one from `process.env`, one from a `.env` file, both
constructing a `DefinePlugin` and handing it definitions. The three move
together. `ProvidePlugin` substitutes an import rather than a value, so it
stays out. Root goes from 45 files to 42.

No re-export is owed: the ecosystem scan records no package importing any of
the three paths, and none of them is serializable or named by a `tsType`.
`DotenvPlugin` reads `schemas/WebpackOptions.json` relative to itself to
validate its own options, so that request moves down a level with it.

* refactor: move the last three stragglers into the directories that own them

`ModuleInfoHeaderPlugin` prints the per-module comment header into the
generated bundle through `Template`, so it joins `lib/template/`.
`LazyBarrel` finds the barrel files worth deferring and is required only by
`Compilation`, and `ConcatenationScope` is the protocol scope hoisting runs
on — `lib/optimize/ConcatenatedModule.js` is the only thing that constructs
one, while generators across `lib/` render through it — so both join
`lib/optimize/`. Root goes from 42 files to 39.

`lib/errors/` needed nothing: `WebpackError` and `ModuleNotFoundError` were
moved there already and the root files are re-exports, as are `InitFragment`
and `DependencyTemplate` for `lib/template/`. No re-export is owed for these
three — the ecosystem scan records no package importing them — and none is
serializable or named by a `tsType`.

* refactor: record every path this moves that the scan finds no importer for

The three-way classification the guide describes leaves a path nobody imports
in neither map, which reads as an omission rather than as a decision. Record
those nine under `removed` with the reason, so the record answers for every
path that left `lib/` root in this change rather than only for the ones an
importer was found for.

The consequence is worth stating: `--write` skips a path `removed` names, so
a package that starts importing one of these later is no longer recorded, and
`--check` no longer learns a re-export is owed.

* test(css): expect an escaped at-keyword to join like the name it spells

`mergeDistantRules` landed with two expectations written before #22201 made
an at-keyword read as the name its escapes spell. They assert that
`@\6d edia` is left as it stands and that an escaped `@keyframes` keeps
`from` — both of which were only true while the tables were keyed by the raw
spelling, so they have failed on `main` since #22174 merged.

The behaviour is the one #22201 established and is what the plain spelling
already gets: the two escaped conditions join, the sheet echoes the spelling
back as written, and the keyframe selector shortens. Added alongside is the
refusal under an escaped name, since reading the name is what reaches it.