@variance-authority/core 0.2.0 → 0.3.0

package/CHANGELOG.md

@@ -1,67 +1,110 @@
 # @variance-authority/core
 
+## 0.3.0
+
+### Minor Changes
+
+- 05d6683: Changes before and beyond reach
+
+  A run reads left to right: the harness starts it, the tests enter your code,
+  your code goes out into the install and never comes back. Selection lives in the
+  middle, and both ends were invisible for opposite reasons.
+
+  The far right already arrived — a package is a node, a bump is a seed, the same
+  backwards walk answers it. The far left is this. Nothing imports a
+  `vitest.config.ts`, a setup module, a CI workflow or a `.nvmrc`, so no walk
+  reaches one and the honest structural answer about a change to one is *no
+  component moved*: a skipped suite over the file that decides how every test in
+  it runs. A diff that was *wholly* outside the graph already widened. The hole
+  was a config edited beside an ordinary source file, where the walk had a seed
+  and answered confidently about a change it never looked at.
+
+  `source.before` names those files, repository-root-relative, and a directory
+  claims everything under it. What a declaration buys beyond its own name is
+  everything below it: `beforeReach` in `@variance-authority/core/relate` walks
+  *along* the arrows from each entry — the one question whose subject has no
+  dependents — and collects the setup module, the fixture only that setup
+  imports, and the packages the environment rests on. The descent stops at the
+  first file `source.dirs` already covers, because that file has dependents and a
+  change to it is answered exactly by walking them; everything below it is
+  reached through it and does not arrive either.
+
+  The two ends meet there. A `jsdom` bump is named by the install comparison,
+  reaches `jest-environment-jsdom`, and reaches a config no file in the
+  repository imports — a change beyond reach arriving before it.
+
+  `scanRelations` takes `before` to seed those paths by name, since a harness
+  lives above every directory a component scan is pointed at. A named path that
+  is absent or has no reader is dropped rather than recorded unreadable: an
+  unknown file seeds every walk forever, so a typo would otherwise widen every
+  run in the repository. A declared entry the graph does not hold contributes
+  only its own name, which is the whole answer for a `.nvmrc` and a symptom for a
+  harness config, so the run reports it as a note rather than guessing.
+
+  `source.before` requires `source.relations: true`.
+
 ## 0.2.0
 
 ### Minor Changes
 
 - e546e21: A collector's complaints travel into the observation
-  
+
   A finding produced at collection reached the CLI's record and nobody else. A
   suite driving this from its own runner asks the observation what is wrong with a
   subject, and the one diagnostic that explained the moved pixels — the host, not
   the stylesheet, chose the typeface — was produced, carried, and dropped one call
   short of the person reading the failure.
-  
+
   `mergeDiagnostics` dedupes on every field, so the CLI, which now meets the same
   list twice, says a shared complaint once.
 - f4b328b: Mainline's evaluation can travel, so the next machine does not derive it again
-  
+
   A suite index is bytes addressed by the commit they were written at, which is
   exactly the shape a cache already wants. What was missing was anywhere to put
   them: every checkout that wanted to know what `main` looked like had to run
   `main`, and in CI that is the whole suite, twice, on every pull request.
-  
+
   `@variance-authority/core/share` is the transport, and it has two backends
   because there are two. A **directory** is what `actions/cache`, `aws s3 sync`, a
   network mount and a laptop all are once the bytes are on disk; an **HTTP
   endpoint** is what a bucket, a presigned URL and a tribunal deployment all are.
   `createDirectoryShare` in `@variance-authority/store/share` is the filesystem
   half, kept out of `core` because `core` takes no platform.
-  
+
   A share never fails a run. `get` answers `null` and `put` resolves, whatever the
   network did — a share that is down, misconfigured or empty is indistinguishable
   from a cold one, and the run derives its own index and continues. The cost is
   deliberate: a typo in an endpoint is a suite that quietly got slower rather than
   a build that went red, and the `behind` count `variance share` prints is the
   signal that publishing has stopped.
-  
+
   The CLI gained a `share` section in its config, a `variance share` command that
   says what the share holds for mainline or publishes what this run derived, and a
   line on `variance run` naming where the index went. A lookup walks the lineage —
   `merge-base` with the configured `mainline`, then first-parent — and asks this
   machine's own cache for every commit before it asks the share, because the run
   that just published is usually the one asking.
-  
+
   `docs/sharing.md` is the arrangement end to end, GitHub Actions first.
 - e546e21: A subject is allowed to reference something that is not there
-  
+
   A fixture that points an `<img>` at a path nobody serves is testing the fallback,
   and the broken state is the subject. Capture had no way to say so: the resolver
   returned bytes or the capture was refused, so Material UI's CardMedia,
   ImageListItem and Avatar suites — which all reference a deliberately absent
   `/fake.png` — could not be captured at all.
-  
+
   `resolveResource` may answer `{ absent: true }`. The resource is recorded with
   its status, and the renderer answers that status instead of counting the request
   as one the document failed to carry.
 - e546e21: A subject with no pixels is a baseline without an image, not a subject that got away
-  
+
   A wrapper whose only child went to a portal, or a conformance mount with no
   children, occupies nothing. Refusing to photograph it is right; refusing the
   subject was not. On Material UI's unit tier that was 1109 of 4371 subjects
   reported as unobserved while the capture held their markup, their rules, their
   component hashes and their accessibility tree — none of which was in doubt.
-  
+
   `Raster` makes the image optional: `bytes`, `width` and `height` are absent
   together or present together, and `pictured` is the one place that narrows all
   three. Absent means *this subject has no pixels*, which is a measurement — it
@@ -72,25 +115,25 @@
   verdict. `promotionOf` promotes the sidecar alone when there is no `after`,
   because the subject reached a verdict and the only missing half is the one a
   camera would have produced.
-  
+
   The file-backed store carries the same nullable pair, and with it the split of a
   baseline's two halves into two path prefixes. `identities` scans the record root,
   because a subject with no pixels has no image directory to be found in and the
   sibling scan would otherwise call another machine's baseline new.
-  
+
   Capture stops handing this to Playwright to fail on. Both screenshot paths used
   to refuse a zero-area subject in terms of their own arguments — the clip path
   with `Expected options.clip.height to be greater than 0`, the element path by
   spending the full actionability timeout and then complaining about visibility —
   so a reader had a component that rendered nothing and a sentence about a
   rectangle. `captureSubject` decides it now, and the renderer names the subject.
-  
+
   **Operators:** this is schema **18**. `baselines` and `render_cache` drop
   `NOT NULL` from `width` and `height`, shipped as `0016_pixel-less-baselines.sql`.
   Apply it before pointing a CLI of this version at the deployment; `GET /version`
   reports the schema a build expects.
 - f4b328b: A run leaves behind what its suite looked like, in bytes another checkout can read
-  
+
   A run report answers about *this run*. Nothing answered about the suite from
   somewhere else: which components exist, which subjects hold them, which of them
   has an example of its own, and every name the run indexed. A branch asking any
@@ -98,7 +141,7 @@
   on a machine that has since gone away — which is why *this component is not
   covered anywhere* was a question with no reader, and why a second run could only
   re-derive what the first already knew.
-  
+
   `@variance-authority/report/suite-index` is that artifact. `suiteIndexOf` takes
   the baseline-bearing half of a report — the census, the subject denominator it is
   counted against, and the lexicon — and `encodeSuiteIndex` writes it as one
@@ -106,18 +149,18 @@
   functions. The other half of a composition does not travel: `movements` is what
   moved since a comparison nobody else made, and `divergences` and `echoes` are
   readings of one commit's snapshots. None of the three is a fact about the suite.
-  
+
   It carries the commit it was written at and nothing else about where it is, which
   is the rule the selection index already settled — a commit answers *what changed
   since this was written* exactly, and a timestamp says when a machine was rather
   than where a tree was. An index that cannot name itself is the honest record of a
   run that could not, and a reader holding one has nothing to diff.
-  
+
   It is not JSON. A lexicon is the same few thousand strings written once per
   subject that holds them, and in JSON the repetition *is* the payload. Interned
   once and referenced by number it stops being one, and equal facts encode to equal
   bytes — a sorted dictionary, so a cache that keys on content actually hits.
-  
+
   `@variance-authority/core/segment` is the arithmetic underneath: named columns,
   alignment, interned strings, and the validation a decode performs before it
   believes a file. A column's width comes from the array's own type, so a mismatch
@@ -127,17 +170,17 @@
   now reads and writes through it and keeps its own bytes; what it still owns is the
   part that is about source.
 - e546e21: Two readings of two places are not a flake
-  
+
   `flake` is an accusation — *this reading is not repeatable* — and it was being
   made about a comparison that never claimed repeatability. Lifting two instances
   out of two subjects at one commit and parting them produces a case where every
   input agrees, the tree holds, and the output moved; the old table had exactly one
   row for that.
-  
+
   Where a component sits is decided by the boxes around it, and no component
   receives its own position as a prop. Two instances that agreed on everything they
   were handed and landed at different coordinates have contradicted nothing.
-  
+
   `partingOf` takes a `PartingPlace`: `same` for one subject read twice, across
   revisions or across two moments of one scenario, and `elsewhere` for two
   instances at one commit. It decides only between `flake` and the new `placed`
@@ -149,7 +192,7 @@
 ### Patch Changes
 
 - e546e21: Four ways a large unit tier lost subjects, none of which said so
-  
+
   **The styling was gone before the capture looked.** A capture taken from a setup
   file runs in the outermost `afterEach` a run has; every hook registered inside a
   `describe` has already finished, and `onTestFinished` runs later still. CSS-in-JS
@@ -162,7 +205,7 @@
   `retainStyles` records style elements as they are inserted and puts the removed
   ones back, in insertion order, for the length of one capture. The next test still
   gets a clean page.
-  
+
   **A subject id can be longer than a filename.** A suite that names subjects after
   the test that produced them — a file path and a full test name — passes 255 bytes
   on ordinary tests, and the point of that convention is that the id says where the
@@ -175,13 +218,13 @@
   `@variance-authority/core` directly rather than reaching it through `react` — the
   host stacks it keeps optional are Playwright and Testing Library, and a filename
   rule is not one of them.
-  
+
   **An inline `url()` is spelled in entities.** `style="background-image:url(&quot;/a.png&quot;)"`
   reaches the scan through HTML serialization, and reading it literally asked the
   caller for bytes at a URL that exists nowhere but in the escaping. The HTML half
   is scanned with the attribute's entities undone; the stylesheet half, which was
   never escaped, is scanned as before.
-  
+
   **A stubbed `getComputedStyle` is not a broken asset scan.** Replacing
   `window.getComputedStyle` with a map of the two properties a component reads is
   the only way to drive some layouts in jsdom. The scan called `getPropertyValue`
@@ -192,12 +235,14 @@
 
 ## 0.1.1
 
+Lockstep release — nothing in this package changed. Every `@variance-authority/*` package shares one version.
+
 ## 0.1.0
 
 ### Minor Changes
 
 - 5c34e6d: Say which input moved, not just which tag did
-  
+
   `partingOf(baseline, candidate)` climbs from a set of deltas to the component
   boundary that owns them and names the input that carried the decision:
   `Cart chose differently — useState #0 moved`, with `Summary` reported as having
@@ -206,7 +251,7 @@
   as lines. Two runs of one page that differ in a `<p>` where the other has a
   `<span>` previously produced a structural delta and nothing to say about it;
   the tag is now the symptom and the hook cell is the report.
-  
+
   `holdingOf` from `@variance-authority/react` supplies the evidence — props,
   context values and hook cells as digests, per boundary, including
   `useSyncExternalStore` snapshots so a store that moved outside React is
@@ -218,28 +263,28 @@
   caller-supplied reader, opt in separately from `wiringOf`, and `capture` in
   `@variance-authority/unit-test` now accepts `provenanceOf`, `wiringOf` and
   `holdingOf` so a unit test can read one.
-  
+
   A boundary whose own input could not be read is reported as `unread` and never
   as nondeterminism: `undetermined` is reserved for a component whose every input
   was read and agreed. A wrapper that roots a component boundary no longer
   collapses, because collapsing it discarded the holding — the cost is that a run
   reading holdings keeps wrappers a run without them removes, which is why both
   sides of a comparison must be read the same way.
-  
+
   A component that ran no hooks is read as having run none, rather than as
   unreadable. React writes `_debugHookTypes = null` on every fiber and fills it on
   the first hook call, so the property's absence is the only silence — and
   collapsing the two reported the one shape most worth calling nondeterministic, a
   component with no props and no hooks that renders differently twice, as
   something nothing could be said about.
-  
+
   `PartedBoundary.moved` names the properties the owned deltas named — `color`,
   `padding-top`, `width` — and `explainParting` spends them on the delta line.
   That is the last joint of the chain the rungs climb: a hook cell moved, a prop
   carried it down, and this is what the prop turned into on the page. A count and
   a band stop one link short of what somebody chasing a visual regression is
   trying to name.
-  
+
   `Parting.slice` answers the question asked before which input moved: is this
   worth opening. Three facts are read independently — did the component tree
   move, did any input move, did the output move — and the combinations collapse
@@ -248,24 +293,24 @@
   since a pixel differ can say the screenshots match and nothing about what was
   rewritten underneath. `flake` is `settled`'s opposite number and is refused on
   silence — a run that read no boundary reports `unread`, never `flake`.
-  
+
   `explainParting` leads with that line, and stops enumerating manifestations
   past three: one input at a fork can put a boundary on every component beneath
   it, and nine lines carrying one decision bury the one line worth reading.
-  
+
   The whole layer is documented in `docs/parting.md`: the six slices, the seven
   rungs, what a holding carries and what it deliberately does not.
 
 ### Patch Changes
 
 - e8fee66: Refuse a value whose state is not in its own enumerable keys.
-  
+
   `shapeValue` documented a `Date` as refused and encoded it instead. `typeof`
   answers `object`, `Object.entries` answers empty, and the canonical text came
   out `{}` — so a snapshot holding a timestamp addressed to the digest of an empty
   object, and a later run comparing a different timestamp reported unchanged. The
   same hole swallowed `Map`, `Set`, `URL`, and `RegExp`.
-  
+
   An object with no own enumerable keys and a prototype other than `Object`'s is
   now refused by name at its pointer. A genuinely empty object is still a value,
   and an instance carrying its own fields still serializes them.