v0.7.0

TypeScript (npm)

• ⚠️ Breaking: Stimulus handlers are typed to return the complete next state, not a partial
  A stimulus handler's return type is now C | void instead of
  Partial<C> | void. The runtime always REPLACED state with what the handler
  returned — there was never a merge — so Partial<C> type-checked the one thing
  the semantics forbid: a return that omits a field silently dropped it while
  every later step kept the full C type. Returns that were partial now fail to
  type-check; spread the current state to keep the rest, e.g.
  (state) => ({ ...state, count: 1 }).
• ⚠️ Breaking: var-vitest: Exported names are Varar-prefixed
  @varar/vitest's exports are renamed to match the product name:
  VarResultsReporter → VararResultsReporter, VarResultsReporterOptions →
  VararResultsReporterOptions, VarVitestPluginOptions → VararVitestPluginOptions,
  varVitestPlugin → vararVitestPlugin (the default export, so import vararPlugin from '@varar/vitest' needs no change), collectVarExamples →
  collectVararExamples, varTestBody → vararTestBody, isVarSpecId → isVararSpecId.
  The old names are gone. CollectPorts keeps its name — it carries no product
  prefix and is parallel to the same type in the other ports.
• Added: var-cli: Varar init makes the project ESM so the scaffolded steps file loads
• Added: var-cli: Varar run explains the two reasons Node refuses a step file
• Fixed: Commit the generated pnpm-workspace.yaml in the synced sample
• Fixed: Published packages report their real version in VERSION, not 0.0.0

Ruby (RubyGems)

• Fixed: Describe the gems as Varar, not Vár

C# / .NET (NuGet)

• Added: Ship the C# sample in varar-examples

VS Code extension (Marketplace & Open VSX)

• Fixed: Highlight only the captured value of a step parameter, not its surrounding notation

Specification (all ports)

• ⚠️ Breaking: Steps receive the state your factory made, never frozen or retyped
  State is no longer deep-frozen at runtime, and TypeScript
  handlers receive C instead of DeepReadonly<C>. Mutating state no longer
  throws at runtime nor fails to type-check — declare your own state type
  readonly if you want the compiler to stop you (in Java and Kotlin the
  record/data class already does). Evolution is still meant to happen by
  returning the next state; the relaxation is what lets state hold live objects
  such as a DB client or a page object, which freezing broke and a mapped
  readonly type made unassignable to its own declared type.
• ⚠️ Breaking: Drift is accepted with VARAR_UPDATE, not VAR_UPDATE
  The environment variable that accepts drift is now
  VARAR_UPDATE in every adapter (vitest, the varar CLI, pytest, unittest,
  minitest, rspec, junit, kotest, gotest and cargotest); VAR_UPDATE is no longer
  recognized. pytest's flag is now --varar-update and the JUnit/Kotest system
  property is now varar.update. Set VARAR_UPDATE=1 (or true) where you set
  VAR_UPDATE=1 before.
• ⚠️ Breaking: A sensor with slots that returns nothing now fails instead of silently passing
  A sensor whose expression has parameters — or a trailing data
  table or doc string — must return one value per slot. Returning
  undefined/None/null/nil previously skipped the comparison entirely and the
  example went green, so a typo in a property access turned an assertion into a
  no-op while the document went on claiming something nobody checked. It is now a
  ReturnShapeError ("a sensor with N slot(s) must return one value per slot, got
  nothing"), in every port. The same applies to a header-bound row step, which
  must return its row object.

Zero-slot sensors are unchanged: they must return nothing, and throwing is still
how they fail. Throwing remains a valid way to fail at any slot count — what is
no longer allowed is answering nothing at all when the document asked a
question. Where a greedy {word} capture made the slot awkward to echo back, the
fix is to match the punctuation in the expression ("The result is {word}.") so
the slot holds just the value being compared.

Rust drops the Asserted IntoSensor impls, which existed only to opt a slotted
sensor out of the comparison: the type system now rejects at compile time what
the other ports reject at run time.

• ⚠️ Breaking: A doc-string mismatch is a cell mismatch
  A doc string is one cell, compared whole, so it no longer has
  an error of its own. DocStringMismatchError / DocStringDiff (and the
  isDocStringMismatchError guard) are gone from every port; a differing doc string
  now throws CellMismatchError carrying a single CellDiff whose column is
  "doc string". The run-result payload drops its doc field — a doc-string
  failure appears in cells like any other. Conformance traces lose the
  doc-string-mismatch failure kind; it is cell-mismatch now.

The message is unchanged: expected/actual stay quoted, so a doc string that
differs only by a trailing newline still shows a visible difference where a bare
cell would render none.

• ⚠️ Breaking: An inline cell is labelled "cell N", not "arg N"
  A mismatched value captured from a paragraph is now labelled
  cell 1, cell 2, … instead of arg 1, arg 2. It shows up in the message
  (CellMismatchError: cell 1: expected £2.55 but was £2.60), in CellDiff.column,
  and in the conformance trace. "arg" named the handler's parameter; what differs
  is a cell in the document.
• ⚠️ Breaking: An example runs until a heading, ---, or a non-matching paragraph (#65)
  Two step-only examples separated by only a blank line now run
  as a single shared-state example — separate them with a --- or a heading, and
  keep prose out of the middle of an example (a non-matching paragraph between
  steps ends the example). Additionally, Varar no longer runs Gherkin .feature
  files and the scannerPlugins config key has been removed (it is now an
  unknown-key error). Migrate .feature suites by translating them to Markdown
  specs and porting the step definitions.
• Added: Add built-in {emph} parameter type for Markdown emphasis
• Fixed: Sensor return errors name slots and cells consistently