fix(samples): repair samples/minimal:nativeTest and guard it in CI (#195)

[monkopedia-coder]

The bug (reproduced)

samples/minimal:nativeTest fails to compile:

e: .../samples/minimal/src/nativeMain/kotlin/Main.kt:33:14 Function invocation 'origin()' expected.
e: .../samples/minimal/src/nativeMain/kotlin/Main.kt:33:14 Variable expected.

Main.kt did rect.origin = Point().apply { ... } — property-assignment syntax. But Rect.origin
is a Point-valued field, and a by-value object getter needs a MemScope to allocate the
returned Point into, so the generator can't expose it as a Kotlin property (there's no way to
thread a MemScope through a property getter). It emits origin() (a
context(scope: MemScope) fun) / _origin(value) (a plain setter fun) instead — confirmed by
reading the actual generated build/krapped-cpp/src/geo_Rect.kt. This is correct generator
behavior; the sample was stale. Pre-existing, caused by the context-parameter migration
(#145/#146), and invisible until now because samples/minimal was only ever exercised as far as
kplusplusSync, never nativeTest.

The fix

• samples/minimal/src/nativeMain/kotlin/Main.kt: rect._origin(Point().apply { x = 0.0; y = 0.0 })
  instead of property assignment, plus a rect.origin() read to demonstrate the getter shape too.
  Scalar fields (width/height/x/y) are untouched — those stay ordinary var properties,
  confirmed against the generated bindings.
• Verified end-to-end: ./gradlew runReleaseExecutableKlinker prints the expected
  Rect area: 6.0 / Rect origin: (0.0, 0.0) / distance ... 5.0 / center: (1.0, 1.5).

The guard

samples/minimal is a from-published consumer (resolves the released plugin from
mavenCentral/Plugin Portal, not the in-tree generator), so it can never catch a generated-shape
change at the moment the generator changes — only the next time this sample itself builds against
a new plugin version. So the guard is: make that next build fail loudly instead of silently.

1. Added src/nativeTest/kotlin/GeometryTest.kt — three kotlin.test cases exercising every
   accessor kind the header produces: scalar var properties, the context(scope: MemScope)
   get/_set pair for the object-valued origin field, the by-value-returning center()
   (also context'd), and the plain const-ref distanceTo. Wired kotlin("test") into the
   nativeTest source set in build.gradle.kts.
2. Added .github/workflows/samples-minimal.yml: runs ./gradlew nativeTest in
   samples/minimal on any push/PR touching samples/minimal/** (there was previously no
   PR-triggered CI in this repo at all — only release.yml on release/workflow_dispatch).
   It's cheap: no in-tree LLVM build, just the published plugin + system clang++/LLVM 22 for
   the runtime front-end, so it's independent of (and much faster than) the main self-hosting
   build.
3. Added a nativeTest step to the release "local dry run" in docs/releasing.md, since that
   checklist already runs samples/minimal by hand before every cut.

Proof the guard actually catches drift: reintroduced the exact rect.origin = … bug
temporarily and re-ran ./gradlew nativeTest:

> Task :compileKotlinNative FAILED
e: .../Main.kt:40:14 Function invocation 'origin()' expected.
e: .../Main.kt:40:14 Variable expected.
BUILD FAILED in 11s

then reverted (diff against the pre-break copy showed byte-identical) and re-ran clean —
green, 3/3 tests pass:

<testsuite name="nativeTest.GeometryTest" tests="3" skipped="0" failures="0" errors="0" ...>
  <testcase name="rectExposesOriginAsAccessorsNotAProperty[native]" .../>
  <testcase name="rectAreaAndCenterMatchTheGeometry[native]" .../>
  <testcase name="pointDistanceToIsAPlainConstRefMethod[native]" .../>
</testsuite>

Also ran a full ./gradlew clean nativeTest runReleaseExecutableKlinker --rerun-tasks to confirm
nothing regressed end to end.

samples/v8

Checked whether samples/v8 has the same rot (source-level audit only — did not run a fresh
kplusplusSync/build; no generated-bindings output existed to diff against, and v8 is much
larger, so this was read-only per the task instructions):

• samples/v8/src/nativeMain/kotlin/Main.kt has exactly one struct-field access
  (createParams.array_buffer_allocator = NewDefaultAllocator(), line 87). The underlying C++
  field (v8-isolate.h:255) is a raw pointer (ArrayBuffer::Allocator*), which — per the
  same migration rule that broke origin here — correctly stays an ordinary var property
  (only by-value object getters need the context(MemScope) treatment). Everything else in the
  file is already call-style (.reference(), .Int32Value(context), .FromJust(), etc.), and
  a couple of in-file comments/issue references ([krapper_gen] Local<PolymorphicBase>.reference() returns an empty marker Api interface — base instance methods unreachable #107, and an older _reference() → reference()
  rename) show it's been kept in sync with codegen shape changes before.
• So samples/v8 does not appear to have this specific bug. It's not build-verified (no
  nativeTest, no fresh generated output to compare against — same "never exercised beyond a
  manual run" rot pattern samples/minimal had), but there's no property-vs-accessor mismatch
  visible at the source level. Not fixing anything there in this PR; flagging in case a
  samples/v8:nativeTest guard is wanted as a follow-up (would be a larger, separate change
  given v8's size and fixup{} complexity).

Scope

Touched only samples/minimal/**, a new .github/workflows/samples-minimal.yml, and
docs/releasing.md (the release checklist that already references samples/minimal) — nothing
under compiler/**, the root build.gradle.kts, or krapper's drop-ledger source, per the
concurrent-work lanes for this cycle.

[monkopedia-reviewer]

Reviewing PR #197 (Reviewed SHA so far: 64be5d5).

Measured test count (independent — the CI console log itself shows no per-test names/count, and no test-report artifact was uploaded, so a green checkmark alone would NOT have been trusted here):

build/test-results/nativeTest/TEST-nativeTest.GeometryTest.xml
<testsuite name="nativeTest.GeometryTest" tests="3" skipped="0" failures="0" errors="0" ...>
  <testcase name="rectExposesOriginAsAccessorsNotAProperty[native]" .../>
  <testcase name="rectAreaAndCenterMatchTheGeometry[native]" .../>
  <testcase name="pointDistanceToIsAPlainConstRefMethod[native]" .../>

3/3 executed, 0 skipped, 0 failures — matches the PR's claim exactly, from my own local build (LLVM 22 / clang++ 22.1.6) run against the actual :nativeTest task, not just BUILD SUCCESSFUL.

Also confirmed CI run 30477106930 on this exact head SHA (branch fix/195-minimal-sample) triggered on pull_request and genuinely executed the pipeline: kplusplusSync ran, compileKotlinNative/compileTestKotlinNative/linkDebugTestNative/nativeTest all executed (none NO-SOURCE/UP-TO-DATE), 10 actionable tasks: 10 executed.

Proceeding to independently reproduce the guard (reintroduce the rect.origin=... property-assignment bug, confirm red, revert, confirm green) and the rest of the checklist.

[monkopedia-reviewer]

Reviewed: 64be5d5

Verdict: APPROVE + MERGE (tier-1, green, independently verified)

Measured test count (not trusting green alone)

CI's console log shows no per-test names/count (default K/N test-task output is terse, no
--info, no report artifact uploaded), so I built locally in my own worktree off this exact
SHA (LLVM 22 / clang++ 22.1.6 on PATH) and read the actual XML report:

build/test-results/nativeTest/TEST-nativeTest.GeometryTest.xml
tests="3" skipped="0" failures="0" errors="0"
  rectExposesOriginAsAccessorsNotAProperty[native]
  rectAreaAndCenterMatchTheGeometry[native]
  pointDistanceToIsAPlainConstRefMethod[native]

3/3 executed, matches the PR's claim exactly. Also confirmed via --rerun-tasks (no
UP-TO-DATE/NO-SOURCE masking anything): 10 actionable tasks: 10 executed, kplusplusSync ran,
compileKotlinNative/compileTestKotlinNative/linkDebugTestNative/nativeTest all genuinely executed.

CI run 30477106930 (branch fix/195-minimal-sample, this head SHA) fired on pull_request as
expected and matches this same shape.

Guard reproduction (independent)

Reintroduced the exact original bug (rect.origin = Point().apply { x = 0.0; y = 0.0 } in place
of rect._origin(...)) and ran ./gradlew nativeTest --no-daemon:

> Task :compileKotlinNative FAILED
e: .../Main.kt:40:14 Function invocation 'origin()' expected.
e: .../Main.kt:40:14 Variable expected.
BUILD FAILED in 17s

Reverted (diff against pre-break content was a single line, exact inverse), re-ran fresh with
--rerun-tasks: BUILD SUCCESSFUL, 10/10 executed, 3/3 tests green. The guard has teeth.

Generated-shape claim verified against the real artifact

Read build/krapped-cpp/src/geo_Rect.kt and geo_Point.kt from my own local kplusplusSync
run: origin() is inline context(scope: MemScope) fun origin(): Point?, _origin(value) is a
plain setter fun, width/height are ordinary var properties, center() is also
context'd, and distanceTo on Point is a plain non-context'd method taking Point by value
(const-ref in C++). Exactly matches Main.kt/GeometryTest.kt's usage and the PR's description.

Test quality

GeometryTest.kt's three cases assert actual values (origin x/y round-trip, area, center
midpoint, distanceTo=5.0 on a 3-4-5 triangle) — not tautological, and would fail to compile
(not just assert wrong) against the original bug, since rect._origin(...) doesn't exist if the
generator ever reverts to a property. Good bidirectional coverage.

v8 audit spot-check

Confirmed Isolate::CreateParams::array_buffer_allocator (v8-isolate.h:255) is
ArrayBuffer::Allocator* — a raw pointer field, correctly still exposed as an ordinary var
property per the same migration rule. samples/v8/src/nativeMain/kotlin/Main.kt:87 uses
property syntax accordingly. Claim checks out; v8 itself remains unguarded by CI (correctly
flagged by the author as a separate, larger follow-up, not this PR's scope).

CI workflow

• Path-scoped to samples/minimal/** + itself, doesn't touch the expensive in-tree LLVM build —
  appropriately cheap and independent of release.yml.
• No -PenableClang anywhere (correctly dead/removed per the require-LLVM contract).
• Installs LLVM 22 via apt.llvm.org and symlinks clang++/clang/llvm-config — matches the hard
  LLVM-22 requirement.
• Minor non-blocking nit: no timeout-minutes on the job (defaults to GitHub's 360-minute cap).
  Given the run takes ~2 minutes and the task is a plain nativeTest invocation with no obvious
  wedge vector, not worth blocking on, but worth adding opportunistically next touch.
• Confirmed genuinely from-published: build.gradle.kts pins plugin 0.3.4, my local mavenLocal
  cache only has up to 0.3.2, and the build still succeeded — so resolution came from
  mavenCentral/gradlePluginPortal for real, both locally and (necessarily, no local cache exists
  on a fresh runner) in CI.

Scope / overlap

Diff touches only .github/workflows/samples-minimal.yml (new), docs/releasing.md,
samples/minimal/**. No overlap with compiler/**, root build.gradle.kts, or krapper/src/**
(the concurrent #194/#196 lanes).

No defects found. Merging.