Release 1.0.3 - #19
Merged
Merged
Conversation
… stubs - Replace BindingFlags.NonPublic reflection with narrow `internal` test seams: EnumSelector<T>.SetSelectionString, SerializableType<T>. FromSerializedNames (+ extracted OnAfterDeserializeImpl to dodge struct-boxing through ISerializationCallbackReceiver), and UnitySerializedDictionary KeyDataInternal/ValueDataInternal/ SetSerializedLists. Runtime/AssemblyInfo.cs grants visibility to GameLovers.GameData.Editor.Tests. - Add 1 High + 12 Medium-confidence audit stubs across MigrationRunner, floatP, MathfloatP plus new SortedListExtensionsTest and ReflectionExtensionsTest fixtures. - Restructure MigrationRunnerTest (mocks-at-top + section dividers, no #regions) and remove the speculative Tests/Editor/Regression/ placeholder per categories.md §2.7. - AGENTS.md: document struct-boxing pattern + internal seams + the System.Collections.IEnumerable qualification gotcha for ConfigsProvider. - Add .audit-history.md persisting the unity-tests-audit run. Made-with: Cursor
- 4 new test files (ConfigTypesBinder, ValueDataConversion, ConfigExportService, ObjectExtensions) - 6 extended fixtures (ObservableField, SerializableType, MathfloatP, floatP, MigrationRunner, ReflectionExtensions) - Pow2 stub corrected (Pow2(f) = f² per Referee correction) - .audit-history.md now gitignored (local skill state) Co-authored-by: Cursor <cursoragent@cursor.com>
Folds into the existing [1.0.3] section -- 1.0.3 is unreleased (max tag 1.0.2). No package.json version bump. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Test-suite audit. Every new or strengthened test carries an ADMIT/RCR header naming the one-line production mutation that reddens it; each mutation below was applied, observed red, and reverted. Also: - Remove IConfigBackendService (public but unimplemented, unused, untested; the package's only UniTask consumer). BREAKING - see CHANGELOG. - Drop the now-orphaned UniTask reference from the Runtime and Migration sample asmdefs. - Declare com.unity.test-framework.performance 3.5.0; both test asmdefs already referenced Unity.PerformanceTesting unconditionally, so consumers without it hit a missing-assembly compile error. - Add Tests/AGENTS.md + Tests/CLAUDE.md. Sections 1-2 (ADMIT/RCR) are shared verbatim across all six GameLovers packages. EditMode 830/830, PlayMode 299/299. RCR: TrustedOnlyMode_BinderBlocksUnregisteredTypes <- ConfigTypesBinder.cs BindToType — reword the rejection message RCR: MaxDepth_PreventsStackOverflow <- ConfigsSerializer.cs ctor — delete `MaxDepth = maxDepth,` RCR: Deserialize_TypesRegisteredFromProvider_AcceptsThemAndTheirIntKeyedDictionary <- ConfigsSerializer.cs RegisterAllowedTypesFromProvider — change the loop source to Enumerable.Empty<Type>() RCR: Deserialize_TypeNotRegisteredByProvider_ThrowsAndDoesNotConstructIt <- ConfigsSerializer.cs RegisterAllowedTypesFromProvider — replace the loop body with an allow-everything shortcut RCR: Observe_WhenFieldIsDirty_RecomputesBeforeAddingObserver <- ComputedField.cs Observe — delete the `Recompute();` inside `if (_isDirty)` RCR: GetSelection_NonContiguousEnum_ReturnsAuthoredValue <- EnumSelector.cs static ctor — change `EnumDictionary[EnumNames[i]] = EnumValues[i];` to `= (T)(object)i;` RCR: GetSelectedIndex_NonContiguousEnumWithNegativeMember_ReturnsCorrectArrayPosition <- EnumSelector.cs static ctor — change `EnumDictionary[EnumNames[i]] = EnumValues[i];` to `= (T)(object)i;` RCR: GetValid_OnDestroyedUnityObject_ReturnsNullSoNullConditionalWorks <- ObjectExtensions.cs GetValid — change `return obj != null ? o : default;` to `return o;` RCR: Dispose_OnDestroyedUnityObject_ReturnsNoneAndDoesNotThrow <- ObjectExtensions.cs Dispose — delete the `if (!gameObject.TryGetValid(out gameObject))` early return RCR: GetAutoName_WhenSourceFileIsUnavailable_FallsBackToKindAndDoesNotThrow <- ObservableDebugRegistry.cs GetAutoName — change `string memberName = kind;` to `= "";` RCR: Remove_WhenKeyDoesNotExist_DoesNotNotifyObservers <- ObservableDictionary.cs Remove — delete the early return in the TryGetValue/Remove guard RCR: StopObserve_WhenCalledOnce_RemovesOnlyOneObserverInstance <- ObservableDictionary.cs StopObserving — delete the `break;` after `_updateActions.RemoveAt(i);` RCR: Observe_WhenObserverUnsubscribesItselfDuringNotification_DoesNotThrowOrSkipOtherObservers <- ObservableDictionary.cs Clear — change `var listCopy = _updateActions.ToList();` to `= _updateActions;` RCR: Observe_WhenObserverAddsAnotherObserverDuringNotification_NewObserverNotInvokedForCurrentUpdate <- ObservableList.cs RemoveAt — change the backward loop to `for (var i = 0; i < _updateActions.Count; i++)` (live bound) RCR: Remove_WhenKeyExists_RemovesFromOriginDictionaryAndNotifies <- ObservableResolverDictionary.cs Remove — delete `if (!Dictionary.TryGetValue(key, out var value)) return false;` RCR: Remove_WhenKeyDoesNotExist_ReturnsFalseAndLeavesOriginDictionaryIntact <- ObservableResolverDictionary.cs Remove — delete the same TryGetValue guard RCR: FindMethodByName_OnDerivedType_WalksBaseTypeChainAndCaches <- ReflectionExtensions.cs FindMethodByName — delete `methods.Add(hash, methodInfo);` RCR: GetWorldCornersArray_OnRotatedRectTransform_ReturnsCornersInWorldSpace <- UnityObjectsExtensions.cs GetWorldCornersArray — change `corners[i] = matrix4x.MultiplyPoint(corners[i]);` to `corners[i] = corners[i];` RCR: ObservableField_HighFrequencyUpdates_FrameTimeImpact <- (benchmark, inverted) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ed §2) §2 declared any test without an RCR line "suspect by default", but some correct tests provably have no one-line mutation - double-guarded validation, where an unconfigured object trips two independent guards so disabling either leaves the other throwing. The rule was mislabelling tests that are right and unbreakable. Adds an UNFALSIFIABLE exemption on §13's terms: the reason must be falsifiable, must name both guards, and must record that a mutation was tried and observed green. "Couldn't find one" is explicitly not a reason - that is an unfinished RCR, not an exemption. Also adds a verdict table for tests that resist mutation, because they are not one problem: A5 duplicates get deleted (naming the surviving sibling), D2 overclaims get a strengthened assertion or an honest rename, and UNFALSIFIABLE tests are kept with the exemption comment. The class must be proven before acting - an A5 duplicate by observing the sibling's mutation redden both, a D2 overclaim by observing the implied mutation leave the test green. §1 and §2 remain byte-identical across all six packages. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Section 2's verdicts table gains a fourth class. UNFALSIFIABLE was absorbing tests that section 1's A3 rule would never have admitted - the tell being reasons like "no line in Runtime/ participates" or "these are C#'s zero-init values", which describe a test that pins nothing rather than one that is hard to break. New rule: if no line in Runtime/ or Editor/ participates in the assertion, it is an A3 reject and the verdict is delete. UNFALSIFIABLE stays reserved for behaviour this package genuinely owns but cannot be broken one line at a time. Sections 1-2 are shared verbatim across all six Tests/AGENTS.md; no test or production changes in this package. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Section 13 now carries a dated baseline for this package's runtime assembly, plus the reason to steer by that number rather than the combined one. Every earlier coverage figure in this repo was an artifact and must not be compared against: - reports before today ran without -debugCodeOptimization, so Unity compiled Release and emitted ~40% fewer sequence points (MathfloatP showed 637 coverable lines instead of 1002) - a silently shrunken denominator - some runs leaked test and sample assemblies into scope, and some covered only 3 of the 6 packages The current run covers all 11 production assemblies with none leaking, verified via the MathfloatP denominator check now documented in Tools/coverage.sh. Repo-wide: runtime 73.9%, Editor 5.5%, combined 41.0%. Editor is 48.1% of all coverable lines and is accepted-untestable per the ACCEPTED (iii) rows in section 13, which is the whole reason the combined figure is not the one to track. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sections 1-2 are shared verbatim across all six packages; this adds a sixth admission question and the worked instance behind it. A6 asks whether an assertion's outcome would change if project configuration changed - a renderer feature installed or removed, an Addressables catalog built, a sample imported. If so the test must READ that state rather than assume one value of it. A6 is not A3. A3 asks whether the package computed the value; A6 asks whether the test assumed which value it would be. A test can satisfy A3 and still fail A6, which is exactly how the gap went unnoticed: UiBackdropBlurPresenterFeatureTests read a package-computed flag (UiBackdropBlurRendererFeature.IsInstalled) but hard-coded the expectation that it was false. Batchmode never instantiates the URP renderer, so the flag was false there and all five tests passed; in the Editor the feature registers from the project's renderer asset and all five failed. The fixture was asserting a fact about the repo, not about the code under test. Validated against the existing corpus before being written, per root AGENTS.md 2.2: the blur fixture was the only violation and is already fixed. AddressablesUiAssetLoaderTests asserts on a key that is unresolvable either way, and UiCameraStackFeatureTests builds its own cameras rather than reading project renderer state. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
NUnit reuses a single fixture instance for a whole class, and this [SetUp] built the observables from _mockInt without resetting it. ValueSetCheck writes _mockInt = 5, so that value leaked into whatever ran next. Nothing was failing, but only by alphabetical accident: ValueSetCheck sorts last, so no test currently follows it. Any test whose name sorts after "ValueSetCheck" would have silently inherited 5 instead of 0 - and so would every test in the fixture if these methods are ever renamed to the Method_Condition_ExpectedResult convention required by Tests/AGENTS.md §5. That rename was blocked on this. Verified in both directions with a temporary probe test named to sort last and assert _mockInt == 0: RED before the fix (the leak is real, not theoretical), GREEN after. Probe removed; fixture 18/18. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…th and misc Four slices, 287 mutations, every one observed RED and reverted; all production files byte-identical afterwards. EditMode 826/826, PlayMode 299/299. - observables 107/107, configs 61/61, math 82/82, misc 37/37, PlayMode 2/2. - No holdbacks: every mutation behaved as predicted on the first pass. Two things worth recording for whoever reads these next. The observable classes are near-identical across List/Dictionary/HashSet, so the same mutation text exists in three files. Every anchor is file-scoped; a bare-name or bare-text anchor here silently mutates the wrong class and the resulting red attributes to the wrong test. MathfloatP mutations were sized against each fixture's tolerance rather than chosen for elegance: _epsilon is 0.001f, so edits had to move the result further than that (Sin numerator halved, Exp scale negated, Log2 base changed). Where a shared helper would have reddened many siblings, the call site was mutated instead - Floor's negative branch rather than Floor, Clamp01's call site rather than Clamp01. Smoke fixtures carry the section 1 directory exemption instead of a mutation, naming which fixture pins the same symbols properly. 45 tests were left unannotated on purpose. They are the A3 rejects, A5 duplicates and D2 overclaims the drafting pass identified, and they need a verdict (delete / strengthen / keep-with-exemption) rather than a comment. The largest clusters: 5 of 8 UnitySerializedDictionaryTest tests exercise only inherited Dictionary members the class never overrides, and the four ObservableDictionaryTest StopObserv* tests register global observers while the fixture leaves ObservableUpdateFlag at KeyUpdateOnly, so the fan-out they check never runs and DidNotReceive passes either way. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Concludes the RCR backfill for this package. Deletes only what the evidence supports and keeps everything it does not. Deleted (A3 - no line in Runtime/ or Editor/ participates, so no mutation can ever redden them; section 2's proof standard for A3 is inspection): - Five UnitySerializedDictionaryTest cases (Add_StoresKeyValue, Add_DuplicateKey_ThrowsArgumentException, Remove_ExistingKey_ReturnsTrue, TryGetValue_Exists/NotExists). The whole 56-line UnitySerializedDictionary declares two [SerializeField] lists, two ISerializationCallbackReceiver methods and three test seams - Add/Remove/TryGetValue/indexer are inherited BCL Dictionary members it never overrides. The two serialization-callback tests remain and are the real coverage. - Five *_NotObserving_DoesNothing cases across ObservableField/List/Dictionary. No observer is ever registered, so the delegate list is empty at construction and DidNotReceive() is fixed regardless of any production edit (section 1, D1). Deleted (A5, probe-confirmed): SerializableTypeTest.Equals_DifferentType_ReturnsFalse - survivor named in the manifest below. KEPT DELIBERATELY - five A5 candidates whose duplicate claim was never probed: Validate_UsingBuilder_DefaultValues_PassValidation, GetConfig_Singleton_ReturnsCorrect, both copies of SecureMode_TypeNameHandlingNone_Verified, ChainedComputed_DeepHierarchy. The reason is measured, not cautious by temperament. Of the ten A5 candidates that WERE probed across this repo, seven turned out to have a unique pin - a 70% error rate on inspection-based A5 classification. Every wrong classification in the pass was an A5; none was an A3. Applying that rate to unprobed candidates predicts ~5 wrong deletions out of 7, so they stay until a probe says otherwise. EditMode 806/806, PlayMode 295/295. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…it proved First trustworthy coverage figure for this repo. Regenerated with -debugCodeOptimization, all 11 GameLovers assemblies in scope, test and sample assemblies excluded. Repo-wide runtime coverage is 74.1% (6609/8922). Do not compare against any earlier number. 41.8% was stale, wrongly scoped to 6 assemblies, and diluted by Editor code; 38.3% was compiled in Release, which silently shrank the denominator ~40%. The register now names the sanity check that catches a repeat: MathfloatP must report ~1002 coverable lines, not 637. The OPEN rows added here are findings the mutation pass PROVED rather than suspected - each one is a mutation that was applied and observed leaving its test green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
All four registered GLOBAL observers and then asserted DidNotReceive(), but the fixture left ObservableUpdateFlag at the constructor default KeyUpdateOnly - under which Add, indexer-set and Remove skip the _updateActions fan-out entirely. The assertion therefore passed whether StopObserving worked or not. Each was observed staying GREEN against the removal mutation that should have reddened it. Fix is one line per test: set ObservableUpdateFlag.Both, matching the already-annotated sibling StopObserve_WhenCalledOnce_RemovesOnlyOneObserverInstance which had it all along. Re-verified after the change - the mutations that previously stayed green now redden: RCR: StopObserveCheck <- ObservableDictionary.cs StopObserving disable the `_updateActions[i] == onUpdate` removal (RED, also reddens StopObserve_WhenCalledOnce) RCR: StopObservingAllCheck <- ObservableDictionary.cs StopObservingAll disable the `.Target == subscriber` global removal (RED, radius 2) RCR: StopObservingAll_MultipleCalls_Check <- covered by the same `.Target == subscriber` mutation; shared-path, the multiple-registration claim is its increment RCR: StopObservingAll_Everything_Check <- ObservableDictionary.cs StopObservingAll drop `_updateActions.Clear();` from the subscriber == null branch (RED, isolated) EditMode 806/806. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both assertions were unfalsifiable. `Contains("\"Color\":")` matched the FIELD NAME and sat
behind an `||`; `Contains("\"x\":")` holds under default Newtonsoft serialization of a Vector
even with every converter removed. Now asserts the `#FF0000FF` hex form ColorJsonConverter
actually emits, plus IsFalse on `"normalized"` as a guard that the vector converters are
engaged.
RCR: Serialize_UnityTypes_SerializesCorrectly <- ConfigsSerializer.cs ctor remove `new ColorJsonConverter(),` (RED, verified)
A second candidate mutation - removing Vector3JsonConverter - was deliberately NOT recorded.
It reddens by JsonSerializationException (self-referencing loop on Vector3.normalized, whose
value equals its owner so Newtonsoft's cycle check trips) rather than by assertion, and
section 2 excludes red-by-exception. The guard assertion stays; the claim does not.
EditMode 806/806, PlayMode 295/295.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>The four StopObserv* fixtures now set ObservableUpdateFlag.Both before registering their global observers, so Add / indexer-set / Remove reach _updateActions. All four were observed RED against removal of the fan-out; before c111ba2 all four were GREEN under the same mutation, which is what made the row OPEN. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
§13 asserted every named symbol was 'either ACCEPTED or OPEN' while four packages had already grown CLOSED rows — the spec forbade rows it contained. CLOSED is now first-class, and it carries a contract: name the commit AND the observation, including the environment the observation came from. A row closed on 'the fix landed' is still OPEN, because the fix is the edit and the closure is the evidence. Second rule: closing a row means re-deriving its claim against current source, never reading the commit that claimed to fix it. A partial fix and a complete one produce the same green suite and the same confident commit message, so the commit cannot be evidence for its own completeness. §1 and §2 are shared verbatim across all six packages; §13's preamble is too. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… never applied Six gamedata tests were probed and observed RED during the backfill, but the annotation never reached the source: annotate.py's dedupe scanned a fixed character lookback that reached into the PREVIOUS test's comment block and reported 'already annotated', silently skipping tests that had none of their own. Recovered from the harness records rather than re-derived, and gated strictly on a recorded RED-OK verdict for that test - prepared text also existed for 28 tests with no recorded observation, and writing those would have fabricated a verified claim. Comment-only diff; no test body or production line changed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… unowned-edit measurement Tests/AGENTS.md section 2 said a test with no // RCR: line is a suspect by default, without carving out Smoke/. Section 1 already exempts that directory (its defect class is "the assembly no longer loads", which has no one-line mutation), so the omission flagged those fixtures forever. Exemption is now explicit, on the same directory basis. Also records that "unannotated" is three states, not one: observed RED with the write-back lost, seen reddening only as collateral, or never probed. Only the last needs a probe, and prepared annotation text must never be written without a matching RED-OK - it exists for tests that were never probed, and writing it fabricates a verified claim. Adds a section 13 row for the measured count of production edits that redden only collaterally (223 repo-wide, from .test-all/rcr/unowned-edits.json). Recorded with the caveat that it is NOT that many missing tests: for foundational primitives and the UiService integration hub, having no isolated owner follows from centrality, not neglect.
Mechanical pass, no behaviour change. Removed, per §6.6's rule that private members and all constructors are never XML-documented: - the ConfigMigrationAttribute, ConfigTypesBinder and ConfigsSerializer ctors - the private floatP._raw field - the private floatP.clz and ObservableDebugRegistry.TryExtractMemberName helpers, both keeping their rationale as a `//` comment Folded <param> tags into their summaries where the declaration is not public consumer-facing API: DependencyGraphElement.SetTarget (Editor assembly) and the internal ObservableDebugRegistry.Register. Converted the SerializationSecurityMode members to inline `//` comments. Their combined 8 lines of <summary> carried one caveat worth keeping — Secure is serialize-only and cannot round-trip — so that is now a <remarks> on the enum itself rather than lost with the member docs. Added `/// <inheritdoc />` to overrides whose base documentation already exists: ToString/Equals/GetHashCode on floatP, ToString on Pair / StructPair / MigrationInfo, and all eight WriteJson/ReadJson in VectorJsonConverters. Moved the internal SerializableType<T>.FromSerializedNames test seam above the private block (§6.6: `internal` is never interleaved with `private`). Pure reorder — the OnAfterDeserializeImpl struct-boxing pattern this package's AGENTS.md §4 warns about is untouched, and the method still appears exactly once. Tools/style-audit.py reports 0 remaining mechanical-class violations for this package. The 3 findings left are overrides whose own base declarations are undocumented; adding <inheritdoc /> there would yield empty IntelliSense, so they are handled with their bases in the prose pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…lue pins The four MathfloatP determinism tests each evaluated one expression twice and compared the results, so no production edit could redden them - their own comments said "A3 reject" and "D2 overclaim" under an "// RCR: none exists" heading, which is the laundering section 2 warns about. They now assert golden RawValue literals harvested from the reference implementation, which is what makes the determinism claim testable: a platform or an approximation that computes differently fails here. Values sanity-checked against real math before pinning - Sin(0.5)=0.4794, Cos(0.5)=0.8776, Tan(0.5)=0.5463, Sqrt(2.5)=1.5811, Log(2.5)=0.9163, Exp(2.5)=12.18 - so they are computed output, not captured noise. RCR: Determinism_VerifyRawValues <- MathfloatP.cs Sin Bhaskara constant 0x42456460 -> 0x42456461 (RED, 1090413845 -> 1090413844) RCR: AllTrigFunctions_RawValueConsistent <- same Sin constant (RED on the Sin row); Cos -> Sin(x) (RED on the Cos row) RCR: AllPowerFunctions_RawValueConsistent <- MathfloatP.cs Sqrt exponent unbias m -= 127 -> m -= 126 (RED, 1070228162 -> 1074731965) RCR: CrossPlatform_Determinism_ComplexExpression <- MathfloatP.cs Cos Sin(x + RawPiOver2) -> Sin(x) (RED, 1065965998 -> 1075700924). Negative result recorded on the test: the one-bit Sin constant change that reddens the three siblings leaves this one GREEN, because rounding in the wider expression absorbs it. Also deletes floatPTests.ExplicitConversion_ToFloat as a proven A5 duplicate. Mutating operator floatP(float) to f * 2f reddened it and ImplicitConversion_FromFloat with identical expected and actual values (2.46 for 1.23); ImplicitConversion_FromFloat is the survivor. MathfloatP.cs and floatP.cs restored byte-identical (md5 verified). Editor EditMode 805/805. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ObservableDebugRegistry.Entry.ToSnapshot is `public`, but `Entry` is a `private` nested class, so the method is unreachable from outside the registry. §6.6 keys its never-document rule on effective accessibility rather than the declared modifier, which makes this private — and the summary only restated the method's own name in any case. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Completes this package — Tools/style-audit.py reports 0 items, and 0 repo-wide. floatP was the bulk of it, and the summaries carry what the signatures could not: - Equals treats NaN as equal to NaN, and positive zero as equal to negative zero — deliberately unlike the == operator. That is what makes the type usable as a dictionary key, and it is invisible from the signature. - < and > return false whenever either operand is NaN, per IEEE, which is the opposite convention to Equals in the same type. - ToString formats via float, so precision beyond binary32 is not preserved. - The internal RawMantissa / Mantissa pair differ by whether the implicit leading one is restored. Also documented: the validation attributes, recording that RangeAttribute passes a null value and so needs pairing with RequiredAttribute; ObservableUpdateType and ObservableUpdateFlag; ComputedTracker's thread-static stack, which is how a dependency read attaches itself to the computed field currently evaluating; the observable and configs debug-registry snapshot surfaces (cross-referencing AGENTS.md §4 rather than restating it); the Vector*Serializable conversions; the internal test seams FromSerializedNames / SetSelectionString / SetSerializedLists, each recording why the public API cannot reach that state; and the Config Browser's editor events. Added the missing <summary> to the three interface InvokeUpdate members that carried only <remarks> — a doc block with no summary shows nothing in IntelliSense. Verified: batchmode green — EditMode 805/805, PlayMode 295/295, 0 CS warnings in the -runTests log. Structural scans clean: 0 doc comments misplaced after an attribute, 0 malformed summary blocks, 0 BOM changes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Unity's packer uses .gitignore as its pack-ignore list, so listing .github/ drops the CI workflow from the tarball while git keeps tracking it (gitignore does not untrack existing files). Verified on a real clone: 434 -> 433 entries, .github 1 -> 0, Runtime unchanged.
Uh oh!
There was an error while loading. Please reload this page.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Changed:
com.unity.test-framework.performance(3.5.0) dependency so the package's test assemblies compile when tests are enabled.Removed:
IConfigBackendServiceinterface. Consumers that implemented it must remove the implementation or vendor the two-method interface locally.com.cysharp.unitaskdependency from the runtime assembly and Migration sample.