Conversation
Object records carry a type name and nothing else, so attribution has always had to guess ownership from the namespace prefix. Namespaces are shared, so that guess cannot be made correct at any precision. The dataset's new `_asm` field closes the gap; this is the runner half that consumes it. Three changes: - `ParseObjectEventAssembly` reads the declaring assembly and type from a new event shape, consulted only when the Method event yielded nothing so the method path is untouched. - `ProbeTypeCandidates` collects every loaded assembly declaring a given type, so an object record whose type is declared more than once is counted by the existing ambiguity metric instead of being normalised away silently. - No change to year handling. `StripConfigSuffix` already collapses the Revit `_20NN` suffix and `IsFromSubjectAssembly` already compares on it, so this adds tests rather than code. Inert until Versioning_Toolkit emits the event: the runner's only view of the dataset is the TestResult tree, so with nothing emitting, every path is unchanged. The existing 135 tests pass unmodified.
This was referenced Sep 14, 2026
The dataset-provenance tests and one comment in DatasetProvenance.cs named the oM assemblies of three private Revit tools. Each of those names is a literal project directory in a private repository, and a path search on it returns that repository alone, so carrying it in a public repository identifies it. Substituted for names that identify nothing, reusing two schemes already in this repository rather than inventing any: Revit_X_oM_20NN, which SerialiserRunner's HostDependencyTests already uses as the placeholder for a year-suffixed Revit oM assembly, and the public Revit_Core_Engine_2022 method-event fixture from ClosureReclassificationTests. Year suffixes, family distinctness and ordinal ordering are all preserved, so every assertion tests what it tested before. DatasetProvenanceTests' header comment claimed the file named only public assemblies while using one of these names five times. Rewritten to a claim the file meets. 193 tests pass, unchanged from before the substitution. Eleven mutation checks ran, five on the map and six on the collector, and each one failed tests and reverted clean.
… was dropped Two reporting defects, both of which let a run say something that is not true. The ambiguity warning describes one population and now covers three. Candidates used to come only from the method signature probe, where more than one match genuinely means enumeration order decided the answer. Reading an object record's declaring assembly from the dataset added two more sources: a closure scan that reports how many assemblies also declare a type the dataset had already settled, and a dataset dispute where nothing was attributed at all. The single message asserted enumeration order for all three, so it could tell someone load order explained a finding it had no part in. The counter now splits by source and each message says what actually happened. The source is recorded on the diagnostic rather than derived. It is inferable today from Path plus a null check on DeclaringAssembly, which is exactly the coupling worth avoiding: that inference breaks silently the first time either field moves for an unrelated reason. Nothing counted findings dropped at attribution. A dropped leaf returns before it can produce a diagnostic row, so a run reporting zero findings because nothing failed and one reporting zero because attribution discarded everything were the same output in every surface. Both counts are now carried on the coverage record, printed in every state including zero, and shown as a summary row. A run whose entire population was dropped now says so, because that result rests on attribution being right rather than on nothing having failed. The counter counts leaves dropped at attribution. It is not a diff against a previous run and the comment says so in both places, because the two coincide on the validation run and would be easy to conflate. Also removed: six citations of an issue number in a repository whose issues are disabled, so the reference resolved to nothing for any reader. One of the sentences carrying it also named a private repository's assembly, and that went with the rewrite rather than being left in a line being edited anyway. 201 tests pass, 193 of them pre-existing and unmodified. Five mutation checks ran against the new behaviour, covering the drop count, its split by basis, and each of the three candidate sources. Each failed tests and reverted clean.
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 free
to 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.
Issues addressed by this PR
The versioning check decides which repository a finding belongs to. For an object record it has only a type name, so it infers ownership from the namespace, and namespaces are shared.
BHoM/BHoMrun35986238826reported 145 findings, every one attributed by guess and none by evidence.An object record's declaring assembly is in the dataset, in the record's own
_asmfield. The runner now reads it from the tree the build stages on the same machine, which an earlier step already validates, and a record without the field keeps today's behaviour.Run
36512880648re-ran that population against a dataset carrying the field. The 145 findings become 47, all attributed by declaring assembly and none by namespace guess, over the same subject set, 30 assemblies and 1,192 types on both. 98 dropped as another repository's, 46 stayed unverified, and one became a real failure:BH.oM.Structure.Loads.LoadCombination, declared byStructure_oM. The run is red for that one finding and is meant to be.Test files
201 tests pass, 135 of them pre-existing and unmodified. Sixteen mutation checks ran, across the map, the collector, the reclassification guard, the drop counter and each candidate source; each failed tests and reverted clean.
Changelog
ci-versioningattributes an object-record finding by the declaring assembly recorded in the dataset, rather than inferring ownership from the namespace.Additional comments
Inert until a dataset carries the field. No version does today, so the map is empty on a current pull request and nothing changes.
A finding is joined to its record on the leaf's description, which for an object record is the
_tvalue. That join is what holds attribution together, and it rests onVersioning_Toolkitcontinuing to describe a failing object record by its type name; if that changes, attribution returns to the guess with no test in either repository going red. Against that: the namespace fallback already reads the same description as a type name, so nothing newly depends on it, and the description is the finding's label in every artefact, so a change shows rather than hides.The map unions every staged version directory rather than the newest, because which version a pull request reads is not visible to the runner. Two versions naming different assembly families for one type leave it unattributed and reported. Only one version carries the field today, so the run above had nothing to disagree with and that path is covered by unit tests alone; comparing the two backfilled versions offline gives 1,711 shared type names and 0 disagreements.
AttributeToSubject,IsFromSubjectAssemblyandStripConfigSuffixare unchanged, as is every dataset. The action's inputs are unchanged: it emits the dataset path its own validation step already resolves, passes that to the runner, and adds two summary rows.The ambiguity metric did change, because this gave it two new sources. Candidates previously came only from the method signature probe, where more than one match means enumeration order decided the answer. An object record adds a closure scan, which reports how many assemblies also declare a type the dataset had already settled, and a dataset dispute, where nothing was attributed at all. Reporting all three as order-dependent would have told a reader load order explained a finding it had no part in, so the counter splits by source and the source is recorded on the diagnostic rather than inferred from two other fields.
Nothing counted findings dropped at attribution, so a run reporting none because nothing failed and one reporting none because attribution discarded everything were the same output. Both counts are now carried, printed in every state, and a run whose whole population was dropped says so.