Repository navigation
Conversation
Typed field members across the bridge boundaryPaired source checkpoints: ObjectState Owner and consumers
Source evidenceNRA Exact search after migration:
Evidence root: Confirmed behaviorReal Qt form/native typing, production Qt main-thread dispatcher, authenticated localhost ZMQ server/client, declaration-aware DTO decoding, mutation/catalog, reset, original-handle history undo, absence and view/server disposal pass. Explicit wire examples:
Receipts: Remaining application blockerOpenHCS tests do not reach collection. The old installed metaclass-registry lacks |
DTO correction checkpoint (not the completed node correction)Removed the eleven Python-class-name DTO declarations. All six named DTO deletion checks pass; MIME metadata is retained. Missing and present previews are now members of the existing preview family and own rendering. The renderer no longer tests preview presence. Preview null/type flags are deleted too. The declaration owner is not derivable from a raw/resolved datum (especially null or absence). Source-paired verification: 28 checks pass (22 projection/codec cases and six declaration gates). The classifier gate remains failing and is explicitly excluded from this DTO-only result. The existing real Qt native-input → main-thread dispatch → authenticated localhost ZMQ → DTO decode → reset/history undo → disposal journey passes, including literal null, empty text and the string Initial failure evidence: Still unfinished: operation-owning configuration/leaf values, removal of classifiers and the full application path. Normal OpenHCS startup still has the previously retained lazy |
Declaration-owned checkpoint (source-paired, not installed)Pushed heads: ObjectState 93eae56; FormBridge b5d3af8; PyQt 8690060; OpenHCS f0c07c4. Dependency python-introspect ffcffaa (PR #25). ValueDeclaration compiles annotations once. ParameterOwnerDeclaration owns parameters, field_types and init_field_names (one definition each); constructor and configuration owners share that implementation. FieldValue.of only preserves meaning members or boxes data; it does not classify runtime payloads. Named configuration access is absent on scalar leaves. Declared collection element types own traversal and serialization. DeclaredValue shares metadata/copy/serialization lifecycle between configuration and collection values. InheritedDefault is the single frozen class-keyed default record; ValueDefault/FactoryDefault materialize it. own_default and inherited_default now index declared fields and raise KeyError for unknown names. Their direct resolver callers validate against the requested declaration; configuration_mro(field_name) filters ancestors by their declared fields before reading defaults. All touched production roots have zero Text-line counts (published baseline -> checkpoint): | production root | is_container | is_config_type | literal The two remaining lazy callback-name lines are its definition and register_type_resolver registration; zero callers, imports or exports remain in the touched production family. The normalization wrapper/cache and per-caller fallback expressions are removed. Archived dirty-before normalization count was 45, now 2. FormBridge FieldValue.of: published 3 -> 2; archived dirty-before 15 -> 2. get_base_type_for_lazy: 8 -> 0. metadata.get: 4 -> 2. Combined requested regex: published 15 -> 4; archived dirty-before 27 -> 4. The remaining FieldValue.of sites preserve DTO meaning after mapping JSON data, not payload kind. The remaining metadata.get sites read canonical runtime code-document tokens, not declaration metadata. widget_creation_type consumers: WidgetCreationPipeline.init selected an enum by the string; it now calls param_info.creation_type(WidgetCreationType). ParameterFormManager._create_parameter_widget only logged it; it now logs the actual info class. All four string ClassVars are deleted. NRA syntax/family evidence: 1,590 production/dependency modules, 9,953 classes, no parse errors. AST calls: ObjectState getattr 28 -> 13, FormBridge 17 -> 15; is_container/is_config_type calls zero, is_dataclass exactly one annotation compiler call. Dynamic resolution is not established by these syntax counts; the actual source-paired journey below checks the affected behavior. Exact deleted Python type-name DTO declarations remain: source_type (1), object_state_path_type (2), raw_value_type (2), resolved_value_type (2), object_type (3), help_target_type (1). Genuine mime_type is retained. Deletion gates cover the exact declarations, not all *_type fields. Verification: python-introspect public resolver/analyzer 64 passed; producer declaration/handle/inspection/scopeless/codegen/deletion cases 51 passed; bridge DTO/deletion cases 29 passed; Qt declared/numeric cases 66 passed. Actual source-paired native Qt input, main-thread dispatch, authenticated localhost ZMQ, DTO decode, reset, undo and disposal passed with null, empty string, text None, inheritance and absence. No provider and no installation. Still unfinished: OpenHCS config startup reaches SourceBindingsConfig.post_init but attempts to iterate Inherit in its source-binding normalization. Earlier removed has_value reads in the shared normalization owner are repaired; the whole remaining post-init normalizer needs a directive-aware owner pass. The old test_subfield_semantics.py still imports deleted MISSING/string APIs and fails collection; this failure is retained, not skipped or xfailed. Full OpenHCS/app acceptance and #32/#46 cutover are not complete. doctrine-segment remains the implementation owner. Dependency: python-introspect ffcffaa, draft PR #25. No shared runtime publication/restart performed. |
|
Checkpoint 560b443 restores real source-paired PipelineConfig construction. Shared source-binding declarations have one base, raw Inherit/Unset survive normalization, compiled plans reject directives, and seven direct application API checks pass. This is not final P44 acceptance: owner review correctly rejects the remaining per-field normalization roster and explicit parent post-init calls. Both are being replaced in this PR by field-declared transforms, complete declaration iteration and cooperative C3 lifecycle. The OpenHCS pytest run is retained as blocked by unbuilt native extensions after pairing PolyStore source; no skip/xfail or installation claim. |
|
Fresh direct-task checkpoint addressing the source-plan roster/C3 review; this does not resume the blocked P44 goal or claim final P44 acceptance. Published source pairing:
Ownership and deletions:
NRA source evidence: 54 relevant production modules across ObjectState, FormBridge, PyQt, python-introspect and OpenHCS parsed; zero parse errors. Before/after source-plan roster calls: source_bindings.py 5→0, config.py 5→0. Explicit named-parent post-init calls: 2→0. Declaration, inheritance, write/check/call and consumer sites are retained at Verification:
|
Companion to ObjectState #62, implementing #32 and then #46. Not ready to merge or install. Constructor-patching scopes are removed; explicit Inherit and Unset replace inheritance through None. Qt input state and native signal subscriptions are being moved into the existing shared input mixin rather than retaining duplicated None decisions. Producer core resolution and six real Qt controls have targeted checks; remaining form-manager, nested/inline input, saved-code, field-handle and installed application paths are unfinished. No complete migration or application-readiness claim. Original sessions and other owners worktrees are untouched.