L6: code generation owned by pycodify and objectstate.codegen

[trissim]

Surface L6 (receipt). Pipelines persist as pycodify-generated .py files, so code generation is the persistence format.

Library PRs to release first, in this order:

1. 0.2.0: generated source loads back every value it names pycodify#4: 0.2.0
2. 2.1.0: objectstate.codegen; drop well_filter debug ObjectState#28: 2.1.0 (depends on pycodify 0.2.0)

Pins are raised here: pycodify>=0.2.0,<0.3, objectstate>=2.1,<3. The submodules point at the two branch heads.

Deleted from OpenHCS (now owned by the libraries)

• openhcs/core/python_source_literal.py: FunctionReference now subclasses pycodify.PythonSourceLiteral, which has one source_fragment method.
• openhcs/serialization/source_path_factoring.py: path factoring is built into generate_python_source, and the six document renderers call it directly.
• EnumMemberFormatter, AxisDeclarationFormatter, OpenHCSImportableTypeFormatter and NameMappingLookup: replaced by pycodify.imported_name and TypeFormatter.
• The LazyDataclass formatter (six classes) and FunctionStepCleanModeFieldPolicy: replaced by objectstate.codegen.
• The repr() catch in OpenHCSCallableFormatter: an unreferenceable callable now raises instead of saving <function f at 0x…>.

Round-trip test (tests/unit/test_source_round_trip.py) runs four checks:

• Every registered config dataclass (all 22 lazy/base pairs minus the abstract StreamingConfig), with every field set to a non-default value derived from its annotation. This covers axis classes, the PartitionFilterMode family, partition filters, nested source-binding records and paths.
• The same configs at their defaults, in both clean and full mode.
• A pipeline document with every function-pattern shape: callable, (callable, kwargs), a declared path kwarg, nested-enum kwargs (CropModule.Shape), hidden runtime kwargs, chain, and a dict by group. It asserts the path_root factoring.

Each check loads the source back and compares raw values.

It found one real bug: base configs that inherit through None (StepPartitionFilterConfig, PathPlanningConfig, the streaming configs and others) were written with their resolved values in full mode, so a saved GlobalPipelineConfig stopped inheriting. The fix is in ObjectState#28. I checked that the test fails without it.

Guards (tests/unit/test_l6_codegen_guards.py):

• The two deleted modules stay deleted.
• No OpenHCS can_format claims Enum, Path, classes, PythonSourceLiteral or lazy dataclasses.
• No __qualname__.partition or _explicitly_set_fields appears under openhcs/.
• No well_filter appears in ObjectState sources.

Deletion certificates

• derived: every OpenHCS formatter and module listed above. The libraries own the same fact through imported_name, PythonSourceLiteral, PathFactoring and objectstate.codegen.
• contract: pycodify FormatContext.extensions and generate_python_source(context=). Their only user was OpenHCS path factoring, which moved into the library.
• contract: the repr() outputs, which never loaded back.

Checks run locally

• Ratchet over openhcs, scripts and benchmark: no increases (5 counts decreased).
• nra-r1: increased: [].
• Unit tests: 365 + 263 passed across the codegen, document, GUI and agent tests listed in the receipt.

L6 · PR #1198 · guards 4/4 · production −591 +43 · tests −215 +419

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.