fix: DbtProcess/DbtColumnProcess parents + typing.get_type_hints shim (SHA-887)

[Aryamanz29]

Three SHA-887 follow-ups that all independently break atlan-publish-app's ordering graph against pyatlan v9.x. Bundling them because the symptoms overlap (publish-app order regressions reproduce until the full set lands) and they share the same generator-level mechanism.

1. DbtProcess extends Process, not Dbt

• Atlas typedef: superTypes=[\"Dbt\", \"Process\"]
• Generator picks super_types[0] → Dbt → publish-app's issubclass(cls, Process) check (used to identify lineage types that must publish after the entities they reference) returned False.
• DbtProcess then landed in Batch 1 alongside its upstream entities → ATLAS-404 on inputs / outputs lookups in live dbt runs.
• Fix: add \"DbtProcess\": \"Process\" to _SUPERCLASS_OVERRIDES in class_generator.py and patch the generated pyatlan/model/assets/dbt_process.py source.

2. DbtColumnProcess extends ColumnProcess, not Dbt

• Same root cause: superTypes=[\"Dbt\", \"ColumnProcess\"].
• Same fix applied to pyatlan/model/assets/dbt_column_process.py and the override map.

3. typing.get_type_hints() shim for core Attributes classes

• Five core asset modules (s_q_l, schema, data_product, data_quality_rule, process) gate cross-module relationship-type imports under if TYPE_CHECKING: to avoid circular imports at module load. With from __future__ import annotations, those forward-ref strings can't be resolved by typing.get_type_hints() against the defining module's __dict__ and the call raises NameError — e.g. name 'DbtTest' is not defined for Table.Attributes.
• Pydantic v1's update_forward_refs(**localns) in core/__init__.py already handles this for the SDK's own use, but reflection-based downstream tooling (notably atlan-publish-app's ordering graph) falls through to a degraded fallback that collapses all SQL types into one publish batch.
• Fix: after every asset has loaded, core/__init__.py injects the resolved classes back into each affected module's __dict__ so typing.get_type_hints(cls.Attributes) succeeds without a caller-supplied localns.
• A new _RUNTIME_INJECTION_OVERRIDES constant + _build_runtime_injection_block helper in class_generator.py plus a corresponding placeholder in templates/core/init.jinja2 ensure the shim survives regeneration.

Tests

• tests/unit/model/dbt_process_test.py — 7 cases (issubclass guards, __bases__[0], type_name immutability, dbt_* + Process attribute round-trip).
• tests/unit/model/dbt_column_process_test.py — same 7 cases for the ColumnProcess parent.
• tests/unit/model/get_type_hints_compat_test.py — parametrised over all 5 affected core classes plus a concrete spot-check that SQL.Attributes.dbt_tests resolves to the real DbtTest class.

Verified

• ./qa-checks — ruff format / ruff check / mypy all pass
• pytest tests/unit -q --ignore=tests/unit/aio — 6033 passed, 2 skipped
• Manual issubclass / get_type_hints checks against installed pyatlan
• Downstream verification in atlan-publish-app once this ships: DbtProcess / DbtColumnProcess move to a strictly later batch than the entities they reference, and typing.get_type_hints() resolves on every core Attributes class without a workaround localns

Notes

Release notes will be added in the next version bump — HISTORY.md is left alone here so the already-shipped 9.7.0 section isn't retroactively edited.

🤖 Generated with Claude Code

[claude]

Bug: Duplicated shim block

Lines 404-428 and 430-454 contain identical runtime injection shim code. The main branch __init__.py ends at line 402 — this PR should add one shim block (26 lines), not two (52 lines).

The duplication suggests either:

1. The file was generated twice without cleaning the intermediate state, or
2. The generated output was manually edited after generation

Suggested change

	# typing.get_type_hints() compatibility shim
	#
	# Some Attributes classes reference cross-module relationship types whose
	# imports must stay under TYPE_CHECKING to avoid circular imports at module
	# load time. Pydantic v1 resolves these forward refs via update_forward_refs()
	# above, but typing.get_type_hints() does its own resolution against each
	# class's defining module __dict__ — which doesn't see TYPE_CHECKING-only
	# names. Inject the resolved classes back into the affected modules now that
	# all assets have loaded so downstream tooling (e.g. atlan-publish-app's
	# ordering graph) can call typing.get_type_hints() without a custom localns.
	from . import data_product as _data_product_module
	from . import data_quality_rule as _data_quality_rule_module
	from . import process as _process_module
	from . import s_q_l as _s_q_l_module
	from . import schema as _schema_module
	_data_product_module.StarburstDataset = StarburstDataset # type: ignore[misc]
	_data_quality_rule_module.Column = Column # type: ignore[misc]
	_process_module.Procedure = Procedure # type: ignore[misc]
	_process_module.BigqueryRoutine = BigqueryRoutine # type: ignore[misc]
	_process_module.FabricActivity = FabricActivity # type: ignore[misc]
	_process_module.Function = Function # type: ignore[misc]
	_s_q_l_module.DbtTest = DbtTest # type: ignore[misc]
	_schema_module.SnowflakeDynamicTable = SnowflakeDynamicTable # type: ignore[misc]
	_schema_module.Table = Table # type: ignore[misc]

Remove this entire block (lines 430-454). The first shim block at lines 404-428 is sufficient.

[Aryamanz29]

Superseded — splitting into two cleaner PRs (atlan-python parent-class fixes + atlan-publish-app switch to pydantic ModelField). The get_type_hints shim added too much pyatlan-side maintenance burden; the better fix is publish-app moving off typing.get_type_hints entirely. New PRs incoming.