refactor: replace df.attrs with typed PipelineContext dataclass

[memadi-nv]

Summary

Replaces the experimental df.attrs mechanism for threading the user-facing
text-column name through the pipeline with a typed ResolvedInput dataclass
owned by the orchestrator. Workflows no longer need to preserve attrs
across pandas operations that silently drop them.

Changes

• New ResolvedInput dataclass (engine/resolved_input.py): frozen
  dataclass holding the input dataframe, the user's requested_text_column,
  and the resolved_text_column (post-collision rename). read_input returns
  this directly.
• Reader (engine/io/reader.py): populates both requested and resolved
  text-column fields; collision rename logic unchanged but now exposed
  explicitly via the dataclass.
• Orchestrator (interface/anonymizer.py): _run_internal accepts
  ResolvedInput and pulls resolved_text_column from it for output-column
  restoration — no more df.attrs reads.
• Workflows (detection_workflow.py, llm_replace_workflow.py,
  rewrite_workflow.py, row_partitioning.py): drop their attrs= plumbing
  / merge_and_reorder no longer takes the attrs parameter.
• Public result types (interface/results.py): AnonymizerResult and
  PreviewResult gain resolved_text_column (replaces original_text_column).
  render_record_html parameter renamed accordingly.

Breaking Changes

• AnonymizerResult.original_text_column → AnonymizerResult.resolved_text_column
• PreviewResult.original_text_column → PreviewResult.resolved_text_column
• render_record_html(..., original_text_column=...) → render_record_html(..., resolved_text_column=...)

Tests

• New tests/engine/test_resolved_input.py: covers construction, immutability,
  the requested-vs-resolved asymmetry, and the headline invariant that
  metadata survives pandas merge/concat/groupby.
• Workflow-level assert "..." not in df.attrs regression guards removed
  (they enforced an implementation detail, not the orchestrator's contract).
• Replaced with a behavioral orchestrator test that injects deliberately
  misleading attrs into mocked workflow output and asserts Anonymizer.run()
  • display_record still produce correct user-facing columns.

Follow-up

• Add renamed_input_columns field to ResolvedInput so users can see all
  collision-driven renames, not just the text column (deferred per review).

[greptile-apps]

Greptile Summary

This PR eliminates the use of pandas' experimental DataFrame.attrs for threading pipeline-level metadata (original_text_column) through workflow stages, replacing it with a typed ResolvedInput dataclass that carries both the requested_text_column (what the user asked for) and resolved_text_column (the post-collision-rename identifier). The orchestrator now owns this metadata explicitly rather than smuggling it through a fragile, drop-prone attrs slot.

• ResolvedInput is introduced as a frozen dataclass in engine/resolved_input.py, combining the role of the old ResolvedInputColumns helper and the old df.attrs[\"original_text_column\"] key. read_input() returns it instead of a bare pd.DataFrame.
• AnonymizerResult and PreviewResult gain a resolved_text_column field; _DisplayMixin.display_record reads from it instead of trace_dataframe.attrs. All manual attrs={**a.attrs, **b.attrs} plumbing in merge_and_reorder, LlmReplaceWorkflow, and RewriteWorkflow is removed.
• Tests are migrated from result.attrs[...] assertions to result.requested_text_column / result.resolved_text_column / result.dataframe.*, and a new integration-level regression guard verifies the orchestrator never consults df.attrs for text-column resolution.

Confidence Score: 5/5

Safe to merge — the refactor correctly threads metadata through an explicit typed field at the orchestration layer, and all test stubs have been updated to match the new API surface.

The removal of df.attrs plumbing is complete and consistent across detection, replace, rewrite, and row-partitioning. The new ResolvedInput dataclass correctly preserves both the user-requested and post-collision-resolution column names throughout the pipeline. _rename_output_columns and _build_user_dataframe receive resolved_text_column explicitly, and the new integration-level regression guard verifies this end-to-end.

No files require special attention.

Important Files Changed

Filename	Overview
src/anonymizer/engine/resolved_input.py	New frozen dataclass replacing the old ResolvedInputColumns + df.attrs pattern; with_dataframe() uses dataclasses.replace correctly; docstring explicitly documents the eq/hash unsupportability with a pd.DataFrame field.
src/anonymizer/interface/anonymizer.py	_run_internal now takes context: ResolvedInput; resolved_text_column is threaded explicitly to _rename_output_columns, _build_user_dataframe, and AnonymizerResult without touching df.attrs.
src/anonymizer/interface/results.py	AnonymizerResult and PreviewResult gain resolved_text_column: str as a required field; _DisplayMixin reads from it instead of trace_dataframe.attrs.
src/anonymizer/engine/io/reader.py	read_input now returns ResolvedInput; the rename of the text column to COL_TEXT uses resolved.resolved_text_column correctly, and the result is wrapped with with_dataframe() cleanly.
src/anonymizer/engine/row_partitioning.py	attrs parameter dropped from merge_and_reorder; the combined.attrs assignment is removed, simplifying the function signature.
tests/engine/test_resolved_input.py	New test file covering construction, immutability, with_dataframe(), and the headline invariant that metadata survives merge/concat/groupby on the wrapped DataFrame.
tests/interface/test_anonymizer_interface.py	All df.attrs setups removed from stubs; new regression guard verifies orchestrator uses ResolvedInput rather than workflow-output df.attrs.

_{Reviews (4): Last reviewed commit: "address feedback" | Re-trigger Greptile}

[lipikaramaswamy]

IMO we should tighten the abstraction before merging. The name PipelineContext indicates something broader than what this object currently is. It is not really passed through the whole pipeline; it is the result of input normalization plus metadata the top-level orchestrator needs when turning internal columns back into user facing columns. A name like ResolvedInput, ResolvedInputFrame, InputContext, or similar would describe the current role better.

Related: now that I think about it, original_text_column is misleading. In collision cases this can become something like final_entities__input, which is not the original requested column name anymore. It is the resolved user facing/output text column. I think the object can capture the full input-resolution result, for example: dataframe, requested_text_column, resolved_text_column, and renamed_input_columns. Ok to punt some of that as follow-up, but if we're introducing the dataclass now, it's probably worth getting the core vocabulary right.

Re tests: assert "original_text_column" not in result.dataframe.attrs enforces an implementation detail that workflows must not preserve attrs. But the real invariant is that the orchestrator no longer depends on attrs. We should instead test that mocked downstream workflow outputs have no attrs and Anonymizer.run() still restores the correct user-facing columns/display behavior.