[binaryaaron]

Detection now runs main's stages on the private kernel from #324, and the old graph SDK is deleted:
engine/graph_sdk, engine/harness, their adapters and their test suite. anonymizer.graph now exports the
kernel's 23 public names. Apart from the model bridge timeout, which again behaves as on main (below), users of
Anonymizer see no change from this branch's base: outputs, errors and failed-record accounting are the same on
pandas 2 and 3, and the root anonymizer exports are unchanged.

This branch stacks on #325 and is rebased onto its base branch when #324 and #325 merge.

What changes

• Detection (engine/detection/detection_workflow.py). EntityDetectionWorkflow.run starts from main's body and
  runs it as up to three kernel stages: detect, tag latent entities (only when enabled) and finalize. Each
  DataDesigner call is one opaque operation. Row correspondence for inputs with duplicated rows is kept:
  source_positions from the NDD adapter are composed across the stages, so each surviving row maps to its input row.
• run_stages moves to the public module engine/graph_core/stages.py, used by detection and both Rewrite
  workflows.
• anonymizer.graph re-exports exactly 23 names from engine/graph_core: Assessment, CancelToken,
  ContextScope, ContextSource, ContextView, Datum, Dependency, GraphRejection, GraphScope,
  InspectionSummary, Limits, Mention, Operation, OutputGroup, PreparedProtection, ProtectionGraph,
  ProtectionResult, Requests, TargetOutcome, Workflow, prepare, protect, register_kind. The previous 228
  names are removed. tests/graph/test_public_api.py pins the list and checks that importing anonymizer.graph loads
  only the kernel.
• Example and docs. examples/graph_sdk/local_workflows.py and docs/concepts/graph-sdk.md are rewritten against
  the 23 names. docs/development/graph-sdk-cutover.md is left as is and still names deleted modules.
• Model bridge timeout. In the detection column plugin, a bridged model call that times out is cancelled and
  raises ModelTimeoutError, as on main. Before this change the handler waited for the call to finish with no bound.
• NDD adapter. The caller-supplied source-frame path, which only the graph SDK used, is removed. The adapter
  always captures its own frame from the input.

Deleted

• engine/graph_sdk/ (47 modules) and engine/harness/ (4 modules).
• Detection's graph route: detection/_semantic_detection.py, _opaque_redact.py, _stage_binding.py, artifacts.py,
  run_arguments.py, ndd/_route.py, engine/source_text.py, and the graph wiring in
  workflow_columns/detection/impl.py (now identical to main).
• tests/graph_sdk/ and the test files main does not have, outside tests/spec_programs/. Test files main has keep
  their current content, minus the individual tests that imported deleted code. The compat pytest marker goes with
  them; pyproject.toml matches main.

Net non-blank, non-comment code lines under src/, tools/ and examples/: −22,407 (49,955 to 27,548). Test code
lines: 69,421 to 24,696.

Known differences from main that this branch keeps

These come from the base branches, not from this change:

• Library log messages no longer include payload text, paths or exception details.
• The root package loads its public names lazily.
• Caller detectors with the older signature are not passed entity_label_examples.
• The adapter does not apply main's DataDesigner scheduler-wait patch.
• Chunked validation waits for sibling tasks to finish after one fails.

Checks

• make check passes.
• make test: 1,557 passed. test_wandb_snapshot_rejects_fifo_socket_and_device fails in the test environment
  with AF_UNIX path too long, as on the base.
• uv run --group dev python -m tests.spec_programs.run_all: 59 pass, 8 retired.
• Parity against main (f4f2b73d) on pandas 2 and 3, over all cases: no blocking difference beyond the
  differences the base already has.
• Executed lines per case against main: median 1.20×, at most 1.22× on cases whose status matches main.

Main merged (5155d6d)

This branch now includes main at 5155d6d. Rewrite's _join_new_columns now uses main's record-id join: surviving rows keep input order. One difference from main remains: when record ids cannot match each surviving row to one input row, the join raises AnonymizerWorkflowError ("Unassociated workflow result has ambiguous source correspondence") where main's raises a pandas ValueError. The test test_run_in_memory_records_tracks_dropped_caller_id_and_survivor_order, added on main, is adapted on purpose: its mocked DataDesigner now returns the source-key column that this branch's adapter adds to each submitted row, for the rows that survive. Its assertions are unchanged.

Main merged (f294dc1)

This branch now also includes main at f294dc1 (#328, recovery from incomplete replacements in Rewrite) through its base branch.

🤖 Generated with Claude Code

[binaryaaron]

Folded into #319: aagonzales/graph-sdk/feat/main was fast-forwarded to 44924e2e, which contains this branch unchanged, so this step (detection on the kernel and removal of the old graph SDK) is now reviewed in #319. Nothing was rewritten; this description stays as the record of the step.