[binaryaaron]

Redact, Annotate, Hash and Substitute now run on a small private kernel, engine/graph_core/, instead of the
graph SDK's replacement dispatch, and the graph SDK code that only those routes used is deleted. Users see no change
from this branch's base: outputs, errors and failed-record accounting are the same on pandas 2 and 3, and the
public anonymizer and anonymizer.graph exports are untouched. Rewrite and detection keep their current paths for
now.

What changes

• Kernel (src/anonymizer/engine/graph_core/). A domain-free kernel (kernel.py: kind registry, static plan,
  reducer over an immutable state, cancel token), the authoring types (authoring.py) and prepare/protect
  (execution.py). Everything runs on the caller's thread with no event loop of its own. A raising callable closes
  its target as failed, and BaseException (Ctrl-C) propagates unchanged. The package exports no names.
• Routes (engine/replace/route.py). run_local_route runs Redact, Annotate and Hash with one target per row.
  run_substitute_route runs the LLM map step (LlmReplaceWorkflow.generate_map_only, whose body is 0.4.0's again)
  as one opaque operation, then applies the returned rows through the same route. ReplacementWorkflow.run and
  apply_local_replace_strategy call them, so a caller's replace_runner subclass still works.
• Facade. For a replace config, _run_internal_impl calls the detection workflow and then the replace runner
  directly. The existing entity_label_examples withholding for 0.4.0-signature detectors is kept.
• Context sources. Two targets that declare a context source under the same name but with a different provider or
  bounds are rejected at build time (duplicate_context_source), so one target can never read another's context.
• Null and non-string text cells. The standalone detection call now publishes the text port only when every cell
  is a str, so these cells behave as in 0.4.0.

Deleted

The following files are deleted:

• engine/replace/_semantic_replace.py
• engine/replace/_semantic_map_profile.py
• engine/replace/_graph_redact.py
• engine/detection/_graph_profile.py
• engine/detection/_graph_invocation.py

These code paths are deleted from files that stay:

• run_map in _semantic_map.py;
• the Redact/Substitute branches of _opaque_redact.py;
• native_map and builtin_replacement in ndd/_route.py;
• apply_replacement_map in strategies.py;
• DelegatedRequestContract.publication_sources, which is not exported;
• the entity-span ports in harness/frames.py.

Tests that exercised only deleted code or pinned the old dispatch topology are deleted. Tests that asserted on the old
route's internal executions keep their behaviour assertions:

• Deleted files: test_all_replace_graph.py, test_detection_redact_config.py, test_substitute_semantic_map.py
  and test_publication_source_admission.py, plus individual functions in the route-resolution, detection-control and
  replace-strategy tests.
• Rewritten: test_injected_redact_graph.py and parts of test_detection_source_frame.py.

The kernel's behavioural checks live in tests/spec_programs/.

Net non-blank, non-comment code lines under src/, tools/ and examples/: +58 (1,228 added, 1,170 removed; the
new kernel package is 869 of the added lines). Most of the Substitute map path stays for now because Rewrite still
reaches it.

Checks

• make check passes.
• make test: 5,442 passed. One test, test_wandb_snapshot_rejects_fifo_socket_and_device, fails in the test
  environment with AF_UNIX path too long, and fails the same way 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 are at most 3.7% above the base on any replace case. Substitute runs with no detected
  entities execute about a third fewer lines.
• Replace cost is linear in row count. It is still several times 0.4.0's plain loop: Redact takes about 7.8 s against
  1.3 s at 40,000 rows. The kernel applies state events for every row and operation, which a plain loop does not.

Main merged (5155d6d)

This branch now includes main at 5155d6d. Two tests that main added in #322 fail on this branch: test_join_new_columns_preserves_surviving_target_order and test_run_in_memory_records_tracks_dropped_caller_id_and_survivor_order. This slice keeps a positional Rewrite join and rejects a record that DataDesigner drops without reporting it; the remove-old-sdk PR in this stack takes main's record-id join and drops that check, and both tests pass from there on.

Main merged (f294dc1)

This branch now also includes main at f294dc1 (#328, recovery from incomplete replacements in Rewrite) through its base branch, with #328 ported as described on the base PR.

🤖 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 (Redact, Annotate, Hash and Substitute on the private kernel) is now reviewed in #319. Nothing was rewritten; this description stays as the record of the step.