chore: upgrade data-designer to 0.6.0

[lipikaramaswamy]

Summary

Upgrades data-designer from 0.5.7 → 0.6.0 to pick up the native async engine and bundles a few cleanups that surfaced while verifying the upgrade by re-running the tutorial notebooks end-to-end.

Changes

• pyproject.toml + uv.lock: bump data-designer to 0.6.0. No code changes
  were required — the async engine is enabled by default in 0.6, so the existing
  NddAdapter paths get the benefit transparently.
• models.yaml: raise gpt-oss-120b max_parallel_requests from 2 → 16
  to actually exercise the new async concurrency. Previously gated the rest of
  the pipeline at 2 in-flight LLM calls; bumping cut tutorial notebook runtime
  roughly in half on gpt-oss-120b-heavy stages.
• Makefile: make convert-notebooks now passes --set-kernel anonymizer-venv
  to jupytext. Previously jupytext picked an arbitrary registered kernel
  (often a stale one from another project), causing notebooks to silently run
  against the wrong Python environment.
• llm_replace_workflow.py: replacement-map debug logs no longer emit raw
  PII. The previous [DEBUG] Replacement map record … line dumped the full
  original → synthetic mapping for every record, leaking the source values
  into log streams whenever DEBUG was enabled. Logs now summarize counts per
  label plus any anomalies (entities unrequested by the prompt, requested
  entities the LLM didn't fill). The warning path for empty/missing maps was
  also tightened to always fire (it previously sat under an elif that swallowed it).
• docs/notebook_source/*.py: consolidate from anonymizer.config.rewrite import PrivacyGoal
  into the top-level from anonymizer import … line. PrivacyGoal is already
  re-exported from the package root, so the deeper import was unnecessary and
  inconsistent with the bundled skill template.

What's NOT in this PR

• Re-rendered notebook outputs (docs/notebooks/*.ipynb) — kept as a separate
  step so the dependency change is reviewable on its own.
• A separate sensitivity-disposition refactor (schema simplification + prompt
  fixes) that surfaced during this verification. Tracking that as a follow-up
  branch since it's a larger change with its own design discussion.

Test plan

• make test — unit tests pass on data-designer 0.6.0
• make convert-notebooks — all five tutorial notebooks execute end-to-end
  against gpt-oss-120b and nemotron-30b-thinking
• Verified detection + replace + rewrite pipelines all run under the async engine
• Confirmed no raw PII in [DEBUG] Replacement map record … log lines
• CI green

[greptile-apps]

Greptile Summary

This PR upgrades data-designer from 0.5.7 to 0.6.0 to pick up the native async engine, and bundles several correctness fixes that surfaced during end-to-end notebook verification.

• pyproject.toml / uv.lock: pin data-designer==0.6.0; the async engine is transparent to existing NddAdapter callers.
• llm_replace_workflow.py: debug logs no longer emit raw PII — counts + labels only; the empty-map warning is promoted from an elif branch (silently suppressed when DEBUG was enabled) to a standalone if, so it always fires. Regression tests added to test_llm_replace_workflow.py.
• models.yaml / Makefile / notebook sources: raise gpt-oss-120b concurrency cap from 2 → 16, pin the jupytext kernel to anonymizer-venv, and consolidate the PrivacyGoal import to the package root (where it is already re-exported).

Confidence Score: 5/5

Safe to merge — all changes are additive or correctness fixes with no breaking surface area.

The dependency bump is a pin-level upgrade with no API changes required in callers. The logging fix removes raw PII from debug output and tightens a previously silenced warning — both are strictly safer than before. The concurrency bump and kernel-pin changes are config-only and exercised by end-to-end notebook runs. The new regression tests directly validate the PII-safe logging invariant.

No files require special attention.

Important Files Changed

Filename	Overview
src/anonymizer/engine/replace/llm_replace_workflow.py	PII-safe debug/warning logging: raw entity values replaced with counts + label counters; elif warning promoted to if so it fires regardless of log level.
tests/engine/test_llm_replace_workflow.py	Three new regression tests covering happy-path debug log, anomaly summaries, and empty-map warning — all asserting no raw PII appears in log output.
pyproject.toml	Bumps data-designer pin from 0.5.7 to 0.6.0; no other dependency changes.
src/anonymizer/config/default_model_configs/models.yaml	Raises gpt-oss-120b max_parallel_requests from 2 to 16 to utilize the new async engine concurrency.
Makefile	Adds --set-kernel anonymizer-venv to the jupytext invocation so notebooks run against the correct Python environment.
docs/notebook_source/04_rewriting_biographies.py	Consolidates PrivacyGoal into the top-level from anonymizer import … line; PrivacyGoal is in __all__.
docs/notebook_source/05_rewriting_legal_documents.py	Same PrivacyGoal import consolidation as notebook 04.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[LLM returns replacement map] --> B[_filter_replacement_map_to_input_entities]
    B --> C{valid dict?}
    C -- no --> D[WARNING: unexpected type\nreturn empty]
    C -- yes --> E[Build allowed_pairs from\nparsed_entities]
    E --> F[Filter: keep only pairs\nin allowed_pairs, deduplicate]
    F --> G{DEBUG enabled?}
    G -- yes --> H[Log counts only:\nrequested / raw / filtered\nunrequested_by_label / unfilled_by_label\nNO raw PII]
    G -- no --> I{filtered empty\nAND allowed_pairs non-empty?}
    H --> I
    I -- yes --> J[WARNING: empty after filtering\ncounts + labels only\nNO raw PII]
    I -- no --> K[Return filtered replacements]
    J --> K

Loading

_{Reviews (2): Last reviewed commit: "test: add caplog regression tests for PI..." | Re-trigger Greptile}

[andreatgretel]

Reviewed the dependency bump, replacement-map logging change, notebook conversion update, and import cleanup. Smoke tests and targeted unit tests pass. Only left one small suggestion to add a committed regression test for the PII-safe debug logging behavior. Otherwise looks good to me!