[Phase 16] S9 — Compliance evidence integrity and GDPR erasure completeness #895

Roadmap · Phase 16, step S9 · size L · 5 unit(s) · planned in docs/roadmap/phase-16-trust-surface-hardening.md:553-581

Summary

The compliance surface claims a stronger evidentiary state than it holds. purge --kind artefacts cannot find the compliance bundle ForgeLM itself writes: the exporter emits fixed basenames while the resolver looks for file names that embed the run id. It fails safe (deletes nothing, emits data.erasure_failed with NoMatchingArtefacts, exits 1), so the defect is completeness, not destruction. The trainer also writes data_governance_report.json there, and the pipeline writes compliance/pipeline_manifest.json, which belongs to the whole pipeline and must not be deleted by one run's purge.

Units

Plan

• Rewrite _artefact_targets_for_run to the allow-list + ownership-proof model the design document already specifies, mirroring the _legacy_staging_owned_by marker pattern in the same file.
• Breaking changes planned here: AuditLogger.log_event becomes fail-closed; reverse-pii and purge move to non-zero exits on states that returned 0.

Exit criteria

• A real bundle produced by generate_training_manifest() + export_compliance_artifacts() is fully removed by a matching run id and untouched by a wrong one, while pipeline_manifest.json and audit_log.jsonl survive — a writer → reader integration test, not a synthetic-filename fixture; the fabricated docstring examples in _purge.py are deleted.
• A malformed line sharing an identifier blocks the completion record.
• reverse-pii returns non-zero when AuditLogger initialisation raises.
• An OSError injected separately at the manifest open, fsync and os.replace steps never yields silent success.
• A required → rejected → required chain is served by one shared ordered-state helper across the listing and mutation surfaces.

How to deliver

• One root-cause family, committed on its own; then an Opus review round and a Sonnet review round, each with verified findings fixed and committed before the next step begins.
• A reproducer that is red before the fix and green after; the full gauntlet from CLAUDE.md passes under ./.venv/bin/python.
• Every English documentation edit ships with its Turkish mirror in the same commit; a new or renamed audit event gets its EN and TR catalog rows in the same PR.
• A step that grows a deferred module carries its budget_history justification in the same diff; a widened guard rolls out non-strict first, then flips.
• Commit bodies cite Absorbs F-W20260729-…; the CHANGELOG [Unreleased] section gets an entry only for a change a user or library consumer can observe (decision C-20).

_{Planned in docs/roadmap/phase-16-trust-surface-hardening.md (read at f94595f). Unit IDs (F-W20260729-…) come from the 2026-07-29/30 full-project review; the issue numbers next to them are the October 2026 review's records of the same defects.}