feat(registry): Phase 2 provenance — attributes, coverage, bad→redispatch

[GondekNP]

Implements Phase 2 of the registry provenance model (REGISTRY_PROVENANCE.md §7, §10, §11.1, §12), building on Phase 1's status/supersession/cell_data_current keystone.

What's in it

1. Run-level attributes (the run_hash slice of the tag store)

tags → attributes turned out to be a whole scoped metadata subsystem (run_hash + session_id + run_id + custom scopes), so per the agreed run-level-slice decision this renames only the run-level surface and leaves the general facility as tags (the run_tags store is unchanged — no migration):

• set_attributes / get_attributes / list_attribute_keys (were tag_by_run_hash / get_tags_by_run_hash).
• Currency-aware find_by_attribute(key, value, current_only=True) and find_current_by_attribute → AttributeMatch(run_hash, label, attributes).
• General tag_by_session_id / tag_by_run_id / tag_custom / find_tagged keep the tags name.

2. "Current by default" for diagnostics

DiagnosticQueries read methods now read through the cell_data_current view. Added _refresh_current_view() (called after dynamic cell_data column additions) so the SELECT c.* view keeps up with variable columns — fixes a latent staleness bug the repoint exposed.

3. Target designs — attribute-based completeness (§12)

Persisted TargetDesign / Requirement / check_design → TargetCoverageReport, over two new tables (target_designs, target_requirements). Completeness is asserted over attributes, not run_hashes, against active runs only (superset match; bad/superseded surfaced but never satisfy). register_design / get_design / list_designs / delete_design.

4. bad → re-dispatch (§11.1)

A run marked bad is no longer a satisfied cell: the sweep reset_runs it (clears executions/results, reactivates status, keeps the config row) and re-dispatches in full. reset_run (not drop_run) is used because the sweep pre-registers configs — an integration test caught the FK violation that full-drop caused. The wrong-inputs case needs no code: a corrected input changes the content hash and dispatches as a fresh run.

Testing

• Unit: new TestTargetDesigns, TestResetRun, TestBadRunRedispatch, attribute-currency tests; full suite green.
• Real-jar integration (test_provenance_integration.py): attribute currency, check_design over real runs, and a real bad→redispatch (drop+rerun verified fresh run_ids + active status + rows in cell_data_current).
• 0 net new ruff/mypy.

Docs

best-practices, project-organization (freeze regenerated with the real jar), architecture, llms-full.txt, objects.json + reference pages for the new classes. Version → 0.0.9.25.

🤖 Generated with Claude Code