07 — AI Review / Action Panel Update

AI review (real, not faked)

19 rows inserted into process_axis_ai_review (subject_kind='candidate', subject_code='WPC-*'), idempotent via ON CONFLICT (subject_kind,subject_code). Review state assigned by confidence of disposition, not rubber-stamped:

• AI_REVIEWED_PASS = 14 — candidates/components/OS/noise where the disposition is well-evidenced (rp_visibility ∈ CANDIDATE_VISIBLE, NOT_PROCESS).
• NEEDS_MORE_EVIDENCE = 5 — WPC-DOT-BIN-RECONCILE, WPC-DB-SCHED-JOB, WPC-DOCKER-RUNTIME, WPC-INFRA-CONFIG, WPC-UNCLASSIFIED (rp_visibility ∈ NEEDS_REVIEW, OWNER_GATED). These are NOT marked PASS — genuine uncertainty is preserved.

Each review carries evidence jsonb (member_count, process_type, classification, 5 sample object_keys), gate_status jsonb (owner/birth/governance/rp/canon), risk_level, recommended_action, evidence_hash, reviewer='ai_agent'. High-risk candidates (PROCESS_CANDIDATE) carry an irreversible_warning that official registration/birth/canon are owner+president gated.

Action panel (fail-closed, no checkbox)

New table wf_candidate_action_vocabulary (11 buttons, is_checkbox=false enforced by CHECK, backend_handler='unimplemented' by default). Buttons: Confirm Process Candidate, Merge, Split, Mark Not Process, Accept OS-Level, Quarantine/Ignore, Assign Owner (owner-gated), Request More Evidence, Reconcile to dot_tools (canon-mutating), Create Birth Request (president-gated), Send to Governance (president-gated).

v_process_axis_candidate_action_queue_v3 (59 rows / 17 candidates) exposes the applicable actions per candidate by classification — e.g. PROCESS_CANDIDATE gets Confirm/Merge/Split/AssignOwner/CreateBirthRequest/SendToGovernance; OS_LEVEL_INFRA gets only Accept OS-Level; BACKUP_OR_NOISE gets only Quarantine. Canon-mutating and president-gated actions are flagged but have no implemented handler — fail-closed by construction.

Companion views: v_process_axis_candidate_review_dashboard (NEEDS_MORE_EVIDENCE sorted first).