FIX6 02 — Native dependency model (SUPERTRACK B) + v2-clone authority removal (SUPERTRACK C)

B — Native callgraph, regex demoted to diagnostic

Key catalog fact (verified): SELECT count(*) FROM pg_depend d JOIN pg_proc p1 ON p1.oid=d.objid AND d.classid='pg_proc'::regclass JOIN pg_proc p2 ON p2.oid=d.refobjid AND d.refclassid='pg_proc'::regclass WHERE p1.proname LIKE '%qt001%' = 0. PostgreSQL does not record function-body call edges, so a native callgraph is authoritative for VIEW dependencies only (5,638 real view edges via pg_depend/pg_rewrite).

Objects:

• v_qt001_native_dependency_edges_v6 — pg_depend/pg_rewrite view→relation/function edges (NORMAL/AUTO deptype).
• v_qt001_native_dependency_closure — recursive, no depth cap (DAG terminates), rooted at qt001 control views (audit scope), traversal follows edges to ANY object incl. non-qt001.
• v_qt001_callgraph_text_scan_diagnostic — wraps the old regex v_qt001_callgraph_edges_v2, flagged is_authority=false, diagnostic_only=true.
• v_qt001_callgraph_unclassified_fail_guard — apply roots (writer/driver/builder) are plpgsql/sql functions with catalog_call_edges=0 → pass=false (BLOCKING; cannot natively certify legacy-free). This is the honest structural answer to B1: native no-legacy proof is IMPOSSIBLE in 2.6A; resolution (drop legacy fns / planner-as-views so pg_depend captures edges) routed 2.6B.
• v_qt001_callgraph_no_regex_authority_guard — proves via native pg_depend closure that the 4 authoritative guards (readiness v9, hardcode v7, no_bypass v4, scorecard) do NOT reach any regex view (authoritative_guards_reaching_regex=0) AND the readiness registry points at no regex view (registry_rows_pointing_at_regex=0) → pass=true.

C — v2-clone detected + blocked (not re-authored)

• v_qt001_v2_logic_clone_detector — compares prosrc of plan_v5 vs plan_v2: normalized_identical=true (after stripping comments + the 'fact_source','PLAN_V5_NATIVE' labels + whitespace) AND all 5 signatures shared (EXECUTE format, ELIGIBLE_SYNTHETIC, apply_safe, legacy_id_single_colon, FROM %I t). clone_detected=true.
• v_qt001_current_planner_not_v2_clone_guard.pass=false (clone + performs apply-scale full scan) — BLOCKER feeding readiness; full-scan/keyset/perf declared NOT_SAFE, routed 2.6B (genuine re-authoring is 2.6B, not faked here).
• v_qt001_plan_facts — reads already-computed facts from qt001_plan_snapshot (PG-driven facts, no rescan) for control reads.

Codex-honesty: detecting a clone via source comparison is a fail-closed DETECTOR (conservative), not a source-text safety CERTIFIER — and it is corroborated by the catalog-level normalized-md5 equality.