03 — Final Clone Executor Rehearsal

Date: 2026-06-03. Objective B. The headline technical result.

3.1 New vs prior revalidation

The release package's revalidate.sql deliberately did NOT reset the golden clone (it argued the clone IS the from-scratch result). This macro proves the unproven bit: the consolidated 99_run_all executor run top-to-bottom against a true greenfield reaches the finalized state in one shot — on a fresh clone of current production, not the 2026-06-02 snapshot.

3.2 Harness (safety + faithfulness)

• Golden clone untouched; re-snapshotted to /tmp/clone_finalized_2026-06-03.dump (95.6 MB) as a net + restored reset path.
• Fresh greenfield clone directus_gov_rehearsal_20260603 via streaming pg_dump --no-owner --no-acl directus | psql (+1.2 GB; 20 GB free after). Greenfield baseline: own=0, gap=210, conflict=0, ospa=0, inventory=35, scopes=6, axis=0, entity_labels=820,183 (exact prod copy).
• Clone-variant executor = canonical tiers in 99_run_all order with exactly two clone-only deltas: (1) gate simulation — one os_proposal_approvals row (ospa 0→1) on the clone only; (2) all 7 db-name guards redirected 'directus'→'directus_gov_rehearsal_20260603' (0 bare guards remain). No DDL/seed/assertion altered.

3.3 Run result — END-TO-END PASS (exit 0, all 8 tiers)

• T0 Preflight PASS: ospa=1 gap=210 own=0 conflict=0 emit=0 axis=absent prereqs=present.
• T1 Structure: 11 tables + 31 views + 1 function.
• T2 Responsibility: required=210 generalized gap=210 == canonical=210 (zero drift) (+1 axis,+6 val,+1 rule).
• T3 Ownership seed: new=30 gap=0 conflict=0 → own 0→30, gap 210→0, gov_emit=0.
• T4 Containment: required=35 gap=0 (+1 axis,+5 val,+1 rule,+35 assign,+35 own).
• T5 Topic: divergent=0 projection_gap=0 (topic axis +8 val,+7 map,+6 assign, 5 auto-decisions).
• T6 Scanner: scan_id=1, critical=0; fail-closed (emittable=0, outbox=0).
• T8 Verify: gap=0 conflict=0 containment=0 divergent=0 projection=0 critical=0 fail-closed.

Tier-8 table: resp_gap=0, conflict=0, resp_axis 210/0, containment 35/0, active_axes=3, divergent=0, projection_gap=0, label_orphans=0, open_drift=14, critical_drift=0, emittable=0, gov_outbox=0, kg-domain-x=active, reconcile_decisions=5. Drift breakdown (14, all info/warning, 0 critical): coverage_rule_missing 1, projection_lag 1, projection_lag_timing 4, topic_without_label 2, ui_visible_unapproved 6.

3.4 Reproducibility vs golden clone — divergences all explained

Every asserted invariant matched exactly. Non-asserted counts differ, all expected:

Count	Golden	Rehearsal	Why
axis_object_ownership	40	35	golden +5 from earlier substrate experiments; canonical is clean 35
topic axis_value	9	8	golden +1 from promote/merge/split experiments
reconcile_decisions	6	5	Part-C auto-derived from live views; current prod yields 1 quarantine+4 orthogonal+0 projection-pending
open_drift	38	14	detectors reflect live entity_labels/taxonomy; critical=0 in both

This is stronger than byte-equality: it proves the executor is environment-adaptive, not clone-coupled.

3.5 Guard + reversal proofs

• Re-run guard: re-applying to the finalized clone → exit 3, PREFLIGHT ABORT: axis substrate already present (6 tables). Refuses second apply.
• Rollback #1: exit 0, ownership=0 gap=210 axis=absent — baseline restored; post-read own=0/gap=210/axis=0/v_axis% views=0.
• Rollback #2: exit 0 — idempotent. Clone then dropped (disk back to 21 GB); golden clone intact.

3.6 Production isolation

Read-only throughout; post-rehearsal prod re-probe == entry baseline (ospa=0, own=0, gap=210, conflict=0, gov_emit=0, inventory=35, axis=0, idle=0, system_issues=204,556). Zero production mutation.

Verdict: Objective B PASS.