IU Core MVP Sandbox — 07 Hardening Patches

Date: 2026-05-21 · Branch 6. Patches authored from sandbox findings.

H1 — seed anchor selection + rollback coupling (FIXED)

Found: the first 010_sandbox_probe_seed.sql selected child/sibling with WHERE id<>v_parent and OFFSET 1/2, double-skipping rows (it picked the 1st/3rd/~5th IU, not 1st/2nd/3rd). The companion rollback recomputed anchors with a different ORDER BY … OFFSET, so it could orphan the depth-1 iu_tree_path row (iu_tree_path has no created_by to match on).

Fix:

• Seed uses one explicit row_number() OVER (ORDER BY canonical_address, id) ranking — ranks 1/2/3 = parent/child/sibling.
• Rollback derives the exact probe anchor IUs from the probe's own iu_structure_operation row (parent_unit_id + unnest(target_unit_ids)), deleting iu_tree_path rows by unit_id before the operation row is removed. No re-derivation of an ordering; cleanup is data-driven.

Verified: the fixed rollback cleaned the buggy seed perfectly; the fixed seed re-applied; re-running is a proven no-op.

H2 — psql -1 conflicts with self-transacted files (FIXED in docs)

Found: every forward/rollback file already wraps itself in BEGIN; … COMMIT;. The apply form recommended by authoring report 10 (psql -1 -f) adds a redundant outer transaction → two misleading warnings per file (there is already a transaction in progress / there is no transaction in progress). Atomicity is unaffected.

Fix: sql/iu-core/README.md gains an explicit Apply command section — use psql -v ON_ERROR_STOP=1 -f and do NOT add -1. The next production approval package (report 10 of this macro) drops -1 from its command block.

H3 — sandbox lacks event tables (RESIDUAL, documented; not a code defect)

Found: restore_test has no event_outbox / event_type_registry, so the gate-OPEN path of fn_iu_emit_event (an actual event_outbox insert) could not be exercised. The gate-CLOSED fail-closed path was proven (verification C5). The production directus DB does have both tables (confirmed read-only), so the dependency is satisfied at the real target. Recommendation carried to the next macro: exercise the gate-OPEN path either on a fuller restore that includes the event tables, or as a post-apply step on the production target after sovereign approval.

Repo changes (commit 4ce2667)

File	Change
sql/iu-core/sandbox/010_sandbox_probe_seed.sql	NEW — additive probe seed
sql/iu-core/sandbox/010_sandbox_probe_seed.rollback.sql	NEW — exact cleanup
sql/iu-core/sandbox/020_sandbox_verification.sql	NEW — non-mutating verification
sql/iu-core/sandbox/030_drill_reparent_behaviour.sql	NEW — non-mutating tree drill
sql/iu-core/README.md	apply-command note + sandbox/ section
tests/test_iu_core_ddl.py	+6 TestSandboxPackage tests

Not changed — and why

The DDL (001..005) and rollback (rollback/001..005) files needed no patch: they applied cleanly, every constraint/guard behaved as designed, the row-loss guards refused correctly, and forward DDL is idempotent. The cutter_agent/iu_core/ helpers were unchanged (covered by the existing + new static tests). No production mutation, no trigger attached, no deploy.