dot-iu-cutter v0.4 — PG-backed Dry-run RERUN: EXECUTION FAIL (honest)

Date: 2026-05-17 · execution_status: FAIL (honest). Rerun on the txn-fix commit 6060e1ae8b958fcb8a61ed45b597dc553b8688be. The transaction-lifecycle fix is validated (the previously-fatal S4 now passes with real writes); a new, distinct accepted-code defect was found at S6 CUT. Orchestrator STOPped at the failing gate, tore down the isolated env exact-name, did not improvise, did not mark PASS. Production untouched.

1. Progress vs prior FAIL (txn-lifecycle fix confirmed working)

Scenario	Result
S1 config load (no connect)	PASS
S2 missing key fail-closed	PASS
S3-neg current_user mismatch → ConfigMismatch	PASS
S4 MARK	PASS ← was the fatal failure before the fix; now writes real rows
SWEEP	PASS (promote + sweep_log)
S5 REVIEW	PASS (envelope + unit_block + review_decision + history)
S6 CUT	FAIL (new defect, §2)
S7 VERIFY … S12	not reached

SQL trace verbs observed: {SET, SELECT, INSERT, UPDATE} — real MARK/SWEEP/REVIEW writes occurred (proving commit 6060e1a resolved the RealPostgresAdapter find()/_begin transaction-lifecycle defect). No forbidden verb (no DELETE/TRUNCATE/DROP/ALTER/GRANT/CREATE/COPY).

2. New root cause — accepted-code: CUT/VERIFY signing-body json.dumps over psycopg3 UUID (NOT harness / NOT r3 / NOT txn-fix / NOT schema-binding-logic)

cutter_agent/phases.py cut() builds the DOT executor signing body:

body = json.dumps({"entry": entry_id, "manifest": manifest_id},
                  sort_keys=True).encode("utf-8")

manifest_id is obtained from self._reviews_for_entry(entry_id) → adapter.find("review_decision", …), i.e. read back from the real Postgres uuid column. psycopg3 returns SQL uuid as a Python uuid.UUID object, which json.dumps cannot serialize → TypeError: Object of type UUID is not JSON serializable. verify() has the same shape (json.dumps({"entry": entry_id, "manifest": cs["manifest_id"]})).

This was latent because the InMemory adapter stores/returns the ids as the str the ledger inserted, so json.dumps worked; the real psycopg3 driver returns uuid.UUID. It is a genuine accepted-code defect surfaced only by a real PostgreSQL connection — exactly the class of latent gap the PG-backed dry-run exists to catch (third such finding: structural schema mismatch → txn lifecycle → now UUID/JSON typing).

• Not a harness defect: the failure raises inside cutter_agent.phases.cut()'s json.dumps, accepted code, when fed real-psycopg UUID values; the harness only drove the accepted pipeline.
• Not an r3 defect: counting never reached.
• Not the txn-fix: 6060e1a is validated working (S4/SWEEP/S5 now pass with real writes).
• Not the schema-binding logic per se: the SB-DEC-1 real-schema lineage read is correct by design; it merely surfaced that cut()/verify() json.dumps DB-sourced UUIDs.

3. Required output

Field	Value
execution_status	FAIL (honest; accepted-code CUT/VERIFY UUID-JSON defect)
accepted commit	6060e1ae8b958fcb8a61ed45b597dc553b8688be (mounted RO; txn-fix validated)
dry-run env / sysid	pg-dry-run-v0.4-db-adapter-2026-05-17 / DR sysid 7640819868740706349
production sysid pre/post	7611578671664259111 / 7611578671664259111 (unchanged)
backup path + sha	…/.dryrun-v0.4-2026-05-17/prod-directus-20260517T112308Z.sql · sha256 e14089005192085b8766d64243e5e8de8fdb3897fc36059bb3517d04440dc3a3 · 680,066,849 B (read-only pg_dump -U workflow_admin; plain-format ⇒ per-run sha)
SQL/harness artefact shas	roles+matrix 2a409696dc3f60cb6328a77afd345e7638685f8d70cb5c0995b40f5841a57584 · harness ddf14a94438a6b8ed621329d2f3b62ca7da2b58724d6fd363136a0f1c8d3aa96 (sha-gated, unchanged across runs) · orchestrator iterated for env-readiness robustness (throwaway tooling)
psycopg3 install scope	ephemeral python:3.12-slim --rm container only (pip install "psycopg[binary]>=3.1,<3.4"); never host/incomex-*; gone with container
scenario S1→S12	S1/S2/S3-neg/S4-MARK/SWEEP/S5-REVIEW PASS; S6 CUT FAIL; S7…S12 not reached
12-table matrix	not asserted — fail at S6 before baseline checkpoint (rows written by MARK/SWEEP/REVIEW were discarded with the env teardown)
final 15-row baseline	not reached
negative Δ0	not reached
no forbidden SQL	✓ trace verbs {SET,SELECT,INSERT,UPDATE} only
no secret leakage	✓ openssl dry-run-only passwords; dr.env/dr.harness.env 0600 then shredded at teardown; orchestrator/harness never echo secrets
teardown	DR container+network removed exact-name (DR_GONE); secrets shredded
protected envs untouched	✓ all 3 (…v0.2-p0-2…, …v0.2-phase-alpha…, …hb05…) running, unchanged
production untouched	✓ only read-only pg_dump; prod postgres StartedAt 2026-04-17T05:35:18.48439927Z unchanged; prod sysid pre==post; no adapter/secret/.env/CUT/VERIFY
git SSOT	branch main; HEAD 6060e1ae8b958fcb8a61ed45b597dc553b8688be (unchanged); git status --short -- iu-cutter clean (throwaway WD excluded); no code change, no commit
hardcode/scale control	no literal IP/DSN/password/container/vector collection introduced; no label/key/metadata-schema change; dry-run config from dry-run-only env; no hardcode/label finding

4. Honest-FAIL discipline

Orchestrator STOPped at G_HARNESS_HAPPY, ran teardown of the exact DR env only, did not improvise (I did not patch phases.py/db_adapter.py to str() the ids, nor patch the harness to mask it — that is a code change forbidden in this dry-run cycle and would hide a real defect), reported honest FAIL, did not mark PASS. Two earlier interim STOPs this rerun were orchestrator-artefact robustness bugs (a stale set -u $PORTS fixed previously; the postgres-image bootstrap transient-ready race in the readiness gate — hardened to require the target DB to answer a real query twice consecutively); both with prod/protected untouched and the isolated env removed exact-name. Those are tooling false-negatives, distinct from this substantive S6 accepted-code defect, which is reported as a true FAIL and not worked around.

5. Remediation (NOT self-applied — needs GPT authorisation)

A separate GPT-gated code-authoring cycle. Candidate minimal fix (design to be reviewed): make the DOT signing-body construction UUID-safe — e.g. json.dumps({"entry": str(entry_id), "manifest": str(manifest_id)}, sort_keys=True) in phases.py cut() and verify(); or normalise uuid.UUID → str in RealPostgresAdapter.find() row mapping (db_adapter.py) so all DB-sourced ids are JSON/equality-stable. Add a real-psycopg (or UUID-returning fake) regression test so the unit suite catches it. Then re-author the PG-backed dry-run (command-review r1 / verification-plan r3 unchanged; binding still count-invariant). No self-advance.

6. Next gate

GPT review of this honest-FAIL report + adjudication of the CUT/VERIFY UUID-JSON remediation cycle. PG-backed dry-run remains blocked/failed until that fix lands under a separate GPT-gated cycle and is re-authorised. The txn-lifecycle fix (6060e1a) is independently validated by this run (S4/SWEEP/S5 PASS with real writes). No production action; no PASS claimed.