10 - T1-Max Adversarial Review (stricter than XHigh)

Max posture (assumed worst case): a future implementer will misread the blueprint; any ambiguity causes production damage; any unlisted dependency will break; any hidden hardcode scale-fails; any missing rollback is costly. Max reviews the already-XHigh-revised blueprint.

#	check	verdict	evidence / finding
1	can a future T1 implement without guessing?	MAX_FINDING (MX-1)	PKG-B must author privilege_set_manifest #21 rows equal to "the existing Directus SELECT set", but that set is referenced, never captured. Without an explicit enumeration artifact, PKG-B guesses the read contract - a no-guess violation and a Directus-read-regression risk
2	can Codex review without reading whole history?	PASS	doc 00 carries status, invariants, sources, and the central refactor finding; each component cites its approved source doc
3	can operator run only allowed parts?	PASS	every package marks who = T1 vs OPERATOR; PKG-A..D have no production mutation
4	can a live apply be split safely later?	PASS	PKG-E..H split create/seal/activate / repoint+neutralize / owner-ACL cutover / freeze+deprecate, each independently gated and reversible
5	all live/refactor changes traceable to approved FIX7 design?	MAX_FINDING (MX-2)	the 27/11/14/7 model, byte-DDL, hashes, #26/#27 repoint are traceable to approved docs. But two operational dispositions - S15 legacy-entrypoint neutralization (REVOKE EXECUTE + fail-closed stub) and S17/S18 legacy freeze/deprecate - extend beyond the literal approved design text. They introduce no new authority surface/gate/hash, but must be explicitly flagged as T1-operationalization requiring Codex confirmation, not presented as already-approved design
6	all non-authority tables clearly non-authority?	PASS	11 runtime-evidence tables; G-RUNTIME-NONAUTH; never counted in 27
7	all authority surfaces exact-set counted?	PASS	27 named; G-AUTH-27 both-EXCEPT vs envelope
8	all future SQL actions staged, not live?	PASS	author/rehearsal/read-only for PKG-A..D; operator-gated for PKG-E..H
9	all destructive actions operator-gated?	PASS	seal/activate/repoint/REVOKE/cutover/freeze all OPERATOR
10	all "routed later" items currently blocking?	PASS	G-17 + apply/permit/REAL_RUN/2.6B blocked across all steps
11	can rollback restore safe-blocked state?	MAX_FINDING (MX-3, minor)	doc 05 requires the S16 prior-ownership+ACL snapshot, but does not require it to be captured, read-back-verified, and rollback-rehearsed before the REVOKE runs. A snapshot taken but not verified could fail to restore exact grants
12	does every package have a no-go condition?	PASS	PKG-A..I each carry an explicit no-go
13	is every gap resolved, planned, or blocked?	PASS	doc 03 disposition guarantee; 7 P0 / 9 P1 / 2 P2 / 0 open
14	is the blueprint ready for Codex independent critical review?	PASS-after-revision	yes once MX-1/MX-2/MX-3 are revised below

MAX_FINDINGs requiring revision

• MX-1 (P1, real - no-guess / read-path): add an explicit artifact "enumerate the current Directus SELECT grant set on business base tables" as an output of S00 (re-baseline) and PKG-D, consumed by PKG-B to author #21 rows. G-DIRECTUS-READ input updated to compare #21 against this captured set. Until captured, PKG-B is blocked.
• MX-2 (P2, traceability/honesty): mark S15 legacy-neutralization and S17/S18 freeze/deprecate as T1 operationalization beyond the literal approved design, consistent with the approved no-bypass discipline and #26/#27 rollback-stub mechanics, introducing no new authority/gate/hash, and explicitly request Codex confirmation of these dispositions in the critical review. Recorded in doc 02 and doc 12.
• MX-3 (minor): strengthen doc 05 / PKG-G precondition: the prior-ownership+ACL snapshot must be captured, read-back-verified, and its restore rehearsed before any REVOKE executes.

Max disposition

3 findings (MX-1, MX-2, MX-3). None alters the 27/11/14/7 invariants, the hard blocks, or the read-only/no-mutation posture. All are revised in doc 11; affected checks (1, 5, 11) are re-run there. After revision, the blueprint can withstand a misreading implementer and is ready for Codex independent critical review.