# T05 zlib85091 release — preliminary, unapproved

This is a distinct proposal/trial for existing INFD-WP-0009-T05. The actual T04 source2 acceptance does not authorize it. Previous R1 copy and R2 verify UNCERTAIN journal states remain untouched. INFD-IN-0010 remains an unpromoted general verifier residual; this new task-specific verifier is a prerequisite of active T05, not an amendment of the old approved helper.

Exact intended image f5588f496b1f267d07cdc9f5208e6f2f241f74b739c23924b03702dec13a5c23 was built by admitted owner CI11534/run4 from 9b4bb25d325698ae543e7e8a8a3c3592db1f58fc. Publication is not installed acceptance. Final packet requires independent published-image scan/runtime/filesystem/source proof and fresh read-only native projection/one-object dry-run evidence. Until all final inputs and independent owner acknowledgments are frozen, execution-plan state remains preliminary and the actuator refuses before any authority, journal or native action.

After a NEW exact confirmed founder acceptance, root freshly authenticates fixed-origin native source-version1 outcomes before each phase. Receipt files are supervised bookkeeping, not signed grants or autonomy. Wrong task, source, memo/version, stale/pending/agent disposition, old T04 acceptance or changed inputs refuse. Root retains exact approved source revision/digests and action results in supervision. Approval does not promote any agent.

Fixed phase order: drill-backup → drill-copy → offline-verify → cleanup-offline → cutover-backup → cutover-copy → release → verify. New private native /data/private/zlib85091-r1, root .private-backups/20261011-infd-zlib-r1 and local volume infd-zlib-r1-restore-20261011 are distinct from all T04 custody. New private journal intent is fsynced RUNNING before dispatch; no attempted phase replay. Fixed output schema contains only counts, hashes, object identity and boolean results; raw errors and private content are withheld. Existing native snapshots/database/sidecars are never removed or restored over current data.

Fresh backup inspection uses copied explicit-close SQLite helpers: live and offline clone readers are mode=ro/query_only and WAL-aware, before any Store/admin open. Immutable reader is used only for freshly closed hash-bound private snapshot with absent/zero WAL frames and stable identity/metadata/digest before/after read/copy. Main and sidecars remain retained. Schema must equal5, integrity ok, and exact snapshot/isolated clone counts match before/admin/after. The detached local clone has network none, readonly root, no service/runtime config/credential mounts; only narrowly scoped UID0 CAP_CHOWN initializer precedes UID10001 reads/admin. Original private copy is hash-stable. Cleanup removes only the disposable volume whose nonce/labels/creation metadata match this trial, before cutover.

Each snapshot is bounded512MiB. At most two new snapshots occupy each native/offcluster custody (1GiB payload each); the disposable offcluster clone adds512MiB while present, plus16MiB sidecar/metadata reserve. New native drill phase checks free space for both; local custody checks two snapshot bounds plus reserve. Existing T04 retention consumes additional pre-existing space. EUR0 recurring infrastructure estimate; human/compute time unpriced.

Cutover takes a fresh native backup and exact offcluster copy. Release then requires latest schema5 readonly proof, sealed Pod/Deployment/config/PVC custody and same baseline image763c/current Application sourcec96, a healthy single target92.205.62.239 node and fresh requested CPU ≤ allocatable capacity. Recreate and resources remain unchanged. Server dry-run/client diff must show only the image. Application UID/resourceVersion/source CAS dispatches manual exact final revision, initiatedBy custodian-codex, explicit prune=false and only apps/Deployment informed-decision in its namespace. No whole-Application sync, namespace/config/credential/resource change or pruning.

Verification reads imageID/readiness/owner chain/exact configured Pod fields, schema5/integrity/metadata counts, installed policy hash and synthetic purpose refusals, health200, installed fixed zlib and exact image proof. Real PDP-gated lifecycle readback is a separate root authenticated observation and not replaced by synthetic checks. Argo3.5.3 omits false prune in operation status: missing or exactfalse is accepted only together with this journal's COMPLETE explicitfalse request, exact source/resources/initiation, operation start after our durable attempt, bounded finish and one Synced Deployment result. True/null/string/prior/foreign/extra-resource results refuse.

Rollback → rollback-verify is separate and permitted only after a recorded release attempt. It switches only Application source back to c96 and Deployment image763c; it never restores an old snapshot over the latest database. Recovery dispatch works even if candidate Pod is absent/unready and never depends on candidate DB inspection. Recovery verification separately requires Ready1/old image/current schema5/data-custody and exact recorded scoped completion. All phases must begin before2026-10-12T23:59:59Z; release must begin at least30minutes earlier. Beyond expiry stop and obtain a new founder decision; no automatic extension or emergency grant. A failed/uncertain attempt requires informed recovery rather than retry.

T05 remains progress until actual installed acceptance and native rescan prove the finding absent. No Risk Engine grade is assigned. Only then may owner close T05/risk and consider finishing WP9, with existing residuals live first.
