sausage/docs/SAU-384-runbook.md

One-off handover runbook — rehearsed; live approval required

Authority: design / Rollout, SAU-356 gates. Results: initial, races/rollback, native package, external driver rework, round-3 recovery/git gates, round-4 failure paths. This is a proposed live plan, not authorization. Workers never execute it. Mediator/orchestrator lands and executes only after Surma's go; every push separately needs a fresh go.

0. Prepare everything while the old world is still available

Do not begin downtime with scripts, caller decisions or launch instructions still to be assembled.

Reviewed selection

• Bundle: /docs/symphony/artifacts/SAU-384/cutover-targets-1389.tar.gz, SHA-256 31dbb95b796c1fedfd8c37a8b2354cb0977face8653345dc026cc1bbdefac18b. It includes final-main cred 0db40101, build 1636 / 274 tests, preserving approved SAU-418 instead of downgrading it. Git 9cc9fa6f1c9f01ce1a53d8042623bde60e026637, approved SAU-418 artifact /state/orchestrator/builds/git/9cc9fa6f1c9f01ce1a53d8042623bde60e026637.wasm (177 tests), is included as deferred_plugins/deferred_artifacts, not an offline replacement. It activates only after target cred is live AND Surma unlocks it in the UI. Earlier bundles are immutable historical evidence, not this selection. Verify all 23 payloads (including deferred git) against its embedded manifest.json. target-artifacts.tar.gz is older debug-host evidence, not this installation.
• Native host: /nix/store/spyd9rix6r38m6vxfppyy9z83yp5dh85-sausage-0.1.0, aarch64-linux Nix release package from unchanged host source e09c4f9f. The embedded native-host-manifest.json pins the NAR, binary and complete runtime closure. Import the reviewed NAR if needed; never overwrite immutable store files or substitute a diagnostic host. Import/deployment is an operator action, never a worker action.
• Old package for this observed baseline: /nix/store/mggvibk8lcgv2qw28jimrl60mi0qm41p-sausage-0.1.0. Re-identify it and the owner before go; observed PIDs are not constants.
• Actual owner: Herdr w7:p6, tab w7:t5, not a systemd Sausage unit. Its loop is /nix/store/pz2b94a2v99r3msljdz82h7n3i81hhh5-sausage-loop/bin/sausage-loop /home/surma/world/. Read herdr pane process-info --pane w7:p6 once to resolve current shell/loop/host identities. Do not close Herdr or unrelated panes.
• SDK copy list is src/lib.rs only. Preserve world-owned Cargo manifests, relocated interface, imports and all other SDK bytes. Resolve a changed baseline before proceeding.
• Web is approved commit 3d76026749a5d21722033bbfe8743ac368946b0f (SAU-375), held by publish:hold. Do not release it before target session and web are active.

External driver and resolved operator manifest

Use external root /home/surma/sau-384-cutover, outside the world and stopped process group. Stage the extracted bundle under targets/. Reconstruct only cutover_once.py, freeze.py, and driver.py under tool/ from the source-as-log headings in /logs/box/2026-10-31/tc-zg0j56-exec-3030415406.stderr; check their hashes against evidence/round4-1392/final/tool-deletion.json. These are the executed round-4 versions (source lines 2–189, 192–263, 266–669 respectively), not earlier scripts. Hashes/ranges are in evidence/round4-1392/final/source-log-roundtrip.json. They have a real CLI entry point. Do not reconstruct rehearsal/test controllers.

The tested interpreter is /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3, with inotify-simple 2.0.1. Resolve its closure availability before go. After go and before starting the driver, pin its package and the inspected Herdr package with host-visible GC roots for the operation (nix-store --realise <exact-package> --add-root <external-root>/gc-controller --indirect, and equivalently gc-herdr). These root writes/imports are included in the approved live plan, not preparation actions authorized to a worker. The driver itself registers and verifies old host, target host and loop roots at <root>/gc-{old,target,loop} before freeze. Use the host namespace, not a box talking to shared Nix state: /rework roots in a box are not host-visible. Do not run GC as a test. Keep old/target/loop roots and both selectors for recovery after tool deletion.

Start from evidence/round4-1392/live-manifest.template.json, a non-executable template, and resolve every placeholder before approval:

• native paths, Herdr binary/context, pane and current shell/loop/host PIDs;
• external root, world, bundle, embedded manifest SHA, offline artifact paths and deferred git, SDK path, exact target boot list; root and world are on the same filesystem in this plan; resolve a different storage topology before go;
• complete old in-flight /proc/*/info.json and native-owner inventory, plus uncertain external effects; a finished-looking assistant tail is not settlement;
• a new continuation UUID, generated once, and the explicit new mediator verification instruction. Keep that UUID on all retries;
• typed executable-caller review and its frozen guards below.

Do not copy sandbox PIDs, gc_store, or any test_* gate into a live manifest. The template deliberately omits those test-only fields. Record/hash the fully resolved operator.json; driver checkpoints bind to its exact bytes. Resolve a schema/selection discrepancy while the old world is available, not during downtime.

Executable caller review and frozen guards

Use predecessor formats: session source f9122585 Paths/queue; cron ba75c102 and current cron README one-off procedure. Do not substring-rewrite prompts or diagnostics. Review current stores as in resume-1373.md, then encode the approved file hashes and deterministic SQL projections in caller_file_guards and caller_sql_guards:

• cron state/cron/jobs.sqlite: use the template's ordered id, json_extract(data,...) projection of method, params_json, cron, at, timezone, detach, call_timeout_secs. Do not hash next_run, last_run, last_result_json or last_occurrence_id;
• board state/board/board.sqlite: ordered id,data for hooks/invariant_target; id,target,session,data for subscriptions; and select id,data from subscriptions union select subscription,data from subscription_batches order by 1,2 for executable subscription snapshots. Ignore last_sent/ended, transient batch IDs/lines/due; a pending delivery of the same active configuration does not change this union. Historical hook-attempt snapshots remain diagnostics and must not replay. Automatic final subscription removal changes the executable set; resolve/refresh before arm, or take the pre-start abort path if it happens after arm;
• boot and other discovered executable configuration/manager registrations. Review typed room/workflow/orchestrator stores too; preserve their histories/settings and non-executable records. Do not mistake prompt text for an RPC template.

A SQL guard hashes sha256(json.dumps(rows,separators=(',',':')).encode()), where rows = connection.execute(sql).fetchall() in the approved ordered projection. JSON strings inside rows retain their exact bytes. These projections, including all column names, live in the current template and were tested against mutable cron/board bookkeeping. Any genuinely new executable store/field still needs review; do not silently omit it. The observed live baseline had six cron templates to convert, already-target-valid board hook/invariant calls, no subscription batches and no live manager registrations. This is an observation, not a future frozen snapshot. Unknown executable references, session.work, session.cancel, ambiguous interrupt:true, invalid text or unexplained notify destinations stop preflight for explicit resolution.

Before start, again before arm, and once immediately before freeze, check/driver checks artifact and file guards, SQL projections on consistent live SQLite backup-API copies, special-file types (without hashing mutable bytes), space, and a dry run of the same cron converter used by plan. Object/text/interrupt refusals therefore occur before downtime, not just in plan. The dry run never rewrites the live DB. After freeze, guards run on disposable complete DB copies before apply; immutable backup evidence is never opened/recovered in SQLite.

Space is an allocated-block measurement, not a guarantee. Root/world must have the same filesystem device in this reviewed layout. Walk regular files, directories (including the root) and non-followed symlinks; charge each at the larger of POSIX st_blocks × 512 and its logical size rounded up to that filesystem's f_frsize. This includes sparse-file expansion by copy2, not just logical file bytes. Add the concrete staged plugin/prev/SDK allocations and the three existing DB-copy roles (recovery, staging, audit). Save source allocation, backup allocation budget, directory/file counts, fragment size and required/free bytes in preflight-live.json; refuse before freeze unless free covers that budget. The fresh sparse/tiny-file/directory fixture measured 12,288 source-allocated bytes and 4,599,808 copied bytes against a 4,603,904-byte budget (see evidence/round4-1392/allocated-space.json for actual measured values). No claim that this can never underestimate: metadata, DB normalization, growth, changed mounts and competing writers can still exhaust space. Resolve storage uncertainty in the approved maintenance plan, not with an invented margin.

Add any genuinely new executable store to the manifest before arm. From resolution until freeze: no caller/boot/SDK/plugin changes; routine cron next/last/result and subscription delivery bookkeeping is permitted. If new intent is needed, cancel unarmed or abort pre-start, re-resolve, and approve the changed manifest; never edit a hash-bound in-progress manifest.

No cron wake-up is required: the external driver publishes the explicit new mediator input. Do not rely on a timed wake-up for controller recovery. If a fallback wake-up is explicitly wanted, add/review it before resolving operator.json, not between resolution and freeze. It is useful only after the target is running and can accept new input. One due during downtime is skipped by cron; one firing immediately before freeze can enter an old pending queue that this approved cut deliberately drops. Neither is migrated/replayed. A guessed wall-clock time after the estimated backup duration is not readiness proof; use the driver's event/completion evidence and stable continuation instead.

Approval and external launch surface

Surma's go must explicitly cover this resolved manifest, the stop-all core/session/cred cut, the detached external controller launch/arm, pinned launcher installation and the post-start verification/release/deletion plan. This card's outside-world-controller scope requires a controller that survives the mediator; it does not authorize executing it now.

The proposed concrete execution surface is two finite foreground ssh.exec calls as surma on citadel.vpn.surma.technology, using the existing approved signing grant, with commands cd /home/surma/sau-384-cutover && <the start command below> and then the same prefix with <the arm command below>. Each command terminates; the reviewed native CLI, not a shell &, creates the detached OS child. Ask Surma's go to explicitly authorize this planned exception to the ssh-exec skill's usual prohibition on creating background work through exec strings. Until that explicit authorization, do not launch. Do not improvise a tunnel, SSH PTY, new worker session or an in-world controller. This is a fully prepared proposed live operation, not a claim that the current worker has permission or has tested a live SSH launch. The same finite foreground surface executes the final verify/delete command after target verification. After arm, loss of the old mediator/SSH RPC result is expected and does not terminate the already-detached controller; inspect the external checkpoint instead of retrying start.

1. Surma monitors; mediator starts/arms the external controller

Close ingress and record accepted first-update cuts before arm. Surma is the external human recovery operator while the world is down. He uses an independent Citadel terminal, never an in-world session/tool process. After mediator start acknowledges, Surma runs the foreground monitor below and leaves it visible. It binds a local Unix datagram socket and records the monitor/controller identities before arm. The controller waits for its own monitor on every start/resume; arm refuses without it. If an old monitor was killed, a new monitor removes its stale socket only after the saved exact process identity is dead; never unlink a live monitor's socket by hand.

The monitor receives inventory begin/completion and completed backup-entry progress without writing disk, plus CUTOVER FAILED and the notification outcome. It also waits on the controller's pidfd: normal exit, SIGKILL or crash prints CONTROLLER EXITED, even when there is no log/failure.json. Progress datagrams are best effort; the independent death event is not. No heartbeat/liveness/endpoint polling or backup-duration timeout. Read driver.json and inspect the pane once after exit; distinguish awaiting_verification from failure instead of blindly restarting.

Disk-backed failure.json, failure-notification.json and driver.log are best effort. Each may be unwritable at ENOSPC; none is a prerequisite for the failure signal or the Herdr notification attempt. One best-effort Herdr notification Cut-over needs Surma is submitted without first saving failure.json. Rehearsal submission returned shown:false/disabled; neither delivery nor a physical display is guaranteed. Keep the external terminal monitor attached: a silent tail of driver.log is not sufficient during backup. Whole-machine power loss can stop both controller and monitor and is not claimed covered.

From the external root, on the approved host execution surface, mediator first runs the finite preflight command, then starts:

PYTHONDONTWRITEBYTECODE=1 /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3 -B tool/driver.py check operator.json
PYTHONDONTWRITEBYTECODE=1 /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3 -B tool/driver.py start operator.json

start launches a new OS session with stdin closed and output in <root>/driver.log. A pipe completion handshake records the child's actual post-Nix-wrapper-exec identity, rather than racing an interpreter exec. It returns an acknowledgement and leaves the child waiting on inotify for its own monitor, then durable arm.json. There is no arm FIFO or reader-dependent write. Check that receipt/checkpoint once. Surma now runs in his independent terminal:

cd /home/surma/sau-384-cutover
PYTHONDONTWRITEBYTECODE=1 /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3 -B tool/driver.py monitor operator.json

Wait for MONITOR attached; arm may now proceed in that terminal. Then mediator runs the finite arm command:

PYTHONDONTWRITEBYTECODE=1 /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3 -B tool/driver.py arm operator.json

arm requires the current monitor, repeats live preflight and atomically writes a manifest-bound marker. Re-arming is idempotent while waiting with that monitor; without it arm refuses finitely, never blocks on a missing FIFO reader. A crash after observing the marker but before saving armed resumes from waiting_for_arm and reads the same marker.

From here the external driver, not the old mediator, owns steps 2–4. The old default run and its tool calls are deliberately cut. They are not recovered from legacy history. The driver uses an external advisory lock to exclude another controller. Rehearsal killed both the invoking caller shim and, separately, the detached controller after freeze; checkpoint resume completed the chain.

2. Driver: freeze → full backup → plan/audit → offline apply

The driver verifies reviewed bundle/plugin/SDK bytes and native closure hashes, registers real GC roots, and creates hash-guarded old and target selector/owner artifacts before freeze. It then:

1. Pins boot ID, start ticks, executable and group for both native processes; confirms controller is outside their group. Signals the outer loop first and awaits its pidfd exit. Saves owner_stopped.
2. Signals the exact host SIGTERM then SIGCONT, so a job stopped by background-PTY SIGTTIN can handle termination, and awaits host exit. Checks the saved group once for unexpected survivors; never blindly kills them. Saves freeze.json phase frozen, world path, boot ID and empty group inventory. No process/endpoint polling, sleeps or stop-request counting. Host shutdown stops all plugin/agent trees and suppresses restart admission. core.restart alone is not stop-all; after the loop is gone, exit 75 cannot respawn it.
3. Makes the complete frozen backup outside the world. Streaming SHA-256 does not read whole files into memory. Directory inventory and copy preserve empty directories and symlinks without following links. An interrupted plan resumes the same immutable input inventory and partially completed backup; it neither overwrites an earlier backup nor requires deleting the plan directory. Progress is sent to the external terminal monitor. A pre-plan abort may discard only that reproducible incomplete backup, as described below. Unknown images refuse.
4. Stages the approved transformations, then checks the frozen caller guards. Both plan and apply independently require the actual frozen checkpoint, same world/current boot, dead pinned owners, empty old group and no subsequently running host for this world. A caller-supplied {remaining:0} literal is not accepted. Changed checkpoint bytes also refuse.
5. Stages plugin.prev.wasm = that plugin's frozen pre-cut plugin.wasm for every offline-swapped plugin, then target plugin.wasm. Both prev/current writes have before/after hashes and participate in restore. It never retains the unrelated older prev as the cut-over fallback. Git is deliberately untouched here; its later standard activation will save its own preceding binary. Applies hash-guarded before/after operations offline. File replacements fsync temporary bytes, rename, and fsync the parent. A third image refuses. If atomic replacement dies before rename, abort removes only that saved operation’s new .cutover-tmp, absent from the frozen inventory and byte-prefix-matching its approved before/after image; unknown temporaries refuse. Resume overwrites/completes its own temporary. migration/plan.json reaches installed only after all operations complete.

Exact logical losses are only old session pending.jsonl, cancel, and unclaimed mail.deliveries rows with state='pending' and recipient not surma. Original messages, Surma copies, done rows and uncertain unknown deliveries stay; do not replay them. Loss records and original bytes/rows remain in backup/evidence. Histories/settings, scheduled SQL IDs/calendar/outcome fields, grants/requests and unrelated files remain intact.

Cron conversion: require an object and nonempty text, refuse interrupt:true, remove false/absent legacy interrupt, inner detach/notify/permanent ID, set absent kind follow-up, outer detach false. Preserve other fields and non-session JSON bytes. Review removed notify destinations; no invented answer-forwarder.

SQLite sidecars: copy the complete DB plus -journal/-wal/-shm family into immutable backup first. Recover only another disposable copy; use SQLite's backup API to stage a consistent DELETE-journal image. Plan removal of old sidecars before replacing the main DB, retaining the original physical family. This is normalization, not additional logical data loss. Tests covered committed WAL data, a genuinely hot uncommitted rollback journal, interruption and exact pre-start restore.

Full-world backup is not instantaneous: the 138,444-file / 12,441,162,164-byte / 45,310-directory size-bucket-matched synthetic replica took 70.55 s inventory + 135.39 s copy/hash verification = 205.94 s in the box. This includes repos/plugins/logs, unlike the earlier session-sized gate. It is warm-cache scale evidence, not an exact live-byte or cold-storage downtime guarantee. Allow that uncertainty in the approved maintenance plan; do not invent a timeout and kill the controller mid-copy. Power-loss recovery is not claimed.

3. Driver: ordered boot installation and pinned owner selection

The offline target boot list is:

[
  {"method":"session.manager","params":{}},
  {"method":"web.start","params":{}},
  {"method":"cred.serve","params":{}},
  {"method":"orchestrator.start","params":{}},
  {"method":"cron.start","params":{}}
]

Manager precedes web, so web-chat becomes best-effort. No mail.start deliverer remains. Resolve any extra actual boot entry before approval rather than deleting it accidentally.

The permanent <root>/target-owner exports SAUSAGE_FLAKE=path:<root>/target-selector itself, then execs the inspected loop with the exact world path. The selector's sole app is the pinned target package. <root>/old-owner similarly selects the explicit old package, never the remote default. These are permanent installation/launcher artifacts, not migration code. Their bytes/hashes are retained as manifest evidence; do not delete them with tool/.

Every later pane/cold restart uses /home/surma/sau-384-cutover/target-owner, with no caller environment required. Never invoke bare sausage-loop or the mutable gitea default for this world. Keep this command in the operator's pane-restart procedure. Rehearsal cold-launched it in the same Herdr shell without ambient SAUSAGE_FLAKE and observed the target executable. Earlier cold rollback exercised the explicit old wrapper against a separate old-backup clone; pre-start abort gates also exercised it against the exactly restored original frozen fixture, never over new target work.

core.restart now restarts the pinned reviewed host; it does not self-update from gitea. Future host updates require a separately approved package/closure/root/selector change. A remote push alone must never silently choose an old host against target state. Keep both GC roots through the rollback window; retire the old root only after an explicit decision, not tool deletion.

Expected approved live Nix side effect: each controller run makes three nix-store --query --roots calls, one each for old package, target package and loop, immediately after rooting those paths and before freeze/resume of offline work. Every resume repeats all three queries and their possible pruning, even if the roots already exist. Monitor/check/arm/continue/verify alone do not make those roots queries. Those queries can prune stale auto-root registrations and stale temporary-root files in Nix state, even though they do not delete store packages. Surma's go must include that side effect (and the five root registrations including controller/Herdr); the earlier shared-store incident is disclosed in rework-1381.md. Rehearsal now uses only a private Nix DB; no live root query was rerun.

4. Driver: started → launch in the same pane → fresh verifier

The driver checks the pinned shell still owns the selected pane and no unexpected foreground process exists. It checkpoints started before any launch attempt, then submits the permanent target-owner path once via herdr pane run to the manifest's pane (live w7:p6). No shell reconstruction or in-world mediator action is needed during downtime.

Readiness is event driven: arm a kernel inotify watch on state/session/manager/registrations, queue one session.subscribe command in the console, and await its unique registration file publication. The wait has one absolute 120 s deadline, from cron src/types.rs AddParams' existing registration/publication budget (also the earlier test_driver completion bound); it is not a periodic scan. Herdr/RPC calls use that existing finite bound. Missing registration or a failed post-start cred.status/subscribe/send produces the disk-independent failure/death signal and exits, keeping started. Never resume, restore or launch again automatically. Herdr wait-output is internally polling and is not used.

Forward completion after a post-start failure (Surma's independent terminal)

1. Read checkpoint/monitor and inspect Herdr pane/native identities once. If target is not actually running, preserve all new work and obtain a narrow forward repair plan; this mode will refuse and does not launch it. Do not relaunch blindly, restore an old snapshot or hand-edit the checkpoint. If only registration/publication failed but exactly one target is running, use:

cd /home/surma/sau-384-cutover
PY=/nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3
PYTHONDONTWRITEBYTECODE=1 "$PY" -B tool/driver.py continue operator.json

2. continue exclusively locks, binds the unchanged manifest, inspects that exact shell/pane, exactly one target host for this world and all offline target plugin hashes. It never freezes, applies, restores or launches. It then performs the skipped exact RPCs, through operator.json's gateway: cred.status {} must report locked and is durably saved as driver.json.locked_cred_status/locked_cred:true before Surma unlocks; session.subscribe uses id cutover:<continuation_id>, session cutover-readiness, method web.session_event, best_effort true; session.send uses session default, the same continuation_id, kind follow-up and exact continuation_text. It saves the receipt and reaches awaiting_verification. A recorded locked-startup proof is reused if a later retry occurs after UI unlock; absent that proof, unlocking first is a discrepancy, not permission to fake the check.
3. If a method/reply fails again, keep target work, inspect the error and fix only that cause; repeat the same continue command, not migration/start. Stable subscribe/message IDs make this safe even when send committed and only its reply was lost. Both started and awaiting_verification are accepted. Tests injected each RPC failure and committed-then-lost send acknowledgement; one continuation run-start/consumption remained. The real 120 s registration timeout also recovered through this command without a second owner.
4. verify still requires awaiting_verification. Run continue first when phase is started; there is no direct started→verified override. The fresh mediator can then verify/ask for unlock/release as §5; if locked cred prevents SSH signing, Surma performs the same approved local external commands. Live SSH/new-cred signing is not claimed tested. No hand-edited checkpoint or new continuation UUID is needed.

Finally the driver publishes the manifest's new explicit default verification message with its stable UUID and records awaiting_verification. The manager, not the driver, runs it. This names the new mediator as verifier without claiming the cut default run was recovered. The test produced one run-start and mailbox-consumed pair for that message. Publication is not immediate model execution: in the last gate it waited 23.753 s to run-start / 24.530 s to consumption behind initial reconciliation. Do not confuse that with publication latency or a completed startup scan.

No automatic relaunch, whole-world restore or reapply is allowed after started. If launch/acknowledgement is lost, inspect the pane/checkpoint once before any repair. Do not launch a second owner blindly.

5. New mediator: locked proof → Surma unlock → git → release web → delete tool

The explicit new verification run reads this runbook and in-world ownership/cred state. While cred is locked, Surma checks the external manifest/checkpoint/migration evidence in his independent terminal; the new mediator can read it through approved SSH only after UI unlock if signing works. Surma can also perform the final local verify command, so no vault-key transfer or untested signer is required to finish. Keep ingress closed until checks pass:

• one target manager and one runner ownership lock per running session; no old owners, mail deliverer or old orchestrator service;
• cred locked, no vault/key transfer; grants/requests/settings preserved. Interrupted callers reuse their existing stable credential-request keys, not automatic retries of secret-backed mutations;
• explicit target run-start without run-end is the only automatic cut-run recovery; old finished-looking history alone is neither settlement nor permission to resume;
• durable event/mailbox catch-up and same-ID default/advisor POST replay evidence holds; only intended loss/conversion/schema/new-activity differences exist. Surface uncertain mail/tool outcomes;
• scheduled jobs and non-target state match the inventory; real new mediator/runner activity is distinguished from migration changes;
• design effective grace 5 s and separate EOF-stop 5 s remain; crash supervision is enabled. Deliberate core.stop/kill disables service registration and is not a crash-supervision test.

Prior post-record live-size samples remain manager ack 0.6078 s, registration catch-up 0.0128 s, mail.send-to-default publication 0.5376 s (integrated 0.5235 s), below existing orchestrator 3 s / cred 10 s caller budgets. Disclose integrated catch-up 12.0419 s and the new initial-scan delay above. They are synthetic/warm-cache observations, not production latency guarantees. Prior real-core watch/upgrade/all-watch readiness, ~0.10% whole-host idle duty, detached error data and real cron warning/stop evidence remain linked in the result docs.

Before any Git network work, record locked-startup proof, ask Surma to unlock the target vault through the existing UI, then check cred.status reports unlocked for that same target service. Wait here if Surma is unavailable; never transfer a key/password. Only after that, mediator/orchestrator activates the reviewed git artifact:

{"method":"core.activate","params":{"plugin":"git","wasm":"/docs/symphony/artifacts/SAU-384/git-9cc9fa6f.wasm"}}

This is a proposed live operator step covered by the go, not a worker action. The permanent world artifact above must be staged/hash-checked from the selected bundle before go; the driver preflights its external deferred_artifacts.git copy too. Verify active git hash equals manifest files["git.wasm"].sha256 before releasing Git callers. See /plugins/git/source/{VAULT-HTTPS.md,VAULT-SSH.md}; both require cred first, live and unlocked. No push is part of verification; every later push still needs a separate fresh Surma go. Final proof includes git_activated_after_unlock.

Interim compatibility: observed live Git f089b56d still speaks new cred's explicit-grant fetch/item_get/public_key/sign contracts. Actual old-WASM + new-cred loopback SSH clone signed/authenticated successfully with a synthetic explicit grant after unlock; locked use returned vault_locked. Granted HTTPS reached new cred, reserved exactly one use, then the deliberately closed loopback port returned network_error with a capture (no successful HTTPS transfer claimed). Old grant-free shared SSH fails asking for a local key; old grant-free HTTPS retains its old config/anonymous path, not new anyone_may_use selection. Therefore do not schedule Git network work during the locked/old-git interval, do not introduce local key/config-secret workarounds, and surface any already-started caller's lock/missing-key error rather than blindly retrying it. Local git status/commit/authoritative-main merges remain available. New Git shared SSH clone was tested after real synthetic unlock/activation.

Only now use orchestrator.release for SAU-375 to release its approved commit. Inspect that method's current typed arguments; do not invent them. Verify published asset hashes and ingress. No worker publishes anything. The rework tested approved asset hashes in temporary output, not an actual live release or physical browser.

The Surma unlock above is a prerequisite for using the existing approved signing grant for the final external SSH verification/deletion command (live signing interoperability is still unverified); it is not a second vault handover. Never ask for Surma's password or transfer the old vault key. Read-only gateway checks and checkpoint inspection do not justify bypassing the lock. If unlock is not available, remain awaiting_verification, keep git activation/ingress/release gated, and do not reconstruct or rerun the migration.

Write external verification.json with evidence links and the required true checks named by the driver's verify mode: one_manager, no_old_owner, locked_cred, non_target_preserved, losses_exact, catchup_and_replay, web_released_and_hashes, ingress_verified, git_activated_after_unlock. They are operator attestations after the checks, not substitutes for them. From the external root, mediator runs the finite command below after target cred is unlocked and signing works; Surma may instead run the identical command in his independent Citadel terminal, avoiding any dependency on in-world SSH signing:

PYTHONDONTWRITEBYTECODE=1 /nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3 -B tool/driver.py verify operator.json verification.json 2>> tool-source-and-delete.log

It marks the saved plan/driver verified, logs source hashes/source-as-log, and deletes the three migration executables and bytecode caches. Confirm tool-deletion.json and absence once. Delete any other reconstructed executable/controller copies too. Retain only manifests, backup/inventory/loss/caller-audit records and logs as migration evidence, plus the separately reviewed target/permanent launcher artifacts. Rehearsal deleted all executed migration/test controllers before exporting evidence. If verification/deletion is interrupted, inspect the verified checkpoint and delete only remaining hash-matched tool copies; do not reconstruct and rerun migration effects.

Failure/rollback boundaries

Boundary	Exact recovery rule
Waiting for arm, including consumed-marker/save race	World untouched. Repeated arm is finite even with no reader. After the saved controller exits, resume accepts waiting_for_arm and the durable marker; attach a new monitor for that replacement before it continues. Or run abort (untouched) after obtaining the external controller lock.
Armed/captured/owner_stopped/frozen; no target attempt	External operator can use ... driver.py resume operator.json only after the saved controller has exited. It detaches a replacement and returns; Surma then runs the same foreground monitor command. The replacement waits for its own monitor receipt before effects. Same boot/identities/manifest/backup/plan must pass. No agent needs to exist. Both caller-killed and controller-killed resume were rehearsed.
Interrupted planning/swap, no start attempt	Resume the same immutable backup/plan with before/after hashes. A caller/config after-image is accepted only if that exact saved operation approved it; third images refuse. Alternatively Surma runs the real driver's abort mode below: it restores every approved changed path from the immutable complete backup, verifies the full frozen inventory, and launches the explicit old-owner in the selected pane. It never restores an incomplete backup. If planning never completed, apply cannot have started: abort first discards only migration/backup (including its partial copy temporaries), retaining planning/inventory/log evidence, then checks space before any checkpoint save. If enough space is available, it verifies any saved inventory and relaunches the untouched old world. If a competing writer still fills the disk it refuses before writes/launch; Surma must free only identified, separately approved non-world data and repeat abort. Never delete a complete backup/plan or unrelated world data to make space. Restore is before/after-hash guarded, resumable, and writes main DBs before their old sidecars without opening them. Never partially start mixed contracts. A frozen audit discrepancy is an offline stop, not permission to improvise a conversion.
started or later, even if boot seems failed	No automatic whole-world rollback/reapply/relaunch. New histories, mail, jobs, grants or external effects may exist. Preserve them. Use the forward-only continue command above for readiness/post-start RPC failure with one healthy target; otherwise plan a narrow forward repair or coordinated revert.
Any plugin core.revert	Offline-swapped prev now means exactly pre-cut current (hash-tested), not pre-pre-cut. Do not use blind individual core.revert across incompatible contracts. Coordinate providers/callers and preserve new work; the session prerequisite below still applies. Later git activation uses normal core prev semantics.
Session revert after target activity	Close gateway/ingress. While new manager is available, unsubscribe web-chat, verify removal, and prevent gateway re-registration while the old session binary runs. Never call web.start against that old manager or restore a registry backup that reintroduces it. Actual old-manager/gateway tests proved unsubscribe safe and ID-only re-key unsafe: the old binary ignores best_effort and replays history once.
Host reboot / changed PID identity / unexpected survivor	Refuse saved native identities; re-inventory externally. Never signal a reused PID, blindly kill a group, or infer quiescence from a stop request.

Exact external failure commands (Surma, not an in-world mediator)

In the independent Citadel terminal, cd /home/surma/sau-384-cutover; read driver.json, monitor output, any available failure.json/driver.log, and the pane once. At ENOSPC those disk-backed diagnostics may be absent or truncated; the terminal failure/death signal is authoritative that the controller ended, not that rollback is safe. Commands use the pinned interpreter:

PY=/nix/store/fixvq6n42pmvy49yh6kfmpax1s3j8md2-python3-3.14.7-env/bin/python3
# Same manifest/boot/owners, known exited controller, pre-start only:
PYTHONDONTWRITEBYTECODE=1 "$PY" -B tool/driver.py resume operator.json
# After resume acknowledges, attach its new foreground monitor before effects:
PYTHONDONTWRITEBYTECODE=1 "$PY" -B tool/driver.py monitor operator.json
# Or abandon pre-start attempt, restore coherent old bytes, submit old-owner once:
PYTHONDONTWRITEBYTECODE=1 "$PY" -B tool/driver.py abort operator.json

Choose one, not both. Both use a nonblocking exclusive controller lock; do not delete the lock file or bypass it. Abort accepts waiting/armed/frozen/installed/restoring/restored only. It checkpoints old_start_attempted before pane launch, then old_launched. If interrupted during restoring/restored, repeat abort on the same manifest/checkpoints; if old_start_attempted/old_launched, inspect the pane instead of repeating it. Abort is forbidden after any target started attempt. Never call Kit.rollback by hand or relaunch a bare loop. If the pinned shell/pane changed, stop for explicit identity resolution; do not guess a replacement.

After abort, Surma verifies exactly one old package owner in the pane and old world readiness before retry. Preserve the frozen evidence; the sole permitted deletion of backup evidence before verified success is the reproducible incomplete pre-plan backup subtree in the guarded abort path above. Complete backups are never deleted by abort. Every abort, including one with a completed plan, checks space before its first checkpoint save; a full disk is a deliberate pre-write refusal, not a failed checkpoint write. Abort rechecks the same measured space budget again immediately before old launch; if it refuses, free only identified/approved competing non-world data, not histories, jobs or complete backup evidence, then repeat abort on the same checkpoint. For a clean retry only after old_launched and verification, archive checkpoints (not owners/selectors/GC roots/tool):

# Pick a fresh, unused archive path and record it; this example is this run.
ARCHIVE=attempt-before-retry
PYTHONDONTWRITEBYTECODE=1 "$PY" -B -c 'import json; assert json.load(open("driver.json"))["phase"] == "old_launched"'
mkdir -- "$ARCHIVE"
for p in driver.json freeze.json arm.json controller-process.json migration driver.log failure.json failure-notification.json preflight-live.json monitor.json poststart-inspection.json; do
  if [ -e "$p" ]; then mv -- "$p" "$ARCHIVE/"; fi
done
cp -- operator.json "$ARCHIVE/operator.json"

Re-identify the now-running old loop/host and pane shell; refresh old-work/caller guards (including any wake-up) and approval for the new resolved operator.json, retaining the stable continuation UUID if no target start was attempted. Then check/start/arm as §1. Do not merely edit a live checkpoint or delete the backup. For aborted_untouched, no old relaunch is necessary; archive only after verifying the original owners are still running. No such reset/archive is allowed for started/awaiting_verification/verified or unknown owners: retain new work and plan a forward repair.

Independent plugin recovery/accounting/provider-cancellation follow-ups remain independent. Accepted first-update cuts and uncertain external effects are not silently solved by this kit, grace=5, or a successful rehearsal.