akyldash/docs/operations/CURRENT_STATE.md

# Infrastructure current state

## Windows ACL SID resolution — 2026-11-03

- Read-only PowerShell 5.1 inspection found that display-name conversion of
  ACL entries can throw `IdentityNotMappedException` even though the ACL itself
  is readable. The installer and external-receiver guards now enumerate access
  rules directly as `SecurityIdentifier` values and read owners as SIDs; unknown
  principals remain visible to the existing fail-closed policy. Regression
  coverage exercises synthetic unmapped SIDs through the actual installer guard
  functions. These repository changes are pending review; no Windows ACL,
  installer, task, VPS, or production state was changed. The native PowerShell
  harness still requires execution on Windows before this behavior is confirmed
  in that runtime.

## External deployment runtime binding — 2026-11-03

- PR #59 is merged to `main`; it contains the root-only read-only status
  endpoint and exact-literal SSH dispatcher for `status` and `deploy`. The
  status path bounds the Gitea run list to ten entries, verifies the selected
  run/workflow/current `main`,
  validates artifact metadata and ZIP bytes, and returns only a release tuple.
  This unmerged change adds a two-SHA compatibility gate: the fixed root helper
  compares protected Compose/Caddy bytes and installed release policy with
  Gitea content pinned to the release SHA, verifying each Git blob hash and
  size. This permits application-only release B after bootstrap A without
  advancing the root checkout or executing new root code. A runtime config or
  policy change fails closed and requires a reviewed bootstrap update. The
  Windows response state is `prepared`, never `candidate`; the `--once` task
  stores a protected local tuple and cannot call deploy. A separate disabled
  auto-apply task and config gate require the tuple to match a fresh status
  response from a previous poll. Status and apply retain current-main,
  successful CI/release provenance, descriptor digest, external backup,
  cutover, canary, and rollback checks. The receiver repeats the config check
  before journaling/stopping services and again after backup immediately before
  switching refs.
- Windows client предназначен для подготовки и auto-apply; обе задачи по умолчанию
  выключены, целевой путь фиксирован: `D:\Dev\akyldash`, SSH alias —
  `akyldash_release`. Примеры config и Task Scheduler XML также выключены; задачи
  запускаются от интерактивного пользователя с минимальными правами, не допускают
  параллельных запусков и не будят компьютер. Windows task, SSH key, VPS command,
  helper и trigger не устанавливались и не запускались. Путь canonical
  `C:\ProgramData\Akyldash\release-trigger.json` известен, но содержимое и
  enabled-состояние не читались. Метаданные Windows подтверждают
  `C:\Python313\python.exe`, действительную Authenticode-подпись Python Software
  Foundation и защищённые ACL интерпретатора.
- Linux fixtures прошли для Python import и command-path poisoning, включая очистку
  унаследованных `GIT_*` и фиксацию путей `ssh.exe`/`git.exe`/`powershell.exe`.
  Installer получает ограниченный bundle из точной вершины `main`: упаковщик
  читает Git objects commit, tree и allowlist-файлов; elevated installer проверяет
  SHA-256 ZIP и цепочку `commit → tree → файлы` из одного read-locked потока до
  записи и не запускает Git. ACL проверяются явными масками мутаций и подмены;
  текущая учётная запись не может владеть кодовыми путями, а право создать
  соседнюю папку не считается подменой предка. Windows ACL/install проверки пока
  только прошли код-ревью; installer и launcher на Windows не запускались.
  Production CD остаётся выключенным.
- Remaining owner prerequisites are installation of the reviewed root helper,
  exact forced-command/SSH host-key setup, a protected Windows client and its
  private state/config ACLs, private D: backup ACLs, and all previously listed
  backup/restore and production readiness gates. Neither the status endpoint nor client advances
  the host checkout or updates root helper code.

- Read-only Gitea check for fresh `main` SHA
  `f19faf628e67015febac22d0ec93ce4782caf4bf`: CI run #706 succeeded; release
  run #708 failed at the explicit registry credential-presence check because
  `REGISTRY_USERNAME` and `REGISTRY_PASSWORD` are absent; handoff run #709 was
  skipped. No secret values were read. Image publication remains owner-blocked.

Owner decision recorded 2026-10-26: a complete backup must not be stored on
the VPS, including temporarily. The local archive/full cold-backup commands in
`BACKUP.md` are superseded and unauthorized. PR #59 now contains a proposed
Windows-controlled external stream receiver and root-side deploy/rollback
code. The Windows rerun passed 12 synthetic client checks and 8 protocol
checks, plus a disposable real-NTFS fixture for the private owner ACL. The
production D: path, SSH connection, VPS installation, full backup, and
separate-target restore remain unverified. No verified external backup or
separate-target restore exists, so production deployment remains blocked by
the backup gate.

## Code-only deployment work — 2026-10-31

- Work is being prepared in the existing PR #59 branch from the then-current
  `main`; no merge, VPS operation, live release, or workflow dispatch was
  performed for this work.
- The release descriptor now binds both the release run and its successful CI
  source run. Host-side source verification uses the installed `gitea-api`
  helper contract; the PR carries the inspected helper fixture plus a bounded
  binary-download patch. The installer verifies the fixture hash before
  applying the patch.
- Backup bytes are streamed through a bounded SSH protocol to the fixed
  Windows destination. Root release state stores only a secret-free backup
  identity. The rollback handshake requires the exact previously pinned
  archive and creates a fresh external backup before image references change.
  Local VPS archive paths fail closed.
- Review fixes in this PR require Gitea `status=completed` with
  `conclusion=success`, resolve `main` from `/opt/akyldash/current`, write a
  durable pending journal before stopping services, and reject an already
  applied release before downtime. The Windows client checks D: capacity,
  NTFS/path/ACL owner and allowed entries, and the local ignore rule before
  SSH; it requires 100 GiB plus 1 MiB free. The Windows rerun passed 12
  synthetic client checks, 8 protocol checks, and a disposable real-NTFS
  private owner-ACL fixture; actual D:, SSH, and VPS remain unverified.
- Receiver cleanup after a post-publication ACL failure is identity-bound to
  the owned staging directory and permits only the expected archive files. A
  replaced, reparse, or unexpected target is left for manual review without an
  ACK; this path is covered by local fixtures and the Windows rerun.
- Synthetic tests cover archive/protocol integrity, helper download rules,
  artifact/run provenance, receiver publication, release schema, and rollback
  refusal. The 12 synthetic Windows client checks, 8 protocol checks, and
  disposable real-NTFS private owner-ACL fixture passed. The actual D: path,
  SSH alias, Gitea helper against
  the live instance, root helper installation, artifact redirect, production
  backup/restore, and actual release/rollback remain unverified.
- Owner bootstrap blockers remain: verify a root-controlled SSH route for the
  fixed commands using existing host-key checking; provision only the
  necessary read-only Gitea token, exact redirect origin if needed, and
  approved RU/KY canaries; confirm registry pull and retention policy; test
  the Windows client and a separate-target restore. No new key, sudo rule,
  Docker access, credential, or VPS setting is added by the code change.

Read-only deployment recheck: 2026-10-25. Source: SSH as `codex`, direct-origin
and public HTTPS probes, local `main` checkout, and Gitea API. No VPS files,
services, firewall, containers, volumes, production data, or credentials were
changed or inspected. Root-only Docker state, UFW rule details, protected file
metadata, current backup/restore validity, and the live corpus/index remain
unverified. The 2026-10-18 inventory below is a historical snapshot where it
conflicts with this dated recheck.

## Latest read-only recheck — 2026-10-25

- `https://akyldash.kg/review` returned HTTP 200 with valid TLS through the
  public Cloudflare edge and through direct HTTPS to origin `206.223.249.118`.
  This verifies that route only; it does not prove search or index readiness.
- TCP listeners now include 22, 80, and 443. The older no-listener/Cloudflare
  521 observations below are historical, not current state.
- VPS reports 4 CPUs, 7.7 GiB RAM (about 3.7 GiB available), and 31 GiB free
  on `/`. Current runner-mount free space was not rechecked.
- `/opt/akyldash/current` exists, is a root-owned directory mode `0755`.
  Reading its Git SHA failed with Git's `dubious ownership` safety guard; no
  `safe.directory` setting was added, so the installed revision is unknown.
- `akyldash-act-runner.service` is active under `akyldash-runner`; installed
  `act_runner` is `0.2.13`. The separate `akyldash-runner-restart.service`
  remains failed since 2026-10-14; its cause could not be read from the
  unprivileged journal view. No restart was attempted.
- `akyldash-verify`, `akyldash-compose`, and `akyldash-rollback` are absent
  from PATH in the unprivileged check; no root helper or Docker operation was
  run.
- Gitea is `1.25.5`; `main` is `b5adc09203cbbbd298054248206b9f47ae4a44ca`.
  CI run 175 and runner smoke 176 succeeded on that commit. Package write
  access and runner inventory remain outside this API principal's permissions.
- PR #59 head `852d639c455ff1c60202405e112d47b8128d66c6` received formal
  review #308 with five requested fixes. CI run 177 failed in
  `test_digest_rollback_fails_before_stopping_or_changing_references` because
  the test's global `Path.stat` mock made `Path.exists()` report a false
  recovery-journal hit under CI Python 3.12; runner smoke 178 succeeded. A
  fix batch in this PR isolates state paths, handles ambiguous alias/cutover
  outcomes, verifies rollback image IDs, and adds a synthetic artifact upload
  check. On 2026-10-25 CI run 179 and runner smoke 180 succeeded at PR head
  `9d8ce6b6`; CI took 2m43s and the synthetic upload step reported success.
  The run API identifies run ID 567 and reports `run_attempt: 0`, Gitea's
  legacy sentinel. The release workflow therefore uses the unique run ID and
  does not depend on the unavailable attempt value. The run's artifact link was
  downloaded from Gitea UI as `artifact-smoke-567-`: the ZIP was 191 bytes,
  SHA-256 `ae0a2262147f222aacb4fdd1d12459f7160c67b1d0ee5d7ae2db1de8958f120f`,
  and contained the expected 51-byte `marker.txt`. The API artifact listing
  available to this principal returned no entries, but the downloaded file
  verifies artifact availability. No live release or VPS change was made.
- The release descriptor SHA-256 sidecar now records `release-manifest.json`
  relative to the `release` directory, matching the flattened layout observed
  for downloaded Gitea artifacts. A local synthetic regression verifies the
  checksum from the artifact extraction root; the credentialed release
  workflow remains unrun.
- Owner-only outstanding evidence: sanitized container/volume and protected
  file metadata; current backup and separate-target restore verification;
  live corpus/index counts and RU/KY canaries; Cloudflare TLS mode; registry
  writer identity and digest retention. Never include secret values in that
  evidence packet. The checklist is in
  [`DEPLOYMENT_OWNER_PACKET.md`](DEPLOYMENT_OWNER_PACKET.md).

The deployment PR adds a manual main-only image-publish workflow and an
owner-run `--no-build` cutover path, but neither has been executed against the
VPS. Its release descriptor is an unsigned trace record; cryptographic image
attestation remains a separate owner decision.

## Historical host snapshot — 2026-10-18

Source: SSH as `codex`, local repository inspection, and Gitea API. Root-only
Docker state, UFW rule details, and protected file metadata were not verified.

## Historical VPS facts — 2026-10-18

- Host: `avprogs.hlab.kz`, `206.223.249.118`, SSH port `22`, user `codex`.
- OS: Ubuntu Server `24.04.5 LTS`; kernel `6.8.0-142-generic` from the prior
  inventory audit.
- Resources: 4 vCPU, 7.7 GiB RAM, 3.8 GiB swap, 99 GiB root filesystem with
  84 GiB free on 2026-10-18. Runner storage is a separate 15 GiB ext4 loop
  mount with 12 GiB free at that check.
- Network: interface `ens3`, IPv4 `206.223.249.118/32`.
- `codex` is UID/GID 1000, groups `codex` and `users`; `sudo -n true` requires
  a password. Docker CLI cannot access `/var/run/docker.sock` as `codex`.
- SSH service is running. The inspected sshd configuration has public-key and
  password authentication enabled and `PermitRootLogin yes`; no SSH changes
  were made.
- Bootstrap completed. Docker CLI `29.9.0` and Compose `5.6.0` were rechecked
  over SSH on 2026-10-18; containerd `2.4.1` and runc `1.5.2` are from the
  prior package audit. The Docker daemon API remains root-only.
- `/etc/docker/daemon.json` has the expected `local` log driver and `10m` / `3`
  rotation limits. Bootstrap paths exist with root ownership: `/opt/akyldash`
  mode `0755`, `/etc/akyldash` and `/var/backups/akyldash` mode `0700`, and
  `/var/lib/akyldash/data` mode `0750`.
- `/opt/akyldash/current`, verifier/rollback binaries, and the sudoers file are
  absent. `codex` has not received sudo or Docker-group access. Metadata for
  `/etc/akyldash/production.env` and `/etc/akyldash/review_secret` could not be
  checked because access is denied. The target
  `/var/lib/akyldash/data/minjust-normalized` and its manifest metadata were
  also denied; their presence, ownership, and contents are unknown. No sudo was
  used.
- TCP listeners observed: SSH `22` and loopback DNS only; no listener on `1500`,
  `80`, or `443`. No ISPmanager package or service unit was found through the
  unprivileged checks; root can confirm with the audit script below.
- Owner reports UFW is active and allows TCP `22`, `80`, `443`, and `1500`.
  Rule details were not independently read; no firewall changes were made.
- Additional read-only runner checks on 2026-10-12 confirmed `/` is ext4 with
  about 92 GiB free at the prior audit, `/usr/bin/dockerd-rootless.sh` exists,
  and `git.vakanaut.ru` resolves to `195.209.48.28`. From the VPS,
  `https://git.vakanaut.ru/api/v1/version` returned Gitea `1.25.5`; runner
  registration and Gitea Actions protocol connectivity are not confirmed.
- `akyldash.kg` resolved to Cloudflare IPv4/IPv6 edge addresses. HTTPS returned
  Cloudflare `521` on this historical 2026-10-18 check; current recheck is
  recorded above. Cloudflare Full (strict) mode remains unverified.

## Historical application and Gitea facts — 2026-10-18

- Repository: `admin/akyldash`, default branch `main`; Gitea version `1.25.5`.
- The repository API reports push permission, Actions and packages enabled, and
  PRs enabled. `main` is protected and requires one approval; the current API
  principal cannot push or merge directly to `main`.
- The repository runner-list API returns `403` because this API principal is
  not the repository owner. The latest verified CI run 75 and smoke run 76 both
  succeeded on `main` head `0e276ae6`. The Actions run-list endpoint returned
  `403` for this API principal during this recheck, so no newer run state was
  independently retrieved. The owner reports the runner is registered and
  Idle; repository runner inventory and package publishing permissions remain
  unverified.
- No public application listener is observed on TCP `80`/`443`. Docker
  container/volume state still requires the root audit. Repository Compose describes
  Caddy `2.11.6` pinned by digest, Python Search API `0.12.2` in this branch,
  SQLite account/review data, and OpenSearch `3.7.0` with `analysis-icu`.
- The target publishes only Caddy on TCP `80`/`443`. Backend and Caddy share a
  proxy network; OpenSearch and backend share a separate internal network, and
  backend has no host port. Caddy certificate storage, OpenSearch, and
  application data use persistent volumes/bind mounts.
- The root-owned repository checkout, production env/secret files, Compose
  volumes/containers, backup, verifier/rollback helpers, and application are
  not confirmed installed. TLS and registry digest deployment for
  backend/OpenSearch are not configured.
- The repository has a cold-backup archive/verification tool and a one-shot
  restore apply tool that requires a marked replacement host. Backup/restore
  tests use synthetic data; neither tool has run against production. The
  archive preflight permits a clean first deployment to skip backup, requires
  backup for existing persistent data even without containers, allows absent
  named volumes only in explicit corpus-only initial-deployment mode, and
  rejects inaccessible/ambiguous mounts and incomplete/running service
  containers. Initial deployment docs require a second verified backup after
  index and admin initialization.
- Recovery bootstrap has an explicit `--recovery-target` mode that refuses the
  inventoried production hostname/IP; its host-identity guard has shell tests.
  The mode has not been run on a VPS.
- The repository defines a root-owned read-only verifier, Compose `config` and
  `status` helper, guarded rollback helper, and installer with exact sudoers
  command rules. None was installed on the VPS at this inventory. The current
  PR adds fixed root-side release/rollback receivers without adding a `codex`
  sudoers rule or Docker access. Neither receiver has been installed or run.
- The former local-VPS rollback path has been superseded and now fails closed.
  The PR rollback path requires a root-pinned external archive identity,
  rechecks its exact bytes through the owner client, creates a fresh external
  backup before changing refs, and preserves a secret-free recovery journal on
  failure. Synthetic tests only; no production rollback was performed.
- The image release workflow now starts after successful CI for a push to
  `main`. It checks the source workflow/event/branch/conclusion and exact source
  SHA, checks out that SHA, and uses both release and source run IDs in the
  schema-v4 descriptor. Registry secrets `REGISTRY_USERNAME` and
  `REGISTRY_PASSWORD` remain unconfigured/unverified; no values were created or
  inspected. PR image builds remain no-push. The descriptor SHA-256 is included
  in the artifact name, and the artifact also retains its checksum sidecar.
  The descriptor remains unsigned trace metadata, not cryptographic
  attestation.
- A separate `deploy-handoff.yaml` runs after a successful release and checks
  the source run and stale-main guards. Its job is explicitly
  `NOT DEPLOYED - handoff status only`; a green result is not a deployment. It
  reads artifact metadata through `gitea-api` only if that helper is available
  in the job image; otherwise it reports metadata and host handoff as
  unverified. It does not download artifact ZIPs, connect to SSH, or apply a
  release. No repository workflow binding to a Windows runtime starts the
  receiver. PR #59 now proposes a separate disabled local `--once` trigger
  that queries a root-verified tuple; it is not installed or active. Image
  release is automated, but production deployment is not yet automatic CD. The original
  audited `gitea-api` helper lacks binary output and safe 302 handling; the
  current PR carries a tested patch and host-side artifact verifier, but
  neither is installed. No deploy key, fixed SSH entry, extra sudo rule, or
  Gitea CD variable has been configured. Production apply remains blocked on
  owner bootstrap, external backup and separate-target restore, and the other
  gates above.
- A repository-scoped official `act_runner` `0.2.13` installation is
  configured in `infra/runner/`: rootless Docker under `akyldash-runner`, 15 GiB
  ext4 loop-backed image/workspace storage, capacity 1, per-job limits (2 CPU,
  3 GiB RAM, 512 processes), account slice limits (2.5 CPU, 4 GiB RAM, 2048
  tasks), rotated file logs, and a systemd health timer. Jobs target the
  `akyldash-ci` label and do not receive either Docker socket. Latest verified
  main CI run 175 and smoke run 176 succeeded on `b5adc092`. On 2026-10-18,
  SSH showed `akyldash-act-runner.service` active under `akyldash-runner`, with
  zero restarts and about 115 MiB current memory. Its rootless Docker store is
  mounted from `/dev/loop0` as ext4 with `rw,nosuid,nodev`; the 15 GiB mount had
  12 GiB free. The separate `akyldash-runner-restart.service` remains failed
  with exit status 1 since 2026-10-14; its journal had no entries, so the cause
  is unknown. The latest verified main CI run 75 and smoke run 76 succeeded on
  head `0e276ae6`. The Actions API returned 403 during this recheck, so no newer
  run state was independently retrieved. A read-only SSH recheck on 2026-10-21
  confirmed the active `akyldash-act-runner.service` runs as
  `akyldash-runner` with `ProtectHome=read-only`; `ReadWritePaths` includes the
  runner runtime/config/log paths and `/run/user/999`. Its `ExecStart` uses
  `/etc/akyldash-runner/config.yaml`. The live unit retains a writable carveout
  for `/run/user/999`; the installer keeps that existing carveout and changes
  `ProtectHome` from `yes` to `read-only` so the runtime path remains visible.
  No VPS unit was changed. Main commit
  `d2fb4b81` passed CI run 115 and smoke run 116. On
  head `3a89dc21`, CI run 63 failed at the host audit root guard and smoke run 64
  succeeded. On head `436806a8`, CI run 65 failed in the Caddy container action
  with `exec: "validate": executable file not found`; Kaniko was skipped and
  smoke run 66 succeeded. On head `d96fc4d3`, CI run 67 failed because the
  Caddy action's fixed `/github/workspace/...` path did not exist; Kaniko was
  skipped and smoke run 68 succeeded. On head `5d72bb09`, CI run 69 and smoke
  run 70 succeeded. CI validated the Caddyfile and built the backend and
  OpenSearch images using `${{ github.workspace }}`. Actual per-action
  resource-limit enforcement and peak usage were not independently measured or
  verified. Image publishing, host-wide resource metrics, and production
  deployment remain unverified. No VPS changes are made.
- An isolated local Compose smoke test built backend/OpenSearch images,
  reached healthy status for OpenSearch, backend, and Caddy, and returned HTTP
  200 for `/review` through Caddy. Its containers, networks, and volumes were
  removed afterward. This does not verify VPS startup, DNS, or public TLS.
- The app has no migration framework or SMTP setup. Repository versions on this
  branch are:
  backend `0.13.0`, frontend `0.2.0`, Telegram bot `0.2.2` (the bot is not
  deployed on this VPS).

## Corpus readiness (2026-10-18)

- The local raw archive and normalized corpus are ignored under `data/`; no
  corpus files were changed. Their read-only SQLite manifests both pass
  `quick_check` and each report 209,958 documents. The raw manifest is
  28,295,168 bytes with SHA-256
  `f8e3dd2fdec607583ca09cfdb5b76171bb7bfa1e41a0e07ee5aca8481e18e17e`; it has
  one fetch error, code `6`, absent from `documents`. The normalized manifest
  is 31,571,968 bytes with SHA-256
  `1679c20e6d0f3e5a9b50737c006c4417a28ebcaf7467d194d3bf8243538f83bf`; all
  entries are `success`, schema `1`, normalizer `1.0.0`.
- The read-only full-file inventory in `/tmp/akyldash-corpus-audit.Mjr6aa`
  stopped before validation in run `b4a4556e30cc464d90bc8d75ef1673e9` with
  `Python int too large to convert to SQLite INTEGER`. At stop it had processed
  2,120,420 raw entries (1,081,428 files, 1,038,992 directories) and 397,856
  normalized entries (246,321 files, 151,535 directories). This is a verifier
  metadata storage failure, not a corpus finding. A full rehash with the
  corrected, versioned verifier is required before readiness can be stated.
- A fresh full rehash is running in `/tmp/akyldash-corpus-audit-final.gtsD87vc`,
  run `f4dd5e53020f442bbbecee2d0e4986a6`, with verifier SHA-256
  `83dac230099fc91972cd14e86e9f8ec72d13bca08f53d402f13a169e096d98f5`.
  `result.json` currently reports `scan_status: running`; its live counters
  must be recorded after completion. Sources remain read-only.
- Through the repository client, bounded MinJust OpenData requests for
  `GetDocumentListByQuery`, `GetDocument(6)`, control `GetDocument(1096)`, and
  `CheckAvailable` each returned HTTP `404`. The list API returned no code list;
  the direct requests do not establish whether code `6` exists. The public
  listing page is reachable, but it does not map the numeric API code. A
  metadata-only `CheckAvailable` request from the VPS failed in HTTP/2 with
  `curl` exit 92 (`INTERNAL_ERROR`), HTTP status `000`, and TLS verification 0;
  the HTTP/1.1 retry was not completed. Treat the code `6` disposition and
  full corpus readiness as unresolved.
- On the VPS, read-only `stat` of `/var/lib/akyldash/data/minjust-normalized`
  and its manifest was denied. Production corpus presence and metadata remain
  unknown. No corpus was transferred and no production service, Docker socket,
  runner policy, or VPS configuration was changed.

## Current MinJust API staging (2026-10-20)

- The public site's current API uses `POST /api/v1/GetDocuments` with page size
  12, `GET /api/v1/GetDocument?DocumentCode=...`, and
  `GET /api/v1/GetEdition?editionId=...&lang=ru|kg`. A bounded page-1 probe
  reported 210,639 filtered records; no full listing was enumerated for this
  recovery. Official alias `1-4` resolves to canonical document code `205064`.
- Document code `6` has 201 official editions. The API's `lang=kg` response
  places Kyrgyz HTML in `contentRu` while returning `contentKg=null`; staging
  maps by requested language (`kg` to archive `ky`) and preserves the edition
  order from the document response. Normalization and search export now retain
  and use that order instead of sorting IDs.
- Without `exact=true`, one probe for edition `920292` returned response field
  `id=920492` despite a different text length/hash. Two bounded `exact=true`
  probes returned matching IDs and names/dates from `GetDocument`: edition
  `920292` / `04.08.1998` (485,487 RU characters, SHA-256
  `2b51c2ff4e9b91cbe2a7ed052a5a9ed9150bda993fc15c120ac948580926b718`) and
  edition `920492` / `13.04.2017` (1,139,908 RU characters, SHA-256
  `e5f0fdbb47766e5a5dbb971183d765e4c1dcaf9c969a218a84206244ce9132e8`).
  Every edition request now uses `exact=true`, requires returned ID equality,
  and checks RU/KY names against `GetDocument` metadata.
- `backend/ingestion/minjust_current.py` stages one document into a separate
  generation, at no more than one request per second, with a 16 MiB response
  limit, a 4 GiB text budget per document, contract-bound checkpoints, and
  atomic publication. Staging requires a private 0700 directory owned by the
  current user and rejects symlinks under checkpoints/document paths; legacy
  corpus exclusions are anchored to the repository root. The corpus verifier
  continues to validate unchanged normalized manifests from versions 1.0.0
  and 1.1.0. The earlier non-exact pass is quarantined at
  `/tmp/akyldash-current-code6.ZAMC37u4ex/qualification.json`: its 240 language
  responses (304,493,520 bytes) remain unverified and are not reusable. The
  retained work-tree SHA-256 is
  `1e874c1d7fab1e00e7e45ff1b16422ea474e22ee13c901ac9a0612f47aa5d24a`.
- Exact-only code-6 staging is complete at
  `/tmp/akyldash-current-code6-exact.L4k3AvnOlP`: 201 editions and 402
  complete language checkpoints, 604,377,885 raw bytes. The raw manifest
  passes `quick_check`; its document checksum matches the staged directory at
  `81d8e1a53b2ebef61f08b56d04b3f0a80077d587a4fadae0152bb4b7f13055ea`.
  Separate normalization at
  `/tmp/akyldash-current-code6-exact.L4k3AvnOlP-normalized` contains 201
  editions in official response order, passes `quick_check`, and records
  normalizer `1.1.0` (2,048,797,024 bytes). The prior `data/minjust-cbd` and
  `data/minjust-normalized` trees and production were not used as input or
  modified. The raw staging manifest still records attachments and embedded
  media as unverified; this single document does not establish corpus
  completeness or production readiness.

## Checks not completed

- Docker Engine version from the server API, container and volume inventory,
  protected env/secret metadata, sudoers, and detailed UFW rules require root.
  `infra/bootstrap/audit-host.sh` is provided for an owner-run read-only audit.
- The backup archive/restore tool was tested with synthetic local data; a backup
  and restore of real production data remain unverified because production is
  not deployed.
- Cloudflare edge response was checked; origin DNS target, origin certificate,
  Full (strict), registry publishing, host-wide runner metrics, and production
  deployment have not been verified. The
  owner-reported Idle state has not been independently checked through runner
  inventory because the API returned 403.
- This audit made no host changes. No package or service was installed by this
  audit; the owner performed the initial bootstrap beforehand.

## Next owner actions

1. Install the approved root-owned checkout after its PR is reviewed, then run
   `sudo bash /opt/akyldash/current/infra/bootstrap/audit-host.sh` and share
   the output. The script prints no secret values and does not mutate the host.
2. After reviewing the audit output and confirming the target files are absent,
   run
   `sudo bash /opt/akyldash/current/infra/deploy/prepare-production-config.sh`;
   it refuses to overwrite existing files. Set the approved `DATA_ROOT` and
   place the approved corpus/index and Cloudflare origin configuration. Do not
   deploy until data/service state and the required backup path are reviewed.
3. Run `sudo bash /opt/akyldash/current/infra/deploy/install-helper.sh` after
   reviewing the checkout; it grants only fixed verification, Compose
   `config`/`status`, and guarded rollback commands.
4. After merging the reviewed PR, confirm a subsequent CI and smoke run on
   `akyldash-ci`. Image publishing and production deployment remain manual and
   are not part of CI.
5. Test backup/restore and rollback only with synthetic data in an isolated
   environment before any production deployment.

## Change in this PR

Adds a read-only, resumable full-file corpus inventory and normalized-data
cross-check, a source/staging comparator requiring nonzero documents and
fragments plus held shared artifact locks, synthetic regression tests, and an
operator handoff runbook. CI's
existing `infra/deploy` unittest discovery includes the new tests. No workflow,
runner, VPS, production data, or Docker settings are changed. The first real
corpus scan stopped on an integer-storage bug; full validation with the
corrected verifier is still required. The source API's code `6` outcome remains
unconfirmed.