feat(deploy): add canonical Gitea recovery flows

This commit is contained in:
Codex
2026-08-22 11:14:33 +03:00
parent a9ea31f00a
commit 502acd3771
14 changed files with 15257 additions and 4 deletions
+338
View File
@@ -81,9 +81,347 @@ Supported components in this source:
- `n8n-private-extension`
- `module-foundry`
- `device-plane`
- `gitea`
- `proxy-contur`
- `dc-amd-proxy`
## Fresh Gitea installation
`gitea` is a one-time, fresh-install-only component. Its artifact contains
only the reviewed Compose file and strict deployment descriptor; it contains
no database, repository, user, token, key, hook, runtime data or secret. Build
the deterministic artifact locally with:
```bash
node infra/deploy-runner/build-gitea-fresh-install-artifact.mjs \
gitea-fresh-install-YYYYMMDD-NNN
```
The runtime is fixed to Gitea `1.27.1-rootless` by exact digest and
`linux/amd64`. The runner never pulls it: the exact image must first be loaded
through a separately reviewed trusted image-acquisition step. Compose is
invoked with `--pull never --no-deps`, uses `network_mode: none`, publishes no
ports, and serves only the host-mounted Unix socket
`/volume1/docker/nodedc-gitea/socket/gitea.sock`. It exposes no SSH port, uses
fresh SQLite, enables no LFS, and mounts two runner-generated secrets by URI.
Registration and password Basic auth are disabled, while MFA is enforced;
hooks, webhooks, migrations, packages, Actions, OAuth2/OpenID, federation and
update checks are disabled. HTTP Git with PAT remains the intended trusted
client path; PAT clone/push acceptance belongs to the later reviewed
user/repository restoration transition, not this empty installation. LFS may
only be enabled by that later transition after its secret and restart behavior
has been accepted against the pinned runtime.
Plan and apply fail closed unless all live prerequisites are already true:
- `/volume1/docker/nodedc-gitea`, the `nodedc-gitea` project, and any stale
`nodedc-gitea_internal` network are absent; TCP/4022 is closed and Docker
publishes neither TCP/3000 nor TCP/4022;
- Docker Server is exactly 24.0.2, Compose is at least 2.20.1, and the exact
pinned image ID and RepoDigest are local for linux/amd64 with image user
`1000:1000`;
- the separate root-owned Nginx bridge
`/usr/local/etc/nginx/conf.d/http.nodedc-gitea-uds.conf` matches its exact
reviewed bytes, owner/mode and SHA-256; Nginx 1.23.1 configuration validates,
and only its root/uid-1023 processes own `127.0.0.1:3000`. The bridge proxies
to the Unix socket, preserves Host/X-Real-IP/XFF/XFP, rate-limits login and
bounds concurrent requests without limiting Git request bodies;
- DSM reverse-proxy UUID `5bc46027-0307-4261-af7e-4f94a3c508c9` persistently
and in generated nginx config routes `git.dcserve.ru:443` to
`127.0.0.1:3000`; the generated vhost must also retain its exact host guard
and canonical Host/X-Real-IP/XFF/XFP forwarding directives, so the UDS
bridge never trusts a client-supplied `X-Real-IP`;
- the emergency broad INPUT DROP for TCP/3000 has been removed, while the
legacy `172.22.0.222:3000` OUTPUT DROP and TCP/4022 INPUT DROP remain;
- any legacy container named `gitea` is stopped with restart policy `no`.
Changing DSM reverse proxy/firewall rules, acquiring evidence from the old
root, restoring trusted users/repositories, and loading the pinned image are
separate reviewed transitions. This component never reads or mounts
`/volume1/docker/gitea` and never starts the legacy container. Failed apply
stops/removes only the `nodedc-gitea` candidate, proves that it is absent, and
quarantines the complete newly created root (including data/config/socket/secrets) for
evidence, and restores the absent source/root state without deleting runtime
evidence. If candidate absence cannot be proven, rollback preserves the root in
place and records reconciliation-required instead of disconnecting live bind
mounts.
Run the policy suite before publishing:
```bash
PYTHONDONTWRITEBYTECODE=1 \
python3 infra/deploy-runner/test_gitea_fresh_install.py -v
```
## Gitea incident salvage (clean-state recovery)
The additive `gitea-incident-salvage` slice is the reviewed recovery design for
incident `gitea-20260814`. It does not upgrade or mount the compromised state.
The candidate starts with a new SQLite database, new config, new runtime
secrets, new email addresses and new password hashes. The exact v2 decision
bundle is embedded as inert, digest-bound data:
- 2 new active identities (`dctouch`, admin; `SILVER`, non-admin), each with a
new root-contained one-time password and mandatory password change;
- 8 new locked identities with random discarded passwords, inactive,
restricted and prohibited from login;
- 45 exact repositories: 32 private `dctouch` repositories and 13 public
`SILVER` repositories. The other 962 legacy users and 2013 legacy repository
rows are not imported into the clean database.
The successor v3 descriptor embeds the owner-confirmed, canonical
`confirmed-disposition-v1.json` and the additive
`confirmed-closure-disposition-v1.json`. Both are bound to the exact snapshot,
database, identity decision, reference manifest, unsupported-state report,
schema catalog and semantic-topics hashes. The original disposition's 105
reference decisions are explicit rows, not namespace wildcards: 85 normal
heads, 4 tags and 4 wiki heads are eligible for later live reconstruction; the
exact 5 pull refs and 7 remote refs are sealed archive-only evidence and can
never be auto-promoted. The closure disposition authorizes only bounded
plan-time inventory and records the target policy for access, collaboration,
issue/PR metadata, attachments, releases, labels, projects, units, Packages
and Actions. Observed hashes do not become activation authority, and the new
closure-report digest is deliberately unpinned until separate review.
Build the deterministic review artifact locally with:
```bash
node infra/deploy-runner/build-gitea-incident-salvage-artifact.mjs \
gitea-incident-salvage-subrelation-closure-20260814-006
```
The original v1 descriptor and `...salvage-20260814-001.tgz` remain immutable.
The exact `...salvage-disposition-20260814-002.tgz` also remains immutable and
must not be retried: its plan failed safely because it assumed every nonempty
`repository.topics` TEXT value was a JSON array. A bounded immutable-database
classification proved all 45 kept rows are instead exact four-byte canonical
JSON `null`, Gitea's semantic empty representation. The `...topics-...-003`
artifact is the forward fix with a new patch id and digest.
That exact `...topics-...-003` artifact is now also an immutable predecessor.
The local pre-audit `...closure-...-004` draft must not be staged: review found
orphan-join and nullable-Actions gaps. The local
`...closure-reviewed-...-005` artifact is also stale and must not be staged:
its primary closure was fail-closed, but it retained a broad polymorphic-
subrelation blocker. The additive `...subrelation-closure-...-006` successor
closes the concrete comment/history/merger relations in the pinned snapshot
schema and is the only artifact produced by the current builder.
The successor builder refuses to overwrite an existing artifact basename and
publishes from an exclusively created same-directory temporary file using an
atomic no-clobber hard link. A failed build removes only its private temporary
file. The disposition is sorted, compact ASCII JSON with one terminal LF.
The runtime pin is
`docker.gitea.com/gitea:1.27.2-rootless@sha256:7de5f49a…3abd2c`, with exact
linux/amd64 OCI image ID `sha256:272085a8…bc5c4c6`. It runs as `1000:1000`,
uses the direct Gitea binary, `network_mode: none`, no Docker ports, no SSH and
no LFS. Candidate restart policy is `no`. The existing reviewed Nginx bridge
cannot reach the candidate while the socket parent is mode `0700`; publication
is an explicit post-acceptance change to `0750`, followed by public TLS and
smart-HTTP checks. Only after those checks may restart policy become
`unless-stopped`. The transition never changes or enforces two-factor
authentication.
The snapshot boundary is exact: UUID
`f5a3fe3a-93ea-bb4d-847f-6221a6bcbc9f`, immutable SQLite SHA-256
`8db9e74a…1a13052`, and the 972/2058 decisions must recompute from the database
row by row. The runner enumerates only each v2 `repo_relative_path` and
`wiki_relative_path`. It rejects symlinks, special files, hard links, NOCOW,
alternates, shallow/replace/grafts/worktrees, promisor state, unknown object
material and incomplete pack/index pairs. One preloaded mount inventory also
guards the `objects` and `refs` roots plus every traversed directory and file:
all must remain on the trusted snapshot device, outside nested mountpoints and,
for directories, outside nested Btrfs subvolumes. Traversal errors fail closed.
Forbidden-path probes (`objects/info/alternates`, grafts, replace refs and the
other listed sentinels) lstat each path component without following links and
validate every ancestor before probing its child; a symlink or mount boundary
therefore stops the plan before any lookup can escape the snapshot tree.
Only three exact derived-file classes observed in the bounded 45-repository
inventory are excluded: dumb-HTTP `objects/info/packs`,
`objects/pack/pack-<oid>.bitmap` with its exact complete pack/index pair, and
monolithic `objects/info/commit-graph`. Each must be a bounded regular,
single-link, non-NOCOW file. The reference manifest records its exact path,
class and byte size, but the runner neither reads nor copies the excluded
bytes. Unobserved `.rev`, MIDX and split commit-graph layouts stay fail-closed;
`.promisor`, `.mtimes`, `.keep`, unknown and temporary lookalikes are not
classified as disposable accelerators.
The same inventory found receive-pack quarantine directories matching only an
exact direct child `objects/tmp_objdir-incoming-[A-Za-z0-9]{6}`. A match must
retain the reviewed uid/gid `1000:1000`, mode `0755`, single-link and non-NOCOW
directory metadata. Its exact path and lstat fields are recorded, then the
subtree is pruned without reading, hashing, copying or reflinking contents; the
sealed read-only incident snapshot retains the forensic copy. Any name,
metadata, mount or subvolume drift remains a hard stop. A reviewed reference/OID
manifest is required; raw HEAD/refs files are never copied. The intended data
path is same-filesystem per-file `--reflink=always`, with no byte-copy fallback
and a bounded exclusive-allocation gate.
This slice is intentionally **NO-GO for apply** at the current review point.
`plan` inventories and prints the exact ref-manifest digest, byte length and
canonical compact JSON. The JSON is the reviewable, hash-bound reference/OID
artifact: it contains approved repository identities, ref names/OIDs, object
totals and stat-bound derived/quarantine exclusions, but no Git object bytes,
hooks or config. The same plan prints the canonical unsupported-state report
and a separate incident-closure report, each with its SHA-256, byte length and
compact ASCII JSON. The closure report is evidence only: its disposition keeps
`expectedSha256` and `expectedBytes` null and retains an explicit review-pin
blocker.
The reviewed plan evidence bound by the disposition is reference manifest
`9cddaf0e…a311` (36,010 bytes), unsupported-state report
`4b2cecf8…106ac` (119,468 bytes), and schema catalog
`b5e3b677…b6db`. The report has exact 45 repository rows with empty
`schema_missing`, `schema_mismatch`, and `anomalies` arrays. These facts close
only the owner's material-state decision; fsck/reachability, physical blob
inventory, sanitized archive creation and candidate acceptance remain separate
verifiers.
The unsupported-state report has one deterministic row for every one of the
45 kept repository IDs, including explicit zeroes. It separates direct
repository-relation counts, directional base/head PR counts, repo-unit type
counts, text-metadata presence booleans and denormalized numeric hints; it does
not add overlapping relations into a misleading grand total. LFS rows are
reported per repository with validated logical byte sums and distinct-OID
aggregates, including size conflicts and sharing with non-kept repositories.
Every association sharing a kept OID must also have an integer repository ID
that resolves to the exact snapshot `repository` table; invalid and orphan
owners make shared/unique byte aggregates unavailable and add a data-anomaly
blocker.
Attachments are reported per repository with validated logical byte sums and
explicit, overlapping issue/comment/release/unlinked link splits. Neither
category claims that a physical blob exists; no title, body, note, comment,
config, URL, token, key, secret, hash payload or blob content is selected or
exported.
The report also embeds the exact `PRAGMA table_xinfo` catalog and explicit
label-to-table-to-repository-column map for a code-owned allowlist of relevant
tables. `sqlite_master` must prove an ordinary, non-virtual table before any
row query; SQLite 3.37+ adds `PRAGMA table_list` ordinary-main-table proof. On
an older compatible SQLite, the weaker object-kind capability is recorded and
adds a coverage blocker. `query_only=ON` and `trusted_schema=OFF` are read back
before the snapshot connection is accepted. Only column identifiers and
structural metadata are exported, never DDL, defaults or row values. The v3
closure inventory follows code-owned, schema-attested primary joins for the
issue/PR dependency tables (comments, assignees, content histories, labels, issue
users/watches, reactions, reviews/state, dependencies, notifications,
stopwatches, tracked time and auto-merge), plus projects/boards, releases,
attachments, Packages and Actions. It emits only IDs required to prove joins,
actor classes, counts, declared sizes and SQLite byte lengths; it never selects
or prints text bodies, names, URLs, package metadata, workflow payloads,
credentials, tokens, secrets or logs. Package and Actions target features stay
disabled and no legacy row is importable. The successor additionally closes
every concrete FK-like comment field in the pinned schema (labels, projects,
milestones, tracked time, assignees, dependency, review and cross-reference),
content-history-to-comment, review-team and pull-merger provenance. Target
repository and local actor IDs are classified through the exact 2058/972
decisions. Migrated external-author IDs are never treated as local users and
are represented only by presence counts and author-name byte lengths. A
nonzero comment-assignee or review-reviewer team must resolve to an exact team
and organization; because no organization/team allowlist exists, its row,
team and organization IDs enter a sealed HOLD inventory and add the narrowly
scoped `issue-pr-team-mapping-hold` blocker. Orphans, conflicting user/team
identities, invalid cross-reference pairs and schema/type drift fail closed.
The report is still not a physical-file or sanitized-archive verifier, so
those separate blockers remain explicit.
The schema-catalog digest is recorded as observed evidence but deliberately
remains unpinned as reviewed activation authority. Schema absence/type drift,
invalid numeric/OID state or aggregate overflow is a hard error. `apply` fails
before creating `/volume1/docker/nodedc-gitea` until all of the following are
registered:
- root-owned Btrfs proof that the named snapshot UUID is read-only;
- exact stopped/restart-`no` legacy container image ID, image ref and sole
`/volume1/docker/gitea:/data` bind inspect;
- a separately verified exact reference/OID manifest digest, fsck,
reachability and selected-object reconstruction;
- immutable sanitized archives for issues, PRs, comments, releases, labels,
projects and repository descriptions, with no legacy row import;
- physical attachment inventory and reachable LFS pointer/OID/size/SHA proof;
- kept-user mapping before any collaboration recreation; and
- package/Actions schema plus physical zero closure and target unit-policy
acceptance.
All 25 direct repository relations have explicit dispositions, including the
15 exact zero-and-drop categories. Pull base/head counts remain directional
views of the same five PRs. All repository numeric hints are dropped and
recomputed. Repo-unit types 1–10 are total: clean types 1–5 and 8 are recreated
once per repository; external wiki/tracker are forbidden; Packages (9) and
Actions (10) are absent in the target and remain globally disabled. Legacy
unit config is never imported. The topics verifier accepts exact JSON `null`
as semantic empty state, or a canonical JSON array of sorted unique lowercase
names matching `^[a-z0-9][-.a-z0-9]*$` with at most 35 UTF-8 bytes. SQL NULL,
whitespace/case variants, quoted `"null"`, objects, numbers, booleans and
noncanonical arrays fail closed. The current snapshot must prove exactly 45
serialized JSON nulls, zero serialized arrays, zero material topics and zero
`repo_topic` rows; target cache is rebuilt from relations.
The activation hook remains frozen until these blockers are closed and the
full clean-database recreation, reflink reconstruction, doctor/fsck,
credential-table zeroing, rollback and public acceptance path has its own
runtime smoke. The current reviewed source identities are:
- runner `nodedc-deploy` SHA-256
`c766985aa02fe911fa5a873717276f3503d9cecded448e41dc6945030dcbce9e`;
- deterministic builder SHA-256
`b92076583aeab854015a05deb9d942c8dfcf3ca475fd3d52115d402d6f61acca`;
- salvage policy test SHA-256
`99b96b6179286a78e7d53ee6a923fc092e3471a8cece382a472449d455ad212c`;
- fresh-install policy test SHA-256
`20653c640142c25e7d2ef712048acc7889c56e1d519e2d03b5cad4127e48947c`;
- confirmed disposition SHA-256
`0a066724bcf6e4933133db6cab6cc273393e3c262dd00dda0bbf9ceebd84f78c`;
- confirmed closure disposition SHA-256
`7ed66d9848268431a703fe24b22c41afbaa7c5ff48949604d6fc448d93e0d243`;
- successor v3 descriptor SHA-256
`9b98eb1a1640fd5569cf051a621837379b167eff4527313a43a0a851e7cc181a`;
- deterministic local review artifact
`nodedc-gitea-gitea-incident-salvage-subrelation-closure-20260814-006.tgz`
SHA-256
`d3e598cf892b1371912dcd9cef64caa991dd23613e48fe979d0f95d493c23d53`.
Only after that exact runner is separately reviewed, promoted to the
root-owned runner path and its installed SHA-256 is re-attested may the exact
artifact be staged into the canonical inbox for a canonical `plan` command.
Staging is evidence collection only; it grants no authority to run `apply`.
The legacy identity is now pinned to the observed ref `gitea/gitea:latest`
and immutable image ID
`sha256:bf95d9a45ce4fe38b027d051cdc4a4bc531513489fa6244af4074efbb1c376d6`.
The mutable tag is provenance only. Every plan re-inspects exact name
`/gitea`, stopped state, restart `no`, and the sole RW bind
`/volume1/docker/gitea:/data`; drift is a hard stop.
The successor plan must report these exact remaining blockers:
- `attachment-physical-verifier-pending`;
- `candidate-root-activation-hard-frozen`;
- `closure-report-review-pin-pending`;
- `collaboration-kept-user-mapping-verifier-pending`;
- `forensic-ref-archive-verifier-pending`;
- `issue-pr-metadata-sanitized-archive-verifier-pending`;
- `lfs-reachable-pointer-physical-verifier-pending`;
- `package-action-physical-closure-verifier-pending`;
- `reference-manifest-fsck-reachability-verifier-pending`;
- `repository-object-reconstruction-verifier-pending`;
- `target-unit-policy-acceptance-pending`;
- `unsupported-schema-catalog-verifier-pending`.
Any other blocker or identity drift is also a stop. `apply` remains prohibited
and hard-frozen before root creation until a subsequent reviewed runner
revision closes every blocker and implements and smoke-tests the complete
activation/rollback path. The successor files must be independently reviewed
before any runner promotion or inbox staging.
Run the bounded policy suite with:
```bash
PYTHONDONTWRITEBYTECODE=1 \
python3 infra/deploy-runner/test_gitea_incident_salvage.py -v
```
The frozen local cutoff passes 54 incident-salvage tests and 26 fresh-install
tests, plus Python compilation and `git diff --check`.
`n8n-private-extension` is a staging-only trust boundary for reviewed offline
n8n private-node releases. Its artifact may contain exactly one digest-bound
`n8n-nodes-ndc` release with `package.tgz`, `release.json` and