feat(vault): add workspace notebook snapshot and revision commits

CHANGELOG.md

@@ -9,6 +9,13 @@ The project publishes 0.x prerelease versions; a stable release line is not yet
 
 ### Added
 
+- Workspace-scoped Markdown Vault sync HTTP API (`#234`): discover Vaults, pull bounded
+  snapshots and atomically commit logical-note deltas with whole-Vault CAS.
+  Independent empty notes, immutable revisions and deletion tombstones preserve
+  identity and concurrent edits; workspace, read/write and token-path boundaries
+  reuse canonical authorization. Additive migration 0027 leaves existing files
+  unchanged. See `docs/VAULT_SYNC.md` for limits, downgrade data loss and the
+  explicit boundary from current workspace-bundle export/import formats.
 - Per-agent memory namespace filter on lexical recall (`#227`): `POST /v1/context`
   and `mem_context` accept `agent_id` plus optional `extra_agent_ids`. When set,
   structured-memory recall hard-filters `producer_agent` before ranking so
@@ -17,6 +24,17 @@ The project publishes 0.x prerelease versions; a stable release line is not yet
   Design: `docs/agent-memory-namespace.md`.
 - `mem put <path> --watch` one-way foreground directory watch: poll `--interval` (default 30s), ingest new files after one quiet interval, report changed files without re-ingesting, and never propagate local deletes. Refs #110.
 
+### Fixed
+
+- Restore Linux/Windows Web dependency audit gates by removing the unpatched
+  braces build-tool chain with official Tailwind/PostCSS 4 and typescript-eslint
+  8, while retaining existing theme utility behavior and browser acceptance.
+  Apply compatible transitive security updates without relaxing audit thresholds.
+  Validation and public advisory references: `docs/acceptance/VAULT_SYNC_VALIDATION.md`.
+- Build isolated lifecycle-test MinIO/mc images from checksum-pinned official
+  sources when former public images are no longer anonymously pullable. Keep
+  production assets and all HTTP/CLI/MCP assertions unchanged. Refs #234.
+
 ### Changed
 
 - The live recall producer (`python3 -m benchmarks.recall produce`) reads engine, profile, provider and embedding dimension from the running memd instead of CLI labels, scores the multilingual v1 set (English and Chinese, including image-description queries), and stays on-demand rather than a CI gate. Refs #175.

SPEC.md

@@ -254,6 +254,28 @@ workspace import 的数据库 commit 若无法确认结果，服务端不得删
 
 ---
 
+### F9A · Workspace Markdown Vault 同步（HTTP v1）
+
+需求：[mem #234](https://github.com/bytefolk/mem/issues/234)。
+
+`GET /v1/vault/list`、`GET /v1/vault/snapshot?vaultId=<UUID>` 与
+`POST /v1/vault/commit` 提供工作区内的逻辑笔记同步底座。每份笔记使用独立
+`noteId`，相同正文和空笔记不会因 blob 去重而合并。commit 传递
+`{vaultId,baseRevision,title?,entries:[{noteId,path,content,deleted?,properties?}]}`
+增量；整个 Vault 的 head 在单个事务内执行 CAS，过期 head 返回 `409`
+及 `currentRevision`，不覆盖其他客户端。删除保留 tombstone 与不可原地更新的
+正文 revision；重命名保留逻辑 ID。只读需 `read`，提交需 `read` + `write`，
+复用现有 workspace membership 和 `/Vaults/<vaultId>` token path 边界。
+properties 仅为有界数据，不能授予工具或身份权限。图片等原件继续使用 files API。
+
+`vault-snapshot.v1` 是独立的当前状态可迁移合同，详见
+[API、限制与错误码](docs/VAULT_SYNC.md) 和
+[JSON Schema](docs/schemas/vault-snapshot.v1.schema.json)。CLI/MCP/Web 本次不新增
+Vault 命令或独立笔记编辑器。现有 workspace bundle v1/v2 不包含 Vault 表和历史；
+完整历史仍需数据库备份。服务不运行桌面文件监听或自动合并客户端冲突；客户端必须
+拉取 head、三方合并并保留冲突副本。迁移 0027 不改变 files/auth/memory 表；降级
+到 0026 会丢弃 Vault 数据，操作前需备份。
+
 ## 4. 非功能需求
 
 | 维度 | 目标 |

docs/TESTING.md

@@ -274,8 +274,10 @@ Required tests:
 - `TestManagedAISettlementOutboxPostgres`
 - `TestReleasedFileStageRetryPostgres`
 - `TestDurableContextPostgres`
+- `TestVaultSyncPostgres` (two-client CAS, distinct empty notes, auth/workspace,
+  immutable history, tombstones, path swap/title and migration 27 down/up)
 
-Expected result: migrations reach the declared current head (currently `20`),
+Expected result: migrations reach the declared current head (currently `27`),
 all rollback-state assertions pass, every named test prints `PASS`, all commands
 exit `0`, and the race run reports no data race.
 
@@ -298,6 +300,15 @@ Run the isolated, model-free process test:
 make test-acceptance
 ```
 
+The script builds temporary MinIO/mc binaries from checksum-pinned official
+source archives because Community Edition is now source-only. It uses the
+existing Go toolchain, `tar` and `sha256sum`, then builds a uniquely named test
+image on a pinned official Alpine runtime. The temporary Compose override,
+image and run directory are removed by its ownership-checked cleanup. This
+does not replace images or credentials in a production deployment. Network
+access to the fixed source archives and locked Go module dependencies is
+required; missing artifacts or checksum mismatches fail the gate.
+
 The script owns a per-run Compose project, random loopback PostgreSQL/MinIO
 ports, an ephemeral database/store and a new memd process on an atomically
 locked, twice-preflighted loopback port. It builds the current CLI, MCP adapter
@@ -449,8 +460,8 @@ Use this table in pull requests and add implementation-specific scenarios:
 | V2 | Worker processing regressions remain hermetic | `make test-worker` | Exit `0`; real-model gate explicitly skipped |
 | V3 | Localization, theme, enrichment, memory, transfer and managed-embedding control surfaces work in a browser | `make test-web` | Typecheck/lint/build, the localization audit, all browser acceptance suites and managed status mapping pass |
 | V4 | High-risk Go paths are race-free | `make test-race` | Exit `0`; no data-race warning |
-| V5 | Fresh schema, rollback and PostgreSQL semantics hold | `make test-integration` | Migration head and eighteen named tests pass, none skipped |
-| V6 | DB concurrency paths are race-free | `make test-integration-race` | The same eighteen tests pass under `-race` |
+| V5 | Fresh schema, rollback and PostgreSQL semantics hold | `make test-integration` | Migration head and all required named tests pass, none skipped |
+| V6 | DB concurrency paths are race-free | `make test-integration-race` | The same required tests pass under `-race` |
 | V7 | Real service boundaries agree | `make test-acceptance` | HTTP, CLI and MCP share one isolated service; memory citation/provenance, bounded checkpoint listing, full checkpoint get, lifecycle and forget redaction pass |
 | V8 | Five config shapes and the real adapter preserve the host-neutral MCP contract | `MEM_MCP_CERT_BINARY=... make test-agent-certification` | All fixtures and current-adapter scenarios pass with no skip |
 | V9 | Multilingual visual quality meets the chosen checkpoint | Opt-in command in section 7 | All fixed ranking assertions pass |

docs/VAULT_SYNC.md

@@ -0,0 +1,109 @@
+# Workspace Vault sync v1
+
+Implements [mem #234](https://github.com/bytefolk/mem/issues/234).
+
+This additive text-note service keeps logical identity independent of the
+existing `/v1/files` content-deduplication model. It never changes that model,
+invokes AI, or treats note properties as permissions.
+
+All endpoints use existing Bearer authentication and workspace membership.
+`X-Workspace-ID`/a bound token selects the workspace. A token must also cover
+the canonical path `/Vaults/<vaultId>`.
+
+* `GET /v1/vault/list` (`read`) returns
+  `{ "vaults": [{ "vaultId", "title", "revision", "updatedAt" }] }`.
+  Out-of-path Vault IDs are omitted, without revealing their count.
+* `GET /v1/vault/snapshot?vaultId=<UUID>` (`read`) returns a full snapshot.
+  A new Vault returns revision `0`, empty title, `updatedAt: null`, and `[]`
+  without creating a row.
+* `POST /v1/vault/commit` (`read` + `write`) accepts
+  `{ "vaultId", "baseRevision", "title"?, "entries": [...] }`.
+  Entries are **deltas**, not a replacement of the full Vault. An entry is
+  `{ "noteId", "path", "content", "deleted"?, "properties"? }`.
+  Unlisted notes remain unchanged. `deleted: true` requires empty content.
+  Omitted properties retain existing metadata; a new note defaults to `{}`.
+  Empty entries are allowed only with a title update.
+
+Both snapshot and successful commit return:
+
+```json
+{
+  "schemaVersion": "vault-snapshot.v1",
+  "vaultId": "aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa",
+  "title": "Knowledge",
+  "revision": 1,
+  "updatedAt": "2026-10-07T12:00:00Z",
+  "entries": [
+    {
+      "noteId": "bbbbbbbb-bbbb-4bbb-8bbb-bbbbbbbbbbbb",
+      "path": "岗位/空白笔记.md",
+      "content": "",
+      "revision": 1,
+      "deleted": false,
+      "properties": {}
+    }
+  ]
+}
+```
+
+IDs are nonzero UUIDs. Vault IDs in requests use canonical lowercase form.
+Paths are canonical relative POSIX paths up to 512 UTF-8 bytes, preserving
+Chinese and spaces. Absolute paths, traversal, backslashes, control characters,
+Windows-forbidden filename characters and trailing dots/spaces are rejected.
+Two active notes cannot share a case-folded path. Logical notes with identical
+content, including empty content, retain separate IDs. Text is UTF-8; only
+newline, carriage return and tab control characters are accepted. Attachments
+continue to use the existing mem file API and references rather than binary
+data embedded in this text protocol.
+
+One transaction locks and compares the whole Vault head. Successful commits
+advance it exactly once and append content revisions; mutable head pointers
+select current versions. Old entry revisions cannot be updated in place.
+Delete creates a retained tombstone; it does not physically erase history.
+Rename/swap changes logical paths while preserving note IDs.
+
+* Stale head: `409 { "error": "vault_revision_conflict", "currentRevision": N }`.
+* Active path collision: `409`, `error: vault_path_conflict`.
+* Invalid body/path/identity: `400`, `error: vault_request_invalid`.
+* Capacity/body/history limit: `413`, `error: vault_limit_exceeded`.
+* Unavailable storage: `503`, `error: vault_unavailable`; success is not assumed.
+
+Clients keep their own base snapshot and local manifest. On 409 they pull the
+current head, merge nonoverlapping edits, and preserve conflicts as explicit
+copies. A lost response is reconciled through snapshot readback, never a blind
+overwrite. The service does not modify local files or promise a sync daemon.
+
+Bounds: 4 MiB HTTP commit body, 1 MiB per note, 16 KiB per properties object,
+256 UTF-8 bytes per title, 512 delta entries, 1000 heads **including tombstones**,
+16 MiB current serialized content/path/properties, 64 MiB retained entry-history
+content/path/properties, 10,000 commits and 64 Vaults per workspace. Limits
+refuse atomically; snapshots are never silently truncated. A full snapshot may
+require roughly 32 MiB on the wire due to JSON escaping; callers must use a
+bounded Vault-specific response budget rather than an 8 MiB file-card budget.
+
+Validation: unit/API tests are hermetic. `TestVaultSyncPostgres` requires an
+explicit `MEM_TEST_DB` whose database name ends in `_test`; it verifies real
+two-client races, distinct empty notes, workspace/auth isolation, tombstones,
+immutable history, swaps, title CAS and additive migration down/up.
+
+## Compatibility and portability boundary
+
+The HTTP API is the canonical implementation. This initial additive surface
+adds no CLI, MCP or standalone mem Web Vault editor; existing adapters and file
+operations retain their contracts. Clients may synchronize ordinary Markdown
+files and upload attachments through the existing file API.
+
+The current workspace-bundle v1/v2 export/import formats do not contain Vault
+heads or revision history. A Vault snapshot is a separate portable current-state
+contract (docs/schemas/vault-snapshot.v1.schema.json). It has no auth tokens,
+workspace membership or actor audit fields, and importing its properties never
+grants permissions. Restoring a snapshot into an authorized target Vault uses
+a normal CAS commit with the target head and a client-selected logical Vault
+ID. Full history remains in PostgreSQL and requires an operator database backup;
+this API does not claim to transport retained history between deployments.
+
+Migration 0027 only creates new Vault tables and indexes. Downgrading to 0026
+drops all Vault state and history, so preserve a database backup or current
+snapshot before an operator downgrade. Existing files, auth and memory tables
+are untouched. No login recovery, deployment credentials or model dependency
+is introduced.