[YongqiYin]

Import & Export: JSONL data exchange + collection snapshots

Adds data import/export to Zvec Studio, no engine changes required:
export reads through the iterator, import writes through the existing
insert/upsert primitives.

Features

• Data export — GET /collections/{name}/documents:export streams the
  whole collection as JSONL with constant memory (snapshot isolation; writes
  stay available during export). Options: includeVector, includeFields,
  outputFields, mode=data|snapshot. A non-finite value detected before
  the response starts answers 422 instead of silently becoming null;
  one further downstream aborts the transfer mid-download (an incomplete
  file is a failed export).
• Data import — POST /collections/{name}/documents:import streams a
  local JSONL file in batches (512/batch, SDK's 1024 write limit absorbed
  internally) and returns a per-row report. mode=replace|insert,
  onError=abort|skip; row failures stay in the 200 body. Under abort,
  rows after the first failure are not kept in insert mode (compensating
  delete); malformed lines count toward totalLines, so
  imported + failed == totalLines holds in skip mode.
• Snapshot mode — mode=snapshot bundles manifest.json (full schema +
  export options) + documents.jsonl into a .tar.gz. Snapshots are
  faithful copies: the export dialog pins snapshot exports to full data
  (trimming options are data-mode only), so every UI-built snapshot can
  rebuild a collection.
• Collection import — POST /collections:import rebuilds a whole
  collection from a snapshot: schema pre-check before any write (409 with
  mismatch list), optional rename. All-or-nothing: a mid-stream error or
  any row failure rolls the freshly created collection back and answers
  422 with the first failing line. Sidebar "+" menu → "Import
  Collection"; target directory pre-filled next to the snapshot file.
• UI — Import/Export live on a dedicated Import / Export tab
  (?tab=data, a sibling of Overview/Browse/Query/Write); file picker via
  new includeFiles/extensions options on /fs/list; zh/en i18n. The
  import mode is labelled upsert-style ("insert new ids, overwrite existing
  ones").

Design notes (per review feedback)

• JSONL first, extensible by design: formats sit behind
  ImportFormat/ExportFormat protocols plus a name/extension registry;
  both endpoints carry a format parameter from day one (unknown values
  answer 400 listing supported formats). Adding CSV/Parquet later means
  implementing the protocol and registering — no pipeline changes.
• Schemas may declare an id column: zvec keeps the primary key beside
  columns, so when a column occupies id the primary key travels under the
  reserved $id row key (the engine rejects $ in column names, so this
  cannot collide). Without a clash, rows stay byte-identical to the existing
  API representation. Enforced by a shared representation module
  (doc_repr.py) mirrored in the frontend; ambiguous bare-id rows are
  rejected with guidance instead of guessing.

Pre-existing defects fixed along the way

• Insert/upsert/update of >1024 documents answered 500 (SDK write-batch
  limit hit) — now batched internally, up to the documented 10,000.
• Collections created via the SDK with an id column could not be opened
  (500), and the flat representation silently dropped the real primary
  key — both fixed by the $id scheme.
• VECTOR_FP64 was offered but rejected by the engine at create time —
  removed from the enum.
• Open collections are now globally unique by name; a clash answers 409
  with the blocking path (previously two same-named collections could be
  open and every name-only API silently resolved to the first).

Docs & tests

• docs/api.md (new endpoints, $id rules, error catalog) and
  CHANGELOG.md updated; all new error codes mapped in the frontend.
• 13 new backend test files + frontend tests covering roundtrip fidelity
  (export → import → byte-identical), the 1024 boundary, iterator release
  on every exit path, corrupt archives (truncated, non-gzip, deflate) →
  400, snapshot pre-check/rollback, abort prefix semantics, and
  name-uniqueness invariants. make verify green.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[egolearner]

1. 导入导出放browse页面有些奇怪，可以考虑单独tab或者放overview页面最下面。
2. 导出时没勾选向量，再导入集合会报错。

4. 无法只导出向量，不导出标量。
5. "导出期间读写仍可用；建索引、优化等维护操作会在导出完成前被暂时拒绝。"位置不对，可以放最上面。

[egolearner]

覆盖/replace无法体现不存在的id将被插入，建议直接用upsert。

[YongqiYin]

1. 导入导出放browse页面有些奇怪，可以考虑单独tab或者放overview页面最下面。
2. 导出时没勾选向量，再导入集合会报错。

4. 无法只导出向量，不导出标量。 5. "导出期间读写仍可用；建索引、优化等维护操作会在导出完成前被暂时拒绝。"位置不对，可以放最上面。

1、设置为了单独tab
2、报错原因是导出的manifest包含原集合的所有列，现在设置为了“快照模式”导出集合副本，数据列裁剪等语义可通过导出文件实现
4&5、done