sync: knowledge-base images, resumable indexing, readable tool results, job progress

[Luhaozhu]

Community-edition sync derived from the upstream tree with scripts/build_ce.py.
74 files changed (+6513 / −242), no deletions.

Knowledge base learns to see

Images are first-class now. The uploader stops rejecting them (the backend had
always accepted them — the file picker was the gate), figures cut out of document
layouts are stored as assets with a vision-generated caption, and retrieval
returns them as structured images (url + caption + locator) rather than a bare
filename. Both the tool-result card and the citation popover render thumbnails,
and a document that is nothing but a picture no longer reports "no content".

Listing datasets deliberately does not carry images — that is a directory call,
and stuffing every library's pictures into it only burns prompt budget. Anything
you want to show has to be retrieved first.

New migration: ce_0007_kb_assets. New settings are documented in
document/{zh-CN,en}/deployment/environment-variables.md (KB_MULTIMODAL_INDEXING,
KB_ASSET_CAPTION_MAX_PER_DOC, KB_ASSET_CAPTION_BATCH, KB_ASSET_TEXT_MAX_CHARS,
KB_ASSET_MIN_BYTES / KB_ASSET_MIN_EDGE_PX, KB_ASSET_URL_PREFIX).

Indexing survives a restart

Parse → chunk → embed used to run on FastAPI BackgroundTasks, which competes for
Starlette's shared request threadpool. A bulk upload saturated it, so every
endpoint queued before reaching its handler — the UI spun and the gateway
returned 502s — and because the work was in-process it was lost on restart, leaving
documents stuck at "indexing" forever.

It now runs on a persistent queue with a dedicated worker (core/kb/index_queue.py,
core/kb/index_worker.py) that adopts whatever the previous process left behind.

Tool results are readable

Raw JSON walls are replaced by a structured data view: long lists page in, long
text is windowed instead of being truncated into a grey brick, and opening a
citation scrolls to the evidence rather than always showing the first 500
characters of the tool output.

Background jobs report honestly

• Plan progress is persisted server side instead of living only in the prompt and
  being stamped by the browser — the plan strip no longer freezes at "2/5".
• Internal wake-up prompts stop leaking into the visible transcript.
• A batch job that produced nothing stops claiming success.
• The delivery turn — artifact cards and plan close-out included — is pulled back
  into the page for users who walked away while the job ran. This works off the
  conversation's updated_at, not SSE, so it survives a closed tab.

Smaller first paint

• Static assets ship pre-compressed at maximum level and are served with
  gzip_static. nginx's implicit gzip_comp_level 1 was costing roughly 250KB on
  the entry bundle and main stylesheet alone — close to a second on a few-Mbps
  link.
• 76 unused Univer hyphenation tables are dropped from the build.
• The frontend image tolerates macOS AppleDouble sidecar files.

Also

MCP servers published by an administrator on an internal address are now
installable by users.

Verification

python3 scripts/build_ce.py --import-check — all release gates pass: brand scan
(0 text hits, full path scan), required-runtime manifest (66 entries), forbidden
artifacts, binary asset allowlist, LICENSE gate, and the CE import / OpenAPI / ORM
contract self-check (no enterprise modules, routers, ORM tables or schema
properties reachable from the derived tree).