sync: preserve knowledge base vectors and refine context accounting

[Luhaozhu]

Community edition sync derived from the upstream development tree.

Knowledge base collections are never dropped implicitly

get_or_create_collection used to drop and recreate the collection whenever the
stored dense vector dimension could not be read or did not match the current
embedding model. A transient Milvus outage was therefore enough to silently
delete every indexed vector.

The collection is now preserved in all three cases — unreadable schema, missing
dim metadata, and a genuine dimension mismatch — and the operation aborts with
an explicit error so an operator can verify the embedding configuration and
re-index deliberately. Four regression tests cover the new behaviour.

Tool definitions are accounted for as system tokens

Tool schemas are sent on every request as part of the fixed instruction surface;
they are not evidence that the model executed a tool. Counting them in the
tools bucket inflated the tool share of the context panel. They now belong to
the system bucket in both the measured and the post-compaction estimated
snapshot.

The snapshot schema is bumped to context-usage.v2, and the frontend parser
rejects v1 snapshots so old and new accounting rules are never mixed in one
panel.

Also included

• Sandbox provider is pinned to the service event loop, fixing repeated sandbox
  rebuilds caused by cross-loop calls.
• SANDBOX_ARTIFACT_MAX_BYTES becomes the single size limit for every sandbox
  file transfer, including /myspace write-back.
• Sandbox myspace directories are created as the sandbox runtime user, so
  agents can write into them.
• Streaming output can be selected and copied while it streams, and scrolling up
  is no longer forced back to the bottom.
• Cancelling a chat run keeps the partial answer already produced.

Verification

Generated with scripts/build_ce.py; brand, LICENSE, required-runtime-file and
binary-allowlist gates all pass. Backend tests for the touched areas pass
upstream.