fix: mark entityContext few-shots so extractors do not store them as facts

[Divyansh151005]

What

AGENT_ENTITY_CONTEXT is sent as entityContext on every capture. Against self-hosted servers with a local extractor, the few-shot Architecture/Conventions/… quotes were periodically stored as atomic facts about the user's own repo (similarity 0.90–1.00), outranking real content — e.g. inventing "uses monorepo with turborepo" for a Bash/Python toolkit.

The server splices entityContext into space context: with no delimiter while the document is wrapped in <chunk position="N">, so bare quoted examples look like facts to extract.

Fix

Text-only change in plugin/hooks/lib/api.js:

• Wrap the EXTRACT examples in <examples note="another project, never extract">…</examples> (same XML idiom the server already uses for chunks)
• Close with </guidance> The document follows. Extract only from it.
• Stay under the server's 1500-char entityContext limit (marked string is 1315)

No logic or dependency changes. Matches the approach validated in #111 (9 consecutive clean runs after marking).

Testing

$ npm test
# tests 25
# pass 25
# fail 0

Includes a new unit assertion that the markers are present, the turborepo/Drizzle quotes sit inside the examples block, and length stays ≤1500.

Fixes #111