Extract people and organizations from memory content, not just metadata

[ali-ulu]

Summary

Entity extraction only ever read connector metadata (attendees, from/to, speakers). That covers calendar/email/transcript imports but left the most common case empty: a note typed by hand, a commit message, a captured decision. Those memories produced zero entities, so the knowledge graph, People/Organizations views, and the trust corroboration signal all stayed blank for anyone not importing a mailbox — the demo corpus was carrying the whole demonstration, and removing it emptied the product.

Adds server/core/text_entities.py with deterministic, high-precision text signals only (no NER model, no LLM, no I/O — consistent with the rest of the entity layer):

Signal	Produces
e-mail address in the body	person (+ organization via domain)
@handle mention	person
Acme Inc / Zephyr Labs	organization
met with X / X ile görüştüm	person

Wired in at a single shared point (people_in_memory), so the graph, People, and Organizations views all pick it up at once.

Precision over recall. A noisy graph is worse than a sparse one, so:

• Prose heuristics are skipped entirely for source code — detected by extension, by connector path, or by the shape of the content itself — so a repo import can't mint an entity per capitalised identifier.
• Generic nouns ("the team", "Support") are rejected as people.
• Company names matched by the "met with X" phrasing are classified as organizations, not people (you meet with Zephyr Labs, but it isn't a person).
• A company matched in prose folds into its e-mail domain when both appear in one memory, so Zephyr Labs doesn't fork a node next to zephyrlabs.io.

Known limitation (unchanged by this PR): entity resolution is still per-key, so a company named in prose in one memory and by e-mail domain in another remains two nodes. Cross-memory alias merging needs its own design and isn't attempted here.

The demo-corpus assertion in test_seed_demo.py moves from 21→22 entities: the one addition is the prose-named Zephyr Labs organization, which has no e-mail domain to key off.

Test plan

• python -m pytest tests/ -q — 620 passed (17 new)
• New negative tests lock the code-safety guarantee: code yields no prose entities whether detected by extension, by path, or by content shape
• End-to-end against a live CLI DB — five hand-typed notes (previously producing nothing) now yield 4 people + 2 organizations, with working entity profiles and trust scoring:

[person] Ayse — 1 mentions          [organization] Contoso Technologies — 1 mentions
[person] Elif — 1 mentions          [organization] Zephyrlabs — 1 mentions
[person] deniz — 1 mentions
[person] mert-dev — 1 mentions

---

Generated by Claude Code