feat(sdk): v0.9.0 — @memmesh/sdk rename, raw-text observe, knowledge graph

[rrader26]

Brings the SDK up to what the API actually supports, and unblocks the
chat-with-memory example, which needs both the raw-text observe pipeline and
the graph counters.

@thinkfleet/memory-sdk tops out at 0.7.1 on npm, so this release also
carries the unpublished 0.8.0 work (procedural memory, precedence policy).

What changed

memory.observe() takes raw text

const { saved, candidateCount } = await tf.memory.observe({
  text: "I just moved to Denver and I'm still vegetarian.",
  role: 'user',
  userId: 'user-123',
  sessionId: 'thread-456',
})

Hand it the turn verbatim — the engine runs extraction, dedupe, graph
wiring, and embedding, and keeps only what is worth remembering.
candidateCount is what extraction proposed, saved is what survived. Filler
comes back as saved: [].

Identity fields are forwarded

userId / agentId / sessionId now reach the server. The route has always
accepted them; the SDK was silently dropping them, so provenance never
arrived on anything written through observe().

They are provenance, not a tenancy boundary — admin.search filters
chatIdentityId IS NULL OR = $1, permissive by design so project-wide memories
stay visible. Isolating one end user's memories from another's still needs a
project per tenant. Documented as such in the README and the type.

New tf.memory.admin.graph

Method	Endpoint
stats()	GET /admin/memory/graph/stats
listEntities(params?)	GET /admin/memory/entities
getEntity(id, params?)	GET /admin/memory/entities/:id
listEdges(params?)	GET /admin/memory/graph/edges
traverse(id, params?)	POST /admin/memory/graph/traverse

There was no graph surface at all before, so the structural half of memory was
unreachable from the SDK. Read-only on purpose: entities and edges are written
by extraction during observe(), and exposing the manual create/retire routes
would invite hand-maintained graphs — the work the engine exists to do.

Use stats(), not listEntities().length, for any "how big is it" question:
the list routes page, so their length is the page size rather than the total.
(That exact bug was live in the example app before this change.)

Also

• MemMesh exported as the brand-consistent client alias; ThinkFleetMemory
  retained for back-compat.
• README documents observe() for the first time — it was the primary write
  path and absent from the resource tables entirely.
• Tag-triggered npm publish workflow (v* → npm publish).

Breaking

• Package renamed to @memmesh/sdk. Consumers of @thinkfleet/memory-sdk
  do not get this via a version bump; they must change the dependency name.
• observe() return type changed from MemoryItem to
  { saved, candidateCount }, on the legacy content path too. test-app.ts
  is updated accordingly.

Verification

• npm run typecheck, npm run build, npm run check:proto (47 RPCs — 42
  covered, 5 internal) all pass.
• Packed the tarball and drove a full chat turn through it end to end:
  admin.search → Claude → observe (both halves) → graph.stats. Confirmed
  on the wire that identity now arrives:
  {"text":"I just moved to Denver.","role":"user","userId":"u-verify-123","sessionId":"s-verify-456"}

Before merging / after

• The @memmesh npm org must exist and NPM_TOKEN must be set in repo
  secrets — I could not verify either (npm is not authenticated locally).
• Publish is a tag push: git tag v0.9.0 && git push origin v0.9.0.
• memmesh-examples has four examples still on @thinkfleet/memory-sdk
  (three ^0.4.0, one ^0.7.1). They keep working, but the repo now spans two
  package names — worth migrating once this is on npm.