feat: knowledge-graph service + observe identity provenance (and a buildable main)

[rrader26]

.NET parity with @memmesh/sdk v0.9.0 — see
thinkfleet-memory-sdk#21.

⚠️ Read the first commit before the second. main does not currently
compile, and commit 1 is the fix. It is a large body of previously-uncommitted
work that I did not write; I kept it as its own commit so it can be reviewed
independently of the feature.

Commit 1 — main is broken; this restores it

A clean checkout of main fails with 25 CS0246 errors. MemMeshClient
and MemoryService reference RequestOptions, MemMeshException, SeekPage<>,
RequestInterceptor / ResponseInterceptor, and six service classes that exist
only as untracked files on a developer machine — 3,069 lines of staged parity
work that had never been committed.

That commit lands it as-is: Services.cs split into per-service files (Alerts,
Brains, Consent, Events, Financial, Lattice, Typed), plus Errors.cs,
Pagination.cs, RequestOptions.cs, an xunit suite with a stubbed
HttpMessageHandler, and the publish workflow. 103 tests pass.

I did not author this and have not reviewed it line by line — I verified only
that it builds and that its tests pass. If it was being held back
deliberately, drop commit 1 and this PR needs rebasing onto whatever does fix
the build.

Commit 2 — the actual feature

New: mm.Graph

Method	Endpoint
StatsAsync()	GET /admin/memory/graph/stats
ListEntitiesAsync(..)	GET /admin/memory/entities
GetEntityAsync(id, ..)	GET /admin/memory/entities/:id
ListEdgesAsync(..)	GET /admin/memory/graph/edges
TraverseAsync(id, ..)	POST /admin/memory/graph/traverse

Use StatsAsync(), not (await ListEntitiesAsync()).Count for any "how big
is it" question. The list routes page, so their length is the page size —
against a real project, 1,000 vs the actual 12,142.

Edges come back hydrated

ListEdgesAsync, TraverseAsync, and GetEntityAsync().Edges return the
server's GraphTraversalEdge: Subject and Object are full entities, not
ids, plus a Hop counter. No subjectId / objectId on the wire. The raw
memory_edge row is deliberately not modelled — no read route returns it.

This bit for real in the Rust port (missing field 'subjectId') and was
silently wrong in the TypeScript PR before being fixed there.

Hop is 0 from ListEdgesAsync (no seed) and 1-indexed from TraverseAsync.

ObserveAsync forwards identity

userId / agentId / sessionId now reach the server. The route has always
accepted them; the SDK was dropping them. Added only when set, so existing
call sites send identical bodies (Observe_text_omits_identity_when_unset).

Provenance, not a tenancy boundary — search filters
chatIdentityId IS NULL OR = $1, permissive by design.

Query encoding

Filter values go through Uri.EscapeDataString. An unescaped & in a search
string would truncate the query server-side and quietly return the wrong page.

Verification

• 113 tests pass (103 existing + 10 new).
• Verified live against app.memmesh.ai: 12,142 entities / 287,698 edges,
  decoding NVIDIA CORP -[reported_metric]-> Cost of Revenue and the
  literal-object case NVIDIA CORP -[ticker_symbol]-> NVDA.