fix(graph): model the edge read routes' actual hydrated shape

[rrader26]

Follow-up to #21. The squash-merge of that PR did not include this fix, so
main currently ships a MemoryEdge type that does not match anything the API
returns.

The bug

listEdges(), traverse(), and getEntity().edges do not return the raw
memory_edge row. They return the server's GraphTraversalEdge:

{
  "id": "g1",
  "subject": { "id": "e1", "canonicalName": "NVIDIA CORP", ... },
  "predicate": "reported_metric",
  "object":  { "id": "e2", "canonicalName": "Cost of Revenue", ... },
  "objectLiteral": null,
  "weight": 0.85,
  "hop": 0
}

subject and object are fully hydrated entities, not ids. There is no
subjectId, objectId, or brainId on the wire at all — which is exactly what
the shipped MemoryEdge declares.

How it was found

By running the Rust port of the same surface against production, where strict
decoding failed outright:

Error: Decode(Error("missing field `subjectId`", line: 1, column: 1066))

TypeScript's structural typing hid it completely — the call compiles, the
request succeeds, and you get back objects whose every declared field is
undefined at runtime. Anyone reading edge.subjectId gets undefined and no
error.

The change

Replaces MemoryEdge with GraphTraversalEdge, matching the server type name.
The raw row shape is dropped rather than kept alongside: no read route returns
it, and a type nothing returns is a trap.

hop is documented for what it actually does — 0 from listEdges (no seed),
1-indexed from traverse.

Verification

Typecheck and build clean, and verified against app.memmesh.ai — the corrected
type decodes real edges:

hop=0 NVIDIA CORP -[reported_metric]-> Cost of Revenue (w=0.85)
hop=0 NVIDIA CORP -[ticker_symbol]-> NVDA (w=0.85)
hop=0 Cost of Revenue -[disclosed_in]-> 10-K (w=0.85)

The Python (#5), Rust (#5), .NET (#5), and Go (#5) PRs all carry the corrected
shape already, so merging this brings TypeScript back in line with the other
four.