fix(memory): use empty content: [] with structuredContent to avoid double-serialization (#2689)

[knoal]

Summary

Fixes #2689: Memory server double-serializes its results.

The memory server's tool handlers returned both content (text array) and structuredContent (object) simultaneously:

return {
  content: [{ type: "text" as const, text: JSON.stringify(result, null, 2) }],
  structuredContent: { entities: result }
};

The MCP SDK serializes both fields into the response payload, producing:

{ "content": [{"text": "{"entities":...}"}], "structuredContent": {...} }

Some clients (specifically Gemini CLI per the bug report) try to parse both halves as a single JSON document and fail with:

MCP error -32603: Unexpected non-whitespace character after JSON at position 102

The fix: when structuredContent is set, content must be an empty array. Per the MCP spec, when a tool has an outputSchema, the structuredContent is the canonical output; text content is optional.

This change updates all 6 affected tool handlers:

• create_entities
• create_relations
• add_observations
• delete_entities
• delete_observations
• delete_relations

It also covers read_graph, search_nodes, search_graph (the 3 JSON.stringify variants).

Tests

• 52 tests pass (50 existing + 2 new).
• Both new tests verify the contract:
  1. Every tool return with structuredContent has content: [] (not duplicated).
  2. Hardcoded success messages appear only in structuredContent, not content.
• Both tests FAIL on pre-fix (verified by git stash).

Files

• src/memory/index.ts: 9 tool return blocks updated to content: [] + structuredContent.
• src/memory/__tests__/no-dual-content-structured.test.ts: 2 regression tests.

Notes

The issue title mentions Gemini CLI specifically, but the same pattern would affect any client that:

• Parses only one of the two fields, or
• Tries to concatenate both halves into a single JSON document.

The fix is the safest of the three options I considered:

1. (chosen) Set content: [] — satisfies the SDK type, eliminates double-serialization.
2. Remove content entirely — fails TypeScript strict-typing (the SDK requires it).
3. Use content: [...] with a JSON-parseable string — text clients see the string, structured clients see the object. Same risk as the original code.

Fixes #2689.

— knoal (via the operator, Alex Knotts)