[MCP T3] Test fixture project + assertion contract

[DvirDukhan]

Phase 1 ticket T3. Depends on #649 (T2 CI workflow).

Context

Every per-tool ticket from T4 onward runs an integration test against the same fixture project. Without a shared fixture and assertion contract, each tool ticket has to invent its own — creating drift, duplicate work, and weak coverage. This ticket builds the fixture once so subsequent tickets can lean on it.

Scope

In:

• New tests/mcp/fixtures/sample_project/ directory containing:
  • A Python module with a known call graph: entrypoint() → service() → repo() → db(), plus a small class hierarchy (e.g. BaseRepo ← UserRepo, OrderRepo).
  • At least one Java file and one C# file with a known call graph (so the multilspy second-pass code paths are exercised).
  • A handful of imports between files so IMPORTS edges exist.
• New tests/mcp/fixtures/expected.yaml declaring the assertion contract:
  • Expected node counts per label (File, Class, Function).
  • Expected callers/callees for named symbols (e.g. service is called by entrypoint and calls repo).
  • Expected paths between named symbols (e.g. entrypoint → ... → db).
  • Expected prefix-search hits for sample prefixes.
• New tests/mcp/conftest.py with:
  • A session-scoped pytest fixture that boots a clean FalkorDB graph (using a unique graph name per test session to avoid stomping on real data).
  • Indexes the sample project once via Project.from_local_directory() + Project.analyze_sources() (api/project.py:79-94).
  • Yields a handle that subsequent tests use to query the indexed graph.
• A trivial sanity test in tests/mcp/test_fixture.py that loads expected.yaml, indexes the fixture, and asserts the actual node/edge counts match the contract — catches drift if anyone edits the fixture without updating the contract.

Out:

• Any tool implementation (T4–T8, T11).
• Per-tool tests (those land with their tools).
• Per-branch graph naming (T17 — fixture will be migrated when T17 lands).

Files to create / modify

• new tests/mcp/fixtures/sample_project/python/entrypoint.py
• new tests/mcp/fixtures/sample_project/python/service.py
• new tests/mcp/fixtures/sample_project/python/repo.py
• new tests/mcp/fixtures/sample_project/python/db.py
• new tests/mcp/fixtures/sample_project/java/Main.java (small known call graph)
• new tests/mcp/fixtures/sample_project/csharp/Main.cs (small known call graph)
• new tests/mcp/fixtures/expected.yaml
• new tests/mcp/conftest.py
• new tests/mcp/test_fixture.py

Acceptance criteria

• Fixture indexes successfully via Project.from_local_directory() / analyze_sources() against the FalkorDB CI service.
• expected.yaml is loaded and validated by a small Python helper (e.g. pydantic or dataclasses + yaml).
• test_fixture.py asserts:
  • Total node count matches contract
  • Total edge count matches contract
  • At least one named caller/callee assertion passes
  • At least one named path assertion passes
  • At least one prefix-search assertion passes
• Test runs green in the CI workflow from [MCP T2] CI workflow with FalkorDB service for MCP tests #649.
• The session-scoped fixture cleans up its graph on teardown (no orphan graphs left in FalkorDB).

Dependencies

• Blocks: T17 (#TBD), T4 (#TBD), and every per-tool ticket downstream of T4.
• Blocked by: [MCP T2] CI workflow with FalkorDB service for MCP tests #649 (T2).

Out of scope (do NOT do in this PR)

• Tool implementations.
• Per-language additions beyond Python/Java/C# (T16 will add fixtures for the new tree-sitter languages).
• Per-branch graph naming (T17).

Notes for the implementer

• The fixture should be small but not trivial — 4-5 functions per language is enough to assert call/callers/paths/search/impact-analysis without bloating CI runtime.
• Use a unique graph name like code:test_fixture:_default (the _default branch suffix anticipates T17 — it will line up with the per-branch convention).
• Document the fixture's call graph in a tests/mcp/fixtures/sample_project/README.md so future contributors don't have to reverse-engineer it.
• Reference: see how tests/endpoints/ boots and tears down test data for an existing pattern.