fix: 2nd-pass audit Major — protocol contract test, wheel fallback coverage, dead modular example, clone-vs-starter boundary

[heznpc]

Resolves all 4 Major findings from the second-pass audit. Companion to PR #32 (which fixed the 2 Critical findings).

[M1] tests/test_tools.py — add test_greet_registered_on_server

Parallels test_server_info.py and test_code_review.py (both already had this). Calls mcp.list_tools() and verifies the greet tool is wired in and that the generated JSON Schema actually reflects Annotated[..., Field(min_length, max_length)] constraints. Without this, FastMCP's schema generation was an unverified link in the MCP protocol contract.

[M2] tests/test_server_info.py — cover wheel-install fallback path

Existing tests walk the source tree and find pyproject.toml. That's the dev/editable-install path. When a user does pip install my-mcp-server from a wheel, pyproject.toml is NOT shipped — _read_pyproject returns None and _server_metadata falls back to importlib.metadata.version(). That branch (server_info.py:51-56) was previously 0% covered.

Three new tests via monkeypatch:

• _read_pyproject returns None → version comes from importlib.metadata.
• _read_pyproject returns None AND importlib.metadata raises PackageNotFoundError → returns "0.0.0" fallback.
• pyproject.toml exists but has no [project] table → covers server_info.py:42 exit.

server_info.py coverage: 79% → 96%.

[M3] src/my_mcp_server/tools/greet.py — delete dead modular example

The file defined greet_formal + a register() function, but server.py never called from my_mcp_server.tools.greet import register to wire it in. 100% dead code, 0% test coverage. Updated server.py:71-74 comment to describe the modular pattern in prose without referencing the deleted file.

[M4] SECURITY.md — distinguish code-level vs runtime GitHub settings

Previous text was misleading for clones: GitHub does NOT copy secret_scanning / push_protection / dependabot_security_updates / branch protection when a template is cloned. Clones inherited the false belief "I'm safe because the starter is."

Added a new section "What clones inherit vs. what they don't" with the exact gh api invocations clones can run to replicate the toggles on their own fork.

Coverage delta

Metric	Before	After
Tests	17	21
Total coverage	70.64%	84.69%
server_info.py	79%	96%
tools/greet.py	0% (10 dead LOC)	— (deleted)

Test plan

• ruff check . / ruff format --check . / mypy src/ all clean
• All 21 tests pass locally with coverage gate
• CI matrix (3.11/3.12/3.13) green
• CodeQL analyze (python) + analyze (actions) green