Adversarial second-pass on commits 01ed305 (PR #26) and ec7045d (PR #27),
both reported as "done / tests pass / CI green". Two Critical findings.

[C1] CodeQL `analyze (actions)` open alert
- Rule: `actions/missing-workflow-permissions`
- File: .github/workflows/cd.yml
- Message: "Actions job or workflow does not limit the permissions of the
  GITHUB_TOKEN."
- Detected by the CodeQL `actions` language matrix that PR #26 itself added.
  PR #26 introduced top-level permissions on most workflows but missed
  cd.yml and setup.yml. The 5 required CI checks didn't include CodeQL,
  so the alert landed silently after merge.
- Fix: add `permissions: contents: read` at workflow top level in cd.yml
  and setup.yml. Jobs continue to opt into the writes they need
  (contents:write / id-token:write / attestations:write for publish;
  issues:write for setup).

[C2] `ok()` and `err()` Response Helpers — dead code marketed as feature
- server.py:41 and server.py:47 defined ok() / err() helpers.
- Zero references anywhere in src/ or tests/ (`grep -rn "\\bok(\\|\\berr("` =
  only the definitions + the introductory comment).
- Coverage report confirmed: server.py:43-44, 49 uncovered.
- README.md:32 and README.ko.md:30 marketed these as "Response Helpers —
  ok() and err() for consistent tool responses" in the "What You Get"
  list — i.e. a public feature claim with no backing code usage.
- The bundled `greet` tool returns plain `str` — the README's marketed
  pattern doesn't match the bundled example.
- FastMCP's idiom is `return value` / `raise on error`; ok/err helpers
  are a raw-MCP-without-FastMCP relic that actively teach the wrong
  pattern.
- Fix: delete ok() / err() / their introductory comment. Replace the
  Tools-section banner comment with one line documenting the actual
  FastMCP idiom. Remove the "Response Helpers" bullet from README.md
  and README.ko.md.

Why this matters beyond cosmetics: a starter is a teaching artifact.
False "What You Get" entries silently propagate the wrong pattern to
every clone. The dead helpers were also reachable code surface that
CodeQL has to scan but no test exercises.

Verification
- All 7 workflows now have top-level `permissions:` (verified via
  `awk` scan of column-1 `permissions:` line).
- ruff / ruff format / mypy strict / pytest 17/17 all clean.
- Coverage 72.12% (improved from 70.64% — uncovered dead code removed).

Out-of-scope for this PR (will be raised as Major findings for user
decision):
- tests/test_tools.py has no `test_greet_registered_on_server` despite
  test_server_info.py and test_code_review.py both having that pattern.
- server_info.py wheel-install fallback path (lines 41-42, 52-56) is 0%
  covered; test only exercises the dev/source-tree read-pyproject path.
- src/my_mcp_server/tools/greet.py defines a `register()` that's never
  called by server.py — fully dead modular example.
- SECURITY.md "branch protection on main" claim is starter-side only;
  clones inherit nothing.
- 4 Dependabot PRs (#28, #29, #30, #31) opened today — #28 is
  attest-build-provenance v3.2.0 → v4.1.0, making our 6-hour-old SHA
  pin already one major behind.