test(mutation): mutate sql_guard live, not just declared

[brownjuly2003-code]

What

The mutation gate (scripts/mutation_report.py) mutated only the duckdb-free
sdk/agentflow/retry.py. Every serving module was declared-only in
[tool.mutmut], documented as blocked by "mutmut x duckdb" — so a surviving
mutant in sql_guard's NL→SQL denylist would have gone unnoticed (the policy was
decorative for the serving surfaces).

The real blocker was narrower than duckdb: mutmut's trampoline asserts a module
name does not start with src., and the serving unit tests also pull the
duckdb-backed query engine into mutmut's mutants/ workspace.

Change

sql_guard imports only sqlglot, so it can be mutated as a top-level
serving package (exactly how retry.py is mutated as agentflow.retry, not
src.*) against a narrow duckdb-free test:

• mutation_report.py: serving-target workspace handling (copy
  src/serving → serving; also_copy = serving/config/scripts) + sql_guard
  in MODULE_TARGETS (threshold 0.90).
• tests/unit/test_sql_guard_mutation.py: narrow duckdb-free test,
  dual-context import (serving under mutmut, src.serving under pytest). Every
  pytest.raises pins the message so error-text mutants die; the
  Anonymous-vs-typed forbidden-function cases both exercise the .lower()
  casing — a surviving mutant there is a denylist bypass, which mutation testing
  surfaced, so the test is stronger than the original suite.
• Correct the now-stale "serving cannot be mutated" notes in pyproject
  [tool.mutmut] and test_mutmut_policy.py.

Verification

Ran the real mutation.yml workflow on this branch (workflow_dispatch, Python
3.11 — the same job the weekly gate runs):

Overall: killed=59, survived=5, total=64
retry.py:     score=75.0%  threshold=75% (killed=15, survived=5)
sql_guard.py: score=100.0% threshold=90% (killed=44)
Mutation scores meet thresholds

sql_guard now has live mutation coverage (44/44 killed). The remaining
serving modules stay declared-only until each gets its own duckdb-free unit
test (the blocker is the test import chain, not the module).

🤖 Generated with Claude Code

[github-actions]

DORA Metrics

• Window: last 30 days
• Branch: main
• Deployment frequency: 160 total / 37.33 per week
• Lead time for changes: avg 0.28h / median 0.0h
• Change failure rate: 58.13% (93/160)
• MTTR: 0.25h across 3 incident(s)