FileSource.__iter__ delegates to the _ffmpeg_chunks() generator with a plain
`for` loop. Closing the outer generator early (the streaming code stops pulling
chunks) raises GeneratorExit at its `yield`, but a plain loop does not forward
that into the inner generator, so ffmpeg's terminate/wait/close cleanup ran only
at GC instead of synchronously — leaking the child process until collection and
defeating the KeyboardInterrupt-during-wait handling. This regressed in the #29
refactor that split _ffmpeg_chunks() out of __iter__.

Wrap the delegation in try/finally and close the inner generator explicitly so
the subprocess teardown happens deterministically on early stop. The two
test_streaming_sources cleanup tests pass again.

Extend the "no new static-analysis escape hatches" diff gate with two test-suite
shortcuts, same net-new-only policy as the existing type-ignore/noqa checks:

- pytest.skip()/xfail() and @pytest.mark.skip/xfail: how a change makes a red
  test disappear instead of fixing it.
- time.sleep() in tests/: the classic flakiness source (prefer events/polling).

The legitimate existing skips guard the env-gated marker suites (e2e, install,
install_script) and the two real time.sleep() synchronizations all live on
origin/main, so they aren't added diff lines and don't trip the gate; a
genuinely-needed new one must update the gate deliberately. Scoped to tests/ so
production sleeps (e.g. real-time pacing) are unaffected. The matcher ignores the
injectable `sleep=`/`self._sleep(` test idiom.

Coverage proves a changed line executed; it can't prove a test would fail if that
line broke (the earlier ffmpeg-cleanup bug passed at 100% coverage). This adds a
mutation gate that closes that gap, scoped to the diff so it stays fast and only
judges changed code.

scripts/mutation_gate.py:
- Finds added lines in aai_cli/*.py vs the merge-base with origin/main.
- Mutates only those lines with a high-signal operator set (comparison negation,
  and/or swap, arithmetic swap, boolean/numeric constants) — `not`-removal and
  statement deletion are omitted to avoid equivalent mutants and parent rewiring.
- Reruns only the tests that cover each mutant, read from per-test coverage
  contexts in .coverage, so it's fast; a hang counts as killed.
- Fails on survivors. A genuinely-equivalent or intentionally-unasserted line can
  be marked `# pragma: no mutate` (scanned across the statement's whole span so a
  formatter-wrapped comment still counts).

check.sh: the main pytest run now writes per-test contexts (--cov-context=test);
the gate runs after diff-cover and self-skips when origin/main is absent, exactly
like diff-cover. coverage becomes an explicit dev dep (used directly by the gate;
DEP004-ignored since scripts/ is dev-only tooling never shipped in the wheel).

Two changed lines in sources.py are marked `# pragma: no mutate`: `produced +=`
is only tested for == 0 (sign is an equivalent mutant), and the real-time pacing
math is deliberately not asserted (tests inject a no-op sleep).

Aikido's license gate flagged refurb as GPL-3.0-only, which conflicts with this
repo's MIT license. refurb was a dev-only lint tool invoked as a subprocess (no
copyleft obligation on our code in practice), but keeping the dependency tree
GPL-free for a distributed OSS CLI is the cleaner call and avoids a recurring
audit flag. Its value over ruff's UP rules was marginal, and the concrete
rewrites it already suggested (contextlib.suppress, dict-union, list.extend)
remain in the code on their own.

Drops the dev dependency, the [tool.refurb] config, and the check.sh step.
zizmor (MIT) and coverage (Apache-2.0) are unaffected.