fix(skills): caption burn works on FFmpeg 9 and silent takes

[TobiOnabolu]

What this does

Fixes two independent faults in caption_burn.py (the ugc-talking-head-video skill's caption step),
each of which aborted a render outright on a current machine. Both are now covered by tests.

FFmpeg 9 removed -filter_complex_script. Every burn died with Unrecognized option 'filter_complex_script' before a frame was encoded. Worth recording why this is not a one-line flag
bump: the replacement the release points at, -/filter_complex, does not appear in
ffmpeg -h full, so no capability probe finds it by inspection — and a naive substring probe
matches -filter_complex inside -/filter_complex and silently returns the wrong flag (I hit
exactly this while fixing it). Passing the graph inline sidesteps the version question entirely:
every version accepts it, the argv list reaches exec directly so there is no escaping step, and a
caption graph is far under ARG_MAX.

A silent take has no audio stream, but the render mapped 0:a unconditionally. ffmpeg aborted
with Stream map '' matches no streams. The sting is that the skill's own silent mode emits exactly
that video-only file, so captions could not be burned onto the output that mode produces. Audio is
now copied only when a stream is actually present.

How it was tested

tests/test_ugc_caption_burn.py — 5 passed:

tests/test_ugc_caption_burn.py::test_graph_is_passed_inline_not_via_a_script_file PASSED
tests/test_ugc_caption_burn.py::test_audio_is_not_mapped_for_a_silent_take PASSED
tests/test_ugc_caption_burn.py::test_audio_is_copied_when_the_take_has_a_stream PASSED
tests/test_ugc_caption_burn.py::test_output_maps_the_final_chained_overlay PASSED
tests/test_ugc_caption_burn.py::test_has_audio_reads_ffprobe_output PASSED

To make the command checkable without ffmpeg, the render argv moves into ffmpeg_command() and the
audio probe into has_audio(); the assertions read the argv directly. Pillow is imported lazily —
it is only needed to draw, and the module-level import made the script unimportable anywhere
without it.

Also rendered a real 12 s 9:16 take end to end and inspected it:

• output duration 12.041667 s, matching the source
• ffprobe confirms a silent input still yields a captioned output (0 audio streams in, 0 out)
• frame at t=3.0s visually checked: header at 12% height, caption at 78%, white Helvetica Bold with
  dark outline, source video otherwise untouched

Not run: the full suite. PyPI's CDN (files.pythonhosted.org) is unreachable from this machine,
so uv sync cannot resolve the dev dependencies and the repo's conftest.py (which imports FastAPI)
does not load. The 5 tests above were run under real pytest 9.0.3 against an isolated copy of the
test file. Treat CI as the authority on the whole suite; if the run is blocked pending approval,
that is the sandbox, not a failure.

Checklist

• uv run --with pytest-xdist pytest -n auto -q passes locally — could not run in full, see above
• Added or updated tests for the change
• Updated the relevant docs/context/ fragment — not applicable; no subsystem changed, the
  script is a standalone skill asset and no fragment documents it
• No secrets in the diff