Follow the API's ducking flip to default-off (0.11.3 / cli 0.9.0) - #26

The server now treats an unset ducking as off on /v1/video-to-music, /v1/video-to-sound, /v1/video-to-video-music and /v1/video-to-video-sound (dashboard PR #233, deployed).

Both packages already omitted ducking when unset and still do, so nothing changes on the wire. What changed is everything that explained the old default — and on the CLI, the only expressible direction was the one that is now the default.

sonilo 0.11.2 → 0.11.3 (docs only)

• The request builders no longer justify omit-when-unset by naming which way each default points. That reasoning is exactly what went stale; the rule is that the server decides, and neither ducking nor keep_original_sound is pinned here.
• build_v2s_parts' docstring gained the thing specific to /v1/video-to-sound: with no keep_original_sound field on that endpoint, ducking picks the voice source as well as the mix style, so leaving it unset keeps the source's own speech out of the deliverable entirely and produces no music_processed stem.
• dubbing's docstring no longer describes its default as "the opposite of the music endpoints'" — every endpoint now defaults off.
• README: the video_to_music ducking bullet and example, the video_to_video_music example, the sound-endpoint paragraph, and the dubbing paragraph.

sonilo-cli 0.8.2 → 0.9.0

The only behaviour change in this PR.

• --ducking added to video-to-sound, video-to-video-sound and video-to-video-music. Without it the CLI cannot request ducking at all now.
• --no-ducking is kept. Removing it would turn every existing script that passes it into a hard argparse failure. It now sends an explicit False, which is what the server would have applied anyway.
• Passing both exits with a clear message rather than silently picking a winner.

Unrelated bug fixed while bumping versions

_version.py was still 0.11.0 against pyproject's 0.11.2, and sonilo_cli/__init__.py still 0.8.0 against 0.8.2. Those are the values sent as x-sonilo-client-version, so releases 0.11.1, 0.11.2, 0.8.1 and 0.8.2 all reported themselves as the older version in telemetry. All four strings now agree.

The existing identity test only compares the header against _version.py, so it compares the drifted value with itself and cannot catch this. Worth a CI check that ties _version.py to pyproject.toml — not added here to keep the diff on topic.

sonilo-cli's sonilo>=0.11.0,<0.12 pin still holds, since the core stays on 0.11.x. No widening needed this round.

Testing

237 core tests and 88 CLI tests pass, run per package as CI does. New CLI coverage pins --ducking → ducking=true, --no-ducking → ducking=false, unset → field absent, and the both-flags rejection.

Note: running both directories in a single pytest invocation fails at collection because each package has its own top-level tests package. That predates this change and reproduces on a pristine main.

Note for video_to_sound callers

This is the one endpoint where the flip is user-visible without any other flag: callers who relied on the default no longer get the source speech in the result, with no error raised.