feat(effects): a vhs tape-tear scene effect

[LeadcodeDev]

Closes #386. Refs #388.

PostEffect had grain, vignette, pixelate, progressive_blur and flash. None displaces bands of the image. The channel-splitting effects act on a cut (chromatic_wipe) or on one node (chromatic_aberration), never on the whole frame as a timed state — so grain plus aberration was the best available approximation, and it does not read as a tape tear.

"effects": [
  { "type": "vhs", "at": 0.4, "duration": 0.8,
    "bands": 12, "offset": 60, "noise": 0.3, "scanlines": 0.2,
    "tracking_line": { "color": "#3DA5FF", "speed": 1.5 }, "seed": 3 }
]

Two decisions worth recording

The band layout is redrawn twelve times a second, not every frame. A tear that changes at 30 or 60 fps shimmers instead of catching; twelve is slow enough to read as a mechanical fault. The seed and the instant decide together, so two renders of the same file are byte-identical.

A slid band leaves the freed edge black, not smeared or wrapped. That gap is what makes the displacement read as a tear rather than a blur. On a light background it is very visible — the rule file says to lower offset rather than hope it blends.

It is a beat, not a filter

at and duration bound it like flash, because the reference holds it for well under a second. Outside the window the buffer is not touched — not lightly, not at all, which one of the tests asserts byte-for-byte.

Verification

Five tests. Three fail without the effect:

assertion `left != right` failed: the frame must be torn inside the window
even rows must be darkened and odd ones left alone: 128 vs 128
the tracking line must paint its own colour across a band of rows

The other two assert equality and pass both ways on purpose — outside the window nothing changes, and one seed at one instant is deterministic.

End to end on a contact sheet at 0.10 / 0.40 / 0.60 / 1.20 s: clean, torn with the tracking line near the top, torn with it lower, clean again.

cargo fmt --all --check and cargo clippy --workspace --all-targets -- -D warnings clean; cargo test --workspace 1632.

Two ffmpeg-dependent tests (integration_audio_track_from_embedded_video, successful_extraction_leaves_no_partial_file_behind) failed once in a full run under heavy parallel load and pass 23/23 in isolation. That is the third distinct wall-clock-dependent test to do this today under agent load; worth its own issue, unrelated to this change.

Written comment-free, per the codebase-wide rule from #345.