feat(timing): overlapping v2 scenes composite instead of being clamped

[LeadcodeDev]

Refs #344 — its point 1, which the issue author ranks 1 ≫ 2 > 3 > 5 > 4 > 6 and describes as "what makes a rustmotion render read as a series of slides instead of a continuous motion design piece."

The groundwork was there, and refused on purpose

build_slide_view_tasks_v2 already resolved every scene's at onto an absolute timeline. It then threw the result away:

warning: scene {i}'s `at` resolves before the previous scene's own window
ends — clamped to avoid an overlap this workstream does not model

That decided the design. #344 proposes two shapes — a layers/persist track, or overlapping scenes on the v2 timeline — and the codebase already points at the second: no new vocabulary, at already parses it, and the refusal is a deferred TODO rather than an objection.

What it does

Scenes whose windows overlap composite instead of replacing. Participants stack in declaration order; the bottom one supplies the background and the scene effects, the ones above contribute only their children over a transparent surface. Each keeps its own clock, so a scene starting at @6b begins its t at 0 when its window opens and its entry animations play on arrival, not at the video's start.

"timing": "v2", "bpm": 120,
"composition": [{ "type": "slide", "scenes": [
  { "duration": 12.0, "children": [ … the backdrop that holds for 12 s … ] },
  { "duration": 3.0, "at": "@0b",  "children": [ … beat 1 … ] },
  { "duration": 3.0, "at": "@6b",  "children": [ … beat 2 … ] },
  { "duration": 3.0, "at": "@12b", "children": [ … beat 3 … ] },
  { "duration": 3.0, "at": "@18b", "children": [ … beat 4 … ] }
]}]

Total is max(at + duration) = 12.0 s, not the 24 s five sequential scenes would give.

The emission model changed with it. It used to walk scenes and append frames, with global_frame implied by tasks.len(). It now computes absolute windows first, then walks output frames and collects whoever is live: one participant emits Normal exactly as before, several emit a new Composite, none holds the most recently closed scene's last frame so a gap never goes black.

The distinction that took the work

Snapping must never create an overlap. snap: "beat" can round a cut earlier than the previous scene's end — at 24 BPM a beat is 2.5 s, so at: "@3.0s" lands on 2.5 s, half a second inside its predecessor.

Reading that as "play both at once" would make migrate + snap silently shorten every file it touches. That is not hypothetical: snap_beat_on_top_of_the_migrated_file_moves_cuts_and_changes_duration failed with 15.0 s → 13.3 s on my first attempt.

So placement resolves twice — once as written, once snapped:

The author's own at	The snapped value	Result
overlaps	—	composite: that is intent
does not overlap	overlaps	pushed forward, with a warning: that is quantisation

A scene that both overlaps and declares a transition is contradictory: a transition composites two finished buffers, an overlap composites live scenes, and they cannot both describe the same frames. The transition is ignored and named on stderr rather than silently half-applied.

Also in here

build.rs declared cargo:rerun-if-changed for skills/ but not its subdirectories, and cargo watches a directory's own mtime. Adding a file under skills/rules/ did not invalidate the build cache, so a new rule could silently fail to reach rustmotion skills install until something else forced a rebuild — and skill_files_match_disk would then fail with no obvious cause. Every walked directory and every collected file is now declared; verified by adding a file and watching the crate recompile. Found independently by two people working in this area today, which is a fair sign it bites.

Verification

Five task-level tests plus a unit group on the compositing buffer:

• an_explicit_at_that_overlaps_composites_instead_of_being_clamped — 10 composite frames, each carrying both scenes bottom-first, total 30 frames not 40
• a_composited_scene_advances_its_own_clock_from_its_own_at — the overlapping scene's frame_in_scene runs 0..10, not the view's frame index
• a_scene_spanning_several_others_stays_in_every_one_of_their_frames — the spanning scene is the bottom participant in all 30 frames
• a_gap_between_overlapping_scenes_holds_the_last_live_frame
• snapping_a_cut_earlier_never_creates_an_overlap — the regression that nearly shipped

End to end, a 1280×720 12 s demo: the backdrop is present in all eight contact-sheet cells while BEAT 1→4 change over it, and audio and video both come out at 12.000 s with sound measured at −7.8 dB at 0.5/3.5/6.5/9.5/11.5 s.

cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (1534) all clean.

Found while building the demo, filed separately

#349 — audio_spectrum and waveform stay pinned at their minimum on a synthesised score. Unrelated to this change and reproduced on main, but it means the two obvious ways to show a soundtrack cannot currently be combined.

Not in scope here

#344's other five points are separate PRs. Point 6 (glossy 3D material and depth of field) is not started: "looks like rendered 3D" needs a spec before it needs code, and the author ranks it last.