t27f: R4d -- mode 5 overlays live inside one scene: no caption straddles a beat crossfade #10614

Part R4d of epic #10560. Follows R4c (#10609, PR #10610).

Boundary

• specs/t27f/lesson_reel.t27 and its seal .trinity/seals/t27f_lesson_reel.json. Nothing else.
• In scope: in LR_MODE=5 only, the time window of each caption, card and tag, measured against the beat clock's crossfades.
• Out of scope: modes 3 and 4 (their output must not change), the scene clips (R3, t27f: scene renderers for the 12 archetypes and the still anomaly gate (R3) #10592), the render driver, and publishing.

User Scenarios

• A viewer watches a mode 5 reel. A caption appears after a crossfade and disappears before the next one starts. It never fades in over one scene and out over another.
• An author keeps LR_MODE=3 or LR_MODE=4 and gets byte for byte the graph they had before.

Requirements

• Mode 5 computes each overlay's window from the beat clock and on whole frames: start no earlier than the end of the crossfade into its scene, end no later than the start of the crossfade out of it.
• An overlay that cannot fit inside its scene, because the scene is shorter than the overlay's minimum time on screen, is dropped and not squeezed. The test names that case.
• The window is a function with tests. Mode 5's enable= expressions print exactly that function's frames.
• The output of modes 3 and 4 does not change.

Success Criteria

• On the Railway lab, t27c test-report passes with 0 vacuous tests, and the seal is saved and verified.
• A test covers every voice the clock accepts: no overlay window overlaps a crossfade. A control shows that mode 3's windows, laid over mode 4's cuts, do overlap one, on at least one voice.
• On the lab, the 96-graph cmp regression for modes 3 and 4 against the t27f: lay the reel's captions, cards and address over the beat video, LR_MODE=5 (Closes #10609) #10610 binary finds 0 differences.
• On the lab, ffmpeg renders mode 5 from solid-colour clips. Across each crossfade, a pixel inside a caption box shows only the clips and their blend, with no text.