scc/writer: delay colliding pop-on captions instead of dropping them

[jason-valenzuela]

Problem

scc_writer.from_model silently drops whole pop-on captions when a caption's line-21 packets cannot be transmitted back-to-back with the previous caption's (the Skipping ISD … due to overlap in line 21 packets path). On dense content this loses caption text with only a log warning.

Fix

When a pop-on caption collides with the previous one, delay its onset to sit right after the previous caption's packets instead of dropping it, carrying the preceding erase (EDM) into the delayed chunk so no stale erase wipes it and the emitted timecodes stay monotonic and non-overlapping. The pull-back of the preceding erase is bounded to the minimum needed, so the previous caption keeps as much of its authored on-screen time as possible (and an author-written short cue is never itself penalised). A caption that still cannot achieve a minimum readable window even after delaying is dropped honestly (the existing warning) rather than emitted as a two-frame flash.

The minimum readable window is exposed as a new SccWriterConfiguration field min_popon_display_frames (default 15 ≈ 0.5 s at 29.97 fps; 0 disables dropping entirely, preserving maximal text at the cost of possible flashes). The common, non-colliding path is unchanged — existing golden-SCC tests are untouched.

Tests

• test_dense_popon_no_line_drop — dense captions previously dropped now survive a write→read round-trip (exact SCC asserted)
• test_dense_popon_emitted_scc_is_ordered — emitted timecodes monotonic and non-overlapping
• test_dense_popon_cascade_all_emitted_captions_displayable — every emitted caption meets the display floor
• test_dense_popon_authored_short_prev_not_dropped — a short authored cue no longer forces an unnecessary drop
• test_dense_popon_drop_only_when_infeasible — a caption is dropped only when no readable window exists
• test_dense_popon_erase_pullback — the bounded erase pull-back path (erase moved minimally, both captions keep a window above the floor)
• test_min_popon_display_frames — config parse/default/validation

Full suite: 617 passed, 0 failures, 0 errors, 6 pre-existing skips.

Notes

• Unrelated pre-existing bug noticed nearby (not addressed here): from_model indexes chunks[-2] under a len(chunks) > 0 guard, which raises IndexError when exactly one chunk exists — happy to file a separate issue.

[palemieux]

@jason-valenzuela I would be tempted to wait until all chunks are collected before moving them.

[jason-valenzuela]

@jason-valenzuela I would be tempted to wait until all chunks are collected before moving them.

Got it! Will adjust my approach here.

[jason-valenzuela]

@palemieux Updated the approach here, let me know what you think! The chunk-moving logic is now run after the all the pop-on chunks are collected.

[palemieux]

Is the issue happening with content that was originally authored for 608, or content that was authored in other formats?

[palemieux]

Why create a popon_captions list when chunks could be sufficient?

[jason-valenzuela]

I added the intermediate list to keep each caption's (paint, erase) pair and its begin explicit for the delay pass and the warning. But I can build the paint/erase chunks straight into chunks in the loop and run the collision pass over chunks directly, pairing them there, and remove the extra list.

[palemieux]

I am not sure there is every a correct value here since the duration of a caption is related to the amount of text present. I would define a simply boolean (by default true) that indicates whether overlapping captions should be dropped.

[jason-valenzuela]

Got it, will switch this over to a boolean config!

[jason-valenzuela]

Is the issue happening with content that was originally authored for 608, or content that was authored in other formats?

This issue is happening for content that was authored in other formats! I've been seeing this issue when we're doing VTT -> SCC conversions.