fix: normalize YouTube caption event boundaries and timestamps - #905

YouTube timedtext events could join Unicode or quoted words without the required spaces, cross punctuation hidden by closing quotes/brackets, and accept invalid JSON3/XML timestamps. Repeated cue normalization could also change the timeline used for SRT export.

This change preserves spacing between event-level words while keeping JSON3 segments concatenated unchanged. It handles apostrophe suffixes as contractions, recognizes terminal punctuation after closing quotes/brackets, and allows common titles to continue only before a reliable capitalized name-like word. Ambiguous abbreviations, domain-like endings, and decimal-like endings keep their event boundaries.

JSON3/XML start and duration validation rejects invalid types, blank values, and negative times; cue normalization discards non-finite ends. Missing or zero durations retain inference from the next valid cue. Incremental-prefix collapse and word-stream merging converge to make repeated normalization and SRT export idempotent.

The branch is based directly on upstream af81332fb598f78b1199d103e0943e4e521a7f96. Only the data module, its tests, and a short validation note change.

Validation on that baseline, on CW with pnpm 9.12.1 and WXT 0.20.18, through the resource-safe wrapper with one worker and serial commands:

• Red: the candidate tests on unchanged upstream source produced 71 failures out of 294.
• Green: all 294 module tests passed; 71 related video subtitle tests passed.
• Strict module coverage: statements, branches, functions, and lines all 100%, with no coverage exemptions.
• Nine independent Occam ablations each caused existing assertion failures; the source was restored exactly. The module, tests, and dependency lock are byte-identical across the upstream update.
• pnpm test:audit, pnpm compile, Chrome/Firefox/Userscript builds, the Userscript verifier, and pnpm docs:build passed.

These are offline fixture and build checks. They do not establish real YouTube response capture, online translation quality, ASR, or GPU behavior.