[Artifact Forest 6/9] Add timed-text formats and embedded subtitle extraction

[szmyty]

Parent: #421

Outcome

Make timed text a first-class static artifact family and extract/convert embedded subtitle streams through bounded local providers.

Initial formats

• SRT
• WebVTT
• ASS/SSA
• TTML
• plain transcript text
• structured timed-transcript JSON

Scope

• Extend format/capability registry with missing timed-text formats.
• Define renderflow.timed-transcript/v1 with stable cue IDs, start/end times, speaker when known, text, language, source stream, and provenance.
• Add FFprobe-backed stream discovery as an explicit provider capability.
• Add FFmpeg-backed extraction/conversion only for combinations proven by fixtures.
• Distinguish embedded-subtitle extraction from speech recognition.
• Preserve stream/language/codec identity and record lossy conversions.
• Validate monotonic timing, overlap policy, encoding, empty cues, and malformed inputs.

Acceptance criteria

• Video/container subtitle streams can be enumerated without decoding the entire media artifact.
• At least SRT and WebVTT extraction/conversion paths are proven.
• ASS/SSA/TTML are advertised only to the extent fixtures prove.
• No subtitle stream is invented when none exists.
• Cue/source provenance survives format conversion.
• Unsupported subtitle codecs/providers produce typed diagnostics.
• Static extraction remains Renderflow-owned; temporal editing/reconstruction remains Aniflow-owned.

Non-goals

Speech-to-text transcription; that is #428.