[Ingestion]: Transcripts — SRT and WebVTT subtitle extraction

[inherent-prime]

Problem

Meeting and video knowledge reaches agents as subtitle/transcript files; SRT and WebVTT are rejected today even though they are plain text.

Proposed contract

• SRT: application/x-subrip, .srt. WebVTT: text/vtt, .vtt.
• Extraction: strip cue numbers and timestamp lines; join cue text into prose. Keep coarse [t=MM:SS] markers every N cues so agents can cite moments.
• chunking_hint default (prose after extraction).
• Surface: rest+mcp (text formats).

Acceptance criteria

• Fixtures sample.srt / sample.vtt; extraction yields clean prose with periodic timestamp markers and zero cue-number noise — test-pinned.
• MCP upload accepted for both, pinned in MCP tests and failure parity.
• Docs and CHANGELOG updated.

Depends on #117. Audio transcription (mp3/wav/m4a) is a separate issue.