feat(transcripts): normalize SRT and WebVTT caption text deterministically

[EloyEMC]

Problem / requested outcome

TubeMaster currently performs only narrow SRT cleanup. Caption payloads can still contain line-ending differences, structural SRT/WebVTT metadata, markup, excess whitespace, and repeated adjacent cue lines. Phase 4 needs a deterministic normalization boundary before future chunking or indexing.

Scope

Add one pure provider-local caption normalizer, based on issue #16, without changing public transcript contracts:

• Normalize CRLF/CR line endings to LF.
• Remove SRT indexes, SRT/WebVTT timestamps, WEBVTT headers, and supported cue metadata.
• Strip basic caption markup and VTT cue tags without becoming a rendering engine.
• Normalize whitespace per caption line and remove empty output lines.
• Deduplicate only adjacent identical normalized cue lines, preserving order and non-adjacent repetition.
• Decode equivalent string, Buffer, and ArrayBuffer payloads identically.
• Tolerate malformed structural lines without throwing.
• Return existing no-captions behavior when normalized text is empty.

Acceptance criteria

• SRT input with CRLF, indexes, timestamps, blank lines, and markup produces clean deterministic text.
• WebVTT input with header, timestamps, cue tags, and metadata produces the same clean text boundary.
• Equivalent string, Buffer, and ArrayBuffer payloads produce equivalent normalized text.
• Empty cue blocks do not create blank output.
• Adjacent duplicate cue lines collapse once; non-adjacent repeated lines remain.
• Malformed timestamp/index lines never throw and have deterministic tested behavior.
• Available output remains { status: "available", text, language? }.
• Existing unavailable/unsupported results and diagnostics remain unchanged.
• API, CLI, and MCP envelopes remain unchanged.
• Tests cover normalization and existing transcript regressions.

Non-goals

• New transcript providers or fallback changes.
• Language detection, language translation, or language normalization.
• Chunking, persistence, indexing, embeddings, vector databases, semantic search, or topic extraction.
• API/CLI/MCP surface changes.
• Global duplicate removal or sentence reconstruction.
• OAuth, quota, analytics, playlist, reporting, or metadata changes.

Planned validation

• Focused transcript provider/service/API/CLI/MCP tests.
• npm test.
• npx tsc --noEmit.
• npm run lint.
• git diff --check.

Dependency: issue #16 (feat(transcripts): deterministically select YouTube caption tracks, commit cab2140).
Roadmap: Phase 4 Transcriptions.

[EloyEMC]

Implemented and committed as 148e568 (feat(transcripts): normalize caption text deterministically). 107 focused transcript/provider/service/API/CLI/MCP tests passed; 177/177 full tests passed; TypeScript, ESLint, and git diff check passed. The build remains blocked by the external node_modules symlink/Turbopack error: Symlink project node_modules is invalid, it points out of the filesystem root. No push or PR was created.