Add modern transcription provider and privacy matrix

[WalksWithASwagger]

Context

Modern speech-to-text options differ on accuracy, streaming, timestamps, diarization, local/offline behavior, vocabulary, cost, and privacy. WhisperForge should make those tradeoffs visible before adding more providers.

Source: docs/WHISPERFORGE-MASTER-PLAN-2026-05-18.md (wf-transcription-matrix, Phase 5).

Linear: BC-225.

Acceptance Criteria

• Document current transcription backends and candidate providers in a matrix covering accuracy, streaming, timestamps, diarization, local/offline, vocabulary, cost, and privacy posture.
• Add the matrix to docs and optionally expose a compact version in the UI.
• Recommend default backend choices for local privacy, long-form accuracy, timestamps, and low-cost runs.
• Identify the smallest next provider integration, if any, with explicit acceptance criteria.
• Do not change runtime defaults without tests and documentation.

Tests/Evals

• Docs/registry validation only unless code changes are needed.
• If provider-selection code changes, add unit tests around routing and settings persistence.

Verification

• python3 -m json.tool ops/roadmap/features.json
• git diff --check
• make test

Agent Instructions

Treat this as a product/technical decision issue first. Use official provider docs and keep claims dated.

Out of Scope

Do not add Deepgram, AssemblyAI, or another provider in the same issue unless the matrix makes the integration trivial and scoped.

[WalksWithASwagger]

PR #29 is ready for review: https://github.com/WalksWithASwagger/spektorAI/pull/29\n\nThis docs-first slice adds the dated transcription provider/privacy matrix, links it from README/changelog, recommends defaults by use case, and identifies OpenAI diarized transcription as the smallest next opt-in integration. Runtime defaults are unchanged. Verified with roadmap JSON validation, git diff --check, and the full test suite (210 passed).