[Artifact Forest 7/9] Add local speech-to-timed-transcript provider

[szmyty]

Parent: #421
Depends on: #427

Outcome

Add an offline-first speech-recognition provider that turns audio or an extracted audio stream into a versioned timed transcript, then allows ordinary transcript/subtitle derivatives.

Provider direction

Evaluate whisper.cpp or another maintained local-first open provider against:

• platform support;
• model licensing/distribution;
• deterministic/configuration-dependent behavior;
• timestamps;
• language handling;
• resource requirements;
• cancellation/progress;
• output formats;
• current maintenance.

Do not hardcode one provider into the domain model; the capability is provider-neutral.

Pipeline

audio/video
   ↓
audio stream selection/extraction
   ↓
speech.transcribe
   ↓
renderflow.timed-transcript/v1
   ├── SRT
   ├── WebVTT
   ├── JSON
   └── text

Requirements

• Provider registry entry, version/model fingerprint, and explicit model artifact identity.
• Offline execution by default.
• Bounded CPU/memory/time/output behavior where practical.
• Progress/cancellation.
• Word/segment timestamp policy.
• Language selection/detection evidence.
• Transcript confidence/limitations.
• No claim that ASR transcript is human-approved copy.
• Preserve exact audio source digest and extraction lineage.

Acceptance criteria

• One redistribution-safe spoken-audio fixture produces deterministic-enough evidenced timed transcript output.
• SRT/WebVTT/JSON/text derivatives are produced from one normalized transcript rather than independent ASR runs.
• Model/provider identity participates in cache/checkpoint keys.
• Missing provider/model/resource requirements fail clearly.
• No network is required by the default provider path.
• Source media remains unchanged.

Non-goals

Speaker diarization, translation, or advanced temporal editing unless separately scoped.

Candidate provider observation — 2026-09-25

Primary current candidate: https://github.com/ggml-org/whisper.cpp

The repository is public and not archived. Evaluate its current CLI/model/runtime/licensing surface during implementation; keep speech.transcribe provider-neutral so another local engine can satisfy the same contract later.