Auto-register all Whisper models with lazy loading

[basnijholt]

Summary

Instead of requiring --model flags, auto-register all available faster-whisper models by default. Models are downloaded lazily on first API request.

Current behavior

agent-cli server whisper --model large-v3 --model small

User must explicitly specify which models to make available.

Proposed behavior

agent-cli server whisper  # All models available by default

• Register all 19 models from faster_whisper.available_models()
• No upfront download - models download on first request
• TTL unloads idle models as usual
• Optional --model flag to limit to specific models

Benefits

• Simpler Docker config (no WHISPER_MODEL env var needed)
• API clients can request any model without server restart
• No wasted disk space - only requested models are downloaded

Models available

tiny.en, tiny, base.en, base, small.en, small, medium.en, medium,
large-v1, large-v2, large-v3, large, distil-large-v2, distil-medium.en,
distil-small.en, distil-large-v3, distil-large-v3.5, large-v3-turbo, turbo

[basnijholt]

Update: Default model auto-preload is now implemented in the feat/tts-server branch.

New behavior:

• Single model: auto-downloads/loads on startup (no --preload needed)
• Multiple models: only the default (first) model is auto-downloaded
• --preload flag: forces all models to be preloaded

The auto-register-all-models feature (where all 19 models are available without specifying --model) is still pending for a future PR.