build(server): split Dockerfiles, make ML model paths optional (#120)

Existing Dockerfile (with LLM/STT/TTS model download) is now `Dockerfile.full`.
The new top-level `Dockerfile` builds a slim image without python, without
huggingface_hub, without the model download step. The slim image is the new
default for users who only want the orchestration surface — the directive
folder UI, the mesh/task system, the API.

## Slim Dockerfile
* No python / huggingface_hub / model downloads.
* Same runtime tooling as `k8s/daemon/Dockerfile` (git, gh CLI, ssh, jq,
  curl, ca-certs, libssl3).
* Embeds the daemon binary at /app/daemon-binaries/makima-linux-x86_64
  for the in-server download endpoint.
* PARAKEET_MODEL_DIR / SORTFORMER_MODEL_PATH / CHATTERBOX_MODEL_DIR are
  intentionally NOT set — Listen and Speak return "ML models not
  configured" if a client tries to use them.

## ML model paths now optional
`ServerArgs.parakeet_model_dir`, `parakeet_eou_dir`, `sortformer_model_path`,
`chatterbox_model_dir` are now `Option<String>` (no defaults). The bin
constructor inspects them: if all four are present, configures
`AppState::new`; if all four are absent, uses the new
`AppState::new_slim()` which leaves `model_config = None`. The lazy load
path in `get_ml_models` already returned a clean error for None.

Speak (TTS) was already optional via `model_config.as_ref()` — still works.

Mixed configurations log a warning and degrade to slim mode.

## Ops note
The old `Dockerfile.full` retains the original behaviour for anyone who
needs STT/diarization/TTS in production. CI still builds the daemon image
from `k8s/daemon/Dockerfile` (untouched).

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

0 files changed, 0 insertions, 0 deletions