docker-laya: multi-checkpoint Laya API with auth, presets and bulk inference

# docker-laya

[](https://github.com/chneau/docker-laya/actions/workflows/publish.yml)
[](https://ghcr.io/chneau/laya)

Dockerized [Laya](https://huggingface.co/convaiinnovations/laya) prediction
service: loads one or more checkpoints behind a router, then serves typed
decisions (choice / score / noul) over HTTP, auto-routed by language or pinned
with `model`.

Built and published for `linux/amd64` and `linux/arm64`.

---

## ✨ Features

- 🔀 **Multi-Checkpoint Routing**: bundles `english`, `multilingual` and
  `typed-decisions`, auto-selected by language or pinned per request.
- 🎯 **Typed Decisions**: `choice` / `score` / `noul` questions with calibrated
  probabilities, confidence and action probability.
- 🔐 **Timing-Safe Auth**: API keys (`X-API-Key` / `Authorization: Bearer`) and
  HTTP Basic, compared in constant time.
- 📑 **Interactive OpenAPI Docs**: Swagger UI (`/docs`), ReDoc (`/redoc`) and the
  raw schema at `/openapi.json`.
- 🧰 **Built-in Presets**: ready-made question sets (`triage`, `email`, `guard`,
  `moderation`, `router`).
- 📦 **Bulk Inference**: `/predict/bulk` over many states, with per-state
  questions/model and isolated errors.
- 🔎 **Detection & Email Helpers**: `/detect` (script/language) and
  `/email/state` (clean + structure an email).
- 🧩 **Flexible State**: string, JSON object, or conversation turns; criteria
  values may be any JSON (dicts/lists/numbers are rendered as compact JSON).
- 🛡️ **Non-Root**: runs as unprivileged `appuser` (uid `10001`).
- 📦 **Multi-Architecture**: supports both `linux/amd64` and `linux/arm64`.
- 🩺 **Healthcheck**: dedicated `/healthz` endpoint and container `HEALTHCHECK`.
- 🪶 **CPU-Only Torch**: uses the PyTorch CPU wheel index (no CUDA), roughly
  `0.35s` per predict on CPU.

---

## 🚀 Quickstart

Pu