[Artmann]

Summary

The API handlers were doing everything: parsing requests, validating, running the use case (locks, idempotency markers, embedding, write order) and building responses. This PR separates HTTP handling from business logic and organises the code by domain instead of by layer.

src/datapoints/   model · ingest · delete · list · http   (DatapointService)
src/recall/       select (pure algorithm) · http          (RecallService)
src/webhooks/     model · service · rate · http           (WebhookService)
src/app.rs        Infra (store, index, embedder, ready) + Services wiring
src/error.rs      ServiceError, with no HTTP in it
src/http/         shared router, JSON extractor, error envelope
src/locks.rs      per-key locks shared by datapoints and webhooks

• Handlers (*/http.rs) only parse the request, call the service and shape the response. Each domain registers its own routes(), and http::router merges them.
• Services take plain commands (DeleteCommand, RecallQuery, ListQuery, …) and return domain results, e.g. Ingested with an outcome of Created, Deduplicated or Replaced. The handler turns that into a status code and response flags.
• Errors: ApiError wraps ServiceError, plus the errors that only exist over HTTP (malformed request, body too large). Error codes, statuses and the envelope are unchanged.
• Webhook events go through authenticate and then receive, so a wrong id or token still returns 404 before the body is checked.
• CONTRIBUTING.md code layout updated.

No change to the HTTP contract or storage layout. The tests only needed import and path changes.