[thealidev]

Summary

Ephemeral training workers and serverless AI containers (Render, Fly.io, Cloud Run) typically operate with strictly capped local disk storage (0–512 MB). When ingesting large dataset dumps or model weights over IPFS, standard pinning or local blockstore accumulation triggers premature ENOSPC errors.

This PR adds an architecture and configuration guide under docs/zero-disk-ml-streaming.md:

• Explains minimal in-memory datastore configurations and aggressive GC watermarks.
• Details in-flight chunked streaming patterns without disk persistence.
• Provides empirical benchmarks from production runs (12+ hours continuous streaming, 1.7M+ tokens, 0.00 Bytes written to host disk).