The shared wrapper over ALLM.Pipeline.Artifacts — stores and retrieves pipeline artifacts (HTML, JSON, extracted text) through the configured adapter, and returns a URL reference for the PostgreSQL step_log.

What this module owns, and what the adapter owns

Here: compression, checksum and size accounting. store/4 gzips (unless compress: false), SHA-256s the ORIGINAL bytes, records the ORIGINAL size, and fetch/1 gunzips on the way back out. An adapter receives an opaque, already-encoded payload plus a ALLM.Pipeline.Artifacts.meta/0 and stores both verbatim — so a new backend never re-derives any of it.

There: capacity. "Does this fit me?" is the adapter's question. An adapter that cannot take a payload returns {:error, :too_large} and this module routes to the next tier. DynamoDB's answer is the interesting one: its 400KB ceiling constrains the item as stored, i.e. gzipped and then base64-encoded, so Dynamo.fits_item?/1 measures that and not the caller's original byte count — gating on the original routed a 600KB HTML artifact that gzips to 40KB down the (unimplemented) S3 path, where it was discarded. size_bytes in the return tuple and in the stored item is still the ORIGINAL uncompressed size; only the routing decision changed.

URL Format

• DynamoDB: dynamo://table_name/artifact_id
• Filesystem: file:///path/to/artifact
• Memory: memory://artifact_id
• S3: s3://bucket_name/key

Tiering is an adapter choice (Phase 7)

This module no longer routes by size or special-cases an s3:// URL. Every URL is handed to Artifacts.impl/0, which for the production wiring is ALLM.Pipeline.Artifacts.Tiered — it measures the post-encode size and routes small→DynamoDB / large→S3 on put/4, and dispatches fetch/delete/exists? by URL scheme. An oversize artifact now has a real home, which is why ALLM.Pipeline.Executor.build_envelope/3 no longer truncates its LLM envelope in two rounds.

The gunzip is bounded (memory-safety)

fetch/1 gunzips a stored payload with an EXPLICIT decompressed-size ceiling (max_decompressed_bytes/0, default 64 MB), returning {:error, :artifact_too_large} rather than inflating an unbounded amount into memory. The ceiling is a fixed memory-safety budget — what one fetch may safely hold — chosen INDEPENDENTLY of any store's capacity: Tiered/S3 removed the 400 KB DynamoDB bound, so "the max stored size" is undefined, and a cap generous enough to inflate a legitimate huge artifact whole would still admit a decompression bomb of that size. A genuinely huge artifact simply cannot be fetched whole into memory — the correct posture.