We need to add a new Spark measurement that tests how many deals (CIDs) can be retrieved from the network using any available retrieval provider, including non-Filecoin nodes running IPFS.

If a piece is stored with multiple SPs but only one of them serves retrievals, the new score should flag this content as retrievable. That matches the experience of retrieval clients: they wanted to retrieve a CID and they got back their content, all was good.

The new score Should demonstrate the real-world benefits of content addressing and IPFS-based retrievals.

Proposed algorithm:

• If the current retrieval check passes, the outcome of the new check is "OK".
• If the current retrieval check fails because of IPNI error (e.g. 404), the outcome of the new check is the same.
• Only when the IPNI lookup fails with NO_VALID_ADVERTISEMENT, we want to try to retrieve the payload CID from a random providers found in the IPNI lookup response. (Potentially de-duplicating entries from the same provider but with different protocols.)
  • Advertisements supporting HTTP retrieval should have higher chance of getting picked over those supporting Graphsync only
  • Advertisements where ContextID starts with ghsA should have higher chance of getting picked
  • We should Ignore Bitswap protocol